FYP Proceedings 2025-2026

Final Year Project Proceedings D E P A R T M E N T O F E L E C T R I C A L A N D C O M P U T E R E N G I N E E R I N G Editors: Amelia Wong Azman Nurul Arfah Che Mustapha 2 0 2 5 / 2 0 2 6VOL 1:

© Department of Electrical and Computer Engineering 2026 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, without any prior written permission of the publisher Perpustakaan Negara Malaysia Cataloguing-in-Publication Data DEPARTMENT OF ELECTRICAL AND COMPUTER ENGINEERING FINAL YEAR PROJECT PROCEEDINGS Edited by: Amelia Wong Azman and Nurul Arfah Che Mustapha e-ISBN: 978-629-92836-0-7 Published by: Department of Electrical and Computer Engineering Kulliyyah of Engineering International Islamic University Malaysia P.O.Box 10 50728, Kuala Lumpur, Malaysia Tel: +603-6421 4530 Email: ece.koe@iium.edu.my

Low-Power VLC Communication for Wearable Health Monitoring Devices Characterisation of Wide Bandgap Perovskite Solar Cells for Failure Analysis and Degradation Visible Light Communication (VLC) for Real-Time Patient Monitoring in Hospital Wards Degradation Mechanism Analysis of Unencapsulated Narrow Bandgap (NBG) Perovskite Solar Cell (PSC) Devices Bi-directional EDFA Pumping Scheme for FSO Receiver Link in 5G Fronthaul System Machine Learning Framework for Text Simplification to Support Dyslexic Readers Detecting Phishing Websites with Transformer-Based NLP and URL Feature Fusion Performance Benchmarking of Explainable AI Methods for Skin Cancer Detection TABLE OF CONTENTS 21 24 27 PREFACE 2 5 9 14 18 TRACK 1: 1 1 2 3 ANTENNA PROPAGATION / PHOTONICS / WIRELESS TRACK 2: ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING

48 53 58 63 66 72 36 42 31 4 5 Secure MQTT Communication Protocols for IoT Devices AI-Powered Intrusion Detection System for Enhanced 5G Cybersecurity Networks Using Deep Learning TRACK 3: COMPUTER AND SYSTEM / IOT Early-Stage Chronic Kidney Disease Prediction Using Interpretable Machine Learning Experimental Testing of a Flexible Inverter Circuit for Printed Electroluminescent Display FPGA-Based Testing Platform for Semiconductor Devices Radio Frequency Energy Harvesting Wireless Communication for IoT Sensors Energy Harvesting System Using Wi-Fi Signals Inverter circuit for printed electroluminescence control display Simulation of a Microfluidic Device for Size-Based Separation of Microplastics Using Finite Element Analysis TRACK 4: ELECTRONICS / INSTRUMENTATION AND MEASUREMENTS

95 98 103 107 79 84 90 6 7 Comparative Analysis of Fixed and Single-Axis Solar Tracking Systems for Residential Photovoltaic Applications Modelling and Analysis of Electric Vehicle Charging Infrastructure Impact on Distribution Networks Enhanced PV Monitoring with Power Quality and Harmonic Distortion TRACK 5: POWER AND ENERGY Rooftop Photogrammetry for Solar Panel Dimensioning Using Drone Speech Enhancement and Noise Reduction via Machine learning Vessel-Enhanced Deep Learning Approach for Intracranial Aneurysm Detection Smart attendance System using face recognition identification TRACK 6: SIGNAL & IMAGE PROCESSING AND ANALYSIS 8 TITLES PUBLISHED IN OTHER PUBLICATIONS 113

PREFACE As part of the graduation requirement, all IIUM engineering students are required to complete a 6 credit hour Final Year Project (FYP). At the Kulliyyah of Engineering, the credits spanned over the students’ last two semesters. Throughout their FYP journey, students are introduced to essential research skills, including experimental design, data analysis, effective presentation, communication skills, and academic writing. This experience not only strengthens their technical foundation but also prepares them to contribute meaningfully to the engineering profession. At the end of every semester, many of these projects were presented and published in local and international conferences. Others, although equally valuable, have not found a formal platform for dissemination. Recognising this, the department has taken the initiative to provide a dedicated platform to showcase the remaining works. In other words, this publication is to celebrate our students’ perseverance and scholarly achievements in completing their FYP. As a bonus, this publication also serves as an archive of FYP works for our Bachelor of Electrical and Electronics Engineering programme. We also hope that this compilation becomes a source of reference and inspiration for future cohorts. 1

Antenna Propagation / Photonics / Wireless T R A C K 1

Low-Power VLC for Wearable Health Monitoring Devices Amira Farahanim Mohd Rizalme and Suriza Ahmad Zabidi Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—Wearable health monitoring systems require reliable and energy-efficient communication technologies to support continuous physiological data transmission. Conventional wireless technologies such as Bluetooth and WiFi suffer from electromagnetic interference, security concerns, and increased power consumption. This paper presents the design and evaluation of a low-power Visible Light Communication (VLC) system for wearable health monitoring applications. The proposed system utilises a MAX30100 sensor to acquire heart rate and blood oxygen saturation (SpO₂) data, an Arduino Mega microcontroller for data processing, a white LED transmitter, and a receiver unit consisting of a black infrared photodiode, Arduino Uno, and OLED display. On-Off Keying (OOK) modulation is employed to transmit physiological data through visible light. Experimental evaluation was conducted to analyse power consumption, communication reliability, transmission accuracy, communication distance, and ambient light effects. The transmitter consumed 0.4071 W during idle mode and 0.4110 W during active transmission, while the total system power consumption increased by only 0.0036 W during communication. The system achieved 100% transmission accuracy and communication reliability up to 2.2 cm under indoor line-of-sight conditions. Results indicate that VLC is feasible as a low-power and interference-free communication technology for wearable healthcare monitoring applications. Index Terms—Visible Light Communication, Wearable Healthcare, VLC, Heart Rate Monitoring, SpO₂ Monitoring, Optical Wireless Communication. I. INTRODUCTION Wearable health monitoring devices have gained significant attention due to their ability to continuously monitor physiological parameters such as heart rate, blood oxygen level (SpO₂), body temperature, and electrocardiogram (ECG) signals [1]. These devices support early disease detection, remote patient monitoring, and preventive healthcare by enabling real-time collection and transmission of patient data [2]. The increasing prevalence of chronic diseases, aging populations, and the growing demand for home-based healthcare solutions have further accelerated the adoption of wearable medical technologies worldwide. As a result, wearable devices are becoming an essential component of modern healthcare systems and Internet of Medical Things (IoMT) applications. In recent years, advancements in miniaturised sensors, embedded systems, and wireless communication technologies have enabled the development of compact and energy-efficient wearable devices capable of operating continuously for extended periods. The integration of wearable devices with cloud computing and Internet of Things (IoT) platforms has also enhanced remote monitoring capabilities, allowing healthcare professionals to access patient information in real time and make informed clinical decisions [3]. However, the effectiveness of these systems largely depends on the communication technology used to transfer physiological data from wearable devices to monitoring stations or healthcare networks. Radio frequency (RF)-based communication technologies such as Bluetooth, WiFi, and ZigBee are commonly employed in wearable healthcare applications due to their widespread availability and ease of implementation. Despite their popularity, these technologies present several limitations. Continuous RF communication significantly increases power consumption, reducing battery lifetime and limiting long-term operation of wearable devices [4]. Furthermore, RF-based systems are susceptible to electromagnetic interference (EMI), particularly in environments where multiple wireless devices operate simultaneously, such as hospitals and healthcare facilities [5]. RF signals also propagate beyond physical boundaries, raising concerns regarding data security, privacy, and unauthorized access to sensitive medical information. Visible Light Communication (VLC) has emerged as a promising alternative wireless communication technology that utilises light-emitting diodes (LEDs) to transmit information through the visible light spectrum ranging from 350 nm to 800 nm [6]. Unlike RF-based systems, VLC is inherently immune to electromagnetic interference and offers enhanced communication security due to its line-of-sight (LOS) propagation characteristics. Since visible light cannot penetrate opaque objects such as walls, data transmission remains confined within the intended environment, reducing the risk of eavesdropping and improving patient privacy [7]. Additionally, VLC systems can leverage existing LED lighting infrastructure, potentially reducing deployment costs while maintaining low power consumption. Numerous studies have investigated VLC for indoor co

II. METHODOLOGY A. System Architecture The proposed VLC system consists of three major components: a wearable transmitter unit, a VLC communication channel, and a receiver unit. The transmitter incorporates a MAX30100 sensor, Arduino Mega microcontroller, and white LED transmitter. The sensor acquires heart rate and SpO₂ measurements and sends the data to the Arduino Mega through I²C communication. The microcontroller processes the data and transmits it using OOK modulation through the LED. The VLC channel consists of a free-space optical link operating under line-of-sight conditions. At the receiver side, a black IR photodiode detects incoming optical signals and converts them into electrical signals. An Arduino Uno processes and decodes the received data before displaying the information on an OLED display. Fig 1. Block diagram of the Proposed Design Architecture. B. Hardware Implementation The hardware implementation uses a MAX30100 sensor, Arduino Mega 2560, Arduino Uno, white LED transmitter, black IR photodiode, 220 Ω transmitter resistor, 10 kΩ receiver load resistor, and a OLED display. The transmitter LED is driven directly from a digital GPIO pin of the Arduino Mega with a 220 Ω current-limiting resistor in series. The photodiode is paired with a 10 kΩ load resistor to convert photocurrent to a measurable voltage sampled by the Arduino Uno analog input. C. Software and Modulation Software was developed using the Arduino IDE. On-Off Keying (OOK) was selected as the modulation scheme for its implementation simplicity and low computational overhead — suited for energy-constrained Arduino-based wearable nodes. The transmitter encodes health data as binary ON/OFF LED pulses; the receiver applies threshold detection and noise filtering to reconstruct BPM and SpO₂ values, displaying results on the OLED in real time. D. Performance Evaluation Metrics Power consumption was measured using a multimeter in series with the circuit at idle and active modes using Eq. (1): P = V × I …(1) where P is power (W), V is voltage (V), and I is current (A). Transmission accuracy was computed using Eq. (2): Accuracy (%) = (Correct Packets / Total Packets) × 100 …(2) Communication reliability was assessed qualitatively at distances from 0 to 3.0 cm. Ambient light robustness was tested under three conditions: dark room, normal room lighting, and strong flashlight interference. III. RESULTS AND DISCUSSION A. Power Consumption Analysis The measured transmitter power consumption was 0.4071 W during idle mode and 0.4110 W during active transmission mode. The increase of only 0.0039 W demonstrates efficient VLC transmission with minimal additional energy requirements. The receiver consumed significantly less power, requiring only 1.696 mW during idle mode and 1.400 mW during active reception. The total system power consumption increased from 0.4088 W during idle operation to 0.4124 W during active communication, representing an increase of only 0.0036 W. These results indicate that the developed VLC system operates with high energy efficiency and is suitable for wearable healthcare applications where battery lifetime is critical. TABLE I POWER CONSUMPTION SUMMARY Mode Tx Power (W) Rx Power (mW) Total (W) Idle 0.4071 1.696 0.4088 Active Tx 0.4110 1.400 0.4124 B. Transmission Accuracy Analysis Transmission accuracy was evaluated by transmitting 20 packets at each distance increment. The system achieved 100% transmission accuracy from 0 cm to 2.2 cm, where all transmitted packets were successfully decoded. Beyond 2.2 cm, communication performance degraded significantly. Accuracy decreased to 70% at 2.3 cm and further dropped to 25% at 2.5 cm. At distances greater than 2.6 cm, no packets were decoded correctly due to insufficient optical signal strength. The results indicate that the developed VLC system is highly effective within short communication distances commonly encountered in wearable monitoring applications. TABLE II TRANSMISSION ACCURACY VS. DISTANCE Distance (cm) Correct / 20 Accuracy (%) 0.0–2.2 20 100 2.3 14 70 2.4 13 65 2.5 5 25 2.6 0 0 2.7–3.0 0 0 C. Communication Reliability Analysis Communication reliability remained at 100% from 0 cm to 2.2 cm. Within this operating range, all transmitted packets were received correctly without packet loss or corruption. Reliability degradation began at 2.3 cm, where increasing decoding errors were observed. At distances above 2.6 cm, packet corruption became severe and complete communication failure occurred at 2.9 cm and beyond. The findings demonstrate that the developed VLC system provides reliable short-range communication suitable for wearable healthcare monitoring environments. D. Effect of Ambient LightDepartment of Electrical & Computer EngineeringFYP Proceedings3

Ambient light testing was performed under three conditions: dark room, normal room lighting, and flashlight interference. Under dark room and normal room lighting conditions, the system maintained 100% communication reliability. However, under strong flashlight interference, communication reliability decreased to approximately 50%. The additional optical noise introduced by the flashlight affected the photodiode's ability to accurately detect transmitted signals, resulting in corrupted and missing packets. These findings indicate that while the system performs effectively under normal indoor lighting conditions, strong optical interference can negatively affect communication performance. TABLE III EFFECT OF AMBIENT LIGHT ON VLC RELIABILITY (2 CM) Lighting Condition Correct Corrupt Rel. (%) Dark Room 20 0 100 Normal Room Light 20 0 100 Flashlight Interference ~10 ~10 ~50 IV. COMPARISON WITH RELATED WORKS Table IV positions the proposed system against representative prior works. The proposed VLC prototype offers a favourable combination of interference immunity, LOS-enforced privacy, and simple OOK implementation at standard LED power levels. VLC- LE [7] achieves microwatt-level power at the cost of a sub- centimeter range using sub-knee-voltage LED biasing, whereas the present design uses standard biasing, trading marginal power reduction for implementation simplicity on commodity Arduino hardware. Bluetooth-based systems [10] provide wider range and multi-device integration but introduce RF interference susceptibility and higher communication-subsystem power. HBC [13] achieves 98% accuracy indoors but suffers signal leakage at higher frequencies. TABLE IV COMPARISON WITH RELATED WORKS Ref. Medium Key Metric Strength Limitation [6] VLC >1 Gbps Secure, no EMI Alignment, mobility [7] VLC-LE ≈μW, PER<1% 100× lower power Sub-cm range, low rate [10] BLE 4.0 Multi-device Real-time, reliable RF interference [12] RF-BCC 60 Mbps, 1.24 mW High speed Decoding complexity [13] HBC 55 kbps, 98% acc. Secure indoors Leakage at high freq. This work VLC OOK 0.41 W, 100% acc. No EMI, simple impl. Short range (2.2 cm) V. CONCLUSION This paper presented a low-power VLC system for wearable health monitoring using OOK modulation over a MAX30100–Arduino– LED–photodiode link. Key contributions and findings are: 1) Total active communication power overhead of only 0.0036 W over idle, confirming minimal energy impact. 2) 100% transmission accuracy and communication reliability maintained up to 2.2 cm under indoor LOS conditions. 3) Full reliability under normal room lighting; strong optical interference reduces reliability to approximately 48–50%, identifying a direction for receiver improvement. Future work should address: (i) extending range via higher- sensitivity photodetectors or optical concentrators; (ii) adopting V- PPM or error-correction coding for improved noise robustness; (iii) implementing AC-coupled or adaptive-threshold receivers to mitigate ambient light; (iv) developing a fully battery-powered prototype for energy-lifetime characterisation; and (v) integrating with IoT cloud platforms for scalable remote health monitoring. REFERENCES [1] M. Mohammad et al., Wearable physiological monitoring: A review, IEEE Access, 2021. [2] M. Ali et al., Wireless communication in wearable healthcare systems, Sensors, 2022. [3] S. Rahman et al., Power management in Bluetooth wearables, J. Med. Syst., 2022. [4] X. Zhang et al., EMI in hospital wireless networks, IEEE Trans. Biomed. Eng., 2023. [5] A. R. Ndjiongue, H. C. Ferreira, and T. M. N. Ngatched, Visible Light Communications (VLC) Technology, Wiley Encyclopedia Elect. Electron. Eng., 2015. doi: 10.1002/047134608x.w8267 [6] S. Caputo et al., ETSI SmartBAN Vision on the use of VLC in Body Area Networks, Proc. IEEE ISMICT, 2024. doi: 10.1109/ismict61996.2024.10738176 [7] K. Agarwal et al., VLC-LE: Energy-Efficient and Secure VLC for the Internet of Bodies, Proc. IEEE SmartNets, 2024. doi: 10.1109/smartnets61466.2024.10577745 [8] J. Li et al., Indoor Health Monitoring with VLC-based Passive Posture Monitoring, Proc. IEEE PIERS, 2024. doi: 10.1109/piers62282.2024.10618552 [9] T.-C. Yu et al., VLC System Technology Review, Crystals, vol. 11, no. 9, 2021. doi: 10.3390/cryst11091098 [10] K. Kuok et al., A Bluetooth-Based System for Real-Time Health Monitoring, Proc. IEEE CSRSWTC, 2024. doi: 10.1109/csrswtc64338.2024.10811563 [11] T.-C. Zhuo et al., Respiration Monitoring During Sleep via WiFi, IEEE Trans. Mob. Comput., 2024. [12] S. Vijayalakshmi and V. Nagarajan, FPGA Implementation of RF- Based BCC with RZ Uni-polar Encoding, Proc. IEEE ICCSP, 2020. doi: 10.1109/iccsp48568.2020.9182289 [13] A. Karunakaran et al., B2BIoT: Efficient Wireless Body-to-Body Communication, Proc. IEEE ICSES, 2024. doi: 10.1109/icses63760.2024.10910264 [14] A. M. Khalid et al., Diffuse IR-Optical Wireless System for Patient Monitoring, Proc. IEEE ICTON, 2013. doi: 10.1109/icton.2013.6602887 [15] R. Karunanith

Characterisation of Wide Bandgap Perovskite Solar Cells for Failure Analysis and Degradation Amir Harith Abd Khalik Khassunah and Nurul Fadzlin Hasbullah Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—The rapid evolution of Perovskite Solar Cells (PSCs) has positioned them as a leading third-generation photovoltaic technology. However, the transition to commercial viability is hindered by critical engineering challenges, primarily long-term instability. This study investigates the electrical characterisation and degradation mechanisms of unencapsulated inverted (p-i-n) perovskite solar cells utilizing a Self-Assembled Monolayer (SAM) modified Nickel Oxide (NiOx) hole transport layer. Testing was conducted under Standard Test Conditions (AM1.5G, 1000 W/m²) and dark conditions using an Ossila Solar Cell Prototyping Platform. An inadvertent measurement polarity anomaly subjected the devices to a 0 V to -1.0 V reverse bias, precluding the extraction of standard forward-bias metrics (Voc, FF, PCE). However, this facilitated a rigorous diagnostic failure analysis. The fabricated devices were categorized into two failure modes. Group A exhibited active but degraded performance, with short-circuit current densities (Jsc) suppressed by >95% (<1.0 mA/cm²), indicating severe bulk hydrolysis. Dark J-V sweeps revealed premature reverse breakdown (>0.5 V), confirming interfacial passivation failure and massive trap-assisted tunneling. Group B suffered catastrophic open- circuit failure with Giga-ohm level resistances, indicating severe halide migration and galvanic corrosion of the top metal electrodes. Ultimately, the study concludes that while interfacial engineering is vital, it is entirely insufficient without rigorous hermetic encapsulation to block moisture and ion volatility. Index Terms—Perovskite Solar Cells, Failure Analysis, Reverse Breakdown, Trap-Assisted Tunneling, Galvanic Corrosion, Hydrolysis. I. INTRODUCTION erovskite Solar Cells (PSCs) have emerged as one of the most promising third-generation photovoltaic technologies, boasting power conversion efficiencies (PCE) that have recently surpassed 26%. Their tunable bandgap, high absorption coefficient, and low-cost solution processability make them highly attractive for both single- junction and tandem commercial applications. However, the commercialization trajectory is heavily bottlenecked by intrinsic and extrinsic stability issues. Unlike mature silicon- based photovoltaics, metal halide perovskites are highly sensitive to environmental stressors such as moisture, oxygen, ultraviolet (UV) light, and thermal fluctuations. To combat these stability challenges, researchers have heavily investigated the inverted (p-i-n) device architecture. Inverted PSCs demonstrate superior operational stability and mitigated J-V hysteresis compared to their conventional (n-i-p) counterparts. Further optimization involves substituting highly reactive organic transport layers with robust inorganic materials, such as Nickel Oxide (NiOx), modified with Self- Assembled Monolayers (SAMs) to improve energy band alignment and passivate interfacial defects. Despite these architectural advancements, unencapsulated devices remain exceptionally vulnerable to rapid atmospheric degradation. This project performs electrical characterisation and rigorous diagnostic failure analysis on unencapsulated p-i-n PSCs. By analyzing both the photo-response and the dark diode leakage characteristics, this study aims to isolate and quantify specific degradation mechanisms, including bulk hydrolysis, interfacial passivation failure, and halide-induced contact delamination. II. METHODOLOGY The fabrication and testing of the perovskite solar cells involved a collaborative effort between the International Islamic University Malaysia (IIUM), Ming Chi University of Technology (MCUT), and Universiti Kuala Lumpur British Malaysian Institute (UniKL BMI). A. Device Architecture The solar cells utilized an inverted p-i-n architecture with the following multi-layer stack: Glass / ITO / NiOx / SAM / Perovskite / Electron Transport Layer / Top Metal Contact. The devices were sourced from the Department of Material Engineering at MCUT. The inorganic NiOx layer served as the primary hole transport material, which was surface-modified with a Self-Assembled Monolayer to suppress non-radiative recombination at the transport interface. Notably, the devices were transported and tested without rigorous hermetic encapsulation, exposing them to ambient environmental conditions. B. Electrical Characterisation Setup Characterisation was conducted at UniKL BMI utilizing the Ossila Solar Cell Prototyping Platform equipped with a high- precision Source Measure Unit (SMU). Devices were tested under Standard Test Conditions (STC), defined as 1000 W/m² irradiance using an AM1.5G solar simulator spectrum at room temperature. Current-Voltage (J-V) swe

from 0 V to 1.0 V in 0.02 increments with 0.01 seconds of settling time and a pixel area of 0.9 cm2. To isolate the intrinsic diode properties from the photo-generated current, a parallel set of J-V measurements was conducted in complete darkness. III. RESULTS AND DISCUSSION A. Identification of Reverse Bias In a standard forward-biased J-V sweep of a functional photovoltaic device, the photo-generated current and the forward diode injection current flow in opposite directions. This opposition yields a characteristic zero-crossing point, defining the Open- Circuit Voltage (Voc) and enabling the calculation of the Fill Factor (FF) and overall efficiency. However, the initial analysis of the eight fabricated samples (D1–D8) revealed that the measured current densities never crossed the zero-current axis. Instead, currents began at a positive plateau at 0 V and increased exponentially at higher applied voltages. Cross-referencing the illuminated data with the dark J-V sweeps confirmed that the dark and light currents possessed identical polarities and added constructively (J_total ≈ J_dark + J_photo). In semiconductor physics, this constructive drift only occurs when a diode is operating in reverse bias. Because the devices utilized an inverted p-i-n structure—which possesses the opposite electrical polarity of a standard n-i-p cell—the default positive voltage sweep of the SMU inadvertently subjected the junctions to a 0 V to -1.0 V reverse electrical stress. While this precluded standard efficiency benchmarking, it successfully isolated the reverse-blocking characteristics, providing a robust dataset for failure diagnostics. B. Device Categorization Based on the magnitude of dark leakage and the presence of a photo-response, the eight samples were classified into two definitive failure profiles. Fig. 1. Dark J-V curves for Group A devices. As shown in Fig. 1, Group A (D1, D3, D5, D7) maintained electrical continuity but exhibited severe degradation, showing measurable photo-response alongside premature reverse breakdown at higher potentials. Fig. 2. Dark J-V curves of Group B devices. As seen in Fig. 2, Group B (D2, D4, D6, D8) exhibited catastrophic open-circuit failure. Across both light and dark sweeps, Group B devices maintained current densities in the sub- nanoampere regime (~10^-8 A/cm²), indicating an entirely severed electrical pathway. C. Bulk Hydrolysis and Moisture Ingress Fig. 3. Extracted Short-Circuit Current Densities (Jsc) for Group A devices. Referring to Fig. 3, Group A devices, the short-circuit current density (Jsc) was successfully isolated at exactly 0 V, yielding values ranging from 0.53 mA/cm² to 0.90 mA/cm². While this confirms the survival of a localized built-in electric field, the magnitude of the current is severely depressed. State-of-the-art inverted perovskite cells routinely demonstrate Jsc values exceeding 20.0 mA/cm². The extraction of less than 1.0 mA/cm² represents a >95% loss in charge-generation capacity. This massive suppression strongly indicates the bulk hydrolysis of the perovskite lattice. Due to the lack of hermetic encapsulation, 0.00E+00 1.00E-03 2.00E-03 3.00E-03 -0.5 0 0.5 1 1.5 Dark J-V Curves of Group A Device 1 J (A/cm^2) Device 3 J (A/cm^2) Device 5 J (A/cm^2) Device 7 J (A/cm^2) 6.00E-09 6.50E-09 7.00E-09 7.50E-09 -0.5 0 0.5 1 1.5 Dark J-V Curves of Group B Device 2 J (A/cm^2) Device 4 J (A/cm^2) Device 6 J (A/cm^2) Device 8 J (A/cm^2) 0.00052 7867 0.00080 6267 0.00089 9567 0.00072 61 0 0.0002 0.0004 0.0006 0.0008 0.001 1 Jsc of Group A Device 1 J (A/cm^2) Device 3 J (A/cm^2) Device 5 J (A/cm^2) Device 7 J (A/cm^2)Department of Electrical & Computer EngineeringFYP Proceedings6

ambient moisture permeated the device stack. Water acts as a powerful catalyst, hydrogen-bonding with the organic cations and causing the perovskite crystal to irreversibly decompose into insulating Lead Iodide (PbI2). Because PbI2 is a wide- bandgap material transparent to most visible light, the absorber layer was effectively blinded, choking off the photogeneration of electron-hole pairs. D. Interfacial Passivation Failure The dark J-V characteristics provided critical insight into the structural health of the NiOx/Perovskite interface. In a well- passivated diode, the reverse current should plateau at a negligible reverse saturation current. Fig. 4. Illuminated vs. Dark J-V curves of Device 3. However, as the applied reverse bias exceeded 0.5 V, the dark current in all Group A devices surged exponentially, reaching the milliampere range as shown in Fig. 4, 1.13 mA/cm² for D3 at 1.0 V. This rapid non-linear escalation signifies premature reverse breakdown, driven by a massive concentration of defect states. It empirically proves that the SAM layer failed to adequately passivate the NiOx interface. The resulting uncoordinated ions created deep-level energy states within the bandgap. Under the reverse electric field, charge carriers bypassed the junction barrier via trap-assisted tunneling. Consequently, these same defects would act as severe non-radiative recombination centers under forward bias, confirming that efficiency would have been fundamentally limited even without the polarity anomaly. E. Halide Migration and Contact Delamination While Group A suffered from chemical breakdown, Group B (D2, D4, D6, D8) exhibited catastrophic contact failure. TABLE I COMPARING THE SHUNT RESISTANCE (RSH) OF GROUP A VS. GROUP B NEAR 0 V. Device Group Device No. Shunt Resistance, Rsh (Ωcm^2) A 1 8.893 M 3 1.212 M 5 5.582 M 7 108.1 M B 2 742.9 M 4 8.803 G 6 1.62 G 8 3.156 G The near-zero dark currents translated to Shunt Resistances (Rsh) exceeding several Giga-ohms (>10^9 Ω·cm²) as shown in Table I. This infinite resistance indicates that the electrical circuit was physically severed. In unencapsulated mixed ionic-electronic conductors like perovskites, halide ions (I-) are highly mobile. Driven by environmental stress, these ions migrated through the electron transport layer to the top metal electrode. Upon reaching a metal contact, migrating iodine induces a galvanic corrosion reaction, forming insulating Metal Iodide. This not only replaces the conductive contact with a highly resistive barrier but also causes severe volumetric expansion, leading to the physical blistering and delamination of the electrode from the device stack. IV. CONCLUSION This research successfully executed a rigorous diagnostic failure analysis of inverted p-i-n perovskite solar cells. The electrical characterisation proved the profound instability of the unencapsulated architecture under ambient conditions. The empirical data conclusively identified three distinct modes of failure: moisture-induced bulk hydrolysis degrading the photo- absorber, SAM passivation failure causing trap-assisted tunneling, and halide ion migration resulting in Giga-ohm contact delamination. To advance the operational viability of perovskite technology and ensure accurate characterisation in future phases, the following engineering protocols are recommended: 1) Testing Protocol Correction: Future testing of inverted p-i- n architectures must explicitly enable the ‘Inverted Device’ parameter on the SMU or define the voltage sweep from -1.0 V to +1.0 V to accurately capture standard forward-bias metrics. 2) Mandatory Hermetic Encapsulation: The integration of UV-curable epoxy and glass coverslips immediately after fabrication is critical. Interfacial chemical engineering is insufficient on its own; physical barriers are required to block moisture ingress and contain internal ion volatility.Department of Electrical & Computer EngineeringFYP Proceedings7

3) Advanced Characterisation: Future failure analysis should incorporate X-ray Diffraction (XRD) to track PbI2 formation and Scanning Electron Microscopy (SEM) with EDX to visually verify electrode galvanic corrosion. REFERENCES [1] N. G. Park, “Perovskite solar cells: an emerging photovoltaic technology” Materials Today, vol. 18, no. 2, pp. 65-72, 2015. [2] Z. Hawash, I. G. Ono, and Y. Qi, “Moisture and Oxygen Degradation Mechanisms in Organometal Halide Perovskite Solar Cells” Advanced Materials Interfaces, vol. 5, no. 1, 2018. [3] Y. Kato et al., “Silver iodide formation in methyl ammonium lead iodide perovskite solar cells with silver top electrodes” Advanced Materials Interfaces, vol. 2, no. 13, 2015. [4] A. Al-Ashouri et al., “Conformal monolayer contacts with lossless interfaces for perovskite single junction and monolithic tandem solar cells” Energy & Environmental Science, vol. 12, pp. 3356-3369, 2019..) [5] K. O. Ogunniran and N. T. Martins, "Humidity and moisture degradation of perovskite material in solar cells: Effects on efficiency," IOP Conf. Ser.: Earth Environ. Sci., vol. 655, Art. no. 012049, 2021, doi: 10.1088/1755- 1315/655/1/012049. [6] J. Suo, H. Pettersson, and B. Yang, "Sustainable approaches to address lead toxicity in halide perovskite solar cells: A review of lead encapsulation and recycling solutions," EcoMat, vol. 7, Art. no. e12511, 2025, doi: 10.1002/eom2.12511. [7] P. T. Mokabane, V. T. Lukong, and T.-C. Jen, "A review of the effect of stability issues and wide-bandgap in the application of perovskite solar cells," Materials for Renewable and Sustainable Energy, vol. 14, art. no. 34, May 2025, doi: 10.1007/s40243-025-00307-9. [8] T. T. Ava, A. Al Mamun, S. Marsillac, and G. Namkoong, "A review: Thermal stability of methylammonium lead halide based perovskite solar cells," Appl. Sci., vol. 9, no. 1, Art. no. 188, Jan. 2019, doi: 10.3390/app9010188. [9] P. V. Nam, D. Q. Thuy, V. Q. Trung, and T. C. Thanh, "Perovskite solar cells: Literature review," NeuroQuantology, vol. 20, no. 15, pp. 717–730, Nov. 2022, doi: 10.14704/NQ.2022.20.15.NQ88069. [10]B. Rijal, K. M. Alzoubi, Z. Chen, A. L. Khammash, S. Lu, and T. P. Dhakal, “Understanding the Long-Term Instability in Perovskite Solar Cells: Mechanisms and Mitigation Strategies,” Electronics, vol. 14, no. 22, p. 4428, Nov. 2025.Department of Electrical & Computer EngineeringFYP Proceedings8

Visible Light Communication (VLC) for Real-Time Patient Monitoring in Hospital Wards Adlina Husna Abdul Rashid and Suriza Ahmad Zabidi Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—Patient monitoring systems in hospital wards continuously monitor vital signs, such as heart rate and SpO2. However, traditional wireless monitoring systems mainly adopt RF-based communication, which may suffer from performance degradation in dense indoor environments due to spectrum congestion. The present work investigates the feasibility of employing Visible Light Communication (VLC) as an alternative communication technology for teal-time patient monitoring. A prototype of VLC was developed using MAX30100 sensor for vital sign acquisition, an ESP microcontroller, a white light- emitting diode (LED) transmitter, On-Off Keying (OOK) modulation, a photodiode receiver and Arduino Uno for signal recovery and display. The system performance was evaluated in terms of signal-to-noise ratio (SNR), communication distance and sensitivity of line-of-sight (LOS) under different lighting conditions. The results indicate that VLC is a feasible short-range communication technique for indoor patient monitoring applications and provide the foundation for future advancements in healthcare communication systems. Index Terms—Healthcare Communication, Line-of-Sight (LOS), On-Off Keying (OOK), Patient Monitoring, Signal-to-Noise Ratio (SNR), Visible Light Communication (VLC) I. INTRODUCTION HE constant assessment of the physiological condition of patients and the ability to intervene in a timely manner has made real-time patient monitoring an integral part of modern hospital ward care. During the diagnosis, treatment, and recovery processes, monitoring vital signs such as heart rate and oxygen saturation (SpO2) provides healthcare professionals valuable information about the health status of patients. Early detection of abnormalities in these physiological parameters allows the medical staff to respond promptly to a possible clinical deterioration, thus improving patient safety and health outcomes [1]. Moreover, the continuous monitoring biosignals such as electrocardiograms (ECG), heart rate and SpO2 can afford more comprehensive assessment of patient conditions and enable proactive medical decision-making [2]. Efficient monitoring systems are demanded to keep the quality of patient in healthcare environments with increasing patient-to-medical-staffs ratios; hence, such capabilities have become increasingly important [3]. Due to these challenges, wireless patient monitoring systems are widely used in healthcare establishments because of their capability to promote patient monility and reduce the complexity associated with traditional wired monitoring systems. These systems allow the acquisition, transmission and management of physiological data without interruption and reduce the workload of health personnel by automating monitoring processes [4], [5]. Consequently, wireless communication technologies have become core elements of modern healthcare systems, especially for continuous monitoring purposes. Although they offer advantages, most of the current wireless patient monitoring systems depend on the radio frequency (RF) spectrum, which has several limitations in crowded indoor environments such as hospital wards. Many wireless devices, medical equipment, and electronic systems operate simultaneously in a limited space within healthcare facilities. The increasing dependence on RF-based communication can lead to spectrum congestion, which leads to an increased susceptibility to noise and communication interference. Thus, the SNR of the communication link may be degraded, which may cause a delay, corruption or unreliability in the transmission of the vital signs data. To overcome the above limitations, Visible Light Communication (VLC) is proposed as an alternative communication technology for indoor environments. This examines the feasibility of implementing a VLC-based patient monitoring system in hospital ward environments. The paper starts by reviewing existing wireless communication technologies for patient monitoring applications and discusses their limitations for indoor healthcare environments. Then, a prototype of a short-range VLC transmitter is built to transmit heart rate and SpO2 information by using optical transmitter and photodetector-based receiver. Finally, the performance of the proposed VLC system is evaluated based on signal-to-noise ratio (SNR) and line-of-sight (LOS) sensitivity analysis under different operating conditions to verify its suitability for real- time indoor patient monitoring applications. II. LITERATURE REVIEW A range of wireless communication technologies have been explored for real-time patient monitoring systems in hospital wards such as Wireless Fidelity (Wi-Fi), Wireless Body Area Networks (WBAN), fifth generation (5G) netwo

(IoT) to allow real-time transmission of measurements of bio signal vital signs [6]. Likewise, a monitoring system was developed using ESP8266 and Wi-Fi transmission with high measurement data accuracy [2]. However, Wi-Fi is still sensitive t RF spectrum congestion. WBAN have also been proposed for healthcare applications due to energy-efficient operation which it utilized Low Power Wide Area Network (LPWAN) technology to transfer physiological variables [9]. However, WBAN is still susceptible to interference from wireless devices. Low-latency and high throughput healthcare communication have become possible with 5G where 5G-enabled patient monitoring platform that can send pulse rate and biosensor data in real time was proposed [8]. Remote healthcare monitoring via long-range (LoRa) communication has attracted much attention because of long communication range was developed to transmit temperature, heart rate and SpO2 up to 3.6 km [7]. From [10], also showed the feasibility of using ESP32 for real-time monitoring physiological parameters using LoRa communication. LoRa provides excellent coverage but has a duty cycle limitation which are less suitable for frequent data transmission. BLE has also been studied where Optical Camera Communication (OCC) combined with BLE to transmit ECG signals via LED modulation method [11]. The proposed system enhances the reliability but has limited range of transmission of receiver angles. VLC on the other hand uses visible light spectrum, for wireless data transmission where the system proposed a Li-Fi telemedicine for bio signals transmission using high-speed LED switching [5]. An improved VLC patient monitoring system was established using OOK modulation and Non-Return-to- Zero Inverted (NRZ-I) encoding to send bio signals data for distances up to 2.5 m [12]. Similarly, there a proposed of Li-Fi infant monitoring platform for a radiation-free healthcare environment through optical wireless communication [13]. However, VLC systems are vulnerable to receiver misalignment, ambient light interference and signal blockage by obstacles in the communication path. III. METHOD A. System Architecture The proposed patient monitoring system consists of a VLC transmitter and receiver designed for short-range indoor healthcare applications. Physiological parameters comprising heart rate and SpO2 are measured using a MAX30100 pulse oximeter sensor connected to an ESP32 microcontroller as shown in Fig.1 The collected data are processed and formatted into a transmission packet and transmitted by a white light emitting diode (LED) using optical wireless communication. At the receiver, a photodiode is used to detect the transmitted optical signal and convert it into an electrical signal. Subsequently, the Arduino Uno demodulates the received data and visualizes data on an OLED display. Fig. 1. Proposed Development System of the VLC B. VLC Communication Protocol The proposed system employed OOK modulation for transmitting digital information over the visible light channel. The LED is switched ON for representing a binary logic ‘1’and OFF for representing a binary logic ‘0’. Fig. 2. Block Diagram of the Proposed VLC Transmitter Fig. 3. Block Diagram of the Proposed VLC Receiver As illustrated in Fig.2 and Fig.3 are the proposed system of a transmitter and receiver subsystem. At the transmitter side, the bio signal received from MAX30100 will go through an embedded Analog-to-Digital (ADC) and processed usingDepartment of Electrical & Computer EngineeringFYP Proceedings10

digital filtering stage to reduce measurement noise. At the receiver side, the transmitted optical signal is detected by photodiode and further sampled by the ADC of the Arduino Uno microcontroller. A threshold-based detection algorithm is implemented to distinguish between logic ‘1’ and logic ‘0’ according to the light intensity. C. Experimental Procedure To evaluate the proposed VLC system, the functional validation and communication performance analysis was performed. Fig. 4. Integrated VLC system Fig.4 shows the integrated VLC system as the experimental setup to evaluate the proposed communication system. Validation of its functionality was performed by comparing the physiological data measured at the transmitter with the values displayed at the receiver. The, the communication was evaluated through SNR and LOS sensitivity measurements. The SNR analysis was performed under different lighting conditions, namely dark-room, ambient-light and flashlight interference by means of oscilloscope voltage measurements. Experiments based on distance were carried out by varying the separation between the transmitter and the receiver, and sensitivity to LOS was studied by imposing the angular misalignment between the optical transmitter and the receiver. The obtained results were analyzed using packet success rate (PSR) and SNR metrics to determine the communication reliability of the proposed VLC system. D. Equations The SNR for each experiment was calculated using the standard equation of SNR. 𝑆𝑁𝑅 (𝑑𝐵) = 20 𝑙𝑜𝑔10 𝑉𝑠𝑖𝑔𝑛𝑎𝑙 𝑉𝑛𝑜𝑖𝑠𝑒 (1) The signal voltage and voltage noise were obtained from the oscilloscope. The higher SNR, the higher reliability of a communication system. IV. RESULTS AND DISCUSSION The results were calculated and evaluated after each of the experiment conducted. Fig. 5. SNR Calculated in Different Lighting Fig.5 demonstrates the SNR variation in three different lighting conditions. The highest SNR recorded was 44 dB obtained under dark room condition. This result shows that the minimal background illumination generates the least optical noise at the photodetector. Therefore, the SNR under dark room was set to be the reference of good SNR value. The SNR ambient-room lighting was reduced to 29 dB where the received data still can be decoded by the receiver even with additional optical noise in the communication channel. The lowest SNR value recorded was 2 dB under the light interference condition. The significant SNR reduction was due to the strong external light source directly illuminating the photodetector, creating a higher noise. The results show that with the increase of light interference, the SNR decreases owing to the increase of optical noise received by the photodetector. However, it was still able to provide reliable communication performance under typical indoor lighting conditions, which showed its suitability in hospital ward environments. Fig.6. SNR Calculated in Different Distances The measure SNR values are shown in Fig. 6 as a function of transmission distances 1 cm to 6 cm. The results show that the SNR decreases gradually with increase of the separation distance between the transmitter and the receiver. 0 10 20 30 40 50 Dark Room Ambient Room Flashlight SNR (dB) Light Condition Light Conditions vs SNR (dB) 0 10 20 30 40 50 1 2 3 4 5 6 SNR (dB) Distance (cm) SNR (dB) vs DistanceDepartment of Electrical & Computer EngineeringFYP Proceedings11

The trend is reasonable as the optical power received by the photodetector decreases with the increasing transmission distance. As the background noise remains relatively steady, the signal receiver is getting becomes weaker, thus reducing the SNR. TABLE I ANGLE ALIGNMENT SENSITIVITY AND SNR Angle (°) SNR (dB) Data reception 0 30.45 Detected 5 27.42 Detected 10 26.30 Detected 15 14.77 Detected 20 4.06 Detected 25 3.24 Corrupted The impact of the receiver angular alignment on the communication performance is summarized in Table 1. The highest SNR value of 30.45 dB was obtained at 0° which means that the LED transmitter and the receiver are perfectly aligned. As the angle of the transmitter increased between 5° to 20°, the SNR value was decreased but the data reception still detected the data sent. At 25°, the received data was corrupted and could no longer be decoded correctly. This was due to the photodetector was not getting enough optical power to reduce the overlap between the transmitter radiation pattern and receiver field of view. The results confirm that the proposed VLC system is highly dependent on the line-of-sight conditions. Therefore, the accurate alignment of the transmitter and receiver is required for acceptable communication performance in practical settings. V. CONCLUSION The findings overall indicate that VLC is a viable communication medium for short-range indoor healthcare environments. However, the current prototype remains limited by its dependence on direct optical paths and vulnerability to ambient light sources. Future works should aim at extending the communication range, improving the receiver sensitivity, and investigating hybrid VLC-RF architectures to improve the communication robustness in realistic hospital wards conditions . REFERENCES [1] M. Cecconi et al., “Unlocking opportunities to transform patient care: an expert insight on limitations and opportunities in patient monitoring,” Dec. 01, 2025, Springer Nature. doi: 10.1186/s40635-025- 00733-z. [2] A. I. Siam et al., “Portable and Real-Time IoT-Based Healthcare Monitoring System for Daily Medical Applications,” IEEE Trans. Comput. Soc. Syst., vol. 10, no. 4, pp. 1629–1641, Aug. 2023, doi: 10.1109/TCSS.2022.3207562. [3] A. K. Khanna, S. Ahuja, R. S. Weller, and T. N. Harwood, “Postoperative ward monitoring – Why and what now?,” Jun. 01, 2019, Bailliere Tindall Ltd. doi: 10.1016/j.bpa.2019.06.005. [4] C. A. da Costa, C. F. Pasluosta, B. Eskofier, D. B. da Silva, and R. da Rosa Righi, “Internet of Health Things: Toward intelligent vital signs monitoring in hospital wards,” Jul. 01, 2018, Elsevier B.V. doi: 10.1016/j.artmed.2018.05.005. [5] M. R. Ezilarasan, M. F. Leung, and X. Dai, “Design and Implementation of Telemedicine System Using Light Fidelity and PIC16F877A Microcontroller,” in 2024 14th International Conference on Information Science and Technology, ICIST 2024, Institute of Electrical and Electronics Engineers Inc., 2024, pp. 938–943. doi: 10.1109/ICIST63249.2024.10805278. [6] S. K., P. A. S., P. U., P. S., and S. P. V., “Patient health monitoring system using IoT,” Mater. Today Proc., vol. 80, pp. 2228–2231, Jan. 2023, doi: 10.1016/j.matpr.2021.06.188. [7] A. P. Murdan and J. S. Ramphul, “LoRa-based Smart Patient Monitoring System,” Journal of Telecommunication, Electronic and Computer Engineering (JTEC), vol. 15, no. 1, pp. 15–21, Mar. 2023, doi: 10.54554/jtec.2023.15.01.003. [8] H. H. R. Sherazi, W. Asif, and Z. A. Khan, “5G enabled Realtime Healthcare System for Heart Patients,” in 2021 IEEE International Smart Cities Conference, ISC2 2021, Institute of Electrical and Electronics Engineers Inc., Sep. 2021. doi: 10.1109/ISC253183.2021.9562935. [9] H. Taleb, A. Nasser, G. Andrieux, N. Charara, and E. M. Cruz, “Energy Consumption Improvement of a Healthcare Monitoring System: Application to LoRaWAN,” IEEE Sens. J., vol. 22, no. 7, pp. 7288– 7299, Apr. 2022, doi: 10.1109/JSEN.2022.3150716. [10] H. F. Jawad, A. Al-Askery, and A. H. Ali, “Design and Implementation of a Healthcare Monitoring System Based on LoRa,” 2022. [Online]. Available: http://journal.mtu.edu.iq [11] M. K. Hasan, M. Shahjalal, M. Z. Chowdhury, and Y. M. Jang, “Real-time healthcare data transmission for remote patient monitoring in patch-based hybrid OCC/BLE networks,” Sensors (Switzerland), vol. 19, no. 5, Mar. 2019, doi: 10.3390/s19051208. [12] T. Adiono et al., “Design of Centralized Oxygen Saturation and Heart Rate Patient Monitoring System using Li-Fi,” in Proceedings of the InternationalDepartment of Electrical & Computer EngineeringFYP Proceedings12

Conference on Electrical Engineering and Informatics, Institute of Electrical and Electronics Engineers Inc., 2021. doi: 10.1109/ICEEI52609.2021.9611106. [13] R. Mohan Kumar, R. Rahul, P. S. Sujith, T. Uvanesh, and R. Krishnakumar, “Li-Fi Based Infant Health Monitoring System,” in 2nd International Conference on Intelligent Cyber Physical Systems and Internet of Things, ICoICI 2024 - Proceedings, Institute of Electrical and Electronics Engineers Inc., 2024, pp. 622–626. doi: 10.1109/ICoICI62503.2024.10696539.Department of Electrical & Computer EngineeringFYP Proceedings13

Degradation Mechanism Analysis of an Unencapsulated Narrow Bandgap (NBG) Perovskite Solar Cell (PSC) Devices Izzuddin Hassanazhari and Nurul Fadzlin Hasbullah Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—This study investigates the terminal environmental degradation and failure modes of unencapsulated Cs0.1FA0.85MA0.05PbI2.95Br0.05 narrow bandgap perovskite solar cells. While pristine devices demonstrate exceptional optoelectronic potential, achieving a benchmark power conversion efficiency (PCE) of 23.04%, exposure to ambient environments (>70% relative humidity) induces rapid, catastrophic structural breakdown. Through comparative dark and illuminated current-density-voltage (J-V) diagnostics, two primary terminal failure mechanisms were identified. Severe Ohmic shunting and total open circuit behavior. Analysis reveals these failures are fundamentally driven by synergistic environmental catalysts. Ambient moisture triggers the thermodynamic phase segregation of the active layer into photo- inactive lead iodide (PbI2), heavily suppressing the short-circuit current (Jsc). Simultaneously, oxygen-catalysed halide migration leads to severe corrosion of the top silver electrodes, forming an impermeable silver iodide (AgI) dielectric barrier. These findings conclusively demonstrate that achieving true operational stability and market viability for this formulation strictly requires robust physical encapsulation and interfacial barrier engineering. Keywords—Perovskite Solar Cells, Environmental Degradation, Failure Analysis, Phase Segregation, Halide Migration, Current-Voltage Characterization. I. INTRODUCTION Advancing high-efficiency photovoltaic (PV) systems is a primary objective in contemporary materials science, aimed at optimizing solar energy capture. In recent years, organic- inorganic hybrid perovskite solar cells (PSCs) have emerged as a highly promising, low-cost alternative capable of rivalling commercial crystalline silicon. The specific narrow bandgap (NBG) formulation evaluated in this study, Cs0.1FA0.85MA0.05PbI2.95Br0.05 exhibits exceptional inherent optoelectronic properties. Utilizing an optimized, planar p-i-n (inverted) heterojunction architecture, pristine devices demonstrate a highly efficient baseline power conversion efficiency (PCE) exceeding 23%. However, operational device stability remains the most critical hurdle for commercialization. The swift penetration of ambient moisture and oxygen triggers severe environmental degradation, permanently disrupting internal charge transport networks and rapidly reducing power output. Understanding the specific physical mechanisms that govern this electrical collapse is essential. The primary objective of this study is to comprehensively evaluate these degradation mechanisms and precisely diagnose the terminal electrical failure modes of unencapsulated NBG perovskite solar cells following ambient exposure, utilizing rigorous current- density-voltage (J-V) failure diagnostics. II. METHODOLOGY A. Device Architecture and Degradation Conditions The devices under test utilized a planar p-i-n heterojunction architecture. The unencapsulated substrates were exposed to standard ambient laboratory conditions with relative humidity exceeding 70%. Due to the absence of physical encapsulation, the active perovskite layers and metal contacts were directly subjected to atmospheric moisture and oxygen, accelerating the natural degradation timeline leading to terminal electrical failure. B. Electrical Characterization Protocol Electrical characterization was conducted using an Automated Solar Cell Testing Kit. To separate the baseline structural integrity of the p-n junction from its photovoltaic generation capabilities, a dual-scan protocol was employed. Diagnostic dark J-V sweeps were utilized to evaluate reverse leakage currents and forward series resistance, effectively isolating internal structural defects. Subsequently, illuminated J-V sweeps were conducted under standard simulated AM1.5G (100 mW/cm²) irradiation to extract the core photovoltaic parameters, Open-Circuit Voltage (Voc), Short-Circuit Current (Jsc), Fill Factor (FF), and Power Conversion Efficiency (PCE). III. RESULT AND DISCUSSION C. Pristine Device Benchmarking To establish a comparative baseline, the ideal performance of the Cs0.1FA0.85MA0.05PbI2.95Br0.05 formulation was evaluated prior to severe environmental ingress. The pristine device exhibited excellent charge extraction under illumination and standard diode behaviour in the dark.Department of Electrical & Computer EngineeringFYP Proceedings14

TABLE I PRISTINE PHOTOVOLTAIC BENCHMARK Device State VOC (V) JSC (mA/cm²) FF(%) PCE(%) PRISTINE BENCHMARK 1.138 24.30 83.32 23.04 As detailed in Table I, the pristine device yielded a benchmark Voc of 1.138 V and a high Jsc of 24.30 mA/cm², resulting in a highly competitive PCE of 23.04%. This baseline confirms the exceptional inherent optophysics of the NBG formulation. B. Electrical Degradation and Failure Diagnostics Following ambient exposure, the unencapsulated substrates exhibited severe, non-uniform degradation. Rather than a singular point of failure, comparative dark and illuminated J-V characterization revealed different stages of structural collapse across the samples. Fig. 1. Standardized J-V characteristics of Sample 1 demonstrating early-stage degradation and significant current loss. As observed in Fig.1, the champion pixel of Sample 1 represents an intermediate degradation state. While the bulk active layer maintained enough structural integrity to generate a measurable photovoltaic response, a direct performance comparison with the pristine device reveals a massive drop in overall efficiency. Specifically, while the pristine baseline achieved a Voc of 1.138 V, Device 1 of Sample 1 experienced a moderate reduction to roughly 0.839 V. This indicates that the core crystal lattice retained enough of its p-n junction characteristics to generate a weakened built-in potential. However, the charge extraction capabilities were critically compromised. The Jsc plummeted drastically from the pristine benchmark of 24.30 mA/cm² down to just 0.072 mA/cm². This >99% catastrophic loss in current density, coupled with a heavily degraded Fill Factor (dropping from 83.32% down to 46.51%), choked the overall PCE to a fractional 0.034%. The corresponding dark diode turn-on curve corroborates this decline, demonstrating that localized structural defects have already begun to form at the interfaces, actively suppressing ideal charge transport. As environmental degradation progressed to a terminal state on other substrates, these suppressed electrical pathways were completely severed. Fig. 2. Standardized J-V characteristics of Sample 4 exhibiting terminal Open Circuit failure due to massive interfacial resistance. As shown in Fig. 2, the terminal failure state of Sample 4 is characterized by immense series resistance. The dark sweep reveals an exceptionally low reverse leakage current, proving the core p-n junction remained physically intact without macroscopic shunting pinholes. Under illumination, the device retained a measurable Voc of 0.885 V, indicating the bulk crystal could still separate photogenerated charges. However, the Jsc experienced a near-total collapse, falling even further than Sample 1 to a mere 0.006 mA/cm². This curve geometry confirms that the unencapsulated interfacial boundaries severely oxidized, creating a massive Series Resistance (Rs) that completely throttles current output before the device fails into a total open circuit. Other localized regions across the substrates like Samples 2 and 3 exhibited alternate terminal failures, such as severe Ohmic shunting. In these instances, structural voids allowed the transport layers to make physical contact, dropping shunt resistance (Rsh) to near zero and bypassing the junction entirely. C. Physical Origins of Degradation The terminal electrical failures diagnosed above are driven by two synergistic physical phenomena highly dependent on the perovskite composition and ambient environment: 1. Phase Segregation: Ambient moisture permeates the film, hydrating the organic cations and forcing the crystalline lattice to undergo a thermodynamic transition into a photo-inactive phase. This subsequently decomposes into electrically insulating lead iodide (PbI2), neutralizing the device's generation capabilities. 2. Halide Migration: Atmospheric oxygen and moisture act synergistically to mobilize unbound iodide ions. These ions migrate rapidly toward the top contact and critically corrode the silver electrodes, forming an impenetrable, wide-bandgapDepartment of Electrical & Computer EngineeringFYP Proceedings15

silver iodide (AgI) dielectric barrier that perfectly correlates with the massive series resistance observed in the Sample 4 diagnostics. C. Terminal Failure Modes in Non-Functional Devices While the champion devices from Samples 1 and 4 exhibited intermediate degradation with partially surviving semiconductor junctions, the vast majority of the unencapsulated pixels deteriorated into completely non- functional states. Comparative J-V characterization of these fully degraded pixels successfully identified two distinct terminal electrical failure mechanisms which are Ohmic Shunt and Total Open Circuit. Fig. 3. Dark J-V characteristics illustrating the two primary terminal breakdown states: Ohmic shunting and total open- circuit failure. Under unilluminated conditions, as shown in Fig. 3, a functional solar cell will actively block reverse charge flow. However, the non-functional devices exhibited two extreme, opposing deviations from ideal diode behaviour. The first terminal state is the Ohmic Shunt. Devices exhibiting this failure mode demonstrated a massive and immediate reverse leakage current. In terms of physical degradation, this indicates that the perovskite absorber layer has severely decomposed, creating macroscopic structural voids or pinholes. These voids allow the top and bottom charge transport layers or the metal electrodes themselves to make direct physical contact. This catastrophic breakdown causes the Shunt Resistance (Rsh) to collapse to near zero, entirely bypassing the p-n junction and transforming the pixel into a passive resistor. The second terminal state is the Total Open Circuit. In direct contrast to the shunted devices, these pixels exhibited a flat, near-zero current profile across the entire dark bias sweep. While the bulk active layer may not have developed shunting pinholes, the unencapsulated interfaces between the perovskite and the transport layers have heavily oxidized. This creates extreme Series Resistance (Rs) that acts as an insulating dielectric barrier, completely severing the internal electrical pathways. Fig. 4. Illuminated J-V characteristics (AM1.5G) of terminal devices exhibiting a 100% loss of power conversion efficiency. The structural devastation identified in the dark sweeps directly correlates to a total collapse of photovoltaic generation under simulated AM1.5G illumination. As illustrated in Fig. 4, both failure modes result in a 100% loss of Power Conversion Efficiency (PCE), but they fail for entirely different electrical reasons. For the Ohmic Shunted devices, the Open-Circuit Voltage (Voc) drops completely to 0 V. Because the shunted junction cannot maintain a built-in electric field to separate photogenerated charges, any generated current instantly leaks back through the physical voids. The illuminated J-V curve forms a purely linear slope passing directly through the origin, confirming the device is merely consuming applied power rather than generating it. Conversely, the Total Open Circuit devices completely throttle the charge extraction, causing the Short- Circuit Current (Jsc) to flatline at 0 mA/cm². Even if the surviving internal perovskite crystal successfully absorbs photons and generates electron-hole pairs, the massive interfacial resistance blocks the charge carriers from reaching the electrodes, trapping them within the active layer and rendering the solar cell completely non-functional. D. Mitigation Strategies and Engineering Countermeasures The terminal failure mechanisms identified in this study highlight the extreme vulnerability of the unencapsulated planar p-i-n architecture to ambient atmospheric ingress. However, because the pristine baseline efficiency (23.04%) proves the exceptional inherent optoelectronic capabilities of the NBG perovskite formulation, resolving these vulnerabilities is fundamentally a packaging and device architecture challenge. To prevent the rapid degradation modes observed specifically moisture-induced PbI2 phase segregation and oxygen-catalysed AgI electrode corrosion, two synergistic mitigation strategies are strictly required for future market viability. First, the integration of robust, hermetic physical encapsulation is mandatory. Utilizing chemically inert, low-permeability encapsulants such as specialized ethylene-vinyl acetate (EVA), polyolefins, or glass-glass bounding with robust edge seals will physically block the penetration of ambient moisture and oxygen. This isolation directly neutralizes theDepartment of Electrical & Computer EngineeringFYP Proceedings16

primary environmental catalysts responsible for macroscopic structural voiding and Ohmic shunting. Second, advanced interfacial barrier engineering must be implemented to suppress internal halide mobilization. The catastrophic open-circuit failures observed in Sample 4 were directly linked to the high reactivity of the top silver contacts. Replacing these reactive metals with corrosion-resistant alternatives, such as carbon-based electrodes, or inserting dense, stable inorganic buffer layers between the perovskite and the transport layers, can effectively halt halide migration. Together, rigorous physical encapsulation and interfacial barrier engineering will protect the delicate internal crystal lattice, bridging the gap between this material's immense theoretical potential and its long-term, practical operational deployment. IV. CONCLUSION This study successfully conducted a rigorous diagnostic failure analysis on the environmental degradation mechanisms of unencapsulated Cs0.1FA0.85MA0.05PbI2.95Br0.05 narrow bandgap perovskite solar cells. Through baseline benchmarking, the pristine planar p-i-n architecture demonstrated exceptional photovoltaic potential, yielding a highly competitive power conversion efficiency of 23.04%. However, comparative dark and illuminated J-V characterization unequivocally confirmed that this theoretical potential is bottlenecked by extreme environmental sensitivity. Exposure to ambient atmospheric conditions (>70% relative humidity) induced rapid, catastrophic structural breakdown, culminating in two distinct terminal electrical failure modes, severe Ohmic shunting and total open-circuit behaviour. The diagnostic evaluation linked these failures to synergistic environmental catalysts. Ambient moisture triggered the thermodynamic phase segregation of the active layer into photo-inactive lead iodide (PbI2), heavily suppressing charge generation and causing macroscopic structural voiding. Concurrently, atmospheric oxygen and moisture mobilized unbound iodide ions, which rapidly migrated to corrode the top silver electrodes, forming an impermeable silver iodide (AgI) dielectric barrier that choked charge extraction. Ultimately, while the inherent optoelectronic properties of this formulation rival commercial silicon, achieving operational stability strictly requires the physical stack to be completely isolated from atmospheric elements. Future Work: To translate this material's immense theoretical potential into a commercially viable technology, future research must pivot from unencapsulated baseline testing to rigorous barrier engineering. Subsequent studies will focus on fabricating identical NBG devices integrated with robust physical encapsulation (such as EVA or polyolefin glass-glass lamination) and evaluating their long-term operational stability under standardized damp-heat testing protocols (e.g., ISOS-D-3). Additionally, future iterations will investigate the integration of chemical inert buffer layers, such as atomic layer deposited (ALD) metal oxides, and explore corrosion- resistant carbon-based electrodes to permanently suppress internal halide migration and ensure long-term interfacial stability. Acknowledgement The authors would like to express their sincere gratitude to the International Islamic University Malaysia (IIUM), Universiti Kuala Lumpur (UniKL), and Ming Chi University of Technology (MCUT), Taiwan. Their collaborative framework and technical assistance were instrumental in the successful execution of this failure analysis study. Conflict of Interest The author declares no conflict of interest. REFERENCES [1] Y. Rong et al., "Challenges for commercializing perovskite solar cells," Science, vol. 361, no. 6408, p. eaat8235, Sep. 2018. [2] R. Cheacharoen et al., "Design and understanding of encapsulated perovskite solar cells to withstand temperature cycling," Energy Environ. Sci., vol. 11, no. 1, pp. 144-150, 2018. [3] H. J. Snaith et al., "Anomalous hysteresis in perovskite solar cells," J. Phys. Chem. Lett., vol. 5, no. 9, pp. 1511-1515, May 2014. [4] K. A. Bush et al., "23.6%-efficient monolithic perovskite/silicon tandem solar cells with improved stability," Nat. Energy, vol. 2, p. 17009, Feb. 2017. [5] M. A. Green et al., "Solar cell efficiency tables (Version 63)," Prog. Photovolt. Res. Appl., vol. 32, no. 1, pp. 3-13, Jan. 2024. [6] D.-H. Kang and N.-G. Park, "On the current-voltage hysteresis in perovskite solar cells: Dependence on perovskite composition and methods to remove hysteresis," Adv. Mater., vol. 31, no. 34, p. 1805214, Aug. 2019. [7] T. A. Berhe et al., "Organometal halide perovskite solar cells: degradation and stability," Energy Environ. Sci., vol. 9, no. 2, pp. 323-356, 2016. [8] C. Besleaga et al., "Iodine migration and degradation of perovskite solar cells enhanced by metallic electrodes," J. Phys. Chem. Lett., vol. 7, no. 24, pp. 5168-5175, 2016. [9] C. C. Boyd et al., "Barrier design to prevent metal-induced degradation and improv

Bi-directional EDFA Pumping Scheme for FSO Receiver Link in 5G Fronthaul System Mohamad Zarif Hakim Mohamad Zamri and Mohd Shahnan Zainal Abidin DDepartment of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract-The growth of fifth generation (5G) mobile networks increases the need for fronthaul links that can provide high capacity, low latency, and flexible deployment. Free Space Optical (FSO) communication is a suitable candidate because it can transmit optical signals through the atmosphere without installing physical fibre. However, the received FSO signal can be weakened by fog, rain, haze, turbulence, beam divergence, and pointing misalignment. This article presents a receiver- side bi-directional Erbium-Doped Fibre Amplifier (EDFA) pumping scheme for an FSO receiver link in a 5G fronthaul system. The setup uses a 1550 nm input signal, a 1 m erbium-doped fibre, and two 980 nm pump lasers injected from opposite ends of the fibre. Forward pumping, backward pumping, and bi-directional pumping are compared using output power, gain, optical signal-to-noise ratio (OSNR), noise figure, and residual pump power. The current report indicates that bi-directional pumping gives the best overall performance among the tested configurations, but the gain remains lower than the expected EDFA performance. The main limitations are related to short EDF length, incomplete pump absorption, residual pump interaction, amplified spontaneous emission noise, and component insertion loss. The final article will be updated when the remaining measurements are fully verified. Index Terms--5G fronthaul, bi-directional pumping, EDFA, Free Space Optical, optical amplifier, OSNR, residual pump power. I. INTRODUCTION The fronthaul link in a 5G network connects the baseband processing unit to remote radio equipment. This part of the network must support high data capacity, low latency, and stable transmission quality. Optical fibre is commonly used for this purpose, but fibre installation can be expensive or difficult in temporary, urban, and hard-to- access locations. FSO communication is considered in this project because it can carry optical signals through free space. The main weakness is that the signal travels through the atmosphere before reaching the receiver. Weather conditions, beam spreading, and alignment error can reduce the received optical power and degrade receiver performance. For this reason, the receiver requires an amplification method that can strengthen the weak optical signal before it is analysed or detected. This study focuses on a receiver-side EDFA using a 1550 nm signal and two 980 nm pump lasers. The project compares forward, backward, and bi-directional pumping. The objective is to identify whether pumping the EDF from both sides can improve output power, gain, OSNR, and residual pump behaviour compared with single-direction pumping. II. RELATED WORK AND RESEARCH GAP Previous studies generally focus on three areas. The first area is the use of FSO for 5G and beyond-5G fronthaul, where FSO is discussed as a high-capacity wireless optical link that can reduce the need for physical fibre installation [1], [2]. The second area is EDFA-assisted FSO, where optical amplification is used to compensate for power loss caused by the atmospheric path [3], [4]. The third area is EDFA pumping optimization. Forward pumping is simple because the pump and signal travel in the same direction, while backward pumping injects pump power from the output side. Bi-directional pumping combines both directions to produce a more balanced population inversion along the EDF. Literature also discusses gain flattening, noise figure reduction, and dual- stage EDFA structures for improving amplifier performance [5]-[7]. Although the literature shows that EDFA can improve FSO links, many studies are simulation-based or focus on system-level performance. Less attention is given to practical receiver-side issues such as short EDF length, residual pump power, insertion loss, and pump interaction in a laboratory bi-directional 980 nm pumping setup. These issues form the main research gap addressed in this project. III. METHODOLOGY A. Experimental Setup The experimental setup follows the receiver-side EDFA arrangement in the final report. The 1550 nm signal from the tunable laser source passes through an optical isolator, WDM 1, a 1 m EDF, WDM 2, a tunable bandpass filter, and finally the optical spectrum analyzer. Pump 1 is connected through WDM 1, while Pump 2 is connected through WDM 2. This arrangement allows the EDF to be pumped from both directions. Fig. 1. Setup for bi-directional EDFA pumping scheme from the report. B. Experimental Flow The work begins with literature review, problem identification, objective definition, and equipment preparation. Both 980 nm pump lasers are then characterized before testing forward, backward, and bi- directional EDFA configurations. The measur

Fig. 2. Research flowchart used in the project report. C. Pumping Configurations and Parameters In the forward pumping test, only Pump 1 is activated and the pump travels in the same direction as the 1550 nm signal. In the backward pumping test, only Pump 2 is activated and the pump travels in the opposite direction. In the bi-directional pumping test, both pumps are activated at the same time. The gain is calculated by subtracting the input optical power from the output optical power, while OSNR is obtained from the OSA spectrum. The gain equation used in the report is G(dB) = Pout(dBm) - Pin(dBm). The conversion from dBm to milliwatts is P(mW) = 10^(P(dBm)/10). These equations are used to compare the measured optical power in a consistent form. IV. RESULT STATUS AND DISCUSSION PLAN The final result section can be updated after the latest measurements are confirmed. At the current stage, the report already provides a comparison of output power, OSNR, and gain for the three pumping configurations. These figures are included only from the report, as requested. Fig. 3. Output power comparison for forward, backward, and bi-directional pumping. Fig. 4. OSNR comparison for the three pumping configurations. Fig. 5. Gain comparison for the three pumping configurations. A. Current Interpretation Based on the current report, the bi-directional pumping configuration gives the highest output power and the only positive gain among the three configurations. This shows that injecting pump power from both sides of the EDF can improve the amplification of the 1550 nm signal compared with using only one pump direction. However, the gain is still lower than the typical expected EDFA performance. The most likely reason is the 1 m EDF length, which may not provide enough interaction distance for the 980 nm pump power to be fully absorbed by the erbium ions. If a large amount of pump power remains unused, the population inversion becomes limited and the signal amplification is reduced. Residual pump power is also important in the discussion. If unabsorbed pump power remains in the system, it may affect the pump distribution in the bi-directional setup. This condition can be explained as residual pump interaction. It may disturb the uniformity of the erbium excitation and reduce the possible gain improvement. Fig. 6. Residual pump power measured in single-direction pumping.Department of Electrical & Computer EngineeringFYP Proceedings19

V. RECOMMENDED IMPROVEMENTS The first improvement is to test longer EDF lengths such as 2 m and 3 m. A longer EDF may allow more pump absorption and stronger population inversion. However, the EDF should not be too long because excessive length can increase reabsorption and ASE noise. Therefore, the best length should be determined experimentally. The second improvement is to optimize the pump current combination. Instead of using only one fixed setting, Pump 1 and Pump 2 should be varied in several combinations. This can show whether the best performance occurs when both pump powers are balanced or when one pump direction is stronger. The third improvement is to measure residual pump power for the bi-directional configuration. This is necessary because the current residual pump discussion is mainly based on single-direction pumping. Measuring residual pump power from both sides will make the analysis stronger and help explain the pump interaction inside the EDF. The final improvement is to reduce component loss. The connectors should be cleaned, WDM and isolator insertion losses should be checked, and the tunable bandpass filter loss should be considered. Reducing these losses can improve the measured output power and gain. VI. CONCLUSION This article draft presents a bi-directional 980 nm EDFA pumping scheme for an FSO receiver link in a 5G fronthaul system. The proposed receiver uses a 1550 nm signal and two 980 nm pump lasers injected from opposite ends of a 1 m EDF. The draft article is formatted in two columns, including the abstract, and uses only figures taken from the project report. The current report indicates that bi-directional pumping is the most effective configuration among the tested schemes because it gives the best output power and the only positive gain. Nevertheless, the gain remains limited. Future work should focus on longer EDF length, optimized pump current, careful residual pump measurement, and reduced optical component loss. These improvements will strengthen the final FYP result and journal discussion. REFERENCES [1] I. A. Alimi and P. P. Monteiro, "Revolutionizing free-space optics: A survey of enabling technologies, challenges, trends, and prospects of beyond 5G free-space optical communication systems," Sensors, vol. 24, no. 24, 2024. [2] M. Kamal et al., "Free Space Optics transmission performance enhancement for sustaining 5G high capacity data services," Micromachines, vol. 13, no. 8, 2022. [3] S. T. Hayle et al., "High-speed FSO-5G wireless communication system with enhanced loss compensation using high-power EDFA," Scientific Reports, vol. 15, no. 1, 2025. [4] A. Abisayo, "Optically amplified free-space optical communication systems," Ph.D. dissertation, University of Nottingham, 2013. [5] F. El-Nahal, T. Xu, D. AlQahtani, and M. Leeson, "A bidirectional WDM-PON free space optical system for fronthaul 5G C-RAN networks," IEEE Photonics Journal, vol. 15, no. 1, pp. 1-10, 2023. [6] S. Sharma and A. Kumar, "Gain flattening and noise figure optimization in bi-directional EDFA for WDM systems," Optik, vol. 272, 2023. [7] H. Lao et al., "Dual-stage double pass extended L-band erbium-doped fiber amplifier with improved gain performance," Photonics, vol. 10, no. 11, 2023. [8] R. Paschotta, "Tutorial: Fiber amplifiers, Part 5: Forward and backward pumping," RP Photonics Encyclopedia.Department of Electrical & Computer EngineeringFYP Proceedings20

Artificial Intelligence and Machine Learning T R A C K 2

Machine Learning Framework for Text Simplification to Support Dyslexic Readers Muhammad Irfan Yushamidi and Huda Adibah Mohd. Ramli Department of Electrical and Electronic Engineering, Kulliyyah of Engineering , International Islamic University Malaysia Abstract— This paper presents Simplifly a Django and Python based web application designed for simplifying academic texts for readers who have difficulty reading due to dyslexia. Many dyslexic individuals experience challenges when attempting to read academic material that uses complicated vocabulary, long sentences, complicated structure and high levels of abstraction. To overcome these reading challenges, Simplifly utilizes an integrated hybrid system that performs lexical simplification, syntactic simplification, semantic safety validation and optional T5-based neural rewriting. In this integrated hybrid system, Simplifly detects difficult words to read and complex sentence structure in academic texts, generates possible simplified forms for each word and sentence, then evaluates the quality of its generated simplifications using a number of criteria including: readability, grammar, semantic similarity, percentage copied from the original text, and whether the term used can be classified as a protected term. In addition, a dyslexia-friendly user environment is created for users by offering several features including adjustable font size, adjustable line spacing, highlighting of difficult words, indicating the readability of the text and providing side by side display of original and simplified text. An evaluation of Simplifly was conducted using three different datasets (development, held-out and stress datasets) with SARI scores above 62 for each of the three datasets, with no invalid simplification outputs, no word fragments, and no grammatical or protected term errors in the simplification outputs produced by Simplifly, as well as an external comparison to determine if Simplifly provided better quality of readability and simplification output compared to ChatGPT, QuillBot and Rewordify at the evaluated reading difficulty levels. Key Words — Dyslexia; Machine Learning; Natural Language Processing; Readability; Semantic Validation; Text Simplification. I. INTRODUCTION Dyslexia is a neuro-developmental disorder. People who have dyslexia have processing problems with written language. Dyslexia is not caused by low intelligence or a lack of motivation, but it can make reading slower, harder, and mentally more exhausting due to problems with decoding, fluency, visual processing, and word recognition [1],[2]. For students with dyslexia, academic texts will often be even more difficult to read because they will typically have long sentences, advanced vocabulary, abstract concepts, and dense visual layouts. Text Simplification (TS) is an approach that attempts to provide greater access to written text. TS is an attempt to modify or change complex text to create a clearer version while keeping the same meaning [3]. TS can involve the use of lexical simplification (i.e., using simpler words as substitutes), syntactic simplification (i.e., dividing long or complex sentences; restructuring sentences), or semantic simplification (i.e., eliminating any unnecessary complexity without removing critical information). Advancements in NLP and ML have made it easier for us to automate the process of simplifying text by using neural methods or transformers to generate more accurate results; however, there are still a number of shortcomings that need to be addressed within current literature. Some systems only focus on substituting one word with another whereas there are others that produce grammatically correct and fluent output but have altered the meaning from the original work, failed to preserve key technical vocabulary, or created text that would still be difficult to read for someone with dyslexia. As such, when creating a simplified dyslexia-friendly system for literature, it is essential for a balance among readability, grammar, semantic preservation, and safety. This research proposes a new hybrid machine-learning based text simplification framework called Simplifly for English academic literature. Simplifly encompasses rule-based lexical and syntactic simplifications, optional T5 based rewriting, precise semantic validation, checking for protected terms, and creating an user- friendly online interface for individuals with dyslexia. The rest of the paper is divided into the following sections: Section II discusses other similar projects; Section III describes the methodology used to create the system; Section IV discusses results and provides an analysis; and Section V summarizes the findings and suggests future improvements. II. RELATED WORK Until recently, simplified text was produced using various approaches, which included rule-based, lexical-based, syntactic- based, statistical-based, and neural-based approaches. In addition to the above, hybrid personalizatio

ratios, grammar, completeness and protected term verification before being presented to users is warranted. III. METHODOLOGY A. System Architecture Simplifly is a web app implemented using Django and Python. The Simplifly application has an architecture with 4 major Layers: The Presentation Layer (provides User Interface); The Application Layer (handles the views and requests of Django); The Processing Layer (where all forms of Simplification and Validation occur); and A Data Layer (stores the Synonym entries, Simplification History and the Evaluations as well as Evaluations in SQLite). The Processing Pipeline of Simplifly is displayed in Fig. 1. Fig. 1. Simplifly hybrid text simplification pipeline. B. Simplification Pipeline Preprocessing and analysis of the complexity of input text is where the pipeline begins. To do this, the input text is cleaned of anything irrelevant to simplification and is checked for difficult words, sentence lengths, grade levels, and readability. It may return the initial input with a readability score if the input text is already simplified. However, if the input text contains significant complexity levels, the Simplifly system performs the simplification using Lexical, Syntactic, and Semantic operations. Lexical simplification is performed by recognizing difficult terms & phrases in the input text and replacing them with easier to understand alternatives found in an intuitive and complete user- generated database of synonyms that is domain-aware. Therefore, when substituting do a phrase or term will occur first so as to not disrupt a multi-word term i.e.. digital literacy would be simplified as digital skills and than essential would be simplified to needed. In addition, the details of each synonym would typically have information such as domain, part of speech, difficulty level, status of approval by the end-user, priority of that entry & cite/reference. Syntactic simplification looks to simplify the structure of long complicated sentences in the input text. All sentences that contain a connector (such as although, however or but) have the independent thoughts separated into multiple smaller concise sentences while retaining clarity of the output and grammatical structure. The reasoning behind this functionality is two-fold: First, to reduce the length of the original input document and second, to minimise the cognitive load associated with reading a long multi-clause, academic-style sentence for readers who may be working to understand their content as well. A finely-tuned T5 (Text-to-Text Transfer Transformer) is an optional "Artificial Intelligence" (AI) layer in the system. It attempts to load the most recent version of this model (fine-tuned v3) first, then fallbacks to previous models as necessary. T5 is not automatically applied to every input by default; it is called to perform additional rewriting after attempting to simplify using rules alone. This helps to prevent excessive generations, allows for more efficient processing and increases the likelihood of safe selections when generating output. C. Validation and Evaluation The candidate outputs are validated prior to final selection. The validation layer evaluates the candidate outputs for improvements to readability, grammatical correctness, sentence completeness, semantic similarity, copy ratio, and preservation of protected terms. Protected terms are domains of significant relevance or importance that cannot be incorrectly modified, including but not limited to carbon dioxide, radiation, inflation, diabetes, mutations, cell division and cardiovascular diseases. The system was evaluated using four different datasets; the development & held out were 35 examples each, whereas the stress dataset was composed of 26 examples that were longer and more complex than the examples in the previous datasets. The metrics that were measured included SARI, FRE, FKGL, BERTScore F1, copy ratio, invalid output count, fragment count, grammar flag, and the number of protected term violations. TABLE I IMPLEMENTED SIMPLIFLY MODULES Module Implementation Purpose Web/UI Django templates, HTML, CSS, JS Dyslexia-friendly interaction and result display Lexical Phrase and word replacement rules Replace difficult terms with simpler alternatives Syntactic Sentence splitting and connector cleanup Reduce long academic sentence complexity AI Fine-tuned T5 with fallback models Generate additional candidates when needed Validation Semantic, grammar, copy- ratio and term checks Reject unsafe or weak simplified outputs Data SQLite history and synonym records Store knowledge, logs and evaluation records IV. RESULTS AND DISCUSSION A. Generalization and Safety Results An analysis of the complete V3 application pipeline was completed on Development (held-out) datasets, a separate Development (stress) set, and Development (main) datasets. Table II summarizes the study results and shows that all six datasets received overall SARI scores over 62. The Stress

Stress 26 64.0734 40.6404 9.5319 0.7085 All three datasets had no invalid outputs, fragmentation, or grammar flags - no protected terms were in violation. As a text simplification system for dyslexics, it should provide output that not only reduces length or makes sentences easier to read, but that does so without giving rise to incomplete output (through dropping elements) or incorrectly replacing with an alternative meaning (through substituting meaning for value). The copy ratio reduced from the base set to the stress set, which suggests that the system's ability to create more complex output improved with the increased usage of input to create the output. B. External Comparison Across Difficulty Levels A comparison was made between Simplifly and three other systems (ChatGPT, QuillBot, and Rewordify) across three levels of difficulty (easy, moderate, and hard) in the academic text category. The results of this comparison are presented in Table III, with Simplifly achieving the highest SARI score at all three levels and the highest FRE score and the lowest grade level for all three tested cases. The exact SARI scores for all three levels are as follows: 64.35 for easy text, 61.31 for moderate text and 67.51 for hard text. TABLE III EXTERNAL COMPARISON SUMMARY Diff. SARI Best other FRE FKGL BERT Easy 64.35 55.00 51.26 7.93 0.948 Moderat e 61.31 54.82 43.88 9.26 0.952 Hard 67.51 60.40 68.88 5.75 0.964 The hardest category of the readability scores where Simplifly obtained a FRE score of 68.88 and an FKGL score of 5.75 was the hardest to achieve an overall score for; therefore this hybrid pipeline is recommended especially where complex academic sentences have been able to be rewritten into much shorter and clearer sentences. Although ChatGPT had some cases that yielded slightly higher F1 BERTSCORE values in relation to the other two technologies, it produced an even higher F1 BERTSCORE on average to that of both other technologies for the remainder of the texts evaluated. C. Qualitative Analysis Many examples illustrate how Simplifly can be used to replace complicated expressions with simpler ones to separate very lengthy sentences into much shorter ones while still providing meaningful information. An example of this is when the system changed "digital literacy" into "digital skills" and shortened a very long sentence about online learning into 3 shorter sentences. In additional scientific examples, the meaning of critical terms, including genetic mutations and cell division, as well as radiation and toxic substances, were preserved. Therefore, it demonstrates that the protected term mechanism can assist in reducing loss of meaning and support continued reduction in meaning. TABLE IV QUALITATIVE SIMPLIFICATION EXAMPLES Original text Simplified output Sleep deprivation can negatively affect concentration and emotional stability. Not getting enough sleep can make it hard to focus. It can also affect mood. Digital literacy plays an important role in effective technology use. Digital skills help with using technology well. Genetic mutations may occur during cell division or result from exposure to radiation and harmful chemicals. Genetic mutations can happen naturally when cells divide. They can also happen after exposure to radiation or harmful chemicals. The overall results support this project's goal of establishing a practical academic simplification tool designed specifically for people with dyslexia. The rule-based layers provide the stable basis of the tool, the T5 module allows for additional flexibility when changing original text, and the validation layer ensures that outputs created by the tool are not weak before they are used. The primary limitation of this project at this time is that the evaluation of the system is primarily performed to date through automated metrics. Additional evaluations of the tool's effect on actual reading comfort, comprehension, and usefulness must be conducted using human subjects who have dyslexia. V. CONCLUSION AND FUTURE WORK In this paper, we presented a hybrid academic text simplification framework (named Simplifly) that was designed specifically for the needs of dyslexic readers by integrating multiple modalities of lexical, syntactic and semantic simplification; T5-based optional rewriting; synonym management from a database; and a dyslexia-friendly web interface. The evaluation of the v3 application pipeline demonstrated that it achieved consistently stable SARI (Simplified Academic Readability Index) scores above 62 across approximately 100 academic clarity measures from three datasets: development, hold-out, and stress. Additionally, it produced zero invalid outputs, no fragments, no grammatical flag- related issues, and no protected-term violations; this demonstrates that the validation layer (i.e., the semantic validity check) was able to maintain safety and meaning. Future work should include expanding the custom simplification dataset to contain approximat

Detecting Phishing Websites with Transformer- Based NLP and URL Feature Fusion Mohammad Amir Imtiyaz Mohd Annuar and Othman Omran Khalifa Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract— Phishing websites remain a major cyber-security threat because new domains, obfuscated URLs, and generative artificial intelligence lures can appear faster than blacklist-based defenses can be updated. This article presents a real-time phishing detection framework that fuses structural URL evidence with semantic webpage metadata. The proposed system uses a one- dimensional Temporal Convolutional Network to model raw URL character sequences and a fine-tuned DistilBERT encoder to represent domain names, page titles, and meta-descriptions. Both streams are integrated through an intermediate cross-modal attention layer and hardened using adversarial samples produced by an Adversarial Autoencoder. A balanced corpus of 100,000 domains was assembled from PhishTank, Kaggle, Cisco Umbrella, and Alexa-based sources, with a separate out-of-distribution evaluation set. The trained model achieved 92.45% accuracy, 91.80% precision, 93.12% recall, 92.46% F1-score, and 1.000 ROC-AUC on the evaluated out-of-distribution setting. The deployed FastAPI and Chrome Manifest V3 prototype completed the main inference pipeline in 44.63 ms, showing suitability for real-time endpoint warning applications. Index Terms—Adversarial autoencoder, browser extension, DistilBERT, feature fusion, phishing detection, temporal convolutional network, URL analysis. I. INTRODUCTION HE rapid growth of online services has made phishing one of the most persistent and damaging forms of cybercrime. Attackers imitate trusted websites to obtain credentials, financial data, and personal identifiable information. Recent studies show that phishing continues to be a major cause of organizational compromise and financial loss, while generative artificial intelligence can produce more convincing social-engineering content at low cost [1]-[3]. Traditional Conventional defenses such as rule-based filters, reputation databases, and blacklists are reactive. They depend on prior knowledge of malicious domains and therefore struggle against zero-day phishing pages, cloaking techniques, and short-lived campaigns. Machine learning detectors have improved this situation, but many systems still analyze either the lexical structure of a URL or the visible content of a page separately. This separation creates a semantic-structural disconnect: a URL may appear benign while the page metadata expresses a different commercial or credential-stealing intent. This work addresses that limitation through a cross-modal deep learning model that jointly learns from URL sequences and webpage semantic metadata. The main contributions are: (i) a dual pipeline combining a 1D Temporal Convolutional Network (TCN) and DistilBERT, (ii) an intermediate attention- guided fusion layer to correlate structural and semantic signals, (iii) adversarial hardening using an Adversarial Autoencoder (AAE), and (iv) a real-time proof-of-concept deployment using a FastAPI web service and a Chrome Manifest V3 browser extension. II. RELATED WORK Early phishing detection relied on manually designed lexical, host, content, and reputation features. Although these approaches are computationally light, they require continuous feature maintenance and may fail when attackers slightly modify URL strings or hosting patterns. Deep learning methods reduce this dependency by learning features directly from raw sequences. For example, CNN- based URL detectors have shown strong performance by learning local character patterns and suspicious token co- occurrences [4], while Texception combines character-level and word-level convolutions to improve generalization to unseen phishing URLs [5]. Sequential URL modelling has also been explored using TCNs. TCNs use dilated causal convolutions and residual connections to capture long-range dependencies without the step-by-step bottleneck of recurrent models. This makes them attractive for browser-centered protection, where latency and CPU overhead are important [6]. Transformer-based natural language processing (NLP) adds another advantage by representing context rather than only token counts. DistilBERT-based phishing detectors and hybrid BERT-CNN models can encode subtle semantic relationships within URLs and page text [7]-[9]. Recent research has moved toward multi-view and multimodal fusion. Some systems combine URL tokens, domain attributes, and behavioral features using transformer routing mechanisms [10], whereas others combine HTML Document Object Model (DOM) graphs with URL features through graph convolutional and transformer networks [11]. However, stronger fusion often increases runtime cost. In addition, adversarial examples can still degrade phishing classifiers unless synthetic evasive samples are include

III. METHODOLOGY A. Data Sources and Preprocessing The study used an empirical quantitative design. A balanced dataset of 100,000 unique domain entries was assembled, consisting of 50,000 validated phishing instances and 50,000 benign domains. Phishing samples were collected from PhishTank verified feeds and Kaggle phishing repositories, while benign samples were drawn from Cisco Umbrella and historical Alexa top-site sources. A separate out-of-distribution (OOD) set of 10,000 samples was reserved for final validation using PhiUSIIL and ISCX 2016 phishing datasets. To maintain consistent experimental conditions, the dataset was partitioned into 80% training, 10% validation, and 10% verification subsets. Raw URLs were transformed into padded token sequences capped at 256 tokens. HTML-derived text, including domain names, page titles, headings, and meta- descriptions, was processed using WordPiece tokenization and capped at 512 tokens. Automated Selenium and BeautifulSoup routines were used in a headless Chrome environment to extract page-level metadata and hyperlink indicators. B. Cross-Modal Model Architecture The proposed classifier contains two parallel feature extraction pathways. The first pathway applies a 1D TCN to raw URL character/subword sequences. This module captures local suspicious patterns, long-range sequence dependencies, URL length irregularities, hostname entropy, and keyword signals. The second pathway uses a fine-tuned DistilBERT encoder to represent semantic evidence from webpage metadata. DistilBERT was selected because it preserves much of BERT's contextual modelling ability while reducing parameter size and inference overhead. The hidden states from both pathways are passed to an intermediate multi-head cross-attention fusion module. Instead of concatenating features directly, the attention module learns dependency weights between URL structural sub-tokens and semantic metadata tokens. The fused latent representation is then passed through dense classification layers with a sigmoid output for binary legitimate/phishing prediction. C. Adversarial Hardening and Deployment To improve resilience against zero-day evasion, an AAE generated structured synthetic mutations that imitate high- evasion phishing URLs. These adversarial instances were injected into the training and evaluation loop at increasing ratios to measure performance degradation and to strengthen the classifier boundary. The final prototype was deployed as a cloud-hosted FastAPI inference engine connected to a lightweight Chrome Manifest V3 extension through secure web-tunnelling. The browser extension acted as a client-side sensor, while the resource-intensive model inference was executed on an NVIDIA T4 cloud runtime. IV. RESULTS AND DISCUSSION A. Classification Performance Table I summarizes the main classification results. The proposed fusion model achieved 92.45% accuracy, 91.80% precision, 93.12% recall, and 92.46% F1-score. The close precision and recall values show that the classifier did not rely on a one-sided threshold that favored either phishing or benign labels excessively. This balance is important for browser protection because false negatives expose users to phishing pages, while false positives reduce user trust in warnings. TABLE I MODEL CLASSIFICATION PERFORMANCE SUMMARY Metric Value (%) Accuracy 92.45 Precision 91.80 Recall 93.12 F1-score 92.46 Fig. 1 further supports the classification performance. The ROC curve indicates strong separability, with the proposed fusion model reaching an OOD ROC-AUC of 1.000 compared with a lower baseline 1D-CNN curve. The confusion matrix also shows that the model correctly identifies most phishing and benign samples, with only a small number of phishing samples misclassified as benign. Fig. 1. ROC curve and confusion matrix of the proposed fusion model. B. Robustness and Operational Latency Robustness was evaluated by increasing the adversarial mutation injection rate. As shown in Fig. 2, the standard model without AAE protection degraded rapidly as the adversarial rate increased. In contrast, the hardened model maintained a much flatter accuracy curve, remaining above approximately 94% even at the highest perturbation level shown. This behavior indicates that adversarial retraining improved the decision boundary against generative URL mutations.Department of Electrical & Computer EngineeringFYP Proceedings25

Fig. 2. Robustness degradation under zero-day evasion tactics. Operational latency was measured across preprocessing, feature extraction, and model inference. Table II shows that the full model-side pipeline required 44.63 ms. Feature extraction was the main bottleneck, contributing 22.55 ms or 50.52% of the total time. The actual forward-pass inference accounted for only 7.88 ms, suggesting that future optimization should focus on faster metadata extraction and token preparation rather than replacing the classifier itself. TABLE II COMPUTATIONAL LATENCY DISTRIBUTION Phase Latency (ms) Contribution (%) Data preprocessing 14.20 31.81 Feature extraction 22.55 50.52 Inference latency 7.88 17.67 Total pipeline 44.63 100.00 C. Interpretability Feature attribution analysis indicated that hostname entropy, keyword hits, and URL length had the strongest influence on the final prediction layer. This result is consistent with cybersecurity domain knowledge because phishing URLs frequently rely on abnormal character distributions, urgent or brand-related lexical cues, and unusually long or complex structures. Therefore, the model is not only accurate, but its strongest decision factors also align with interpretable phishing indicators. V. CONCLUSION This article presented a real-time phishing website detection framework that fuses URL structural features with transformer- derived semantic metadata. By combining a 1D TCN, a DistilBERT encoder, intermediate cross-modal attention, and AAE-based adversarial hardening, the system directly addresses the semantic-structural gap and zero-day weakness of many conventional phishing detectors. The final model achieved 92.45% accuracy, 92.46% F1-score, and 1.000 OOD ROC-AUC, while maintaining a model-side processing latency of 44.63 ms. The main limitation is that the evaluation focuses on static URL and textual metadata from English-dominant datasets. Future work should integrate visual webpage evidence using Vision Transformers to counter Browser-in-the-Browser and visual cloaking attacks. Model compression, quantization, pruning, and hardware-aware knowledge distillation should also be explored to enable fully local browser inference without dependency on cloud tunnelling. Finally, multilingual datasets and internationalized domain name homoglyph attacks should be included to improve global deployment robustness. REFERENCES [1] R. Zieni, L. Massari, and M. C. Calzarossa, "Phishing or Not Phishing? A Survey on the Detection of Phishing Websites," IEEE Access, vol. 11, pp. 18499-18519, 2023, doi: 10.1109/access.2023.3247135. [2] B. Naqvi, K. Perova, A. Farooq, I. Makhdoom, S. Oyedeji, and J. Porras, "Mitigation Strategies against the Phishing Attacks: A Systematic Literature Review," Computers & Security, vol. 132, 2023, doi: 10.1016/j.cose.2023.103387. [3] M. Schmitt and I. Flechais, "Digital deception: generative artificial intelligence in social engineering and phishing," Artificial Intelligence Review, vol. 57, no. 12, 2024, doi: 10.1007/s10462-024-10973-2. [4] E. A. Aldakheel, M. Zakariah, G. A. Gashgari, F. A. Almarshad, and A. I. A. Alzahrani, "A Deep Learning-Based Innovative Technique for Phishing Detection in Modern Security with Uniform Resource Locators," Sensors, vol. 23, no. 9, 2023, doi: 10.3390/s23094403. [5] F. Tajaddodianfar, J. W. Stokes, and A. Gururajan, "Texception: A Character/Word-Level Deep Learning Model for Phishing URL Detection," in Proc. IEEE ICASSP, 2020, doi: 10.1109/ICASSP40776.2020.9053670. [6] M.-L. E. Alorvor and S. Dadkhah, "Real-Time Phishing Detection for Brand Protection Using Temporal Convolutional Network-Driven URL Sequence Modeling," Electronics, vol. 14, no. 18, 2025, doi: 10.3390/electronics14183746. [7] A. Aljofey, Q. Jiang, A. Rasool, H. Chen, W. Liu, Q. Qu, and Y. Wang, "An effective detection approach for phishing websites using URL and HTML features," Scientific Reports, vol. 12, no. 1, 2022, doi: 10.1038/s41598-022-10841-5. [8] A. Aljofey, S. A. Bello, J. Lu, and C. Xu, "BERT-PhishFinder: A Robust Model for Accurate Phishing URL Detection with Optimized DistilBERT," IEEE Transactions on Dependable and Secure Computing, vol. 22, no. 4, pp. 1-15, 2025, doi: 10.1109/tdsc.2025.3545771. [9] I. Altan, A. Bachir, Y. Parbhulkar, A. M. Rizvi, and M. Farazi, "Dual-Path Phishing Detection: Integrating Transformer-Based NLP with Structural URL Analysis," in Proc. IEEE/ACS AICCSA, 2025, pp. 1-6, doi: 10.1109/aiccsa66935.2025.11315219. [10] Y. Wang, W. Ma, H. Xu, Y. Liu, and P. Yin, "A Lightweight Multi-View Learning Approach for Phishing Attack Detection Using Transformer with Mixture of Experts," Applied Sciences, vol. 13, no. 13, 2023, doi: 10.3390/app13137429. [11] J.-H. Yoon, S.-J. Buu, and H.-J. Kim, "Phishing Webpage Detection via Multi-Modal Integration of HTML DOM Graphs and URL Features Based on Graph Convolutional and Transformer Networks," Electronics, vol. 13, no. 16, 2024, doi: 10.3390/electronics13163344. [12] H. Shirazi, S. R

Performance Benchmarking of Explainable AI Methods for Skin Cancer Detection El Alewi Tijani and Teddy Surya Gunawan Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—Skin cancer is one of the most common forms of cancer worldwide, and early diagnosis plays a crucial role in improving treatment outcomes. Although deep learning models have demonstrated high performance in skin lesion classification, their black-box nature limits clinical trust and adoption. This study proposes an explainable deep learning framework for skin cancer detection using the HAM10000 dataset. Three convolutional neural network (CNN) architectures, namely AlexNet, ResNet18, and DenseNet121, were evaluated for binary and multiclass skin lesion classification. Model performance was assessed using accuracy, F1-score, ROC-AUC, confusion matrix analysis, and inference time. To improve transparency, explainable artificial intelligence (XAI) techniques including Grad-CAM, LIME, and SHAP were applied to visualise and interpret model predictions. Experimental results showed that DenseNet121 achieved the highest classification performance, followed by ResNet18 and AlexNet. The XAI visualisations successfully highlighted clinically relevant lesion regions and improved model interpretability. Faithfulness evaluation demonstrated that Grad-CAM generally produced the most reliable explanations for the best-performing models. The findings indicate that combining deep learning with XAI can improve both classification performance and transparency, supporting the development of trustworthy AI-assisted dermatological diagnostic systems. Index Terms—Deep learning, DenseNet121, explainable artificial intelligence, Grad-CAM, HAM10000, LIME, SHAP, skin cancer detection. I. INTRODUCTION Skin cancer is one of the most common cancers worldwide and poses a significant challenge to healthcare systems. Early detection is crucial for improving treatment outcomes and reducing mortality rates. In recent years, artificial intelligence (AI), particularly deep learning (DL), has shown great potential in assisting dermatologists with skin lesion classification using dermoscopic images. Convolutional neural networks (CNNs) such as AlexNet, ResNet18, and DenseNet121 have demonstrated promising performance in extracting discriminative features and achieving high classification accuracy on publicly available datasets such as HAM10000 [1], [2]. Despite their success, CNN-based models are often considered black-box systems because their decision-making processes are difficult to interpret. This lack of transparency limits trust and hinders the adoption of AI-assisted diagnostic systems in clinical practice. To address this issue, explainable artificial intelligence (XAI) techniques such as Gradient- weighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-agnostic Explanations (LIME), and SHapley Additive exPlanations (SHAP) have been proposed to provide visual explanations that help users understand model predictions [3], [4]. Although several studies have applied deep learning and XAI methods to skin lesion analysis, comprehensive benchmarking of multiple CNN architectures together with different XAI techniques remains limited. Therefore, this study evaluates the performance of AlexNet, ResNet18, and DenseNet121 for skin lesion classification using the HAM10000 dataset and benchmarks Grad-CAM, LIME, and SHAP in terms of interpretability, faithfulness, and computational efficiency. The findings aim to identify a reliable and transparent framework for AI-assisted skin cancer diagnosis. II. RELATED WORK Convolutional neural networks (CNNs) have been widely studied for skin lesion classification due to their strong feature- extraction capability. Albahar [1] proposed a CNN with novel regularisation techniques to improve classification on dermoscopic images. Their findings demonstrated that regularisation can reduce overfitting and enhance accuracy on datasets such as HAM10000. Similarly, Alsahafi et al. [4] introduced Skin-Net, a deep residual network that uses multilevel feature extraction for skin lesion classification. Their work highlighted the importance of deep, densely connected architectures in capturing discriminative lesion features and improving classification performance. Several studies have also reviewed the broader use of deep learning for medical image analysis. Alzubaidi et al. [2] surveyed deep learning approaches across healthcare and medical imaging and reported that CNN-based models consistently outperform traditional methods in classification tasks. Nawaz et al. [3] reviewed deep learning-based skin cancer classification and emphasised the growing reliance on transfer learning and publicly available datasets. However, these studies primarily focused on classification accuracy and treated the models as black boxes, with limited attention to interpretability or clin

classification. Jan et al. [7] reviewed the integration of XAI in skin lesion classification and reported that Grad-CAM, LIME, and SHAP are the most widely adopted techniques. Jadhav et al. [8], Fiaz et al. [9], Alghamdi et al. [10], and Rahman et al. [11] further applied or benchmarked these methods to enhance interpretability in medical image classification. Although previous studies have addressed CNN-based classification and XAI integration separately, limited research has benchmarked multiple CNN architectures together with several XAI techniques within a single framework. Most existing studies focus on only one aspect, such as model accuracy or a single explanation method. Therefore, this study combines multiple CNN architectures with Grad-CAM, LIME, and SHAP, and evaluates them in terms of interpretability, faithfulness, and computational efficiency to provide a more comprehensive and trustworthy framework for skin cancer detection. III. METHODOLOGY The overall research methodology is summarised in Fig. 1. It consists of three main objectives: CNN model development, XAI integration, and benchmarking and evaluation. B. Data Preprocessing Images were resized to 224 x 224 pixels and normalised prior to training. Data augmentation techniques including random rotation, flipping, and image transformations were applied to improve dataset diversity and reduce overfitting. The dataset was divided into training, validation, and testing sets. C. CNN Architectures Three CNN architectures were developed and trained for the classification task: ● AlexNet ● ResNet18 ● DenseNet121 Transfer learning was employed using pretrained ImageNet weights to improve convergence and classification performance. D. Explainable AI Techniques Three XAI techniques were integrated with the trained CNN models: ● Grad-CAM ● LIME ● SHAP These techniques were used to visualise the image regions contributing to model predictions and to improve interpretability. E. Evaluation Metrics Model performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and inference time, while the XAI methods were assessed through faithfulness and interpretability analysis. A. Dataset Fig. 1. Methodology flowchart. IV. RESULTS AND DISCUSSION A. Classification Performance The classification performance of AlexNet, ResNet18, and DenseNet121 was evaluated using accuracy, F1-score, ROC- AUC, and inference time. Among the evaluated models, DenseNet121 achieved the best overall performance due to its dense connectivity structure, which improves feature reuse and gradient propagation [10]. ResNet18 also demonstrated competitive performance, while AlexNet achieved satisfactory results but was limited by its relatively shallow architecture [11]. Table I summarises the performance comparison of the three CNN models. TABLE I PERFORMANCE COMPARISON OF THE CNN MODELS The HAM10000 dataset was used as the primary dataset in this study. The dataset contains 10,015 dermoscopic images categorised into seven skin lesion classes: melanocytic nevi (NV), melanoma (MEL), benign keratosis-like lesions (BKL), basal cell carcinoma (BCC), actinic keratoses (AKIEC), dermatofibroma (DF), and vascular lesions (VASC). Model Binary Acc. (%) Multiclass Acc. (%) Binary ROC- AUC Multiclass ROC- AUC Inference (ms) AlexNet 84 70 0.92 0.90 0.12 ResNet18 89 80.69 0.95 0.97 0.15 DenseNet121 91 82.50 0.97 0.98 0.18 Fig. 2 shows the training and validation accuracy of DenseNet121 for the multiclass classification task, whichDepartment of Electrical & Computer EngineeringFYP Proceedings28

illustrates stable convergence and limited overfitting across training epochs. Fig. 2. Training and validation accuracy of DenseNet121 for the multiclass task. B. Explainability Analysis Grad-CAM, LIME, and SHAP were successfully applied to visualise the decision-making behaviour of the CNN models. The generated explanations highlighted lesion regions that contributed significantly to the classification outcomes. Grad-CAM produced intuitive heatmaps focusing on lesion boundaries and abnormal structures. LIME provided local region-based explanations by identifying image segments influencing predictions. SHAP offered detailed feature- attribution explanations but required higher computational cost. DenseNet121 combined with Grad-CAM generated the most clinically meaningful visualisations and demonstrated strong alignment with lesion regions. Fig. 3. Grad-CAM explanation of DenseNet121 (multiclass). Fig. 4. LIME explanation of DenseNet121 (multiclass).Department of Electrical & Computer EngineeringFYP Proceedings29

Fig. 5. SHAP explanation of DenseNet121 (multiclass). C. Faithfulness Evaluation Faithfulness analysis was conducted to assess the reliability of the generated explanations using deletion and insertion AUC. Grad-CAM achieved the highest faithfulness scores for DenseNet121 and ResNet18, indicating better alignment between the highlighted regions and the model predictions. As shown in Fig. 6, Grad-CAM and LIME produced larger insertion-minus-deletion gaps than SHAP, confirming that their explanations are more faithful to the model behaviour. These results demonstrate the effectiveness of XAI techniques in improving model transparency and trustworthiness. Fig. 6. Faithfulness comparison of Grad-CAM, LIME, and SHAP for DenseNet121. V. CONCLUSION AND FUTURE WORK This study developed and evaluated an explainable deep learning framework for skin lesion classification using the HAM10000 dataset. Among the evaluated CNN models, DenseNet121 achieved the best overall classification performance in terms of accuracy and ROC-AUC. Furthermore, Grad-CAM, LIME, and SHAP successfully improved model interpretability, with Grad-CAM providing the most reliable explanations based on the faithfulness evaluation. Future work may focus on larger datasets, advanced architectures such as EfficientNet and Vision Transformers (ViT), and the deployment of the proposed framework for practical clinical applications. REFERENCES [1] Y. A. Albahar, "Skin lesion classification using convolutional neural networks with novel regularization techniques," IEEE Access, vol. 11, pp. 1-14, 2023. [2] A. Alzubaidi, M. A. Fadhel, O. Al-Shamma, J. Zhang, and Y. Duan, "Review of deep learning approaches in healthcare and medical imaging," J. Big Data, vol. 8, no. 53, 2021. [3] H. Nawaz, M. Usman, and S. Latif, "Deep learning-based skin cancer classification using dermoscopic images: A review," Healthcare, vol. 11, no. 5, pp. 1-20, 2023. [4] Y. S. Alsahafi, M. A. Kassem, and K. M. Hosny, "Skin-Net: A novel deep residual network for skin lesion classification using multilevel feature extraction," J. Big Data, vol. 10, no. 105, 2023. [5] S. A. H. Shah, M. A. Khan, T. Saba, and A. Rehman, "Explainable AI- based skin cancer detection using CNN, particle swarm optimization and machine learning," J. Imaging, vol. 10, no. 12, pp. 1-24, 2024. [6] G. Munjal, A. Sharma, and P. Gupta, "SkinSage XAI: An explainable deep learning solution for multiclass skin lesion classification," Health Sci. Rep., vol. 7, no. 4, 2024. [7] M. B. Jan, M. Rashid, and R. Vavekanand, "Integrating explainable AI for skin lesion classifications: A systematic literature review," Stud. Med. Health Sci., vol. 2, no. 1, pp. 1-14, 2025. [8] D. Jadhav, P. Patil, and R. Kulkarni, "Enhancing skin cancer diagnosis with XAI: Integrating Grad-CAM, SHAP, and LIME for model interpretability," in Proc. Int. Conf. Emerging Technologies (INCET), 2025. [9] M. Fiaz, S. Khan, M. Ahmed, and N. Ali, "An explainable hybrid deep learning framework for precise skin lesion segmentation and classification," Front. Med., vol. 12, 2025. [10]A. M. Alghamdi, M. Alotaibi, and H. Alharbi, "Explainable deep learning framework for skin lesion classification using Grad-CAM and SHAP," Diagnostics, vol. 14, no. 2, pp. 1-18, 2024. [11]M. H. Rahman, S. Islam, and M. R. Islam, "Benchmarking explainable AI techniques for medical image classification," IEEE Access, vol. 12, pp. 45678-45692, 2024.Department of Electrical & Computer EngineeringFYP Proceedings30

Early-Stage Chronic Kidney Disease Prediction Using Interpretable Machine Learning Ghazi Iqbal and Malik Arman Morshidi Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract— Chronic Kidney Disease (CKD) is a progressive condition that affects kidney function and impacts approximately 10% of the global population. Early detection is crucial to reduce treatment costs and prevent disease progression. This study presents an interpretable machine learning framework for CKD prediction using three machine learning algorithms: Support Vector Machine (SVM), Random Forest, and K-Means Clustering. A clinical dataset obtained from the Kaggle repository, consisting of 400 patient records with 25 clinical attributes, was used. Data preprocessing techniques, including missing value imputation and feature encoding, were applied to improve data quality by reducing missing entries from 1,012 to 152. Feature selection identified Albumin (al), Sugar in Urine (su), Hypertension (htn), Specific Gravity (sg), and Serum Creatinine (sc) as key predictors for CKD diagnosis. The results showed that supervised learning models achieved superior performance, with SVM and Random Forest obtaining accuracies of 98.33% and 97.50%, respectively. In contrast, K-Means Clustering achieved only 17.50% accuracy, indicating its limitations for this dataset. The findings demonstrate the potential of supervised machine learning approaches, particularly SVM and Random Forest, in supporting early and accurate CKD diagnosis in clinical applications. Index Terms— Chronic Kidney Disease (CKD), Machine Learning, Random Forest, Support Vector Machine, Feature Selection. I. INTRODUCTION HRONIC Kidney Disease (CKD) is a serious threatening disease that also is a progressive disease condition that have multiple stages in its lifetime and affected many people around the globe with the estimated people affected by it is around 10% of the total global population [1]. One of the main issues with the disease is the asymptomatic nature of the CKD in the early stages where the patient may have not seen the early symptoms of the CKD disease in their body until the disease has progressively grown into latter stages where it became a serious condition to the patient. The patient may seek out medical attention by going to the doctor only after the condition has become worsen and the disease advanced into the latter stages. Late – stage of the CKD has a life-threatening condition to the patient in the advanced stages of the disease with the complete failure of the renal organ in filtering waste in the body and significantly increase the medical cost in treating the diseases with cardiovascular issues as the side effect of the CKD diseases. Several key test in treating the CKD diseases that also include blood test, urine tests, imaging test (MRI), and also Glomerular Rate (GFR) measurement test may be effective in keeping the CKD diseases at bay but may been less accessible to people that live in inadequate condition where the advanced medical facilities may located far away like in the rural village areas or rubber plantation estate farm. Rural regions of the developing nations like Bangladesh may have these issues where people live in rural areas and lack of government funds for medical facilities [2]. Machine learning techniques has emerged as reliable tools in dealing with complex nature of the patients’ medical records that may mix several numerical and categorical attributes of the medical features in the clinical dataset and records. Recent studies showed several machine learning algorithms that have high accuracy when dealing with the clinical records which include Random Forest, Gradient Boosting and Catboost can exceed the threshold of accuracy and have high accuracy ranging from 90 % to 100 % [3], [4], [5]. However, clinician may encounter several problem and challenges when implementing the machine learning techniques with its black box nature that may decrease the trust of the clinical professionals [6], [7]. This study will try to address the challenges by developing a ML pipeline for this early-stage CKD detection. The framework will integrate the data preprocessing phase of the dataset, feature selection techniques that also include coefficient correlation method and Chi- Square method test and also to evaluate significant metrics of each ML algorithms in terms of F1 score, Recall and also Precision of each machine learning classifier algorithms. II. METHODOLOGY A. Dataset Acquisition The dataset were acquired from the Kaggle repository through two month period data collection that were conducted in the hospital setting which is available publicly [8]. The dataset were comprised by 400 patient records and also 25 clinical attributes that includes numerical attributes which is blood pressure (bp), blood glucose (bgr), serum creatinine (sc), hemoglobin (hemo), and packed cell volume (

diabetes mellitus (fm) and CKD classification status. Table 1 showed the selected attributes that has been used in this study. Table 1 . Attributes in the CKD test B. Data Preprocessing The raw clinical dataset of the patient record present many hurdles during the study and one of the main hurdles is the missing value and null values of the dataset and also a present of incomplete diagnostic test of the patient in the clinical records. The missing values of the dataset were represented by Missing Completely at Random (MCAR), Missing at Random (MAR) and Missing Not at Random (MNAR) for the imputation method strategies [9] . The numerical attributes that were present in the original clinical dataset were replaced by the median values of the representative of the variables that have the missing dataset values (e.g.. Sodium: 140 mEq/L; Potassium: 4.2 mEq/L), while the categorical features attributes of the dataset were handled through mode imputation and KNN – based methods. This method managed to reduce the missing values that were present in the original dataset from 1012 to 152 and solved 860 missing entries in the dataset. C. Feature Selection The feature selection involved two methods that were being used to choose the top and relevant features attributes from the dataset . One of the method that being used is the Chi-Squared (χ²) test that was being applied in measuring the statistical dependency between each categorical clinical attribute and the CKD class label with the implementation of the SelectKBest function from the Scikit-learn Python library. The Chi – Squared formula in (1), where Oi that represent the observed frequency and Ei that represent the expected frequencies: Another method were also being used which is the Pearson Correlation Coefficient method that was being used to measure the linear relationship between the feature and the target variable as in (2). D. Machine Learning Classifier Three machine learning classifiers were being chosen and evaluated in this study which is (1) Random Forest (RF), which is an ensemble machine learning classifier that combine multiple decision trees through bootstrap sampling and majority voting which each attributes will send a vote in terms of finding the accuracy of the ML classifier. Another machine learning classifier that were being evaluated is (2) Support Vector Machine (SVM) that employed a Radial Basis Function (RBF) kernel and also (3) K – Means Clustering algorithm that will act as an unsupervised algorithm that function as the comparative baseline. The dataset was split 70/30 into training (280 samples) and testing (120 samples) sets, maintaining stratified class distribution. E. Performance Evaluation Model performance was evaluated using Accuracy (3), Precision (4), Recall (5), and F1-Score (6), derived from the confusion matrix elements True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN): III. RESULTS AND DISCUSSION A. Data Preprocessing Results The initial result of the raw dataset revealed a pattern of missing and null values across the important features attributes of the clinical records. Red blood cells (rbc) exhibited the highest missingness with 152 missing entries (38% of 400 records), followed by red blood cell count (rc) with 131, white blood cell count (wc) with 106, and hemoglobin (hemo) with 52 missing entries. Fig. 1 illustrates the distribution of missing values prior to preprocessing. Following the imputation procedure, 14 out of 15 primary attributes achieved 100% data completeness, as shown in Fig. 1. The total missing values were reduced from 1,012 to 152, corresponding to a successful imputation rate of 85%. Table 2 summarizes the missing value handling outcomes. Table 2. Missing values summaryDepartment of Electrical & Computer EngineeringFYP Proceedings32

Fig. 1. Preprocessing phase of the CKD dataset The successful imputation that of important markers of the CKD diseases particularly hemoglobin and serum creatinine, which is important in the CKD detection and in assessing the patient condition has decrease the total missing values of the dataset and ready for feature selection phase. B. Feature Selection Results Chi – Squared Test The Chi – Squared Test were being applied in this study where the patient identification (id) were being excluded in this case as it carry no predictive value for this study. As shown in Table 2, Albumin (al) emerged as the attribute that scored the highest score in this case with a score of 216.43, which is more than the second placed attributes which is Sugar in Urine (su). Albumin presence in this case was a key indicator of glomerular dysfunction in human body and also a sign of early kidney injury [10]. Sugar in Urine (su) emerged as the second place with a score 96.84 and are strongly associated with diabetes mellitus that are major factors in the CKD diseases. Hypertension (htn) follow up in the third place with a score of 89.13 and being followed with Specific Gravity (sg) and Serum Creatinine (sc) with both scores 80.62 and 73.59 respectively. Table 3. Top five features of Chi – Squared Test Correlation Coefficient Analysis Weak results were showed when we implement this analysis as all features exhibited weak linear correlation with CKD status with a value less than 0.3 . Appetite (appet) showed a value of r = 0.1396, followed by Hypertension(htn) at 0.0930, hemoglobin (hemo) with a value of 0.0491, packed cell volume (pcv) at 0.0470 and specific gravity (sg) at 0.0354 as shown in Table 4. This showed the superiority of the Chi – Square test over Coefficient Correlation analysis in dealing with the CKD dataset as it has a strong result in categorical features and also showed the strong result of the non – linear ML classifier of the Random Forest and also SVM with RBF kernel. Table 4. Top five features of Correlation Coefficient test C. Classification Results Train Test Split The dataset that were retrieved from the Kaggle repository were being split into a ratio of 70:30, resulting in 280 samples of the patient records (174 CKD, 106 non – CKD; 62.1 % CKD prevalence) and also 120 testing samples(74 CKD, 46 non – CKD; 61.7 % CKD prevalence). Random Forest Classifier The Random Forest classifier was being executed with the usage of the seven features clinical attributes that selected through the Feature Selection process : Albumin (al), Sugar in Urine (su), Specific Gravity (sg), Serum Creatinine (sc), Hemoglobin (hemo), Appetite (appet), and Packed Cell Volume (pcv).The model was employed with the bootstrap sampling with 100 decision trees and majority voting for the final classification. The model achieved a high accuracy with a score of 97.50 % by correctly classify 117 out of the total available 120 test samples. Fig. 2 shows the structure and the operation of the Random Forest architecture. Fig. 2. Simplified architecture of the Random Forest classifier showing bootstrap sampling with a combination of multiple datasets. Final classification are through majority voting across 100 decision trees. Seven clinical attributes were being used as the input features in this context.Department of Electrical & Computer EngineeringFYP Proceedings33

Support Vector Machine (SVM) The SVM classifier employed RBF kernel and achieved the highest accuracy in this study achieving 98.33 % , using 54 support vectors in defining the optimal separation of the hyperplane. The non – linearity concept in the classifier of the Radial Basis Function (RBF) kernel has successfully handled the intricate , non – linear relationship between each features of the clinical dataset as evidenced by the curved decision boundary that separate the CKD (red circles) and non – CKD (blue circles) classes. Fig. 3 showed the SVM classification visualization. Fig. 3. SVM classification visualization showing the RBF kernel that are separating the optimal hyperplane which are suited with non-linearity of the clinical feature space of the CKD dataset K-Means Clustering K-Means Clustering, an unsupervised ML algorithm managed to achieved only 17.50 % only in terms of the score. The Euclidean distance metrics were poorly suited with clinical significance of the clinical attributes of the patients. D. Performance Evaluation Confusion matrix analysis across all three classifier showed Random Forest and SVM has identified 74 True Negatives (non – CKD) and 44 True Positives (CKD) out of 120 samples. K – Means has mistakenly identify 53 samples, majority are classified as false positives. Fig. 4 shows the comparative of the confusion matrices. Fig. 4. Comparative confusion matrices of the three ML classifier algorithms across 120 – sample test as Random Forest (left) and SVM (center) showed identical performance of both classifier with 74 True and 44 True Positives. K- Means (right) showed degraded performance and showed on how the unsupervised methods of the classifier were being applied to supervised classification tasks. IV. CONCLUSION This study successfully developed and validated an interpretable machine learning framework for early-stage Chronic Kidney Disease (CKD) prediction. A robust preprocessing pipeline reduced dataset missing values by 85%, from 1,012 to 152 entries, ensuring high-quality data for model training. Multi-method feature selection using Chi-Squared analysis and Pearson Correlation identified albumin, sugar in urine, hypertension, specific gravity, and serum creatinine as the five most discriminative clinical predictors, consistent with established clinical understanding of CKD pathophysiology. Among the three classifiers evaluated, both Random Forest and SVM achieved exceptional performance with test accuracies of 97.50% and 98.33% respectively and F1-scores of 0.9792, demonstrating their strong suitability for CKD diagnosis from clinical tabular data. The weak linear correlations identified during Pearson analysis validated the use of non-linear classifiers, particularly SVM with the RBF kernel. K-Means Clustering's poor accuracy (17.50%) confirmed that unsupervised methods are unsuitable for this supervised clinical classification task. The findings demonstrate significant potential for machine learning to support clinicians in early CKD detection, particularly in resource-constrained settings where traditional diagnostic methods are expensive or inaccessible. Future work should investigate the integration of SHAP (SHapley Additive exPlanations) for model interpretability, incorporate larger and geographically diverse datasets for improved generalizability, explore real-time clinical deployment architectures, and develop user-friendly mobile interfaces to facilitate direct adoption by healthcare professionals. REFERENCES [1] J. Qin, L. Chen, Y. Liu, C. Liu, C. Feng, and B. Chen, “A machine learning methodology for diagnosing chronic kidney disease,” IEEE Access, vol. 8, pp. 20991–21002, 2020, doi: 10.1109/ACCESS.2019.2963053. [2] R. Sawhney, A. Malik, S. Sharma, and V. Narayan, “A comparative assessment of artificial intelligence models used for early prediction and evaluation of chronic kidney disease,” Decision Analytics Journal, vol. 6, p. 100169, Mar. 2023, doi: 10.1016/j.dajour.2023.100169. [3] D. Gudadhe, M. Gulhane, M. Kaur, M. N, and N. Rakesh, “Optimizing Chronic Kidney Disease Diagnosis with Machine Learning: A Comprehensive Approach to Feature Selection and Classification Models,” in 2024 International Conference on Cybernation and Computation (CYBERCOM), IEEE, Nov. 2024, pp. 82– 87. doi: 10.1109/CYBERCOM63683.2024.10803235. [4] C. T. Su, Y. P. Chang, Y. T. Ku, and C. M. Lin,“Machine Learning Models for the Prediction of Renal Failure in Chronic Kidney Disease: A Retrospective Cohort Study,” Diagnostics, vol. 12, no. 10,Oct.2022,doi: 10.3390/diagnostics12102454. [5] S. R. Kalupukuru and K. Natarajan, “Machine Learning Methods for Predicting the Prognosis of Chronic KidneyDepartment of Electrical & Computer EngineeringFYP Proceedings34

Disease,” in Procedia Computer Science, Elsevier B.V., 2025, pp. 1372–1382. doi: 10.1016/j.procs.2025.04.370. [6] J. Saurabh, A. Shrivastava, V. V. Labde, A. Badhoutiya, L. Hussein, and A. Dutt, “Predictive Modeling with Machine Learning in the Management of Chronic Kidney Disease,” 2025, pp. 123–132. doi: 10.1007/978-3-031- 93673-9_12. [7] M. Sharifuzzaman et al., “Optimized Machine Learning Models and Stacking Hybrid Approach for Chronic Kidney Disease Prediction,” 2025, pp. 196–212. doi: 10.1007/978-3-031-96235-6_15. [8] P. S. and P. E. L. Rubini, “‘Chronic Kidney Disease,’ UCI Machine Learning Repository, 2015,” 2015. [9] A. F. Ahmad, S. Sayeed, K. Alshammari, and I. Ahmed, “Impact of Missing Values in Machine Learning: A Comprehensive Analysis”, doi: 10.32604/cmc.202x.0xx [10] A. Ashitani, T. Ueno, A. Nakashima, S. Doi, K. Yamane, and T. Masaki, “High-normal albuminuria and incident chronic kidney disease in a male nondiabetic population,” Clin. Exp. Nephrol., vol. 22, no. 4, pp. 835–842, Aug. 2018, doi: 10.1007/s10157-017-1522-6.Department of Electrical & Computer EngineeringFYP Proceedings35

Computer and System / IoT T R A C K 3

Secure MQTT Communication Protocols for IoT Devices Muhammad Faez Ahmad Shahriman and Othman Omran Khalifa Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia from eavesdropping, data tampering and injection-based attacks [2], [4]. Traditional security mechanisms such as TLS introduce computational and memory overhead that is not practical for the resources constrained nature of IoT devices [1]. The earlier works included many mitigative approaches, including dynamic lightweight key sharing during MQTT handshakes [3], mutual authentication using ECDHE along with SM4 symmetric encryption [4], decentralized authentication based on blockchain [5], and multilayer security architectures [6]. Significantly more gaps exist, however many of the works do not simulate actual attacks, tend to use legit diminished primitives such as MD5, or leave digital signatures out altogether, leaving unverified security properties in their wake. The present paper seeks to fill these gaps by proposing a hybrid cryptographic framework in which regular RSA-2048 asymmetric cryptography is combined with AES-Fernet symmetric encryption, RSA-PSS digital signatures, and HMAC-SHA256 integrity verification, chained via MQTT. Three objectives guided this research: (i) identification of security challenges in IoT communication environments; (ii) design and implementation of a secure MQTT-based framework for resource-constrained IoT devices; and (iii) evaluation of the effectiveness of the framework against replay, DoS, and MITM attacks by way of live simulation. The rest of the paper is arranged as follows: Section II describes related work; Section III gives system design and methodology and Section IV presents and discusses simulation results while Section V concludes the paper with a mention of future recommendations/works. II. RELATED WORK Kaganurmath et. al. [3] developed the DLKS-MQTT method of embedding ephemeral 128-bit keys generated by a Linear Congruential Generator (LCG) into standard MQTT CONNECT/CONNACK messages. While this method uses very little power (an average of 0.0014 mJ) and executes very quickly (0.40 seconds), the 128-bit key ceiling limits the achievable long-term security margin for sensitive implementation. Zhang and Yu [4] proposed a scheme using ECDHE with PSK for mutual authentication, SM4 as the symmetric encryption algorithm, and SHA-256 for both integrity checks and preventing replay attacks through timestamps. The total execution time was approximately 195 ms. The key weakness Abstract—Internet of Things (IoT) is rapidly growing in various industries such as healthcare and smart homes. Therefore, the need for secure communication between them has increased. Many existing IoT protocols remain vulnerable to attacks such as replay, man-in-the-middle (MITM), and denial-of-service (DoS) because of the limited security measures in place. This paper outlines a secure MQTT-based communication framework for resource- constrained IoT environments. It utilizes RSA-2048 for mutual authentication and session key establishment. In addition, AES-Fernet is used for data encryption, RSA-PSS for message authentication, and HMAC-SHA256 for integrity verification. Timestamp validation and nonce checks are included to mitigate replay attacks. The proposed framework was implemented in a controlled environment consisting of a simulated IoT device, a Mosquitto MQTT broker, and a secure backend server. Results from experiments indicate that the proposed framework successfully creates secure communication and detects/rejects malicious packets when subjected to simulated MITM, replay, and DoS attacks. The proposed framework improves confidentiality, integrity, and authentication while remaining appropriate for resource-constrained IoT devices. Keywords—Internet of Things (IoT), MQTT, Secure Communication, Authentication, Data Integrity, Cybersecurity, Replay Attack. I. INTRODUCTION HE Internet of Things (IoT) consists of a dataset of physical objects (devices), such as home appliances, industrial T machines and medical devices, which contain built in sensors, software and communications technology, which allows them to share data over the internet without the need for human intervention. All IoT communication must ensure three essential security requirements - confidentiality (only authorized parties can see transmitted data), integrity (data is not altered while transmitted), and authentication (two parties communicating can authenticate each other) [1]. MQTT (Message Queuing Telemetry Transport) is a lightweight publish/subscribe protocol which has become one of the leading protocols in the IoT market because of its low bandwidth/overhead requirements, as well as its ability to operate effectively in unreliable networks. However, MQTT was not designed with security in mind and therefore does not support built-in cryptography. Consequently, MQTT deployments can be comp

in this design stems from using the MD5 hash function on session key derivation, which is no longer considered secure by cryptographic standards. The use of this primitive greatly reduces the integrity guarantees of the entire design. Dumitrache and Anghelescu [2] implemented mutual certificate-based authentication via EAP-TLS/EAP- MSCHAPv2 using a Raspberry Pi hardware and built an application with a very strong security level while using a very low amount of power. Their implementation has a scaling limit of approximately 20 concurrent devices because of hardware limitations. Tukur et al. [14] created a multi-layer security architecture that uses TLS/ECDHE at the perception layer, fog computing with blockchain at the network layer, and Role- Based Access Control (RBAC) at the application layer. Blockchain approaches [5], [8], [9] provide increased trust and non-repudiation in a decentralised way but impose high computational and deployment costs that do not suit heavily resource-constrained devices. Nadaf et al. [7] developed a hybrid RSA+AES scheme to solve the key exchange problem for symmetric ciphers but suffer from greater RSA processing cost. None of the above work validated their frameworks with real attempted attacks and quantifying their detection rate, which is a significant open gap addressed in this study. III. SYSTEM DESIGN AND METHODOLOGY A. Study Design and Environment This project utilizes software for experimentation, so everything was simulated. The simulation included all the pieces necessary for an MQTT (Message Queue Telemetry Transport) protocol (e.g., Connection, Device, Broker) and were implemented entirely on localhost. These components communicated with each other via local TCP on the MQTT topic iot/secure. The development of all components was carried out using Visual Studio Code with Python 3.14 and the additional libraries of paho-MQTT, cryptography and hashlib. The main variables under study are the accuracy of detecting attacks (whether an attack was detected), validation of packets (pass/fail for each check) and communication integrity (whether a valid message can be sent through the system without being affected by an attack). To ensure the results of the experiments would be reliable and reproducible, five separate runs of each attack scenario were conducted. B. System Architecture Fig. 1 illustrates the three-component architecture. The IoT devices publish sensor data in encrypted form to the broker via a Mosquitto broker, which solely acts as a message broker, without being given access to any plaintext data therefore immobilizing any possible attacks. The IoT device has an established security authority in the shape of a backend server that can authenticate the device, maintain the keys for encryption/decryption, and verify the contents of the payload received. The communication channel contains attack simulation modules (replay.py, dos.py, mitm.py) that allow for the simulation of various attacks, and an overall attack logger component (attack_logger.py) to record all security events. Fig. 1. Block diagram of the proposed secure MQTT communication framework. C. Secure Communication Algorithm and Flowchart Fig. 2 shows the complete communication process from initialisation to session termination. The system starts up and registers the device by using pre-configured credentials and the unique device identifier. RSA-2048 is used for mutual authentication to verify both parties prior to protocol allowing for the exchange of any data. An asymmetric key exchange provides a shared session key. The session is established and will encrypt sensor data with AES-Fernet; sign the data with RSA-PSS; and publish it; The back-end server will apply a 5- step verification on every packet received. Any packet failing the verification checks will be discarded, logged, and reported as a security incident. Sessions and keys are terminated and discarded once the time-out or key lifetime has been reached.Department of Electrical & Computer EngineeringFYP Proceedings37

Fig. 2. Flowchart of the proposed secure MQTT communication protocol. D. Hybrid Encryption and Key Management Fig. 3 summarises the three-phase hybrid encryption strategy. RSA-2048 asymmetric encryption is used for mutual authentication of devices in Phase One and secure session key exchange, meaning that only registered devices may participate in the session and no key material is exposed on the communication channel. In Phase Two, an established AES session key is used to encrypt all bulk data using AES-Fernet encryption, which significantly reduces the computational cost of each encryption operation. Every encrypted payload includes an RSA-PSS signature that covers the ciphertext, and Fernet's built-in HMAC-SHA256 signature guarantees the integrity of the encrypted payload. Each packet contains a unique nonce and the Fernet-native timestamp. The backend server verifies each packet using the five-step process illustrated in Figure 4: (1) unpack_packet() checks if packet format and length are correct; (2) verify_signature() verifies the RSA-PSS signature; (3) decrypt_session_key() decrypts the session key using RSA- OAEP; (4) decrypt_data() decrypts the recipient's COP using AES-Fernet and verifies the integrity of the ciphertext using HMAC-SHA256; and (5) validate_message() rejects all packets with nonces previously seen and timestamps greater than 30 seconds old. Session keys are transient and discarded at the end of the session, and new sessions require a new full authentication and key exchange cycle. PHASE 1 — SESSION ESTABLISHMENT • RSA-2048: IoT Device ⇔ Server mutual authentication • Server public key used to encrypt AES session key • Session key securely exchanged over MQTT channel PHASE 2 — DATA TRANSMISSION • Sensor data encrypted with AES-Fernet (session key) • RSA-PSS digital signature generated over ciphertext • HMAC-SHA256 integrity tag embedded via Fernet • Nonce + timestamp appended to each packet PHASE 3 — SERVER VERIFICATION (5-Step Chain) • Step 1: unpack_packet() — validate packet format & length headers • Step 2: verify_signature() — RSA-PSS signature verified • Step 3: decrypt_session_key() — RSA-OAEP decrypts session key • Step 4: decrypt_data() — AES-Fernet decrypts payload + HMAC-SHA256 check • Step 5: validate_message() — Nonce uniqueness + timestamp freshness (<30s) Fig. 3. Hybrid encryption strategy: three phases of the proposed framework. Step 1: unpack_packet() Validate packet format and length headers ➤ DoS → FAIL here (invalid format) Step 2: verify_signature() — RSA-PSS Confirm message came from registered IoT device, not attacker ➤ MITM → FAIL here (signature mismatch) Step 3: decrypt_session_key() — RSA-OAEP Unlock AES session key using server’s RSA private key Step 4: decrypt_data() — AES + HMAC Decrypt sensor data and verify Fernet HMAC-SHA256 integrity tag Step 5: validate_message() — Nonce + Timestamp Reject if nonce seen before or message older than 30 seconds ➤ Replay → FAIL here (duplicate nonce / expired timestamp) Fig. 4. Five-step server-side verification chain (iot_server.py) with attack detection points. E. Threat Mitigation Design Replay attack mitigation: The server records a unique nonce for each packet in the form of a 128-bit random number. Fernet enforces a strict timestamp window. Packets that are outside of this window or use a nonce that has previously been used will be discarded immediately and logged with the message Duplicate nonce detected — replay attack? MITM / tampering mitigation: The method for computing RSA-PSS signatures uses the complete ciphertext, so changing even one byte in the payload will result in an invalid signature or verification failing at Step 1 of the server’s verification chain, before decryption occurs. DoS mitigation: If any packets are rejected because they are malformed or do not have a valid RSA-PSS signature, they will receive an ERROR: Invalid Packet Format response as soon as they are completed; this process will use minimal amounts of server resources and allow the server to reserve its processing capabilities for legitimate packets.Department of Electrical & Computer EngineeringFYP Proceedings38

IV. RESULTS AND DISCUSSION A. System Implementation Verification Prior to the attack simulation, all system components were initialized successfully. The backend server was loaded with the RSA-2048 key pair, connected to the Mosquitto broker, and subscribed to the iot/secure topic. The IoT device was also registered and mutually authenticated, creating an AES session key between them. The three-step transmission cycle which included the RSA-OAEP encryption of the session key, AES- Fernet encryption of the sensor reading, and RSA-PSS signature generation, worked as expected. There were successful verifications for each of the three verification steps (signature verification, session key decryption, and HMAC- SHA256 confirmation) on the server for each of the legitimate packets, verifying that the sensor had established baseline functionality. B. Normal Communication Baseline (No Attack) Under purely observation conditions (no_attack.py in observer mode), the framework completed the full secure communication cycle without errors. The no_attack script subscribed to the MQTT topic and monitored traffic without interference. Each packet received at the server passed all four verification steps, and the server marked the transmission as STATUS: SECURE (no attack detected). The justification for this baseline affirms two properties: (1) the system correctly performs the entire authentication-encryption-verification cycle under normal conditions; and (2) an eavesdropper attempting to observe the channel will only obtain the ciphertext in fully encrypted form and would have no opportunity to access actionable plaintext. C. Replay Attack Results The script which is used to replay the packet captured an actual live encrypted packet over the iot/secure topic and sent the exact same byte stream back after a 5 second (as defined by the REPLAY_DELAY constant). The reason for a 5 second delay was to ensure that the original nonce would be recorded at the server prior to the replayed packet arriving at the server and that the Fernet timestamp window would already have expired or nearly expired. Table I summarises all five attempts. TABLE I Replay Attack Detection Results — 5 Attempts Attempt Size (bytes) Detection Trigger Server Status 1 792 Duplicate nonce ATTACK DETECTED (Replay) 2 792 Timestamp expired ATTACK DETECTED (Replay) 3 792 Duplicate nonce ATTACK DETECTED (Replay) 4 792 Timestamp / nonce ATTACK DETECTED (Replay) 5 792 Duplicate nonce ATTACK DETECTED (Replay) There were five attempts to replay. All were detected and rejected. There were two independent mechanisms that caused detection (duplicate nonce for injects 1, 3 and 5; expired Fernet timestamp for injects 2 and 4). The log at the server contained the following log entries: ERROR = Replay attack detected (invalid timestamp or nonce); STATUS = ATTACK DETECTED (Replay). Legitimate device messages were untouched and demonstrated that the use of dual nonce and timestamp provides a strong method of preventing replays and is consistent with prior work using timestamps [4], but without using any deprecated hash functions. D. Denial-of-Service Attack Results The DoS script (dos.py) functioned in a mode referred to as "live-triggered". It injected 20 spam packets (17 bytes each, 340 bytes total) into the message stream immediately upon receiving a legitimate message from the base station. The script used this strategy to overload the base station with additional processing at the same time as handling a legitimate packet. In total, 100 spam packets were produced over five bursts throughout the course of the simulation. Table II summarises the results. TABLE II DoS Attack Detection Results (5 Bursts × 20 Spam Messages) Burst Spam Msgs Bytes/Burst Detection Server Status 1 20 340 DETECTED ATTACK DETECTED (Malformed Packet) 2 20 340 DETECTED ATTACK DETECTED (Malformed Packet) 3 20 340 DETECTED ATTACK DETECTED (Malformed Packet) 4 20 340 DETECTED ATTACK DETECTED (Malformed Packet) 5 20 340 DETECTED ATTACK DETECTED (Malformed Packet) TOTAL 100 1,700 100% All 100 rejected Each of the 100 spam packets was rejected during Stage 1 of the server verification chain, specifically during the RSA-PSS Signature Validation phase, prior to doing any expensive computational work to decrypt them. Each spam packet was logged as ERROR: Invalid Packet Format and STATUS: ATTACK DETECTED (Malformed Packet). Importantly, the actual IoT device message from each “burst" of packets was successfully received and processed. This demonstrates that the early-reject method allowed the system to remain available to legitimate users. This also conforms to the principle articulated by Tukur et al. [6] that access control and authentication filtering should be done prior to performing resource extensive processing. E. Man-in-the-Middle Attack Results The MITM (mitm.py) intercepted live device payloads (792 byte each), overwriting 9 bytes at a random position of the ciphertext (not including the first fo

payloads were then republished to the iot/secure topic. There were five attacks made against different bytes(682, 738, 140, and two random positions). Table III summarises the results. TABLE III MITM Corruption Attack Detection Results — 5 Attempts Attempt Original (bytes) Injected (bytes) Detection Mechanism Server Status 1 792 9 (byte 682) RSA-PSS sig. failed ATTACK DETECTED (Forgery) 2 792 9 (byte 738) RSA-PSS sig. failed ATTACK DETECTED (Forgery) 3 792 9 (byte 140) RSA-PSS sig. failed ATTACK DETECTED (Forgery) 4 792 9 (random) RSA-PSS sig. failed ATTACK DETECTED (Forgery) 5 792 9 (random) RSA-PSS sig. failed ATTACK DETECTED (Forgery) Each of the five packets that had been tampered with caused a failure in the signing verification process as described in Step 1. The result produced was ERROR: Signature verification failed and STATUS: ATTACK DETECTED (Forgery/Tampering). A change of only nine bytes within a 792 byte ciphertext caused the RSA-PSS signature to be rendered invalid. This clearly indicates that the integrity protection mechanism is very sensitive to cryptographic adjustments to the ciphertext. No decryption(s) would be attempted on any rejected packets. F. Security Event Logging Analysis The centralised logging module (attack_logger.py) captured all security events across all simulation runs. Each log entry records timestamp, attack type, detection result, description, payload size, and server outcome. Table IV provides the session-level summary across all simulation runs. TABLE IV Security Event Logging Summary — All Simulation Runs Attack Type Total Logged Detected Detection Rate No Attack (baseline) 3 N/A N/A Replay Attack 5 5 100% DoS Attack 100 100 100% MITM Attack 5 5 100% TOTAL (attacks only) 110 110 100% (110/110) A total of 52 event data entries (3 baseline + 110 attack events) generated during attack event detection and classification resulted in all 110 attack events being detected and accurately classified as attacks. No false-positive entries were recorded. All legitimate devices generated messages were recorded at COMMUNICATION SUCCESS. All security events that occurred were accurately classified and none missed; therefore, validating the integrity of the security solution and its logging system. G. Comparative Analysis Table V compares the proposed framework against the two most closely related works on key performance and security criteria. TABLE V Comparison of Proposed Framework Against Related Works Criteria Kaganurmath et al. [3] Zhang & Yu [4] Proposed Framework Asymmetric Algorithm LCG (128-bit) ECDHE + PSK RSA-2048 Symmetric Algorithm BLAKE2s SM4 AES-Fernet Integrity Mechanism Lightweight hash SHA-256 HMAC-SHA256 Digital Signature Not stated Not stated RSA-PSS Hash Weakness None MD5 in key gen. None (MD5 avoided) Replay Protection Stated Timestamp + nonce Timestamp + nonce Attack Simulation Not tested Not tested Replay, DoS, MITM Detection Rate Not reported Not reported 100% (49/49 events) Execution Time 0.40 s 195 ms Software simulation Among the three frameworks discussed, only this proposed framework uses RSA-PSS for digital signatures, explicitly avoids deprecated cryptographic primitives such as MD5 and provides quantified detection rates that are verified through simulations. Of these frameworks, the basis of Kaganurmath et al.’s work [3] provides better energy efficiency and execution times than both of the other two frameworks because Kaganurmath et al., use lightweight LCG-based keys, while Zhang and Yu [4], provide lower latency by using ECDHE. While the proposed framework results in a higher computational cost when using RSA-2048, it also provides significantly stronger guarantees for authentication and non- repudiation through use of digital signatures and verified resistance against three different types of attacks. For IoT deployments, where assurance of security is more important than stringent power consumption limitations, such as in the case of remote monitoring in health care or in industrial control systems, the trade-offs represented by using RSA-2048 against lower costs through using either of the other two frameworks is acceptable. V. CONCLUSION This paper presented a secure MQTT communication framework for resource-constrained IoT devices, integrating RSA-2048 mutual authentication, AES-Fernet symmetric encryption, RSA-PSS digital signatures, HMAC-SHA256 integrity verification, and timestamp-nonce replay prevention.Department of Electrical & Computer EngineeringFYP Proceedings40

All three project objectives were achieved: security challenges in IoT communication were identified through a comprehensive literature review; a secure MQTT-based framework was designed, implemented, and verified functional under normal operating conditions; and the framework was evaluated against replay, DoS, and MITM attacks which managed to successfully detect all the attack events without disrupting legitimate communications. The framework has several limitations that must be acknowledged. First, the implementation is entirely software- simulated where no physical hardware was used, meaning real- world computational overhead, memory constraints, wireless interference, and network instability were not characterised. Second, attack patterns in the simulation were predictable and of fixed size, which may not reflect the sophistication of real- world adversaries. Third, scalability was not evaluated — the framework was tested with a single IoT device and broker, and performance in large-scale deployments with multiple devices and brokers remains unknown. Future work can address these limitations by: (1) deploying and testing the framework on physical IoT hardware platforms (Raspberry Pi, ESP32) to measure actual energy consumption and latency; (2) evaluating scalability in multi-device, multi- broker deployments; (3) incorporating more sophisticated attack vectors including distributed DoS, credential-stealing, and insider threats; and (4) integrating a machine learning- based intrusion detection system (IDS) with the centralised logging mechanism for adaptive, automated threat response. REFERENCES [1] X. Huo, "Design and analysis of secure communication protocol in the internet of things environment," Discover Internet of Things, vol. 5, no. 1, Dec. 2025, doi: 10.1007/s43926-025-00187-5. [2] C. G. Dumitrache and P. Anghelescu, "Design and implementation of a secure communication architecture for IoT devices," J. Sensor Actuator Netw., vol. 14, no. 4, Aug. 2025, doi: 10.3390/jsan14040064. [3] S. Kaganurmath, N. G. Cholli, and M. R. Anala, "Integrating MQTT protocol to improve the performance using dynamic lightweight key sharing protocol for secure IoT communications," Cluster Comput., vol. 28, no. 16, Dec. 2025, doi: 10.1007/s10586-025-05689-z. [4] S. Zhang and Z. Yu, "Secure communication scheme of Internet of Things based on MQTT protocol," in Proc. 5th Int. Conf. NNICE, IEEE, Jan. 2025, pp. 1126-1130. [5] Q. Wu, T. Zhang, and Z. Yang, "A secure communication protocol based on blockchain and MQTT for IoT," in Proc. IEEE IAEAC, 2024, pp. 1419-1423. [6] Y. M. Tukur, D. Thakker, and I.-U. Awan, "Multi-layer approach to Internet of Things (IoT) security," in Proc. 7th Int. Conf. FiCloud, IEEE, Aug. 2019, pp. 109-116. [7] R. Nadaf et al., "Symmetric and asymmetric cryptographic approach based security protocol for key exchange," in Proc. ICAISC, IEEE, 2023. [8] S. Kumar et al., "Blockchain-based optimization algorithm for secure IoT communication using AMQP," in Proc. ICAIHC, IEEE, 2025. [9] P. Pawar et al., "Improving security of IoT device using blockchain technology," in Proc. IEEE ICCUBEA, 2023.Department of Electrical & Computer EngineeringFYP Proceedings41

AI-Powered Intrusion Detection System for Enhanced 5G Cybersecurity Networks using Deep Learning Nurul Nadiah Ameer Ali and Farah Diyana Abdul Rahman Electrical and Electronics Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—The rapid growth of 5G networks, Internet of Things (IoT) devices, and fog computing environments has increased the need for reliable intrusion detection systems that can detect attacks efficiently in distributed networks. Denial-of-Service attacks are a major security threat because they can disrupt network availability and affect time-sensitive applications. This study develops a hybrid Convolutional Neural Network and Long Short-Term Memory model for DoS attack detection using the UNSW-NB15 dataset. The dataset was preprocessed through categorical encoding, feature scaling, train-test splitting, and reshaping for model input. A baseline CNN-LSTM model and a class-weighted CNN-LSTM model were implemented to examine the effect of class imbalance on detection performance. The models were evaluated using accuracy, precision, recall, F1-score, and confusion matrix. The baseline model achieved high overall testing accuracy but showed limited DoS detection, while the class-weighted model failed to improve minority attack detection. The findings show that accuracy alone is insufficient for evaluating intrusion detection performance on imbalanced datasets. Further optimization is required to improve minority attack detection and support the development of more effective fog-based intrusion detection systems. Index Terms—Class Imbalance, CNN-LSTM, DoS Attack, Fog Computing, Intrusion Detection System, UNSW-NB15. I. INTRODUCTION he emergence of 5G networks, internet of things (IoT) devices, and fog computing has brought new challenges to network security. These technologies offer low- latency communication and massive device connectivity but also increase the attack surface of cyber threats such as Denial-of-Service (DoS) attacks. DoS attacks are highly harmful as they attack the availability of services by flooding network resources, impacting time-critical applications such as smart healthcare, autonomous systems, and industrial automation [9], [10], [11]. Traditional intrusion detection systems (IDS) usually work in centralized environments where the network traffic is collected and analysed at the central server. While this approach is useful for large-scale monitoring, it can introduce to latency, bandwidth overhead and delayed response in distributed 5G and IoT environments. This problem is solved in fog computing. Data is processed near the source. This allows faster detection and eliminates the need for continuous transfer of large amounts of raw traffic to cloud servers [1], [2], [15]. However, fog nodes are usually limited in terms of processing power, memory and energy capacity, thus making the design of intrusion detection models more challenging. IDS based on artificial intelligence and deep learning are extensively studied as they can learn complex traffic patterns and detect attacks in more adaptive manner than traditional signature-based methods [3], [6]. Deep learning models are highly valuable in contemporary network environment as they can handle high-dimensional traffic data and discover the hidden relation between features. Convolutional Neural Networks (CNNs) are effective in extracting important feature patterns among deep learning architectures, and Long Short- Term Memory (LSTM) are suitable for learning sequential dependencies in data [6], [11]. A hybrid CNN-LSTM model extracts features by CNN layers and learn sequences by LSTM layers, combining both advantages. The UNSW-NB15 dataset is a popular choice for IDS research due to its contemporary network traffic features and multiple attack classes such as DoS attacks [7], [12]. Previous studies have used different deep learning models such as CNN, RNN, LRCN, ensemble approaches and hybrid CNN-LSTM architectures to explore the dataset [7], [12], [20]. The studies indicate the potential of deep learning models for intrusion detection, but the class imbalance issue still remains important. In IDS datasets, normal traffic is also more frequent than attack traffic, which can result in high accuracy but poor detection of the minority attack classes [6], [21]. Therefore, in this work, a CNN- LSTM-based IDS model is developed for DoS attack detection and evaluated based on the UNSW-NB15 dataset. The model is trained under two settings: a baseline CNN- LSTM and a CNN-LSTM with class weights. The class weighting is used to test if giving a higher weight to the minority DoS class can improve detection performance. For evaluating, accuracy, precision, recall, F1-score, and confusion matrix are used to give more complete evaluation, particularly in the case of imbalanced class conditions. II. METHODOLOGY This study follows an experimental research design to develop and evaluate a CNN-LSTM-bas

The overall research methodology is illustrated in Fig. I. The process begins with the review of related studies to support the selection of the dataset, preprocessing techniques, model architecture, and evaluation metrics. The UNSW-NB15 dataset is then prepared through data cleaning, categorical encoding, feature scaling, and train-test splitting. After that, a CNN- LSTM model is developed and trained under two configurations: baseline CNN-LSTM and class-weighted CNN-LSTM. Finally, both models are evaluated by accuracy, precision, recall, F1-score, and confusion matrix to compare their performance of DoS attack detection. Fig. I. Project Methodology Flowchart A. Dataset Description The UNSW-NB15 dataset was used as the benchmark dataset in this study. It contains modern network traffic records with both normal and attack samples, making it suitable for intrusion detection research. For this experiment, the dataset was prepared for binary classification, where DoS traffic was treated as the attack class and all other traffic was grouped as Normal/non-DoS traffic. This binary setup was selected because the study focus specifically on detecting DoS attacks. B. Data Preprocessing Data preprocessing was performed to convert the raw dataset into a suitable format for deep learning model training. First, unnecessary attributes were removed and the target labels were prepared for binary DoS and Normal classification. Categorical features such as protocol, service, and state were encoded into numerical values so that they could be processed by the model. Numerical features were then scaled using StandardScaler to ensure that all input values were within a comparable range. Feature scaling is important because network traffic attributes may have different value ranges, and unscaled features can affect the training stability of the model [5], [12]. After preprocessing, the dataset was divided into training and testing subsets using an 80:20 split. The training subset was used to train the CNN-LSTM model, while the testing subset was used for validation and final performance evaluation. Since CNN-LSTM models were require sequential input format, the processed feature data was reshaped into three-dimensional tabular format of (-1, 43, 1). In this structure, -1 represents the number of samples, 43 represents the number of input features, and 1 represents a single channel. This reshaping was performed to make the data compatible with the Conv1D and CNN-LSTM model structure. However, for the Random Forest feature importance analysis, the reshaped data was converted back into a two-dimensional format of (-1, 43) because Random Forest operates on standard tabular input. C. Feature Importance Exploration A Random Forest classifier was also explored to identify feature importance and rank the input features. The Random Forest model was trained using 100 estimators and a random state of 42. Its purpose was to rank the importance of the input features and observe which traffic attributes contributed more strongly to classification. However, the selected features were not used in the final CNN- LSTM training because the final model was trained using the full set of 43 features. Therefore, Random Forest feature selection was treated only as an exploratory step and not as part of the final model implementation. D. CNN-LSTM Model Development The proposed model uses a hybrid CNN-LSTM architecture. The CNN component is used to extract important feature patterns from network traffic data, while the LSTM components learns dependencies from the extracted features. After the CNN and LSTM layers, dense layers are used to perform the final binary classification. A sigmoid activation function is applied at the output layer because the task involves two classes: Normal and DoS. The model was trained using the Adam optimizer and binary cross-entropy loss function. The training process was conducted for 20 epochs with batch size of 64. Two model configurations were implemented: a baseline CNN-LSTM model and a class-weighted CNN-LSTM model. Table 1 presents the experimental configuration used in this study. It summarises the dataset, classification type, model architecture, training parameters, and evaluation metrics applied during model development. Those settings were kept consistent for both baseline CNN-LSTM and class-weighted CNN-LSTM models to ensure a fair comparison of their DoS detection performance. TABLE I EXPERIMENTAL CONFIGURATION Item Description Dataset UNSW-NB15 training set Classification type Binary classification Classes Normal/non-Dos and DoS Programming environment Python Libraries TensorFlow/Keras, Scikit-learn Optimiser Adam Loss function Binary cross-entropy Output activation function Sigmoid Train-test split 80:20Department of Electrical & Computer EngineeringFYP Proceedings43

Epochs 20 Batch size 64 Model configurations Baseline CNN- LSTM and CNN- LSTM with class weights Model parameters 47,809 trainable parameters Evaluation metrics Accuracy, precision, recall, F1-score, confusion matrix E. Class Weighting Strategy Class imbalance is a common issue in intrusion detection datasets because attack samples may appear less frequently than normal traffic. To address this issue, class weighting was applied during model training. Instead of modifying the dataset distribution, class weighting assigns a higher penalty to the misclassification of the minority class. This encourages the model to pay more attention to DoS samples during training. The performance of the class-weighted CNN-LSTM model was then compared with the baseline CNN-LSTM model to evaluate whether class weighting improved DoS detection. F. Performance Evaluation The models were evaluated using accuracy, precision, recall, F1-score, and confusion matrix. These metrics were selected because accuracy alone may be misleading in imbalanced IDS datasets. Precision and recall are important for analysing false alarms and missed attacks, while F1-score provides a balanced measurement between both metrics. The formula used for performance evaluation are shown in Table II. TABLE II PERFORMANCE EVALUATION METRICS Metrics Formulas Accuracy (𝑇𝑃 + 𝑇𝑁) (𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁) Precision 𝑇𝑃 (𝑇𝑃 + 𝐹𝑃) Recall 𝑇𝑃 (𝑇𝑃 + 𝐹𝑁) F1-score 2 × (𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙) (𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙) TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative respectively. In this study, TP refers to Normal correctly classified as Normal, while TN refers to DoS traffic correctly classified as DoS traffic.FN occurs when Normal is incorrectly classified as DoS, which may cause false alarms. FP occurs when DoS traffic incorrectly classified as Normal, which is more critical in IDS because it means an attack is missed by the detection system. III. RESULTS AND DISCUSSION This section presents the performance of the baseline CNN- LSTM model and the class-weighted CNN-LSTM model for DoS attack detection. The models were evaluated using accuracy, precision, recall, F1-score, and confusion matrix. Since the dataset is imbalanced, this dicussion focuses not only on overall accuracy but also on the ability of each model to detect the minority DoS class. Table III shows the performance comparison between the baseline CNN-LSTM and the class-weighted CNN-LSTM models. The baseline CNN-LSTM achieved a training accuracy of 93.46%, validation accuracy of 93.12%, and testing accuracy of 93.12%. However, although the overall the accuracy was high, the model showed weak performance in detecting DoS traffic. The DoS recall was only 5.36% meaning that most actual DoS samples were misclassified as non-DoS. This shows that the model was strongly influenced by the majority class which is Normal. TABLE III PERFORMANCE COMPARISON OF CNN-LSTM MODELS The class-weighted CNN-LSTM model achieved a training accuracy of 85.91%. although its testing accuracy was 93.24%, this result does not indicate better attack detection. The DoS precision, recall, and F1-score were all 0.00%, showing that the model failed to correctly classify any DoS samples. The finding highlights that a high testing accuracy can be misleading when the majority class dominates the dataset. The confusion matrix results provided a clearer explanation of the model behaviour. As shown in Fig. II, the baseline CNN-LSTM correctly classified 32,530 Normal samples and 127 DoS samples. However, it misclassified 2,243 DoS samples as Normal. This translated that although the model worked well in recognizing Normal sample, it failed to detect most DoS attacks. In intrusion detection, this is a serious issue because false negatives represents attacks that pass through the system without being detected.Department of Electrical & Computer EngineeringFYP Proceedings44

Fig. II. Confusion Matrix of Baseline CNN-LSTM Fig. III shows the confusion matrix for the class-weighted CNN-LSTM model. The model correctly classified all 32,699 non-DoS samples, but it misclassified all 2,370 DoS samples as non-DoS. This indicates that the class weighting strategy did not improve minority class detection in this experiment. Instead, the model became fully biased towards the Normal during testing. Fig. III Confusion Matrix of Class-Weighted CNN-LSTM Based on these results, the baseline CNN-LSTM model performed better than the class-weighted model in terms of DoS detection because it was able to correctly detect some DoS samples. However, both models showed limited ability to identify the minority DoS class. The baseline model detected only 127 out of 2,370 DoS samples, while the class-weighted model detected none. This shows that class imbalance significantly affected the learning behaviour of the CNN-LSTM models. The results also shows that accuracy alone is insufficient for evaluating IDS performance. Both models achieved more than 93% testing accuracy, but their DoS detection performance were poor. This happened because testing set contained a much larger number of Normal samples rather than DoS samples. As a result, the models could achieve high accuracy by correctly predicting the majority class while failing to detect minority attack samples. From an IDS perspective, recall is the most important as it measures how many actual attacks are successfully detected. A low recall value means that many attacks are missed, which reduces the reliability of the intrusion detection system. Precision is also important because it indicates whether predicted attack alerts are correct. The low DoS F1-score in both models confirms that the balance between precision and recall were weak for the attack class. Overall, the experiment shows that the CNN-LSTM model has potential for learning network traffic patterns, but further optimization is required to improve DoS detection. Class weighting alone was not sufficient to solve the imbalance problem in this study. Future improvements may include threshold tuning, resampling techniques, hyperparameter optimization, improved feature selection, or alternative lightweight architectures that can support minority attack detection in fog computing environments. IV. CONCLUSION This study developed and evaluated a CNN-LSTM-based intrusion detection model for DoS attack detection using the UNSW-NB15 dataset. The model was tested under two configurations: a baseline CNN-LSTM and a class-weighted CNN-LSTM model. The baseline CNN-LSTM achieved high overall testing accuracy of 93.12%, but its DoS detection performance was limited, with a DoS recall of only 5.36%. The class-weighted CNN-LSTM achieved a slightly higher testing accuracy of 93.24% but failed to detect any DoS samples, resulting in 0.00% for precision, recall, and F1-score. This study is limited to offline evaluation using the UNSW- NB15 dataset and focuses only on binary DoS and Normal classification. The model was developed and tested in a simulated Python environment, without actually being deploy on fog devices. Therefore, real-time performance, memory usage, inference time, and deployment feasibility were not fully evaluated. Future research should focus on improving minority attack through threshold tuning, resampling techniques, hyperparameter optimization, improved feature selection, or alternative lightweight deep learning models. Further work may also include multiclass attack classification and actual deployment on fog devices to evaluate the practicality of CNN-LSTM-based IDS in real network environments. REFERENCES [1] Mishra, R., Singh, D., & Thakur, A. (2025). A Structured Framework for Selecting Cloud or Fog Computing in Application-Specific Contexts. Proceedings - 3rd International Conference on Self Sustainable Artificial Intelligence Systems, ICSSAS 2025, 1729–1734. https://doi.org/10.1109/ICSSAS66150.2025.11081303 [2] Sabir, M., Soni, H., Chhawcharia, P., Sharma, I., & Agarwal, P. (2025). Performance Evaluation of Fog Computing for Health Monitoring Systems: A Comparative Study on Latency Reduction and Energy Efficiency with Cloud-Based Solutions. 2025 International Conference on Artificial Intelligence and Machine Vision, AIMV 2025. https:// doi.org/10.1109/AIMV66517.2025.11203662Department of Electrical & Computer EngineeringFYP Proceedings45

[3] Dandriyal, M., & Gupta, N. (2025). Enhancing Cybersecurity with Deep Learning: Proactive Threat Detection and Mitigation Techniques. 2025 IEEE 2nd International Conference on Advances in Modern Age Technologies for Health and Engineering Science, AMATHE 2025 - Proceedings. https://doi.org/10.1109/AMATHE65477.2025.11081305 [4] Aswini, J., Rekha, K. S., Rosaline, R. A. A., & Sivaneshkumar, A. (2025). Enhancing security in cloud computing systems using hybrid feature selection and ensemble-based machine learning for intrusion detection. Evolving Systems, 16(3). https://doi.org/10.1007/s12530-025- 09725-6 [5] Putrada, A. G., Alamsyah, N., Fauzan, M. N., & Oktaviani, I. D. (2024). Pearson Correlation for Efficient Network Anomaly Detection with Quantization on the UNSW-NB15 Dataset. 11th International Conference on ICT for Smart Society: Integrating Data and Artificial Intelligence for a Resilient and Sustainable Future Living, ICISS 2024 - Proceeding. https://doi.org/10.1109/ICISS62896.2024.10751550 [6] Rizwani, N., Jamil, A., & Hameed, A. A. (2024). Deep Learning-based Techniques for Intrusion Detection Systems. 2024 2nd International Conference on Artificial Intelligence, Blockchain, and Internet of Things, AIBThings 2024 - Proceedings. https://doi.org/10.1109/AIBThings63359.2024.10863292 [7] Chen, D., Song, Q., Zhang, Y., Yu, Q., Li, L., & Yang, Z. (2024). Evaluating the Effectiveness of Various Model Combinations for Network Intrusion Detection on UNSW- NB15. 2024 6th International Academic Exchange Conference on Science and Technology Innovation, IAECST 2024, 312– 317. https://doi.org/10.1109/IAECST64597.2024.11117865 [8] Zhukabayeva, T., Ahmad, Z., Adamova, A., & Satybaldina, D. (2024). Penetration Testing-Aware IDS for Wireless Networks in Urban Digital Ecosystems Integrated with Edge Computing and Machine Learning. Global Congress on Emerging Technologies, GCET 2024, 300–307. https://doi.org/10.1109/GCET64327.2024.10934644 [9] Pons, M., Valenzuela, E., Rodríguez, B., Nolazco-Flores, J. A., & Del-Valle-Soto, C. (2023). Utilization of 5G Technologies in IoT Applications: Current Limitations by Interference and Network Optimization Difficulties—A Review. Sensors, 23(8). https://doi.org/10.3390/s23083876 [10] Heidari, A., & Jabraeil Jamali, M. A. (2023). Internet of Things intrusion detection systems: a comprehensive review and future directions. Cluster Computing, 26(6), 3753–3780. https://doi.org/10.1007/s10586-022-03776-z [11] Yadav, N., Pande, S., Khamparia, A., & Gupta, D. (2022). Intrusion Detection System on IoT with 5G Network Using Deep Learning. Wireless Communications and Mobile Computing, 2022. https://doi.org/10.1155/2022/9304689 [12] Rathod, N. A., Gupta, T., Sharma, N. v., & Sharma, S. (2021). Model Comparison and Multiclass Implementation Analysis on the UNSW NB15 Dataset. 2021 International Conference on Computational Performance Evaluation, ComPE 2021, 549–555. https://doi.org/10.1109/ComPE53109.2021.9751832 [13] Zhu, M., Gong, W., Peng, F., & Qin, H. (2021). OpenStack Oriented Networking-VPP Network Optimization Method. Proceedings - 2021 Asia-Pacific Conference on Communications Technology and Computer Science, ACCTCS 2021, 187–191. https://doi.org/10.1109/ACCTCS52002.2021.00045 [14] Okelmann, P., Linguaglossa, L., Geyer, F., Emmerich, P., & Carle, G. (2021). Adaptive Batching for Fast Packet Processing in Software Routers using Machine Learning. Proceedings of the 2021 IEEE Conference on Network Softwarization: Accelerating Network Softwarization in the Cognitive Age, NetSoft 2021, 206–210. https://doi.org/10.1109/NetSoft51509.2021.9492668 [15] Kumar, P., Gupta, G. P., & Tripathi, R. (2021). A distributed ensemble design based intrusion detection system using fog computing to protect the internet of things networks. Journal of Ambient Intelligence and Humanized Computing, 12(10), 9555–9572. https://doi.org/10.1007/s12652-020- 02696-3 [16] Ali, M., & Zhang, J. (2024). Exploring the Effectiveness of Synthetic Data in Network Intrusion Detection through XAI. 2024 Cyber Awareness and Research Symposium, CARS 2024. https://doi.org/10.1109/CARS61786.2024.10778756 [17] Zambre, P., Panchal, M., & Chouhan, A. (2023). Blockchain Unleashed: Empowering Information-centric Network Computing in Cloud, Fog, and Edge Services. 2023 14th International Conference on Computing Communication and Networking Technologies, ICCCNT 2023. https://doi.org/10.1109/ICCCNT56998.2023.10308013 [18] Nugraha, B., Jnanashree, A. V., & Bauschert, T. (2024). An Efficient Explainable Artificial Intelligence (XAI)-Based Framework for a Robust and Explainable IDS. Proceedings of the 8th Cyber Security in Networking Conference: AI for Cybersecurity, CSNet 2024, 173–181. https://doi.org/10.1109/CSNet64211.2024.10851760 [19] Dang, P. S., Duy, P. T., & Pham, V. H. (2024). Investigating the Robustness Against Transferable Adversarial Attacks of Learning-Based Network Intrusion Detection System. Proceedings - 2024 RIVF Internat

on Electrical Engineering, Computing Science and Automatic Control, CCE 2024. https://doi.org/10.1109/CCE62852.2024.10770969 [21] Hnamte, V., & Hussain, J. (2021). An Extensive Survey on Intrusion Detection Systems: Datasets and Challenges for Modern Scenario. ICECIE 2021 - 2021 International Conference on Electrical, Control and Instrumentation Engineering, Conference Proceedings. https://doi.org/10.1109/ICECIE52348.2021.9664737 [22] 2020 International Conference on Computing and Information Technology (ICCIT-1441). (2020). IEEE. [23] V. Shanmugam, R. Razavi-Far, and E. Hallaji, “Addressing Class Imbalance in Intrusion Detection: A Comprehensive Evaluation of Machine Learning Approaches,” Electronics, vol. 14, no. 1, p. 69, 2025, doi: 10.3390/electronics14010069. [24] A. Telikani, N. E. Rudbardeh, S. Soleymanpour, A. Shahbahrami, J. Shen, G. Gaydadjiev, and R. Hassanpour, “A Cost-Sensitive Machine Learning Model With Multitask Learning for Intrusion Detection in IoT,” IEEE Transactions on Industrial Informatics, vol. 20, no. 3, pp. 3880–3890, 2024, doi: 10.1109/TII.2023.3314208. [25] M. Dener, S. Al, and A. Orman, “STLGBM-DDS: An Efficient Data Balanced DoS Detection System for Wireless Sensor Networks on Big Data Environment,” IEEE Access, vol. 10, pp. 92931–92945, 2022, doi: 10.1109/ACCESS.2022.3202807.Department of Electrical & Computer EngineeringFYP Proceedings47

Electronics / Instrumentation and Measurements T R A C K 4

Abstract—Printed electroluminescent (EL) displays are widely used in wearable and flexible electronics due to their lightweight structure and low power consumption. However, EL displays require high-voltage alternating current (AC) excitation, while most portable devices operate using low-voltage direct current (DC) sources. Conventional inverter circuits are often bulky and unsuitable for flexible applications. Therefore, this project aimed to develop and evaluate an inverter circuit for printed electroluminescent display applications. The project involved PSPICE simulation and experimental testing. An H-bridge inverter was simulated at frequencies of 60 Hz to 2 kHz using capacitance values ranging between 10 nF and 100 nF to represent the EL load. Experimental testing was conducted using a MIC4832 driver circuit with input voltages ranging from 1.5 V to 5 V. The simulation results showed that 500 Hz produced the most stable square-wave output approaching ±100 V, while higher frequencies weakened the output due to insufficient capacitor charging and discharging time. Experimentally, the output RMS voltage increased from 71.1 V to 94.1 V, with operating frequencies between 343.8 Hz and 613.7 Hz. Keywords—Electroluminescent display, H-bridge inverter, inverter circuit, MIC4832. I. INTRODUCTION RINTED electroluminescent (EL) displays have gained increasing attention in wearable and flexible electronic systems due to their lightweight structure, flexibility, low power consumption, and uniform illumination[1], [2]. Unlike conventional light-emitting devices, EL displays can be integrated into deformable substrates, enabling applications in smart textiles, wearable safety devices, signage, automotive dashboards, and display backlighting [2]. However, despite these advantages, EL displays require high-voltage alternating current (AC) excitation to activate the phosphor layer and emit light effectively [3], [4]. Most portable electronic systems operate using low- voltage direct current (DC) sources, typically between 1 V and 5 V [5]. Therefore, a DC-AC inverter circuit is required to convert the available low-voltage DC supply into the high- voltage AC signal needed by the EL display. Conventional EL inverter circuits are often bulky and rigid, making them unsuitable for flexible and compact applications [5], [6]. In addition, the use of transformers and complex switchingconfigurations increases system size and power consumption [6]. Previous studies have investigated various inverter topologies for EL applications, including half- bridge and H-bridge inverters. H-bridge inverters have shown advantages in reducing control complexity and power consumption while maintaining the required AC output voltage and operating frequency for EL devices [5]. For example, Sidiropoulos and Siskos proposed an integrated H-bridge inverter capable of converting 1–5 V DC input into 42.22–311.8 V peak-to-peak AC output at frequencies ranging from 60 Hz to 2 kHz [5]. Similarly, Boonyaroonate and Mori developed a compact inverter capable of generating 200 V peak-to-peak output at 500 Hz for electroluminescent lamp applications [6]. The objectives of this study are to simulate the inverter circuit using PSPICE and experimentally evaluate the output voltage and operating frequency to determine the suitability of the inverter for printed electroluminescent displays. II. LITERATURE REVIEW Electroluminescent (EL) displays have attracted considerable attention in flexible and wearable electronics due to their lightweight structure, uniform illumination, low heat generation, and low power consumption [1], [7]. Unlike light-emitting diodes (LEDs), EL displays operate through electroluminescence, where phosphor materials emit light when subjected to a strong alternating electric field [3]. The brightness and color performance of EL devices are highly dependent on the magnitude and frequency of the applied alternating current (AC), with operating frequencies typically ranging between 300 Hz and 2 kHz [2]. Due to their thin and flexible structure, EL displays have been widely applied in smart textiles, safety devices, automotive dashboards, signage, and wearable electronics [8]. Despite these advantages, EL displays require high-voltage AC excitation, whereas portable electronic systems commonly operate using low-voltage direct current (DC) power sources. Consequently, inverter circuits are required to convert low- voltage DC into high-voltage AC suitable for EL operation [5]. Conventional EL inverter circuits commonly employ transformer-based architecture, but these designs are often bulky, rigid, and unsuitable for compact or flexible systems [5], [6]. Furthermore, high-frequency switching and large Experimental Testing of a Flexible Inverter Circuit for Printed Electroluminescent Display Ajlaa Fareesya Jamaludin, Ahamd Fairuz Mohd Mansor and Aliza Aini Md Ralib Department of Electrical and Computer Engineering, Kulliyyah of Engineering, Intern

passive components increase circuit complexity and power consumption, limiting their integration into printed and wearable electronic devices [6]. Several inverter topologies have been investigated for EL applications, including half-bridge, full-bridge, and H-bridge configurations. Boonyaroonate and Mori [6] proposed a compact DC-AC inverter for large electroluminescent lamps based on a push-pull inverter concept capable of generating 200 V peak-to-peak output at 500 Hz with approximately 78% efficiency. However, the design relied on transformer- based switching, which may increase implementation complexity for flexible systems. Similarly, Sidiropoulos and Siskos [5] developed an integrated H-bridge inverter for electroluminescent lamps using sinusoidal pulse width modulation (SPWM), producing output voltages between 42.22 V and 311.8 V peak-to-peak at frequencies from 60 Hz to 2 kHz. The H-bridge topology demonstrated lower power consumption and reduced control complexity using complementary PMOS and NMOS switching devices. In addition to inverter circuit development, previous studies have also investigated printed and flexible electroluminescent displays. Yang et al. [2] reported that alternating current electroluminescent devices provide excellent flexibility and mechanical durability, making them suitable for deformable displays. However, challenges remain in designing compact inverter circuits capable of generating stable high-voltage AC outputs while maintaining compatibility with low-voltage portable systems. Therefore, this study focuses on the experimental of a compact inverter circuit for printed electroluminescent display applications. An H-bridge-based inverter approach is investigated through simulation and experimental validation to evaluate its capability in generating suitable AC voltage and frequency for printed EL operation. III. METHODOLOGY The methodology of this work consisted of literature review, inverter circuit simulation, design specification analysis, and experimental testing. Initially, a literature review was conducted to study the operating principles of electroluminescent displays and suitable inverter topologies for generating high-voltage AC output. A. Simulation The first stage involved simulation in PSPICE to evaluate the inverter circuit behavior and analyze its performance under varying operating conditions. An H-Bridge inverter topology was selected due to its reduced control complexity and lower power consumption for electroluminescent applications [5]. The EL display was represented as an equivalent capacitive load, since previous studies reported that electroluminescent devices exhibit dominant capacitive characteristics during AC operation [3], [9]. The simulation was conducted by varying the capacitance values between 10 nF, 50nF and 100 nF to represent the electroluminescent load. The operational frequency was also varied between 60 Hz, 500Hz, 1kHz, and 2 kHz to evaluate the influence of frequency on the output waveform and voltage generation. The inverter performance was analyzed in terms of output voltage waveform stability and operating frequency suitability for EL applications. Fig. 1 PSPICE simulation circuit of the H-Bridge inverter used to drive the electroluminescent load. TABLE 1 SIMULATION RESULTS OF H-BRIDGE INVERTER PERFORMANCE AT DIFFERENT CAPACITANCE AND FREQUENCY VALUES B. Hardware Implementation After the simulation stage, a hardware prototype was developed to validate the inverter operation experimentally. The MIC4832 electroluminescent driver IC was chosen because it integrates the voltage boosting stage, oscillator, and inverter function into a single device, reducing the need for complex external circuitry [5]. Once the circuit produced the expected output, the design was transferred to a board and soldered to obtain a more stable and reliable setup for testing. The completed hardware implementation is presented in Fig. 2.Department of Electrical & Computer EngineeringFYP Proceedings49

Fig. 2 Hardware prototype of the MIC4832-based inverter circuit with labelled external components used for experimental testing. The circuit was implemented using external components based on the recommended driver configuration, including a 220 μH inductor for voltage boosting, a 2.2 μF input capacitor, a 2.2 nF output capacitor, a BAV20 diode, and resistor values of 330 kΩ, 1.5 MΩ, 200 kΩ, and 2 kΩ to support the operating characteristics of the inverter. Due to component availability, a 2.2 nF capacitor was used in place of the recommended 2.7 nF output capacitor. Despite this adjustment, the circuit was able to generate measurable AC output suitable for experimental evaluation. C. Experimental Testing After the simulation stage, a hardware prototype was developed to validate the inverter operation experimentally. The MIC4832 electroluminescent driver IC was selected as the main component because it integrates the voltage boosting stage, oscillator, and inverter function into a single device, reducing the need for complex external circuitry. The developed inverter circuit was evaluated through experimental testing under different DC input voltages. The input supply was adjusted between 1.5 V and 5 V to investigate the effect of input voltage variation on the inverter output characteristics. The testing focused on measuring the generated AC output voltage and operating frequency, as both parameters are important for the proper operation of electroluminescent displays. Fig. 3 Experimental setup used for inverter circuit testing, consisting of the DC power supply, multimeters, and testing circuit. The output waveform and operating frequency were monitored using an oscilloscope, while the output voltage was measured in RMS form for analysis. The experimental findings were subsequently compared with the simulation results to assess the performance of the proposed inverter circuit for printed electroluminescent display applications. IV. RESULT AN DISCUSSION A. Simulation Results The inverter circuit was first evaluated through PSPICE simulation to investigate the effect of capacitance and operating frequency on the output waveform. Since electroluminescent displays exhibit capacitive behavior during AC operation, capacitance values of 10 nF, 50 nF, and 100 nF were used to represent the EL [10], [11] The simulation results showed that the output response varied significantly with operating frequency. At 60 Hz, the output waveform demonstrated gradual charging characteristics, while frequencies of 1 kHz and 2 kHz resulted in a significant reduction in output voltage, approaching nearly 0 V. This behavior occurred because the capacitor had insufficient charging and discharging time at higher frequencies. Among the tested conditions, 500 Hz produced the most stable square-wave output for all capacitance values, generating an output voltage close to ±100 V. In addition, larger capacitance values exhibited stronger charging behavior due to increased charge storage capability. The overall simulation findings indicate that 500 Hz is the most suitable operating frequency for stable inverter performance in electroluminescent display applications. Fig. 4 Stable square-wave output waveform generated by the H-Bridge inverter at 500 Hz using a 10 nF capacitive load. Fig. 5 Stable square-wave output waveform generated by the H- Bridge inverter at 500 Hz using a 50 nF capacitive load. Fig. 6 Stable square-wave output waveform generated by the H-Bridge inverter at 500 Hz using a 100 nF capacitive load.Time 0s 2ms 4ms 6ms 8ms 10ms 12ms 14ms 16ms 18ms 20ms V(M2:DRAIN,V3:-) V(VDC:+,0) -100V 0V 100VTime 0s 2ms 4ms 6ms 8ms 10ms 12ms 14ms 16ms 18ms 20ms V(C1:1,V3:-) V(VDC:+,0) -100V 0V 100VTime 0s 2ms 4ms 6ms 8ms 10ms 12ms 14ms 16ms 18ms 20ms V(M2:DRAIN,V3:-) V(VDC:+,0) -100V 0V 100VDepartment of Electrical & Computer EngineeringFYP Proceedings50

Figs. 4–6 show the simulated output waveforms at 500 Hz using 10 nF, 50 nF, and 100 nF capacitive loads, respectively. The output waveform remained stable and reached approximately ±89.5 V, indicating effective inverter operation within this frequency range. In addition, the capacitance value influenced the charging characteristics of the simulated load. Higher capacitance values exhibited stronger charging behavior due to increased charge storage capability. However, despite the difference in capacitance, all tested values demonstrated stable inverter performance at 500 Hz. B. Experimental Results Experimental testing was conducted to evaluate the output performance of the developed inverter circuit under different DC input voltages. The generated output waveform was observed using an oscilloscope, while the RMS voltage and operating frequency were recorded for analysis. The results showed that increasing the input voltage caused a corresponding increase in the generated RMS output voltage. The measured RMS voltage ranged from approximately 71.1 V to 94.1 V, while the operating frequency varied between 343.8 Hz and 613.7 Hz. Fig. 7 shows the measured square-wave output waveform captured using the oscilloscope. Fig. 7 Oscilloscope display showing the square-wave output, RMS voltage, and operating frequency of the inverter circuit. The generated operating frequency falls within the practical range commonly used for electroluminescent devices, which typically operate between 60 Hz and 2 kHz [5]. Although the experimental frequency varied with input voltage, the circuit consistently generated measurable high-voltage AC output suitable for EL operation. The measured RMS output voltage and operating frequency are summarized in Table II. Overall, the experimental findings demonstrate that the inverter circuit was capable of generating high-voltage AC output from a low-voltage DC source, confirming its suitability for printed electroluminescent display applications. TABLE II EXPERIMENTAL RESULTS OF INVERTER OUTPUT PERFORMANCE UNDER DIFFERENT INPUT VOLTAGES Fig. 8 Effect of input voltage on the RMS output voltage and operating frequency of the inverter circuit. Fig. 8 shows the effect of input voltage on the RMS output voltage and operating frequency of the inverter circuit. As the input voltage increased, the RMS output voltage generally showed an increasing trend, rising from approximately 85.7 V at 2 V input to 94.1 V at 5 V, indicating improved voltage generation at higher supply levels. In contrast, the operating frequency did not follow a consistent trend and fluctuated throughout the testing range, with values remaining between approximately 343.8 Hz and 613.7 Hz. Despite these variations, the measured frequencies remained within the practical operating range required for electroluminescent devices. V. CONCLUSION AND FUTURE WORK This study presented the simulation and experimental evaluation of an inverter circuit for printed electroluminescent (EL) display applications. PSPICE simulation was conducted to analyze the effect of capacitance and operating frequency on inverter performance using capacitive loads ranging from 10 nF to 100 nF. The results demonstrated that an operating frequency of 500 Hz produced the most stable square-wave output, reaching approximately ±89.5 V, while higher frequencies reduced output performance due to insufficient capacitor charging and discharging time. Experimental testing further showed that the developed inverter circuit was capable of generating high-voltage AC output from a low-voltage DCDepartment of Electrical & Computer EngineeringFYP Proceedings51

source, with RMS output voltages ranging from 71.1 V to 94.1 V and operating frequencies between 343.8 Hz and 613.7 Hz. These findings are consistent with previous studies reporting that electroluminescent devices operate effectively within the frequency range of several hundred hertz to a few kilohertz and exhibit dominant capacitive characteristics during AC operation [10]. Future work may focus on improving output voltage stability and frequency consistency through circuit optimization and component refinement. Further investigation may also include testing with an actual printed electroluminescent panel to evaluate brightness performance and power efficiency under practical operating conditions. REFERENCES [1] U. Shukla, “Opto-electro & Optical Communication Electroluminescence,” vol. 11, no. 1, p. 2021, 2021, doi: 10.37591/TOEOC. [2] Z. Yang, W. Wang, J. Pan, and C. Ye, “Alternating Current Phosphors for Deformable Displays,” Oct. 21, 2020, Cell Press. doi: 10.1016/j.xcrp.2020.100213. [3] W. Kaiser, R. P. Marques, and A. F. Correa, “Light emission of electroluminescent lamps under different operating conditions,” IEEE Trans. Ind. Appl., vol. 49, no. 5, pp. 2361–2369, 2013, doi: 10.1109/TIA.2013.2260315. [4] P. Harris, D. den Engelsen, G. Fern, and J. Silver, “AC electroluminescent lamps: shedding some light on their mysteries,” Journal of Materials Science: Materials in Electronics, vol. 28, no. 10, pp. 7006–7012, May 2017, doi: 10.1007/s10854-016-5965-4. [5] Design of Circuits and Integrated Systems. IEEE, 2014. [6] I. INVERTER FOR LARGE ELECTROLUMINESCENT LAMP.” [7] D. Routlege, “INSTITUTE OF PHYSICS SERIES IN OPTICS AND OPTOELECTRONICS HANDBOOK OF ELECTROLUMINESCENT MATERIALS.” [8] “a-complex-multilayer-screen-printed- electroluminescent-watch-305g2lqbcm”. [9] R. T. Pogulaguntla and A. Karuppaswamy, “Design and Experimental Validation of a Half-Bridge Gate Driver for SiC MOSFETs.” [10] H. Patangia and E. Dervishi, “A Switched-Capacitor Inverter for Driving a Highly Capacitive Load.” [11] Y. H. Chang, “Design and analysis of power-CMOS- gate-based IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 51, no. 10, pp. 1998–2016, 2004, doi: 10.1109/TCSI.2004.835681.Department of Electrical & Computer EngineeringFYP Proceedings52

FPGA-Based Testing Platform for Semiconductor Devices Nur Ainul Natasya Izham and Amelia Wong Azman Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract — This project proposes an FPGA-based testing platform for semiconductor device characterization, specifically focusing on diodes for educational laboratory use. Traditional precision instruments like curve tracers are often cost-prohibitive for undergraduate labs, leading to a reliance on manual, error-prone measurement techniques. The developed system utilizes a hybrid hardware-software architecture on an Altera DE1 development board, interfaced with a custom stripboard and an ADC0808 converter. The platform enables rapid data acquisition across multiple test points and identifies diode states, including healthy, open, and short- circuit conditions. Experimental validation confirms that the platform can distinguish between these states with a measurement error of less than 5% compared to manual bench-top readings. This framework establishes a scalable, low-cost foundation for semiconductor diagnostics in teaching environments, offering students practical exposure to automated validation workflows. Index Terms — Altera DE1, Diode characterization, FPGA, Nios II, Semiconductor testing. I. INTRODUCTION Semiconductor performance testing is vital for ensuring circuit reliability. While industry utilizes high-speed automated test equipment (ATE), academic laboratories often rely on time-consuming manual point-by- point measurements using digital multimeters [1]. Professional parameter analyzers are typically too expensive for undergraduate access, creating a gap between academic training and industrial standards [2]. This article proposes an FPGA-based testing platform to address these limitations. By leveraging the parallel processing capabilities of FPGAs, the platform automates data acquisition and fault identification [3][4]. The primary objectives are to design a custom interface stripboard, implement processing logic on an Altera DE1 board [5], and validate system accuracy against manual measurements. II.· LITERATURE REVIEW The necessity for automated semiconductor testing arises from the complexity of modern electronic systems. While industrial environments utilize high-speed Automated Test Equipment (ATE), educational laboratories must often operate within significant budgetary and hardware constraints. A. Educational Testing Environments and Hardware Gaps The accessibility of semiconductor validation tools in academic settings reveals a substantial gap between theory and industry-standard practice. As summarized in Table I, undergraduate laboratories primarily rely on basic laboratory instruments such as multimeters and oscilloscopes for fundamental measurements [1]. While these tools are essential for teaching core concepts, they fail to provide exposure to automated data logging or real-time fault detection. Low-cost alternatives, such as Logic Built-In Self-Test (BIST) systems [3] and PSoC-based embedded instrumentation [6], have been proposed to introduce students to automated workflows. However, these often focus on digital logic or limited parameter extraction rather than high-speed parallel diagnostic tracking. TABLE I SEMICONDUCTOR TESTING TOOLS IN EDUCATIONAL ENVIRONMENTS Source Tool / Platform Measurement / testing Capability [9] Basic Laboratory Instruments Fundamental measurements (voltage, current, resistance) [13] Logic Built-In Self-Test (BIST) System Digital IC fault detection, logic verification, open/short detection [11] Low-Cost Embedded Instrumentation for Automated Semiconductor Device characterization, I-V curve measurement, semiconductor parameter extraction Hardware Software / Interface Accessibility in Education Multimeter, Oscilloscope, Power Supply, Breadboard Manual or simple data logging Widely available but limited to basic concepts; no automation or fault detection NI platform connected to laptop, DUT, PRBS generator, MISR Likely LabVIEW; allows configuration of test patterns and automated signature comparison Low-cost alternative for teaching logic testing; introduces automation and structured fault detectionDepartment of Electrical & Computer EngineeringFYP Proceedings53

PSoC 5LP development board with op- amps, VDAC, ADC Custom Diode Characterization Software for data Affordable platform that replicates some commercial lab capabilities; suitable for B. Comparative Analysis of FPGA-Based Testing Platforms Field-Programmable Gate Arrays (FPGAs) serve as a versatile solution for emulating industrial validation systems due to their parallel processing capabilities and reconfigurability. Previous research, detailed in Table II, has demonstrated the feasibility of using FPGAs as stand-alone ATE for ASIC manufacturing and out-of-production test flows [4][7]. Unlike sequential microprocessors, the FPGA’s internal architecture allows for timing-critical ADC handshaking to occur simultaneously with diagnostic processing. This project builds upon these frameworks by developing a hybrid FPGA- based platform specifically tailored for the educational stage, bridging the gap between manual multimeter probing and high- end industrial ATE. TABLE II COMPARISON OF FPGA-BASED TESTING PLATFORMS FROM PREVIOUS WORK Supporting Paper Stage Conventional method A flexible stand- alone FPGA-based ATE for ASIC manufacturing tests [20] Manufacturing Test / Validation High-End Automatic Test Equipment (ATE) A flexible FPGA- based Test Equipment for Enabling Out-of- Production Manufacturing Test Flow of Digital Systems [21] Test Flow Prototyping / Research Industrial ATE FPGA Alternative Context / Detail Low-Cost Stand- Alone FPGA-Based ATE Too expensive and inaccessible for university labs. Low-cost solution (~$1K) using commercial off-the-shelf components. Stand-alone systems collect results with minimal computer interaction Flexible FPGA- Based Test Equipment Prohibitive cost and space requirements limit use in university research labs. Combines structural tests (scan- based, BIST control) and functional applications (SLT). Modular design supports Python scripting and Hardware Abstraction Layer, can be ported to affordable boards like the PynqZ2 III. METHODOLOGY The research methodology for this project employs a systematic, bottom-up approach to develop a reliable diagnostic tool for diode characterization, progressing from theoretical modeling to hardware-in-the-loop validation. A. Overall Research Framework The systematic procedural flow adopted to achieve the project objectives is illustrated in Fig. 1. The process begins with a comprehensive literature review and component selection to establish the design constraints. As shown in the flowchart, the development then branches into two parallel paths which are the hardware fabrication of the custom stripboard and the digital logic design using Verilog HDL and Nios II. This parallel workflow ensures that the software processing layer is ready for immediate integration once the physical analog front-end is troubleshoot and validated. Fig. 1. Project methodology flowchart showing parallel hardware and software development tracks. B. Proposed System Architecture The structural division of the instrumentation setup, conceptually influenced by the functional requirements of automated testing, is detailed in the system block diagram in Fig. 2. The architecture is divided into three distinct operational stages: 1. Stage 1: Input Source: A stable 5V DC baseline. 2. Stage 2: Custom Analog Front-End: Houses the 1N4002 Diode Under Test (DUT), an ADC0808 converter, and an LC low-pass filter for signal smoothing. 3. Stage 3: Digital Processing: The Altera DE1 board implements a Verilog-based Finite State Machine (FSM) to manage the ADC clock and data bus, while a Nios II processor executes diagnostic firmware.Department of Electrical & Computer EngineeringFYP Proceedings54

Fig. 2. System block diagram illustrating the three-stage integration of power, analog conditioning, and digital processing. C. Analog Front-End and Schematic Design The complete schematic for the custom interface is presented in Fig. 3. Rather than a simple diode circuit, the design implements a multi-stage conditioning network to ensure signal integrity. The 1N4002 Diode Under Test (DUT) is placed in series with a 10Ω sensing resistor (R1) to facilitate precise current measurement (If) [8]. To simulate various device-level failure modes without physical desoldering, a modular fault injection network consisting of jumper headers JP1, JP2, and JP3 is integrated. This allows the platform to mimic "Healthy," "Open," and "Short" conditions by manipulating the electrical path of the DUT. Furthermore, an LC low-pass filter (100μH inductor and 100μF capacitor) is utilized to smooth the rectified voltage before it reaches the ADC sensing channels [9] (Vin, Vtp2, Vout), ensuring stable digital conversion. Fig. 3. Complete schematic diagram of the analog front-end, featuring the fault injection jumper network and ADC0808 interface. D. Fault Injection Mechanism and Jumper Configuration The versatility of the diagnostic platform lies in its ability to simulate various device-level failure modes without the need for physical component desoldering. This is achieved through a modular fault injection network integrated directly into the analog front-end. As shown in the schematic (Fig. 3), the electrical path of the 1N4002 Diode Under Test (DUT) is manipulated using three jumper headers: JP1 (Series), JP2 (Shunt), and JP3 (Filter). By manually configuring these jumpers, the user can create specific electrical signatures that mimic real-world semiconductor defects. The mapping of these physical jumper states to the corresponding diagnostic conditions is consolidated in TABLE III. TABLE III DUT FAULT INJECTION AND DIAGNOSTIC STATES E. Circuit Signal Flow During operation, the input voltage flows through the following stages: 1. Fault Injection Network: The signal enters through JP1 and JP2, which determine if the DUT is active, bypassed, or disconnected. 2. Current Sensing: A low-value series resistor (R1, 10Ω) limits inrush current and allows the Nios II processor to calculate the loop current (If) by measuring the voltage drop across it. 3. Conditioning and Filtering: The signal passes through an LC low-pass filter (100μH inductor and 100μF capacitor) to remove high-frequency switching noise or power supply ripple, ensuring a stable DC level for the ADC. 4. Digital Conversion: The ADC0808 converter samples analog potentials at three critical test points (Vin, Vtp2, and Vout) and transmits this data via an 8- bit parallel bus to the Altera DE1 board. To ensure hardware safety, all jumper configurations must be performed while the power supply is OFF to prevent transient voltage spikes from damaging the sensitive FPGA I/O pins.Department of Electrical & Computer EngineeringFYP Proceedings55

IV. RESULT AND ANALYSIS The platform's tracking accuracy was evaluated across simulation, manual benchmarking, and centralized FPGA hardware tracking stages. A. Numerical Validation and Accuracy The integrated platform successfully captured the non-linear forward bias region of the 1N4002 diode. As detailed in TABLE IV, the Altera DE1 platform tracked the diode voltage drop (Vd) at 0.62V and loop current (If) at 4.31mA. When compared to the manual benchmark of 0.65V and 4.33mA, the platform [5] exhibited a maximum measurement error of only 4.62%. This falls within the established 5% tolerance limit, proving that the hybrid hardware-software architecture functions reliably for multi-point data acquisition. TABLE IV TABLE 1 MASTER VALIDATION MATRIX (NORMAL OPERATION) Test CondiƟon Parameter SimulaƟon Preliminary Baseline (V / mA) Normal OperaƟon VTP1 5.00 VTP2 4.22 VTP3 4.22 IF 4.22 VD 0.74 MulƟmeter Benchmark (V / mA) DE1 Altera Plaƞorm (V / mA) Plaƞorm vs Benchmark Error (%) 5.04 5.00 0.79 4.34 4.33 0.23 4.33 4.31 0.46 4.33 4.31 0.46 0.65 0.62 4.62 B. Real-Time Hardware Status and User Interface To provide immediate diagnostic feedback, the Nios II processor [10] manages a multi-layered user interface. As illustrated in Fig. 4, the 4x 7-segment HEX display provides a dynamic readout of system health. Under healthy conducting conditions, the firmware verifies the structural integrity of the P-N junction and outputs a "Good" status string. Due to board- level hardware constraints where embedded decimal points cannot be illuminated, floating-point values are shifted and presented as integer strings (e.g., "d062" for 0.62V) to ensure readability. Simultaneously, the onboard green LED bank illuminates to signify a verified healthy state, while specific red LEDs are reserved for fault flags. Fig. 4. Physical DE1 hardware output displaying a verified "Good" status and active LED indicators during healthy diode conduction. interface. C. Visual Telemetry Dashboard Beyond the physical board displays, the system provides a comprehensive Visual Telemetry Dashboard streamed via a JTAG UART interface to a host console. As shown in Fig. 5, this console provides a real-time parameter breakdown, including the raw 24-bit parallel bus hex code and individual channel potentials. This centralized logging capability allows for rapid diagnostic checks across multiple internal nodes without the need for manual probe relocation, effectively accelerating the extraction of device characteristics in a laboratory environment. Fig. 5. Serial terminal interface console window displaying real-time system parameters and diagnostic status. V. CONCLUSION This project has successfully demonstrated the feasibility and performance of an FPGA-based diagnostic platform for automated diode characterization. By achieving a measurement accuracy comparable to bench-top instrumentation with a deviation of less than 5 percent, the proposed system effectively addresses the limitations of manual probing in educational laboratory settings. The integration of hybrid hardware- software architecture, leveraging Verilog HDL for timing- critical data acquisition and a Nios Il processor for real-time diagnostic computation, provides a robust, scalable, and cost- effective solution for semiconductor testing. These findingsDepartment of Electrical & Computer EngineeringFYP Proceedings56

confirm the platform's potential as a valuable educational tool, offering a repeatable and efficient alternative to high-cost industry-standard analyzers. Future work will focus on expanding the diagnostic capabilities to include active semiconductor devices such as BJTs and MOSFETs, as well as optimizing the modular logic for implementation on more compact and portable FPGA development platforms. REFERENCES [1] T. Nethra Devi and T. Raghuram, “Investigating semiconductor device characterization in educational labs,” Int. J. Electr. Eng. Educ. [2] T. V. S. N. Vuddagiri, C. Baldwin, and S. Das, “Low-Cost Embedded Instrumentation for Automated Semiconductor Diode Characterization,” in Proc. SoutheastCon 2025, 2025, pp. 944–949. [3] S. Demidenko, M. T. Chew, B. T. Nguyen, M. P.-L. Ooi, and Y. C. Kuang, “Logic Built-In Self-Test Instrumentation System for Engineering Test Technology Education,” in Proc. 2020 IEEE Int. Instrum. and Meas. Technol. Conf. (I2MTC), 2020, pp. 1–6. [4] D. de Carvalho et al., “A flexible stand-alone FPGA-based ATE for ASIC manufacturing tests,” in 2018 IEEE 19th Latin-American Test Symposium (LATS). [5] P. Cabuk and T. Girasek, “Education tools for PCB thermal analysis,” in 2015 38th Int. Spring Seminar on Electronics Technology (ISSE). [6] N. di Gruttola Giardino et al., “A Flexible FPGA-Based Test Equipment for Enabling Out-of-Production Manufacturing Test Flow of Digital Systems,” in Proc. 2024 IEEE Int. Symp. Defect and Fault Tolerance in VLSI and Nanotechnology Systems (DFT), 2024, pp. 1–6. [7] Vishay General Semiconductor, “1N4001 through 1N4007 General Purpose Plastic Rectifier Datasheet,” Apr. 2020. [8] Texas Instruments, “ADC0808-N 8-Bit Analog-to-Digital Converter Datasheet,” 2013. [9] “Nios II Processor Reference Guide,” Intel Corp. [Online]. Available: https://www.intel.com. Accessed: Dec. 12, 2025. [10]“Altera DE1 Board,” Terasic Technologies. [Online]. Available: https://www.terasic.com.tw. Accessed: Dec. 12, 2025.Department of Electrical & Computer EngineeringFYP Proceedings57

Radio Frequency Energy Harvesting Wireless Communication for IoT Sensors Anas Abdulsalam Omar Abdulnabi and Khaizuran Abdullah Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract— The widespread deployment of Internet of Things (IoT) sensor nodes across smart building, environmental monitoring, healthcare, and industrial automation domains has created an urgent need for sustainable, battery-free power solutions. Battery dependency imposes significant operational costs, maintenance burdens, and environmental concerns, particularly in large-scale or remote deployments. Radio frequency (RF) energy harvesting has emerged as a technically viable alternative, capable of scavenging ambient electromagnetic energy radiated by existing Wi-Fi and Bluetooth Low Energy (BLE) wireless infrastructure and converting it into usable direct current (DC) power. This study presents a systematic MATLAB and Simulink simulation framework to evaluate RF energy harvesting performance for IoT sensor applications. Signal propagation is modeled using the Friis transmission equation, and RF-to-DC conversion efficiency is applied to estimate harvested DC power and output voltage across varying load resistances and transmission distances. Three representative operating frequencies (100 MHz, 2.4 GHz, and 5 GHz) and multiple load resistance values are examined. Results demonstrate that a maximum harvested DC power of 4.0 μW is achievable at 1 metre with a 50 Ω load from a 2.4 GHz source, sufficient to sustain the standby operation of ultra-low-power IoT microcontrollers. However, output voltage levels remain in the millivolt range across all tested conditions, necessitating voltage boosting and energy storage strategies for full sensor operation. The findings confirm the technical feasibility of ambient RF energy harvesting for battery-free IoT sensing and identify practical directions for future hardware implementation. Index Terms— RF Energy Harvesting, Battery-Free IoT, Wireless Sensor Nodes, MATLAB Simulation, Friis Transmission Equation, RF-to-DC Conversion, Bluetooth Low Energy, Wi-Fi Power Harvesting. I. INTRODUCTION ` The widespread deployment of IoT sensor nodes in smart buildings, healthcare, environmental monitoring, and industrial automation has created a pressing need for sustainable, battery-free power solutions. Batteries impose finite operational lifespans, high maintenance costs in large- scale networks, and environmental concerns from improper disposal — challenges that become increasingly severe as IoT deployments scale [1]. Radio frequency (RF) energy harvesting has emerged as a promising alternative, scavenging ambient electromagnetic energy radiated by existing Wi-Fi and Bluetooth Low Energy (BLE) infrastructure and converting it into usable DC power. Unlike solar harvesting, RF energy harvesting is independent of lighting conditions, making it particularly suitable for indoor environments [2][3]. This study implements a MATLAB and Simulink simulation framework to evaluate RF energy harvesting performance for IoT sensor applications. Signal propagation is modeled using the Friis transmission equation, and an RF-to-DC conversion efficiency parameter estimates the harvested DC power and output voltage across varying frequencies, distances, and load resistances. Results are compared with published literature to assess feasibility and identify practical limitations. The remainder of this paper is organized as follows: Section II reviews related literature; Section III describes the methodology; Section IV presents the results; and Section V concludes with recommendations for future work. II. LITERATURE REVIEW RF energy harvesting for IoT applications has been the subject of extensive investigation over the past decade. Foundational analytical work by Lu et al. [2] established that ambient RF signals from wireless communication infrastructure can be harvested using rectenna architectures and that received power levels at realistic indoor distances follow the inverse square law described by the Friis transmission equation. That survey characterized harvested power levels in the nanowatt to microwatt range and reported RF-to-DC conversion efficiencies of 10–50% at low input power levels, providing critical reference benchmarks for subsequent studies. Moloudian et al. [3] provided an extensive review of RF energy harvesting strategies relevant to Industry 4.0 and IoT applications, highlighting that Wi-Fi (2.4 GHz and 5 GHz) and Bluetooth Low Energy (BLE) at 2.4 GHz represent the most practically accessible ambient RF sources in indoor environments due to their high deployment density and near-continuous signal presence. That work also documented that performance is fundamentally constrained by low ambient RF power density, nonlinear rectifier behavior, and impedance mismatch losses at low input power levels, reinforcing the need for optimized circuit d

Ali et al. [7] demonstrated a dual-band 802.11 RF energy harvesting system targeting both the 2.4 GHz and 5 GHz Wi-Fi bands using an optimized patch antenna and impedance matching network. Their hardware prototype achieved harvested DC power in the 10–100 μW range under favorable Wi-Fi proximity conditions, highlighting the importance of multi-band rectenna design for maximizing energy extraction from co-located wireless infrastructure. Md Jamil et al. [8] developed a rectenna system using harmonic balance analysis and S-parameter optimization for a 2.4 GHz RF energy harvesting application. Their Villard voltage multiplier circuit achieved conversion efficiencies between 25% and 89% depending on the number of rectifier stages and input power level, demonstrating that multi-stage topologies can significantly improve performance in ambient RF environments. A regulated output voltage of 3 V was obtained using a power management integrated circuit (BQ25570), confirming the feasibility of producing sensor- compatible voltage levels from harvested RF energy. Commercial implementations further illustrate both the promise and limitations of ambient RF harvesting. Wiliot battery-free IoT Pixel sensors harvest energy from surrounding Wi-Fi, BLE, and cellular signals to power BLE backscatter communication for asset tracking and logistics applications [9]. Everactive industrial sensors combine thermal and RF harvesting for continuous batteryless monitoring of equipment health in industrial environments. These platforms demonstrate that full RF-powered IoT operation is achievable, though typically constrained to very short communication ranges and intermittent duty cycles. Fig. 1. General architecture of an RF energy harvesting system for IoT sensor applications (adapted from Ibrahim et al., 2022). Taken together, the literature confirms that RF energy harvesting from ambient Wi-Fi and BLE sources is technically feasible for ultra-low-power IoT applications at short distances. However, existing studies predominantly focus on hardware prototypes or specific rectifier topologies, leaving a gap in systematic simulation-based analysis that independently evaluates the interplay of frequency, distance, and load resistance on harvested power and output voltage. The present study addresses this gap through a comprehensive MATLAB and Simulink simulation framework. Table I. Comparison of RF Energy Harvesting Studies for IoT Sensor Applications. III. METHDOLOGY This study adopts a simulation-based methodology to evaluate RF energy harvesting performance for IoT sensor applications. The research workflow follows a structured six- step process as illustrated in Fig. 2, implemented using MATLAB and Simulink as the primary analysis tools. Fig. 2. Flowchart of the simulation methodology for RF energy harvesting system evaluation.Department of Electrical & Computer EngineeringFYP Proceedings59

A. Step 0 — MATLAB Simulation Environment Setup The simulation environment is initialized by installing the required MATLAB toolboxes, including the RF Toolbox, Antenna Toolbox, and Signal Processing Toolbox. The simulation workspace is configured and output logging is enabled to systematically record results across all simulation runs. B. Step 1 — System Parameter Definition All key simulation variables are defined at this stage: operating frequency, transmitted RF power, transmission distance, antenna characteristics, load resistance, and RF signal type. Two representative ambient RF sources are modeled — Wi-Fi (2.4 GHz and 5 GHz ISM bands, transmit power 10–100 mW) and Bluetooth Low Energy (BLE) (2.4 GHz, transmit power 1–10 mW) — chosen for their widespread availability in modern indoor environments. C. Step 2 — RF Propagation Modeling RF signal propagation is simulated using the Friis transmission equation to estimate the received RF power at the harvesting node: Pᵣ = Pₜ Gₜ Gᵣ (λ / 4πd)² where Pᵣ is the received RF power, Pₜ is the transmitted RF power, Gₜ and Gᵣ are the transmitting and receiving antenna gains (assumed unity, 0 dBi), λ = c/f is the signal wavelength, d is the transmission distance, and c = 3 × 10⁸ m/s. This model assumes free-space propagation and provides a tractable upper- bound estimate of received power. D. Step 3 — Received Power Sufficiency Check A decision checkpoint evaluates whether the calculated received RF power exceeds the minimum threshold required for harvesting. If the received power is insufficient, simulation parameters — specifically transmitted power and transmission distance — are adjusted and the propagation model is re- executed. This iterative loop ensures that only operationally relevant conditions are analyzed. E. Step 4 — RF-to-DC Conversion If received power is sufficient, the RF-to-DC conversion model is applied to estimate harvested DC power: P_DC = η × Pᵣ where η is the RF-to-DC conversion efficiency, varied between 10% and 50% consistent with reported rectifier performance at low input power levels [2][8]. The output voltage across the equivalent IoT sensor load is then calculated as: Vout = √(P_DC × R_L) where R_L is the load resistance. Load resistance values of 50 Ω, 500 Ω, and 5000 Ω are evaluated to examine their effect on output voltage and harvested power. F. Step 5 — Output Power and Voltage Sufficiency Check A second decision checkpoint verifies whether the estimated output power and voltage meet the basic operational requirements of ultra-low-power IoT sensors (typically 1.8–3.3 V operating voltage). If the output conditions are unsatisfactory, simulation parameters are revised and the process restarts from Step 1, enabling systematic exploration of alternative system configurations. G. Step 6 — Performance Measurement and Analysis Once acceptable output parameters are confirmed, the three primary performance metrics are recorded and analyzed: received RF power (Pᵣ), harvested DC power (P_DC), and output voltage (Vout). Separate simulations are conducted for Wi-Fi and BLE sources across transmission distances of 1– 1000 m (propagation analysis) and 1–10 m (DC power analysis). The Simulink circuit-level model additionally evaluates system behavior at representative frequencies of 100 Hz, 1000 Hz, and 10,000 Hz under four load resistance gain values (0.5, 1, 2, and 5). All results are compared with published literature benchmarks to assess consistency and practical feasibility. IV. RESULTS AND ANALYSIS A. Simulink Circuit Model and Output Voltage The RF energy harvesting system was modeled and simulated in the Simulink environment within MATLAB. The circuit (Fig. 3) consists of the following blocks connected in series: RF Source, Transmit Antenna Gain, Free Space Path Loss, Receive Antenna Gain, RF-to-DC Rectifier, Capacitor Filter, Load Resistance, and Output Voltage measurement scope. Auxiliary measurement blocks calculate and display the Received RF Power, Harvested DC Power, and instantaneous output voltage for real-time observation. Fig. 3. Simulink RF energy harvesting circuit model. Fig. 4 shows the output voltage waveform recorded by the Voltage Scope. The waveform exhibits a characteristic rising response that stabilizes within the 5-second simulation duration, confirming that the RF-to-DC rectifier and capacitor filter successfully convert the incoming RF signal into a stable DC output. This validates the correct functional operation of the Simulink model.Department of Electrical & Computer EngineeringFYP Proceedings60

Fig. 4. Simulink output voltage waveform showing successful RF-to-DC conversion. B. Frequency Analysis Table I summarizes the simulation results of the Simulink frequency analysis performed at three representative frequencies with fixed system parameters. As operating frequency increases from 100 Hz to 10,000 Hz, the received RF power increases from 0.0004735 W (−3.25 dBm) to 0.001094 W (0.39 dBm), consistent with the Friis equation relationship in which smaller wavelengths at higher frequencies result in higher received power for a given antenna aperture and distance. TABLE I: RF Energy Harvesting Frequency Analysis Results Harvested DC power increases from 8.286 μW at 100 Hz to a peak of 9.840 μW at 1000 Hz before declining marginally to 9.687 μW at 10,000 Hz. Output voltage remains stable in the 20–22 mV range across all tested frequencies, indicating effective DC smoothing by the capacitor filter. Critically, output voltage levels across all frequencies remain well below the 1.8 V–3.3 V threshold required for conventional IoT sensor operation, confirming that voltage boosting circuitry is a prerequisite for practical deployment. C. Load Resistance Analysis Table II presents the load resistance analysis results obtained at 10,000 Hz with varying R_L gain values. Both output voltage and harvested DC power increase proportionally with load resistance, consistent with the relationship Vout = √(P_DC × R_L). Output voltage increases tenfold from 11.0 mV at R_L gain = 0.5 to 110.0 mV at R_L gain = 5, while harvested DC power increases from 2.422 μW to 242.2 μW across the same range. TABLE II: Load Resistance Analysis Results at 10 kHz R_L Gain Rx RF Power (W) Output Voltage (mV) Harvested DC Power (μW) 0.5 0.001094 11.0 2.422 1 0.001094 22.01 9.687 2 0.001094 44.02 38.75 5 0.001094 110.0 242.2 This behaviour underscores the critical importance of impedance matching in RF energy harvesting circuit design. Although higher load resistance yields greater output voltage in this simplified model, real-world maximum power transfer theory dictates that matching the load impedance to the source impedance maximizes power delivery. The standard 50 Ω reference impedance in RF engineering thus represents a practically significant operating point for rectenna design. D. Received RF Power vs. Distance at Multiple Frequencies The third analysis evaluates received RF power as a function of transmission distance for three operating frequencies: 100 MHz (FM broadcast), 2.4 GHz (Wi-Fi/BLE), and 5 GHz (Wi- Fi 5 GHz), simulated using the Friis equation over distances of 1 to 1000 m. Results are presented in Fig. 7 in both dBm and μW scales. Fig. 7. Received RF power versus transmission distance at 100 MHz, 2.4 GHz, and 5 GHz. Received power decreases rapidly with distance for all three frequencies following the inverse square law. Higher frequencies experience steeper attenuation due to the quadratic frequency dependence of free-space path loss. The 100 MHz source maintains substantially higher received power at distances beyond 200 m compared to 2.4 GHz and 5 GHz. At short indoor distances (1–10 m), the 2.4 GHz and 5 GHz sources yield received power in the range of tens to hundreds of microwatts, confirming their practical relevance for indoor IoT energy harvesting. Frequency Rx RF Power (W) Rx RF Power (dBm) Output Voltage (mV) Harvested DC Power (μW) 100 Hz 0.0004735 −3.25 20.35 8.286 1000 Hz 0.0004102 −3.87 22.18 9.840 10,000 Hz 0.001094 0.39 22.01 9.687Department of Electrical & Computer EngineeringFYP Proceedings61

E. Harvested DC Power vs. Distance for Varying Load Resistances The fourth analysis presents the harvested DC power as a function of transmission distance at 2.4 GHz for three load resistance values: R_L = 50 Ω, 500 Ω, and 5000 Ω, across 1 to 10 m. Fig. 8 and Table III present the complete results. Fig. 8. Harvested DC power versus distance for R_L = 50 Ω, 500 Ω, and 5000 Ω at 2.4 GHz At 1 m, the maximum harvested DC power of 4.0 μW is achieved at R_L = 50 Ω, sufficient to sustain the standby or sleep-mode operation of ultra-low-power IoT microcontrollers such as the Texas Instruments CC2640 and STMicroelectronics STM32L0, which typically consume 1–10 μW in low-power states. By 10 m, harvested power drops to 0.02 μW at R_L = 50 Ω, below the operational threshold even for standby mode. The 10 dBm difference between each decade of load resistance confirms the logarithmic power-impedance relationship. These results are consistent with ambient RF power levels reported by Moloudian et al. [3] for Wi-Fi and BLE sources, validating the simulation approach. V. CONCLUSION This study presented a comprehensive MATLAB and Simulink simulation framework to evaluate the feasibility of RF energy harvesting as a sustainable power source for battery-free IoT sensor nodes. Using the Friis transmission equation for RF propagation modeling and a parameterized RF-to-DC conversion efficiency model, the system performance was characterized across multiple operating frequencies, transmission distances, and load resistance values. The simulation results demonstrate that a maximum harvested DC power of 4.0 μW is achievable at 1 metre from a 2.4 GHz Wi-Fi source with a 50 Ω load, sufficient to sustain the standby operation of ultra-low-power IoT microcontrollers at very short distances (1–3 m). The received power decays sharply with distance, limiting practical harvesting to near- source deployments. While higher frequencies (2.4 GHz and 5 GHz) exhibit greater path loss relative to lower frequencies (100 MHz), they represent the most widely available ambient RF sources in modern indoor environments. Output voltage levels remain consistently in the millivolt range under all simulated conditions, far below the 1.8–3.3 V threshold required for direct IoT sensor operation, highlighting the critical need for voltage boosting and energy storage circuits. The primary limitations of this study include the assumption of ideal free-space propagation, abstraction of rectifier behavior through a fixed efficiency parameter, and the absence of hardware validation. Future work should incorporate realistic indoor propagation models (e.g., log-distance path loss), detailed circuit-level rectifier simulation using RF Blockset in MATLAB, integration of voltage boosting and power management ICs such as the TI BQ25570, and physical prototype fabrication for experimental validation. Additionally, multi-source hybrid harvesting combining RF, solar, and thermal sources presents a promising avenue for improving the reliability and power output of fully autonomous, battery-free IoT sensor nodes. REFERENCES [1] T. Sanislav, G. D. Mois, S. Zeadally, and S. C. Folea, "Energy harvesting, wireless communications and rechargeable sensors for Internet of Things," IEEE Access, vol. 9, pp. 39530–39549, 2021. [2] X. Lu, P. Wang, D. Niyato, D. I. Kim, and Z. Han, "Wireless networks with RF energy harvesting: A contemporary survey," IEEE Commun. Surveys Tuts., vol. 17, no. 2, pp. 757–789, 2015, doi: 10.1109/COMST.2014.2368999. [3] G. Moloudian, M. Hosseinifard, S. Kumar, R. B. V. B. Simorangkir, J. L. Buckley, C. Song, G. Fantoni, and B. O'Flynn, "A study of RF energy harvesting strategies for battery-free wireless sensing, Industry 4.0, and the Internet of Things," IEEE Sensors J., vol. 24, no. 5, pp. 5732–5745, Mar. 2024, doi: 10.1109/JSEN.2024.3352402. [4] H. H. Sherazi, L. A. Grieco, and G. Boggia, "A comprehensive review on energy harvesting MAC protocols in WSNs," Sensors, vol. 18, no. 12, p. 4097, 2018. [5] M. Ibrahim, H. Ndoye, and T. Alquthami, "Ambient RF energy harvesting for IoT applications," IEEE Internet Things J., vol. 9, no. 15, pp. 12400–12415, 2022. [6] D. Mishra, S. De, S. Jana, S. Basagni, K. Chowdhury, and W. Heinzelman, "Smart RF energy harvesting communications: Challenges and opportunities," IEEE Commun. Mag., vol. 53, no. 4, pp. 70–78, Apr. 2015, doi: 10.1109/MCOM.2015.7081078. [7] A. Ali, R. Eid, D. E. Manaseer, H. K. AbuJaber, and A. Ware, "Dual-band 802.11 RF energy harvesting optimization for IoT devices with improved patch antenna design and impedance matching," Sensors, vol. 25, no. 4, Art. 1055, 2025, doi: 10.3390/s25041055. [8] M. N. B. Md Jamil, M. Omar, R. Ibrahim, K. Bingi, and M. Faqih, "Rectenna system development using harmonic balance and S- parameters for an RF energy harvester," Sensors, vol. 24, no. 9, Art. 2843, 2024, doi: 10.3390/s24092843. [9] Wiliot, "Wiliot IoT Pixel – Battery-free Bluetooth sensing platform," [Online]. Ava

Energy Harvesting System using Wi-Fi Signal Muhammad Aiman Zulhakim Mohd Hider and Ahmad Fadzil Ismail Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract — This paper presents the design and empirical evaluation of an ambient radio frequency (RF) energy harvesting system optimized for the 2.4 GHz Wi-Fi spectrum to address the chemical waste and maintenance constraints of traditional Internet of Things (IoT) batteries. The proposed system features a custom-engineered rectenna architecture utilizing fast-switching Schottky diodes configured as a voltage doubler network to capture and convert low-density electromagnetic waves into usable direct current (DC) electricity. Experimental methodology focused on a comparative performance analysis before and after integrating a high-gain directional antenna across varying distance intervals from a standardized Wi-Fi source. The empirical results demonstrate that the inclusion of the directional high-gain antenna yielded a fivefold increase in raw DC output voltage, boosting the baseline performance from 0.02V to a peak maximum of 0.11V. However, because the achieved peak voltage remained below the 1.6V–2.0V forward voltage threshold required to illuminate the system's LED indicator, the ambient RF power density proved insufficient for direct hardware operation without secondary amplification. Consequently, future iterations of this energy harvesting design strictly necessitate the integration of a dedicated power management integrated circuit (PMIC) or an ultra-low-power DC-DC step-up booster to bridge the current voltage gap and enable autonomous sensor operation. Index Terms — Agentic AI, autonomous network management, fault remediation, large language models, N8N, network operations, workflow orchestration. I. INTRODUCTION he global proliferation of smart technologies is driving an exponential surge in low-power Internet of Things (IoT) architecture and Wireless Sensor Networks (WSNs), creating an urgent need for sustainable power paradigms. Traditional reliance on finite chemical batteries introduces severe long-term challenges, including toxic electronic waste, high maintenance costs, and logistical impracticality when deploying sensors in remote or inaccessible environments. To mitigate these limitations, ambient radio frequency (RF) energy harvesting has emerged as a compelling alternative, aiming to reclaim the vast sea of invisible electromagnetic background noise generated by modern telecommunications infrastructure. Specifically, the ubiquitous 2.4 GHz Wi-Fi band represents a continuous, highly localized source of ambient radiation that remains largely unexploited due to the low power density inherent to wireless transmissions. Current low- power harvesting technologies frequently struggle to tap into this resource because standard rectification circuits lack the sensitivity required to efficiently process weak, high-frequency signals. This paper addresses these technical bottlenecks by proposing, prototyping, and testing a dedicated RF-to-DC conversion system optimized for ambient Wi-Fi capture. Fig. 1: Radio Frequency Energy Harvesting System The primary contribution of this work lies in the empirical evaluation of frontend hardware optimization to maximize weak energy capture. By comparing a baseline rectification circuit against a system integrated with a high-gain directional antenna, this study quantifies the direct impact of directional gain on power conversion efficiency across controlled distance intervals. The practical testing framework utilizes an LED indicator as a pass/fail threshold to analyse the real-world viability of powering low-voltage nodes strictly from ambient Wi-Fi signals. Through this empirical approach, the paper maps out the precise energy deficits encountered during ambient harvesting and provides a technical roadmap for secondary power management integration. The remainder of this paper is organized as follows: Section II provides a comprehensive review of related literature and circuit topologies; Section III details the hardware design and experimental methodology; Section IV evaluates the empirical findings and system limitations; and Section V concludes with strategic recommendations for future engineering optimizations. II. METHODOLOGY A. System Architecture and Circuit Design The proposed Wi-Fi energy harvesting system was designed to intercept ambient 2.4 GHz radio frequency (RF) signals and convert them into usable direct current (DC) electricity. The architecture consists of a straightforward, cable-free conversion path beginning with a UT-106UV antenna. This antenna was configured to capture electromagnetic waves from standard Wi-Fi routers, generating a small alternating current (AC) voltage at its base. To minimize signal attenuation, the antenna was interfaced directly to the main circuit board via an RP-SMA connector. Following signal ca

KBPC610 bridge rectifier, which incorporates high-speed SR3100 Schottky diodes capable of handling high-frequency oscillations. The rectifier is paired with 100pF ceramic capacitors, effectively straightening the AC signal into DC while accumulating electrical charge to multiply the ultra-low input voltage into a stable output. Fig. 2: Rectenna Circuit B. Power Management and Energy Storage To address the highly fluctuating and low-power nature of ambient Wi-Fi signals, a dual-path power management stage was integrated into the circuit. The first pathway directs the rectified current into a storage capacitor to pool the harvested energy over time. A one-way protection diode is positioned before this storage unit to prevent reverse current leakage, ensuring the accumulated charge does not dissipate back through the antenna when the ambient RF signal drops. The second pathway connects directly to a low-current LED, which acts as a practical load and functional indicator. This LED provides immediate visual confirmation of the system's operational success once the storage capacitors reach the necessary forward voltage threshold. C. Testing Procedure To systematically evaluate the real-world performance and efficiency of the rectenna prototype, a three-phase testing procedure was implemented. Initially, distance and signal source testing was conducted by positioning the system at controlled distances of 0.5m, 1.0m, 2.0m, and 3.0m from an active 2.4 GHz wireless router, where raw DC voltage output immediately following the rectifier stage was measured using a digital multimeter to assess spatial signal degradation. Next, to evaluate the charging rate, the 100pF storage capacitors were fully discharged to 0V before placing the prototype at an optimal distance from the RF source, recording voltage accumulation at 10- and 30-second intervals to calculate energy pooling efficiency. Finally, functional output verification was performed by leaving the circuit within router range to allow voltage to naturally build up to illuminate the LED indicator, followed by intentionally deactivating the RF source to verify the efficacy of the protection diode in trapping stored voltage. III. RESULTS AND DISCUSSION A. Experimental Setup The empirical evaluation of the ambient radio frequency (RF) energy harvesting system was conducted within a controlled indoor environment to preserve structural consistency and isolate testing variables. The physical setup establishes a line-of-sight (LOS) propagation path between a standardized 2.4 GHz Wi-Fi transmission source and the harvesting hardware prototype. To map the spatial performance and geometric attenuation of the system, a linear testing track was constructed using structured distance intervals set precisely at 0.5 meters, 1.0 meter, 2.0 meters, and 3.0 meters away from the emitting source. At each interval point, the harvesting hardware was securely isolated from inductive surfaces to prevent external grounding interference, ensuring all captured energy originated strictly from the background electromagnetic. The core experimental pipeline evaluates two primary operational modes to quantify the benefits of frontend hardware optimization: a baseline phase and an augmented phase. During the baseline phase, the custom-engineered rectification multiplier circuit—comprising high-speed SR3100 Schottky diodes and 100pF ceramic storage capacitors—was exposed to the Wi-Fi field without external antenna amplification. In the augmented phase, a UT-106UV high-gain antenna was interfaced with the rectifier stage using low-loss RP-SMA connectors. Real-world validation was monitored concurrently using two distinct diagnostic tools: a high- impedance digital multimeter connected across the storage terminals to record precision direct current (DC) voltage metrics, and an integrated LED indicator acting as an empirical "pass/fail" visual threshold for standalone sensor operation. B. Output Voltage Comparison A comparative analysis of the system's performance reveals a stark contrast in direct current (DC) output generation between the baseline rectifier configuration and the antenna-augmented harvesting system. Across all tested distance intervals, the integration of the UT-106UV high-gain directional antenna consistently outperformed the raw rectifier circuit, validating the theoretical necessity of frontend wave focus to combat free-space path loss. At the closest testing perimeter of 0.5 meters, the baseline rectifier circuit captured minimal ambient radiation, yielding a negligible output of approximately 0.02V. Upon connecting the high-gain directional antenna at this same distance, the system experienced a dramatic surge in energy collection, achieving its maximum peak output voltage of 0.11V, which represents a definitive fivefold increase in raw voltage generation. As the physical distance from the 2.4 GHz Wi-Fi source widened to 1.0 and 2.0 meters, geometric attenuation caused a rapid deca

C. LED Testing To evaluate the practical, real-world utility of the harvested ambient energy, a light-emitting diode (LED) was integrated into the circuit to serve as a qualitative "pass/fail" visual indicator. Unlike high-impedance digital multimeter that measure open- circuit voltage without drawing significant energy, an LED acts as a real-world application load, requiring a specific minimum forward voltage—typically between 1.6V and 2.0V—and sufficient current density to initiate photon emission. During the testing phase across all established distance intervals, the integrated LED remained completely unlit in both the baseline and antenna- augmented configurations. This absolute lack of visual activation empirically confirmed that despite the fivefold voltage boost achieved by the directional antenna, the system could not independently bridge the immense energy deficit present in the ambient 2.4 GHz Wi-Fi spectrum. The failure of the LED to illuminate provides critical diagnostic insight into the behaviour of ambient RF scavenging systems. While the open-circuit peak measurement reached 0.11V at proximity, connecting the LED load instantly exposed the low current capacity and high internal impedance of the rectification network under ambient conditions. This result mathematically demonstrates that the raw, unregulated energy harvested from everyday background Wi-Fi signals is far too weak to directly power standard electronics. Consequently, this testing phase highlights a vital engineering boundary, proving that for an ambient RF harvesting system to transition from a conceptual prototype to a functioning power supply, it must incorporate an intermediate power management layer designed to accumulate micro-scale energy over time before discharging it in regulated bursts. TABLE II LED TESTING D. Discussion The empirical findings of this study highlight a critical performance gap in ambient 2.4 GHz radio frequency (RF) energy harvesting systems. While the integration of a directional, high-gain antenna successfully demonstrated a fivefold increase in raw direct current (DC) output voltage—boosting the baseline from 0.02V to a peak of 0.11V at a proximity of 0.5 meters—the absolute power captured remains highly constrained. The rapid decay of voltage across the 1.0-meter to 3.0-meter intervals underlines the severe real-world impact of free-space path loss and geometric attenuation on ambient wireless signals. Furthermore, at the maximum perimeter, the ambient wave amplitude completely failed to overcome the internal barrier potential of the high-speed SR3100 Schottky diodes, rendering the raw rectification network entirely inactive at extended distances. IV. CONCLUSION This paper successfully demonstrated the design, prototyping, and empirical validation of an ambient radio frequency (RF) energy harvesting system optimized for the 2.4 GHz Wi-Fi spectrum, addressing the crucial need for sustainable, battery-free IoT power solutions. Through systematic indoor testing, the study verified that frontend hardware optimization plays a decisive role in weak electromagnetic wave capture. The integration of a UT-106UV high-gain directional antenna effectively counteracted free-space path loss at proximity, delivering a definitive fivefold increase in raw direct current (DC) output voltage by boosting the baseline rectifier performance from 0.02V to a peak maximum of 0.11V. However, the rapid voltage decay observed over extended distance intervals highlighted the severe geometric attenuation and physical limitations inherent to scavenging non-cooperative, low-density ambient wireless signals. Future research avenues will focus on bridging this energy gap by incorporating specialized ultra-low-power power management integrated circuits (PMICs) and DC-DC step-up boosters capable of cold-starting from sub-volt inputs. Additionally, subsequent work will explore multi-stage voltage multiplier network scaling, the implementation of hardware-level Maximum Power Point Tracking (MPPT) algorithms for optimal impedance matching, and the deployment of advanced metamaterial antennas. By combining these structural refinements with hybrid multi-source energy harvesting mechanisms (such as integrating solar or thermal transducers), future iterations can achieve the power stability required for truly autonomous wireless sensor nodes. REFERENCES [1] S. N. Arinze, E. R. Obi, S. H. Ebenuwa, and A. O. Nwajana, “RF Energy Harvesting Techniques: Applications, Recent Developments, Challenges, and Future Opportunities,” Sep. 01, 2025, Multidisciplinary Digital Publishing Institute (MDPI). doi: 10.3390/ telecom6030045. [2] Y. Luo, L. Pu, G. Wang, and Y. Zhao, “RF energy harvesting wireless communications: Rf environment, device hardware and practical issues,” Sensors (Switzerland), vol. 19, no. 13, Jul. 2019, doi: 10.3390/ s19133010. [3] M. A. Halimi, T. Khan, Nasimuddin, A. A. Kishk, and Y. M. M. Antar, “Rectifier Circuits for RF Energy Ha

Inverter Circuit for Printed Electroluminescent Control Display Sofea Nasihin Abdul Sukor, Aliza Aini Md Ralib and Ahamd Fairuz Mohd Mansor Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia, Kuala Lumpur, Abstract— Printed electroluminescent (EL) control displays are increasingly used in flexible and wearable electronics because of their thin profile, low power consumption, and uniform illumination. These displays, however, require high-voltage alternating current (AC) excitation at specific operating frequencies, while most portable systems rely on low-voltage direct current (DC) sources. To address this mismatch, this study developed an inverter circuit for printed EL applications using the MIC4832 electroluminescent driver integrated circuit. The methodology combined literature review, MATLAB/Simulink simulation, hardware prototyping, and experimental testing. Simulation work examined pulse width modulation (PWM) strategies, H-Bridge inverter operation, and the influence of capacitance and frequency on waveform quality. A hardware prototype was then implemented on a printed circuit board (PCB) and evaluated using a DC supply, digital multimeter, and oscilloscope. Experimental results confirmed that the circuit converted low-voltage DC into high-voltage AC, producing RMS voltages between 85.7 V and 94.1 V with operating frequencies from 344 Hz to 614 Hz. These findings validate the feasibility of the MIC4832-based inverter and highlight its potential for integration into future printed electroluminescent display systems. Keywords— Printed electroluminescent display, inverter circuit, MIC4832, H-Bridge inverter, pulse width modulation (PWM), high-voltage AC generation. I. INTRODUCTION Printed electroluminescent (EL) control displays are widely used in automotive dashboards, industrial control panels, consumer electronics, and wearable devices due to their thin structure, flexibility, low power consumption, and uniform illumination characteristics [1]. Unlike conventional light-emitting technologies, EL displays generate light through the excitation of phosphor materials using an alternating electric field [2]. Consequently, EL displays require high-voltage alternating current (AC) excitation at specific operating frequencies to achieve stable illumination and satisfactory brightness performance [3]. One of the main challenges associated with EL display systems is the generation of the required AC excitation voltage. Most portable electronic devices and embedded systems operate using low-voltage direct current (DC) power sources such as batteries, typically ranging from 1.5 V to 5 V. Therefore, an inverter circuit is required to convert the low-voltage DC input into a high-voltage AC output suitable for EL operation [4]. Conventional inverter circuits often require multiple power conversion stages and numerous discrete components, resulting in increased circuit size, design complexity, and manufacturing cost. These limitations reduce their suitability for compact and portable EL applications. Previous studies have investigated various inverter topologies for electroluminescent and low-power display applications, including half-bridge and H-Bridge configurations [5]. Among these topologies, the H-Bridge inverter is widely adopted because it can generate a bipolar AC output voltage while maintaining relatively simple control implementation [6]. In addition, pulse width modulation (PWM) techniques have been extensively applied in inverter systems to control switching operation and improve output waveform characteristics [7]. Commercial electroluminescent driver integrated circuits such as the MIC4832 further simplify inverter design by integrating a boost converter, oscillator, and H-Bridge inverter within a single device [8]. Despite the availability of integrated EL driver solutions, studies focusing on the practical implementation and experimental evaluation of MIC4832-based inverter circuits for printed electroluminescent control display applications remain limited. Therefore, this study aims to develop and evaluate an inverter circuit based on the MIC4832 electroluminescent driver integrated circuit. MATLAB/Simulink is used to investigate PWM generation, H-Bridge inverter operation, and the effects of capacitance and operating frequency on output waveform characteristics. The developed circuit is subsequently implemented on a printed circuit board (PCB) and experimentally evaluated by measuring its output voltage, operating frequency, and waveform characteristics. The remainder of this paper is organized as follows. Section II presents the literature review. Section III presents adopted for simulation, hardware implementation, and experimental testing. Section IV discusses the simulation and experimental results obtained from the developed inverter circuit. Finally, Section V presents the conclusions and recommendations for future work.

used to control switching devices and generate the required output waveform [7]. Commercial EL driver integrated circuits have simplified inverter design by integrating voltage boosting and AC generation functions. The MIC4832 is one such device specifically developed for EL applications [8]. The IC incorporates an internal boost converter, oscillator, and H- Bridge inverter, enabling the generation of high-voltage AC output from a low-voltage DC source using a minimal number of external components. III. METHODOLOGY The methodology adopted in this study consisted of three main stages: simulation, hardware implementation, and experimental testing. These stages were designed to evaluate the performance of an inverter circuit intended to drive a printed electroluminescent (EL) control display. A. Simulation The first stage involved simulation in MATLAB/Simulink to model the inverter circuit behavior and assess its performance. The inverter was configured as an H-Bridge with four IGBT switches, and the EL panel was represented as an equivalent RC load. Unipolar Sinusoidal Pulse Width Modulation (SPWM) was implemented to generate gate signals for each leg of the H-Bridge independently. The PWM signals were derived by comparing two 180° phase-shifted sinusoidal reference signals against a high-frequency triangular carrier wave. The simulation evaluated the effects of operating frequency and load capacitance on the inverter output waveform. Operating frequencies of 50 Hz, 60 Hz, 500 Hz, and 2000 Hz were tested using a fixed load capacitance of 10 nF, which was selected based on preliminary capacitance analysis. Key simulation outputs, including AC voltage, RMS voltage, peak voltage, and waveform characteristics, were recorded and analysed. Fig. 1. Simulink block diagram of the unipolar SPWM-based H-Bridge inverter. TABLE I. EXPERIMENTAL OUTPUT RMS VOLTAGE AND OPERATING FREQUENCY AT DIFFERENT INPUT VOLTAGES. Input Voltage (Vdc) Frequency (Hz) RMS Voltage (V) Output Peak Voltage (V) Observation 100 60 69.54 98.34 Stable waveform 100 500 93.32 131.97 Slight distortion 100 2000 73.28 103.63 Increased distortion B. Hardware Implementation Following simulation, a hardware prototype was constructed to verify circuit functionality in practice. The MIC4832 electroluminescent driver IC was used as the main component, providing integrated voltage boosting, oscillator, and H-Bridge functions. External components including a 220 μH inductor, 2.2 μF input capacitor, 2.2 nF output capacitor, diodes, and resistor network were initially assembled on a breadboard for testing. It is important to note that the output capacitor value was 2.2 nF rather than the datasheet-specified 2.7 nF due to component availability at the time of assembly. This substitution allowed verification of the IC functionality and circuit operation. The final hardware will use the recommended 2.7 nF capacitor for optimal performance. Despite this temporary adjustment, the breadboard prototype operated correctly and produced measurable AC output. The breadboard prototype enabled verification of component connections and circuit operation. Once verified, the circuit was transferred to a printed circuit board (PCB) to reduce wiring complexity, improve reliability, and provide a stable platform for experimental testing. C. Experimental Testing The final stage involved experimental testing using a DC power supply, digital multimeter, and oscilloscope. The input voltage to the inverter was varied from 2.0 V to 5.5 V. Measurements focused on output RMS voltage, operating frequency and waveform characteristics. Experimental data were recorded to evaluate the inverter’s ability to convert low-voltage DC to high-voltage AC suitable for EL driving applications. Comparisons with simulation results were made to verify model accuracy and ensure the hardware replicated the expected inverter performance. The output peak voltage was subsequently calculated from the measured RMS voltage using Equation (1). The RMS voltage values were obtained from MATLAB simulation for different frequencies and capacitor valuesDepartment of Electrical & Computer EngineeringFYP Proceedings67

Fig. 2. Experimental setup showing DC supply, oscilloscope, and PCB. This methodology ensures a comprehensive evaluation of the unipolar SPWM-controlled H-Bridge inverter, from theoretical modelling through hardware verification and experimental validation, providing a solid foundation for interpreting the results in the following sections. IV. RESULTS AND DISCUSSION A. PWM Analysis The unipolar Sinusoidal Pulse Width Modulation (SPWM) technique was implemented to generate switching signals for the H-Bridge inverter [12]. As shown in Figure 1, unipolar SPWM was used to generate gate signals. The generated gate signals control the four IGBT switches of the inverter independently, producing alternating AC voltage across the load. The PWM pulses vary in width according to the amplitude of the sinusoidal reference waveform. The gate signals produced by the SPWM controller are illustrated in Fig 3. Each inverter switch receives a complementary gate signal to prevent simultaneous conduction within the same leg of the H-Bridge. The waveform shapes confirm the correct timing sequence for AC generation. Fig. 3: Generated gate signals for IGBT switches S1–S4. B. H-Bridge Inverter Operation The H-Bridge inverter converts the DC input voltage into a high-voltage AC output [6]. Figure 4.3 shows the H-Bridge schematic with labelled current paths for the positive and negative half-cycle operation. During the positive half-cycle, switches S1 and S3 conduct current, whereas S2 and S4 remain OFF. During the negative half-cycle, S2 and S4 conduct while S1 and S3 remain OFF. This alternating switching produces the AC output required for EL excitation. Fig. 4. H-Bridge inverter schematic and current paths during positive and negative half-cycles. Fig 5. represents the inverter output waveform at a DC input of 100 V and an operating frequency of 60 Hz. The AC waveform alternates between positive and negative voltage levels, confirming the correct operation of the H-Bridge inverter. Fig. 5. Simulated H-Bridge output waveform at 60 Hz. C. Effect of Capacitance The effect of different capacitance values on the output waveform was investigated at frequencies of 50 Hz, 60 Hz, 500 Hz, and 2000 Hz. Capacitance values of 1 nF, 10 nF, 50 nF, and 100 nF were used in the simulation. The results showed minimal changes in the waveform with varying capacitance, indicating that the inverter performance is largely independent of the selected capacitor within the tested range. Consequently, 10 nF was selected for subsequent frequency analysis, as it falls within practical values used in EL driver circuits. (a) (b) (c) (d) Fig. 6. Output waveforms for different capacitance values at 50 Hz.Department of Electrical & Computer EngineeringFYP Proceedings68

(a) (b) (c) (d) Fig. 7. Output waveforms for different capacitance values at 60 Hz. (a) (b) (c) (d) Fig. 8. Output waveforms for different capacitance values at 500 Hz. (a) (b) (c) (d) Fig. 9. Output waveforms for different capacitance values at 2000 Hz. The output waveforms shown in Fig. 6–9 illustrate the impact of varying capacitance values on the H-Bridge inverter’s performance at different operating frequencies. Each subfigure (a–d) corresponds to a capacitance of 1 nF, 10 nF, 50 nF, and 100 nF, respectively. Across all tested frequencies, the waveform shape and amplitude remain largely unchanged, indicating that the inverter output is stable and not significantly affected by the load capacitance within this range. Minor variations may appear at higher frequencies due to interactions between switching and load impedance, but these are negligible. Based on these results, a 10 nF capacitor was selected for further frequency-dependent analysis, as it represents a practical and standard value for electroluminescent driver circuits. D. Effect of Operating Frequency With a fixed capacitance of 10 nF, the inverter output was simulated at frequencies of 50 Hz, 60 Hz, 500 Hz, and 2000 Hz. Fig. 10-–13 illustrate the corresponding output waveforms. The results demonstrate that increasing frequency introduces greater waveform distortion, particularly at 2000 Hz. Similar observations have been reported in inverter studies where switching frequency significantly influences output waveform characteristics and inverter performance [13]. Stable and symmetrical waveforms were observed at lower frequencies, indicating that operating frequency has a stronger effect on waveform quality than capacitance. Fig. 10. Output waveform at 50 Hz. Fig. 11. Output waveform at 60 Hz. Fig. 12. Output waveform at 500 Hz. Fig. 13. Output waveform at 2000 Hz. E. Hardware Implementation and Experimental Results The MIC4832-based inverter prototype was developed and tested with a DC input voltage range of 2.0–5.5 V. Fig. 14 shows the breadboard prototype, and Fig. 15 illustrates the assembled PCB.Department of Electrical & Computer EngineeringFYP Proceedings69

Fig. 14. Breadboard prototype of MIC4832 inverter. Fig. 15. Assembled PCB of MIC4832 inverter. Experimental measurements of RMS voltage and operating frequency are summarized in Table II. TABLE II. EXPERIMENTAL OUTPUT RMS VOLTAGE AND OPERATING FREQUENCY AT DIFFERENT INPUT VOLTAGES Input Voltage (V) Output (Vrms) Frequency (Hz) 2.0 85.7 604.1 2.5 89.3 606.9 3.0 88.0 597.4 3.5 93.3 613.7 4.0 93.6 343.8 4.5 93.9 453.8 5.0 94.1 434.8 5.5 93.9 416.1 Fig. 16. Effect of Input Voltage on Output RMS Voltage and Operating Frequency The results confirm that the inverter successfully boosts low-voltage DC input into high-voltage AC output. RMS voltage remained approximately 85.7–94.1 V, while frequency varied from 344 Hz to 614 Hz as shown in Fig. 16. Oscilloscope observations verified stable AC waveforms, suitable for EL panel operation. F. Equation The output peak voltage ( 𝑉𝑝 ) of the inverter can be determined from the measured RMS voltage (𝑉RMS) using the standard relationship for a sinusoidal waveform. This allows the comparison of peak voltage under different operating conditions. The RMS-to-peak conversion is expressed as: 𝑉𝑝 = 𝑉RMS ⋅ √2 (1) where 𝑉RMS represents the root-mean-square voltage measured at the inverter output. This equation was used to calculate the peak output voltage for the simulated H-Bridge inverter at various operating frequencies. V. CONCLUSION AND FUTURE WORK The development and evaluation of a MIC4832-based inverter circuit for printed electroluminescent (EL) control display applications have been successfully completed. The study focused on converting a low-voltage DC input into a high-voltage AC output suitable for EL operation. MATLAB/Simulink simulations demonstrated the effectiveness of unipolar Sinusoidal Pulse Width Modulation (SPWM) in generating stable gate signals for the H-Bridge inverter. Simulation results indicated that operating frequency significantly affects waveform quality, while variations in load capacitance produced minimal impact on the inverter performance. The hardware prototype, built around the MIC4832 EL driver IC, was verified using a breadboard before PCB fabrication. Experimental evaluation with a DC power supply, digital multimeter, and oscilloscope confirmed the successful generation of high-voltage AC output, with RMS voltages ranging from 85.7 V to 94.1 V and operating frequencies between approximately 344 Hz and 614 Hz across input voltages of 2.0 V to 5.5 V. These results validate the inverter’s ability to provide the necessary electrical excitation for EL displays while maintaining waveform stability and reliability. Future work can address several areas for improvement and deeper investigation. Direct testing with a functional EL panel would allow measurement of luminance response and correlation with voltage and frequency. Additional studies could explore optimizing component values, such as the frequency-setting resistors and timing capacitors, to further improve output voltage and efficiency. Incorporating a microcontroller-based control system could enable dynamic adjustment of output parameters to suit various EL panel requirements. Finally, refining the PCB layout for compactness, noise reduction, and thermal management will enhance practical implementation and reliability in commercial EL display systems.Department of Electrical & Computer EngineeringFYP Proceedings70

VI. REFERENCES [1] X. Zhang and F. Wang, “Recent- advances in flexible alternating current electroluminescent devices,” APL Mater., vol. 9, no. 3, Mar. 2021, doi: 10.1063/5.0040109/892505. [2] L. Wang, L. Xiao, H. Gu, and H. Sun, “Advances in Alternating Current Electroluminescent Devices,” Adv. Opt. Mater., vol. 7, no. 7, Apr. 2019, doi: 10.1002/ADOM.201801154. [3] I. Saisuksaard and P. Chansri, “An Electroluminescent driven with Class E Resonant Inverter,” Proceedings of the 2022 International Electrical Engineering Congress, iEECON 2022, 2022, doi: 10.1109/IEECON53204.2022.9741573. [4] I. Sidiropoulos and S. Siskos, “An integrated DC-AC inverter for electroluminescent lamps,” Proceedings of the 2014 29th Conference on Design of Circuits and Integrated Systems, DCIS 2014, 2014, doi: 10.1109/DCIS.2014.7035564. [5] I. Boonyaroonate and S. Mori, “Half-bridge inverter using hysteresis current control for electroluminescent lamp,” IEEE Region 10 International Conference on Electrical and Electronic Technology, pp. 796–798, 2001, doi: 10.1109/TENCON.2001.949702. [6] G. C. Diyoke, I. K. Onwuka, and O. Oputa, “Development of a single-phase h-bridge inverter based on microcontroller sinusoidal pulse-width modulation scheme for residential load applications,” 2023. [7] H. Attia, H. S. Che, T. K. S. Freddy, and A. Elkhateb, “Bipolar and unipolar schemes for confined band variable switching frequency PWM based inverter,” International Journal of Electrical and Computer Engineering, vol. 11, no. 5, pp. 3763–3771, Oct. 2021, doi: 10.11591/ijece.v11i5.pp3763-3771. [8] L. Noise and E. L. Driver, “MIC4832 Typical Application Circuit Functional Block Diagram Switch Oscillator V IN EL LAMP.” [9] M. Bredol and H. S. Dieckhoff, “Materials for Powder-Based AC-Electroluminescence,” Materials 2010, Vol. 3, Pages 1353-1374, vol. 3, no. 2, pp. 1353–1374, Feb. 2010, doi: 10.3390/MA3021353. [10] A. Mishra, “Alternate Energy Sources & Technologies http://engineeringjournals.stmjournals.in/index.php/ JoAEST/index Advancing Light Emission: A Comprehensive Review of Recent Electroluminescence Material Developments,” JoAEST STM JOURNALS Journal, vol. 14, no. 2, p. 2023, 2023, doi: 10.37591/JoAEST. [11] M. G. Batarseh, Y. Samer Al-Jizawi, F. S. El-Faouri, and L. Nidal Bulbul, “A Novel Unipolar Pulse-Width Modulation Scheme for Single-Phase H-Bridge Inverters With Balanced Switch Loss Distribution,” IEEE Access, vol. 13, pp. 163235–163245, 2025, doi: 10.1109/ACCESS.2025.3608602. [12] M. Z. Aihsan et al., “Performance Analysis of Unipolar SPWM Inverter: Resistive load and Inductive load,” in IOP Conference Series: Materials Science and Engineering, IOP Publishing Ltd, Dec. 2020. doi: 10.1088/1757-899X/932/1/012074. [13] G. Ertasgin and E. Nikocevic, “Development and Application of Comprehensive Simulation Models for Current-Source Inverter Modulators,” Applied Sciences (Switzerland), vol. 15, no. 11, Jun. 2025, doi: 10.3390/app15116148.Department of Electrical & Computer EngineeringFYP Proceedings71

Simulation of a Microfluidic Device for Size-Based Separation of Microplastics Using Finite Element Analysis Hilman Hafeez bin Mohd Azmee and Nor Farahidah Za'bah Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—Microplastic pollution has emerged as a critical environmental challenge, as conventional wastewater treatment plants fail to effectively remove particles smaller than 100 μm. This paper presents the design and finite element analysis (FEA) simulation of a passive spiral microfluidic device for size-based separation of microplastic particles. A single spiral microchannel with a width of 500 μm and depth of 155 μm was designed using SolidWorks 2023 and simulated in COMSOL Multiphysics 6.1. The device leverages inertial lift forces and Dean drag forces to separate polystyrene particles of 16 μm and 20 μm in diameter without any external energy input. Validation against analytical Poiseuille flow confirmed laminar flow conditions (Re << 2000) throughout the channel. Particle tracing results demonstrated that 95% of microplastics reaching the outlets were correctly separated. However, a significant limitation was identified: approximately 80% of the 2000 released particles became trapped inside the channel, attributed to simulation mesh coarseness and geometric design constraints. Three additional channel geometries—4-spiral, 8-spiral, and serpentine—were modelled and verified for laminar flow but could not complete full particle tracing simulation due to hardware and time limitations. These findings confirm the viability of inertial spiral microfluidics as a passive, energy-free approach for microplastic separation and provide a computational basis for future experimental realization. Index Terms— microplastics, spiral microchannel, finite element analysis, inertial microfluidics, Dean flow, COMSOL Multiphysics, particle separation. I. INTRODUCTION icroplastic pollution has become one of the most pervasive environmental threats of the 21st century. Defined as plastic particles ranging from 1 μm to 5 mm, microplastics have been detected across virtually every ecosystem on Earth—from the depths of the ocean to Arctic snow, the peaks of Everest, indoor air, drinking water, seafood, and table salt [1], [2]. Their ubiquitous presence and persistence pose fundamental risks to biodiversity and human health, primarily through the bioaccumulation of toxic chemical additives along the food chain as microplastics are ingested by plankton, fish, seabirds, and mammals [3]. A critical challenge compounding this problem is the inadequacy of existing water treatment infrastructure. Conventional wastewater treatment plants (WWTPs) were designed primarily to remove dissolved nutrients such as ammonia, nitrate, and phosphorus—not solid microparticles. As a result, particles smaller than 100 μm routinely bypass primary and secondary treatment stages and enter clean water supplies [4]. Analytical instruments such as Fourier transform infrared (FTIR) spectroscopy face a minimum detection threshold of approximately 20 μm, further underestimating the true extent of smaller microplastic contamination [3]. Microfluidics offers a compelling engineering solution to this gap. By exploiting the unique physics of fluid behaviour at the microscale—particularly the dominance of viscous forces, predictable laminar flow, and size-dependent hydrodynamic interactions—microfluidic devices can achieve precise, continuous, and label-free separation of micron-scale particles [5]. Among the available techniques, passive separation methods such as spiral inertial microchannels and Deterministic Lateral Displacement (DLD) arrays are especially attractive because they require no external energy fields (electrical, magnetic, or acoustic), are low-cost to fabricate, and are highly amenable to miniaturization [6], [7]. This work presents the computational design and FEA simulation of a passive spiral microfluidic device for the size- based separation of polystyrene microplastic particles of 16 μm and 20 μm diameters. The device geometry was modelled in SolidWorks 2023 and simulated using COMSOL Multiphysics 6.1 with coupled Laminar Flow and Particle Tracing modules. The study evaluates separation efficiency, validates the laminar flow regime, and documents the performance and limitations of four candidate channel geometries: single spiral, 4-spiral, 8- spiral, and serpentine. This work serves as a proof-of-concept computational study and a foundation for future physical fabrication and experimental validation. The remainder of this paper is organized as follows. Section II reviews relevant literature on microfluidic separation techniques and microplastic characteristics. Section III describes the device geometry and simulation methodology. Section IV presents and discusses the results. Section V concludes the paper. II. LITERATURE REVIEW A. Microplastics: Characteristics

morphology, and varying density—makes them inherently difficult to capture using mechanical filtration alone [3], [7]. Once ingested by aquatic organisms, microplastics act as vectors for persistent organic pollutants and heavy metals, which bioaccumulate and biomagnify through the food web, ultimately threatening ecosystem stability and human health [9], [10]. Conventional WWTPs achieve only partial removal of microplastics, and removal efficiency drops sharply for particles below 100 μm [4], [11]. This has driven significant research interest in microfluidic separation as a complementary, high-precision downstream polishing technology for environmental water treatment. B. Principles of Microfluidics Microfluidic devices operate on channel dimensions typically between 10 μm and 500 μm, where the Reynolds number (Re = ρvD/μ) is invariably low [12]. In most microchannels Re << 2000, ensuring laminar, highly predictable flow described by the Navier-Stokes equations for incompressible Newtonian fluids [13], [14]. At this scale, surface-dependent forces (surface tension, viscosity, electrokinetic effects) dominate over bulk forces such as gravity, enabling exquisitely precise manipulation of particles without the turbulent mixing that characterizes macroscopic flow. C. Passive Separation Techniques Inertial microfluidics exploits fluid and particle inertia at intermediate Reynolds numbers (Re ≈ 1–100). In a straight microchannel, competing shear-gradient lift forces and wall- induced lift forces drive particles to size-dependent equilibrium positions across the channel cross-section. When the channel is curved into a spiral, Dean vortices—two counter-rotating secondary flow patterns generated by centrifugal pressure imbalances—superimpose a transverse drag force on the particles. The balance of the inertial lift force and the Dean drag force determines a size-dependent lateral equilibrium, allowing larger and smaller particles to migrate to distinct transverse positions and exit through separate outlets [5], [6]. The approach offers high throughput (mL/ min range), low clogging risk, and no requirement for external fields. Deterministic Lateral Displacement (DLD) arrays use periodic micropillar arrangements angled slightly to the primary flow direction. Particles larger than a geometrically determined critical diameter (DC = 1.4 × g × ε^0.48) are displaced laterally by repeated pillar interactions, enabling continuous size-based sorting with resolution down to 10 nm [13]. However, DLD devices are prone to clogging from particles larger than the inter-pillar gap and have low throughput compared to spiral microchannels [4]. Active techniques such as dielectrophoresis (DEP) offer real- time tunability by using non-uniform electric fields to polarize and deflect particles according to their dielectric properties and are particularly effective for separating microplastics from biological samples due to the negative DEP response of electrically insulating plastics in conductive buffers [17]. However, active devices require additional power infrastructure and are generally less suitable for high-volume environmental applications. TABLE I Comparative Summary of Microplastic Separation Techniques Feature Inertial (Spiral) DLD DEP Throughput High (mL/min) Low (μL/min) Low Separation Basis Size Size Dielectric constant Clogging Risk Low High Low Energy Input None (passive) None (passive) Required (active) Best Application Bulk water filtration Precision bead sorting MPs from bio-samples D. Selected Prior Studies Sajeesh and Sen [5] provided a comprehensive review of particle separation in microfluidic devices, establishing the theoretical framework for inertial and Dean-flow-based sorting that underpins the present work. McGrath et al. [13] reviewed DLD technology and demonstrated sub-10-nm size resolution. Fan et al. [6] reported optimized DLD channel geometries for improved trajectory control. Gong et al. [7] demonstrated label- free microfluidic identification of small-sized microplastics in seawater. Sun et al. [15] showed successful microplastic enrichment using derivative electric field gradients. These studies collectively establish the technical foundation and performance benchmarks against which the present simulation is validated. III. METHODOLOGY A. Project Workflow The study followed a four-phase systematic workflow. Phase 1 comprised a literature review to identify the most appropriate passive separation mechanism for the target particle size range (1–50 μm). Phase 2 involved the geometric design and 3D modelling of candidate microfluidic channels using SolidWorks 2023. Phase 3 implemented the FEA simulation in COMSOL Multiphysics 6.1, coupling Laminar Flow and Particle Tracing modules. Phase 4 evaluated separation performance and benchmarked results against published literature.Department of Electrical & Computer EngineeringFYP Proceedings73

Fig. 1. Four-phase project methodology flowchart from literature review through simulation to result evaluation. B. Device Geometry A spiral inertial microchannel was selected as the primary separation architecture based on its favourable combination of low clogging risk, passive operation, and high throughput relative to DLD and DEP alternatives (Table I). The channel was designed with a rectangular cross-section of width w = 500 μm and depth h = 155 μm, giving an aspect ratio AR = h/w = 0.31 (Equation 4). This low aspect ratio concentrates particle equilibrium positions into two stable streams rather than four, simplifying outlet collection. The channel incorporates one inlet and two outlets. Three-dimensional geometry was created in SolidWorks 2023 and exported as a STEP file for import into COMSOL. Three additional geometries were modelled for comparative study: a 4-spiral channel, an 8-spiral channel, and a serpentine channel with a turn angle of 1.75° per bend. All designs share the same cross-sectional dimensions and inlet/outlet configuration as the primary design. C. Governing Equations The complete physical framework integrates four coupled dimensionless and differential equations: Reynolds Number: Re = ρvD/μ, where ρ = 998 kg/m³ (water density), v is inlet velocity, D is hydraulic diameter, and μ = 1.002 × 10⁻³ Pa·s (dynamic viscosity). A value Re << 2000 confirms laminar flow and validates use of the Laminar Flow module. Aspect Ratio: AR = h/w, where h is channel height and w is channel width. AR determines the velocity profile shape and the number of stable particle equilibrium positions. Confinement Ratio: Cr = dp/h, where dp is the particle diameter and h is the channel's smallest dimension. Cr governs the relative magnitude of wall-induced lift forces. For dp = 20 μm, Cr = 0.129; for dp = 16 μm, Cr = 0.103. Fluid Flow: The continuity equation (∇·u = 0) and the steady- state Navier-Stokes momentum equation ρ(u·∇)u = ∇·[−pI + K] + F are solved by COMSOL to obtain the velocity and pressure fields prior to particle injection. Particle Dynamics: Particle trajectories are computed under Stokes drag force (fluid drag acting on the particle) and wall-induced lift force (repulsion from channel walls that increases with particle size and confinement ratio). D. COMSOL Simulation Setup Two COMSOL physics modules were employed: (1) Laminar Flow, solved with a no-slip wall boundary condition, a specified inlet velocity, and zero-gauge pressure at the outlets; and (2) Particle Tracing for Fluid Flow, which uses the computed velocity field to integrate particle trajectories under drag and lift forces. Material properties were sourced from the COMSOL built- in library: water (ρ = 988 kg/m³, μ = 1.002 × 10⁻³ Pa·s), polystyrene particles (ρ = 1050 kg/m³), and PDMS channel walls. The inlet velocity was set to 0.2 m/s. A total of 2000 particles (1000 per size class) were released over a simulation period of 1 second (100 particles per 0.1 s interval). Mesh density was set to coarse/coarser owing to RAM constraints; finer mesh settings exceeded available memory and produced excessive computation times. E. Validation Strategy Validation was conducted in two steps. First, the velocity field computed by COMSOL for a straight rectangular microchannel was compared with the analytical Poiseuille (fully developed laminar) parabolic profile. Agreement between simulated and analytical profiles confirmed the correctness of the Laminar Flow solver. All four proposed channel geometries must produce a parabolic velocity profile consistent with this baseline. Second, particle separation results were benchmarked against published values from the inertial microfluidics and spiral microchannel literature [5], [13]. IV. RESULTS AND DISCUSSION A. Flow Field Validation The simulated velocity field for the straight rectangular reference channel exhibited the expected standard parabolic profile in both the xy-plane and the yz-plane, confirming that the Laminar Flow module correctly solved the Navier-Stokes equations for this geometry. Maximum centreline velocity was 0.0961 m/s, consistent with analytical prediction for the specified inlet conditions. The computed Reynolds number ReDepartment of Electrical & Computer EngineeringFYP Proceedings74

= 29.8, well below the laminar-to-turbulent transition threshold of 2000, confirming laminar flow throughout. Fig. 2. xy-plane slice of the velocity magnitude field in the reference straight rectangular microchannel, showing the characteristic parabolic Poiseuille profile (max. 0.0961 m/s). Fig. 3. yz-plane view of the standard parabolic velocity field in the reference straight microchannel, confirming laminar flow and correct solver behaviour. All four proposed spiral and serpentine channel geometries also produced parabolic velocity profiles, confirming that laminar flow was maintained throughout each design. The velocity magnitude varied between designs owing to geometric path-length differences, but the parabolic form was consistent in every case. B. Single Spiral Microchannel — Separation Results Fig. 4 shows the overall 3D geometry of the single spiral microchannel as rendered in COMSOL, with the single inlet at the channel centre and two outlets at the channel periphery. Figs. 5 and 6 display the xy- and yz-plane velocity fields for this design. Fig. 4. Three-dimensional rendering of the single spiral microchannel design in COMSOL Multiphysics, showing the inlet at the channel centre and two peripheral outlets. Fig. 5. xy-plane view of the fluid velocity magnitude field in the single spiral microchannel. The parabolic profile confirms laminar flow and passes the first validation step. Fig. 6. yz-plane cross-sectional view of the velocity field in the single spiral microchannel, showing consistent parabolic distribution across the channel depth. Fig. 7 presents the particle separation trajectories for the 20 μm and 16 μm polystyrene particles. As expected from inertial microfluidic theory, the 20 μm particles (Cr = 0.129, above the 0.07 threshold) experienced a stronger net inertial lift force, migrating toward the inner wall of the spiral and consistently exiting through Outlet 2. The 16 μm particles (Cr = 0.103, closer to the 0.07 threshold) were subject to a weaker lift force: some migrated to Outlet 1 and others to Outlet 2. This differential lateral migration constitutes successful size-based separation. Fig. 7. Particle trajectory plot showing the separation of 20 μm (inner-wall migration) and 16 μm (intermediate migration) polystyrene particles in the single spiral microchannel.Department of Electrical & Computer EngineeringFYP Proceedings75

TABLE II Particle Count at Inlet and Outlets — Single Spiral Microchannel Table II summarises the particle distribution. Of 2000 particles introduced at the inlet, only 332 (16.6%) reached the outlets: 15 at Outlet 1 and 317 at Outlet 2. Critically, 83.4% of particles became trapped within the channel—at the inlet-to- channel junction, within the spiral body, or at the approach to the outlets. Among the 332 particles that did reach the outlets, the separation purity was 95%: almost all particles were directed to the correct outlet relative to their size class. The high proportion of trapped particles (>80%) is a significant limitation. The most probable causes are: (i) the coarse simulation mesh creating numerical errors in particle- wall interaction forces, particularly at the curved inlet junction and the outlet bifurcation; (ii) potential geometric discontinuities at the inlet-to-spiral transition that create non- physical stagnation zones in the particle tracer; and (iii) the simplified particle dynamics model, which does not include particle-particle interactions, Brownian motion, or accurate hydrodynamic lift force coefficients for finite-Reynolds- number inertial migration. Increasing mesh resolution and refining the inlet geometry are identified as the primary corrective measures for future work. C. Additional Channel Geometries The 4-spiral, 8-spiral, and serpentine channel designs were all successfully simulated for laminar flow and each passed the first validation step, producing parabolic velocity profiles consistent with the analytical solution. Figs. 8–13 show the xy- and yz-plane velocity fields for these three designs. Fig. 8. xy-plane fluid velocity field for the 4-spiral microchannel design, confirming laminar flow. Fig. 9. yz-plane cross-sectional velocity field for the 4-spiral microchannel design. Fig. 10. xy-plane fluid velocity field for the 8-spiral microchannel design, confirming laminar flow throughout the extended spiral path. Fig. 11. yz-plane cross-sectional velocity field for the 8-spiral microchannel design. Fig. 12. xy-plane fluid velocity field for the serpentine microchannel (turn angle 1.75°), confirming laminar flow. Location Particle Count Percentage (%) Inlet (released) 2000 100% Outlet 1 15 0.75% Outlet 2 317 15.85% Trapped in channel 1668 83.4%Department of Electrical & Computer EngineeringFYP Proceedings76

meshes are used, reinforcing the importance of mesh sensitivity studies in future work. Fig. 13. zx-plane cross-sectional velocity field for the serpentine microchannel design. Full particle tracing simulation for these three geometries could not be completed within the project timeline. As shown in Fig. 14, a single coarse-mesh particle tracing run on the 8-spiral channel required approximately 9 hours 26 minutes of computation time. Given that design iteration, error correction, and parameter sweep typically require multiple runs, the total computation demand for the three additional geometries would have spanned several days—exceeding the available project schedule. If a 2D simplification had been adopted from the outset, computation time would have been reduced by an estimated one to two orders of magnitude, potentially enabling a full comparative study. Fig. 14. COMSOL Multiphysics interface showing a last computation time of 9 h 26 min 26 s for the 8-spiral microchannel particle tracing simulation on a coarser mesh, illustrating the hardware limitation constraint. D. Comparison with Literature The 95% separation purity achieved among particles reaching the outlets is broadly consistent with reported performance of inertial spiral microchannels in the literature. Sajeesh and Sen [5] documented high-throughput separation of particles in the 10–50 μm range using spiral inertial microfluidics with comparable channel dimensions, reporting reliable lateral migration of larger particles toward the inner wall under Dean flow conditions. The direction and magnitude of particle migration observed in the present simulation—20 μm particles migrating to the inner wall and 16 μm particles exhibiting intermediate behaviour—aligns with the force- balance theory described in that work. The particle retention problem observed here is consistent with limitations noted in other 3D FEA simulations of spiral microchannels when coarse V. CONCLUSION This paper has presented the computational design and FEA simulation of a passive spiral microfluidic device for the size- based separation of microplastic particles. A single spiral microchanne l (width 500 μm, depth 155 μm) was successfully designed in SolidWorks 2023 and simulated in COMSOL Multiphysics 6.1. The simulated velocity field confirmed laminar flow (Re ≈ 29.8) consistent with the analytical Poiseuille profile, validating the correctness of the flow solver. Particle tracing simulation demonstrated that 95% of particles reaching the outlets were correctly separated by size, with 20 μm particles concentrating at the inner-wall outlet and 16 μm particles distributed across both outlets, consistent with inertial lift and Dean drag force theory. The principal limitation identified is that approximately 80% of the 2000 injected particles became trapped inside the channel, most likely due to coarse mesh settings and geometric discontinuities at the inlet-to-spiral junction. This severely constrains overall collection efficiency, and resolving it through finer meshing, refined inlet geometry, and more accurate inertial lift force coefficients is the primary priority for future work. The three alternative channel designs (4-spiral, 8-spiral, serpentine) all confirmed laminar flow but could not complete full particle tracing within the project timeline due to prohibitive computation times on the available hardware; future work should explore 2D simplification or cloud computing resources to enable comparative geometric study. Despite these limitations, the results establish that passive inertial spiral microfluidics is a computationally viable approach for microplastic separation at sub-100 μm scales—the size range that conventional wastewater treatment infrastructure fails to address. This work provides a computational foundation for the physical fabrication and experimental validation of microfluidic devices targeting environmental microplastic remediation.Department of Electrical & Computer EngineeringFYP Proceedings77

REFERENCES [1] Y. Li et al., "Potential Health Impact of Microplastics: A Review of Environmental Distribution, Human Exposure, and Toxic Effects," ACS, Oct. 2023, doi: 10.1021/envhealth.3c00052. [2] P. Lalrinfela et al., "Microplastics: Their effects on the environment, human health, and plant ecosystems," Environ. Pollut. Manag., Oct. 2024, doi: 10.1016/j.epm.2024.11.004. [3] R. C. Hale et al., "A Global Perspective on Microplastics," J. Geophys. Res. Oceans, Jan. 2020, doi: 10.1029/2018JC014719. [4] A. Mohammadimehr et al., "Review: Impact of microfluidic cell and particle separation techniques on microplastic removal strategies," J. Agric. Food Res., vol. 16, Jun. 2024, doi: 10.1016/j.jafr.2024.101124. [5] P. Sajeesh and A. K. Sen, "Particle separation and sorting in microfluidic devices: A review," Microfluid. Nanofluid., 2014, doi: 10.1007/s10404-013-1291-9. [6] D. Fan, Y. Liu, and Y. Liu, "The Latest Advances in Microfluidic DLD Cell Sorting Technology," Biosensors, Feb. 2025, doi: 10.3390/bios15020126. [7] L. Gong et al., "A microfluidic approach for label-free identification of small-sized microplastics in seawater," Sci. Rep., vol. 13, no. 1, Dec. 2023, doi: 10.1038/s41598-023-37900-9. [8] A. L. Andrady, "The plastic in microplastics: A review," Mar. Pollut. Bull., vol. 119, no. 1, pp. 12–22, Jun. 2017, doi: 10.1016/J.MARPOLBUL.2017.01.082. [9] M. Cole et al., "Microplastics as contaminants in the marine environment: A review," Mar. Pollut. Bull., vol. 62, no. 12, pp. 2588– 2597, Dec. 2011, doi: 10.1016/J.MARPOLBUL.2011.09.025. [10] B. E. Oßmann, "Microplastics in drinking water? Present state of knowledge and open questions," Curr. Opin. Food Sci., vol. 41, pp. 44–51, Oct. 2021, doi: 10.1016/J.COFS.2021.02.011. [11] M. N. Puteri et al., "Technologies to eliminate microplastic from water: Current approaches and future prospects," Environ. Int., May 2025, doi: 10.1016/j.envint.2025.109397. [12] H. S. Santana et al., "Review on microfluidic device applications for fluids separation and water treatment processes," SN Appl. Sci., Mar. 2020, doi: 10.1007/s42452-020-2176-7. [13] J. McGrath, M. Jimenez, and H. Bridle, "Deterministic lateral displacement for particle separation: A review," Lab Chip, Nov. 2014, doi: 10.1039/c4lc00939h. [14] H. A. Stone, A. D. Stroock, and A. Ajdari, "Engineering flows in small devices: Microfluidics toward a lab-on-a-chip," Annu. Rev. Fluid Mech., vol. 36, pp. 381–411, 2004, doi: 10.1146/annurev.fluid.36.050802.122124. [15] Z. Sun et al., "Microplastic separation and enrichment in microchannels under derivative electric field gradient by bipolar electrode reactions," Sci. Rep., vol. 14, no. 1, Dec. 2024, doi: 10.1038/s41598-024-54921-0. [16] H. Warraich et al., "Microfluidic Sensor for Microplastic Detection in Saline Freshwater," in Proc. Transducers, 2025, pp. 1271 –1274, doi: 10.1109/transducers61432.2025.11110378. [17] E. Ece, N. Hacıosmanoğlu, and F. Inci, "Microfluidics as a Ray of Hope for Microplastic Pollution," Biosensors, Mar. 2023, doi: 10.3390/bios13030332.Department of Electrical & Computer EngineeringFYP Proceedings78

Power and Energy T R A C K 5

Comparative Analysis of Fixed and Single Axis Solar Tracking Systems for Residential Photovoltaic Applications Shakira Shazrul Kahar and Suriza Ahmad Zabidi Department of Electrical and Computer Engineering , Kulliyyah of Engineering, International Islamic University Malaysia Abstract—The increasing adoption of residential photovoltaic (PV) systems has highlighted the need to improve solar energy utilization and overall system efficiency. Conventional fixed solar panels are unable to continuously align with the sun throughout the day, resulting in reduced solar irradiance absorption and lower energy generation. This study presents the development and evaluation of a single-axis solar tracking system for residential photovoltaic applications. The proposed system utilizes two Light Dependent Resistor (LDR) sensors, an Arduino Uno microcontroller, and an SG90 servo motor to automatically adjust the orientation of a photovoltaic panel based on sunlight intensity. A prototype was developed and experimentally tested under outdoor conditions. The performance of the solar tracking system was compared with a conventional fixed solar panel system by measuring voltage, current, and power output at hourly intervals from 8:00 AM to 6:00 PM. Experimental results showed that the solar tracking system generally produced higher electrical output than the fixed solar panel configuration, demonstrating improved solar energy collection throughout the day. The findings indicate that a simple and low-cost single-axis solar tracking mechanism can enhance the performance of residential photovoltaic systems. The proposed system provides a practical solution for maximizing solar energy utilization while maintaining affordability and implementation simplicity for residential applications. Index Terms—Arduino Uno, Light Dependent Resistor (LDR), photovoltaic system, residential solar energy, single-axis solar tracker, solar tracking system. I. INTRODUCTION he growing demand for sustainable energy sources and increasing concerns regarding environmental pollution have accelerated the adoption of renewable energy technologies worldwide. Among the various renewable energy sources available, solar energy has emerged as one of the most promising alternatives due to its abundance, environmental friendliness, and long-term sustainability. Photovoltaic (PV) systems enable the direct conversion of solar radiation into electrical energy and have become increasingly popular for residential applications. Residential photovoltaic systems are widely installed on rooftops to reduce electricity costs and decrease dependence on conventional fossil-fuel-based electricity generation [1]. 1 2 However, most residential PV installations utilize fixed solar panels that remain stationary after installation. As the sun moves from east to west throughout the day, the angle of incidence between incoming sunlight and the photovoltaic panel continuously changes. Consequently, fixed solar panels are unable to maintain optimal alignment with the sun at all times, resulting in reduced solar energy absorption and lower electrical energy generation. To overcome this limitation, solar tracking systems have been developed to automatically adjust the orientation of solar panels according to the position of the sun. By maintaining a more favorable angle between the photovoltaic panel and incoming sunlight, solar tracking systems can improve solar irradiance absorption and enhance overall system performance. Previous studies have reported that solar tracking systems can significantly increase energy generation compared to fixed solar panel systems. However, many existing tracking systems involve complex mechanisms and high implementation costs, making them less suitable for residential applications. Therefore, this study focuses on the development of a low- cost single-axis solar tracking system using an Arduino Uno microcontroller, LDR sensors, and a servo motor. The proposed system aims to improve the efficiency of residential photovoltaic systems by automatically adjusting the orientation of the solar panel based on sunlight intensity. The performance of the developed system is evaluated by comparing its voltage, current, and power output with those of a conventional fixed solar panel system. II. LITERATURE REVIEW Solar photovoltaic (PV) systems convert sunlight directly into electrical energy through the photovoltaic effect. The output performance of a PV system is strongly influenced by solar irradiance, panel orientation, temperature, and shading conditions. In residential applications, PV panels are commonly installed in a fixed position due to their simple structure, lower installation cost, and minimal maintenance requirement[2]. However, fixed solar panels cannot continuously align with the sun’s movement throughout the day, which reduces solar energy absorption during morning and evening periods. Previous studies have shown that solar tracking systems can improve

orientation of the solar panel according to the sun’s position. Solar trackers are generally classified into single-axis and dual- axis systems. Single-axis trackers rotate along one axis, usually from east to west, while dual-axis trackers rotate along both horizontal and vertical axes[2], [3]. Although dual-axis trackers generally provide higher energy output, they are more complex, costly, and require additional actuators and control mechanisms. Therefore, single-axis trackers are often preferred for low-cost and small-scale applications. Several studies have implemented solar tracking systems using Light Dependent Resistors (LDRs)[4], microcontrollers, and servo motors. LDR sensors are commonly used to detect sunlight intensity, while microcontrollers such as Arduino Uno process the sensor readings and generate control signals for the actuator. This approach is simple, affordable, and suitable for prototype development. Based on previous research, solar tracking systems generally produce higher voltage, current, and power output compared to fixed solar panel systems[5], [6]. However, many existing studies focus on large-scale solar farms, dual- axis mechanisms, or advanced control techniques. Therefore, this study focuses on a simple single-axis solar tracking system for residential photovoltaic applications[7], [8], [9], [10], [11], [12]. TABLE I COMPARISON BETWEEN SINGLE-AXIS AND DUAL- AXIS SOLAR TRACKERS III. METHODOLOGY A. System Design The proposed solar tracking system consists of a photovoltaic panel, two Light Dependent Resistor (LDR) sensors, an Arduino Uno microcontroller, an SG90 servo motor, a solar charge controller, a battery storage system, and a DC watt meter. The LDR sensors are positioned on opposite sides of the solar panel to detect differences in sunlight intensity. The sensor readings are continuously processed by the Arduino Uno, which determines the direction of maximum sunlight and commands the servo motor to rotate the solar panel accordingly. The tracking mechanism operates along a single axis, allowing the photovoltaic panel to follow the sun’s movement from east to west throughout the day. The use of a single-axis configuration was selected due to its simplicity, lower cost, and suitability for residential photovoltaic applications. Fig. 1. Block Diagram of the Proposed Single-Axis Solar Tracking System. B. Experimental Setup A prototype of the proposed solar tracking system was developed and tested under outdoor conditions. The performance of the solar tracking system was compared with a fixed solar panel system operating under the same environmental conditions. Both systems utilized identical photovoltaic panels to ensure a fair comparison. The experimental evaluation was conducted from 8:00 AM to 6:00 PM. Voltage and current readings were recorded hourly using a DC watt meter. The power output was calculated using: P=V×I (1) where: • P = Power Output (W) • V = Voltage (V) • I = Current (A) The collected data were subsequently analyzed to determine the effectiveness of the solar tracking mechanism in improving photovoltaic performance.Parameter Single-Axis Solar Tracker Dual-Axis Solar Tracker Number of Rotational Axes 1 2 Tracking Direction East-West East-West and Vertical Energy Output Higher than fixed panels Highest among tracking systems System Complexity Moderate High Number of Actuators One Two or more Installation Cost Lower Higher Maintenance Requirement Lower Higher Power Consumption Lower Higher Residential Suitability High Moderate Selected for This Project ✓ ✗Department of Electrical & Computer EngineeringFYP Proceedings80

Fig. 2. Prototype Design of Single-Axis Solar Tracking. IV. RESULTS AND DISCUSSION A. Experimental Results The performance of the proposed solar tracking system was evaluated by comparing its voltage, current, and power output with a conventional fixed solar panel system. Measurements were recorded hourly from 8:00 AM to 6:00 PM under outdoor environmental conditions. Table I presents the collected experimental data. TABLE II EXPERIMENTAL DATA FOR FIXED SOLAR PANEL AND SOLAR TRACKING SYSTEM Based on the collected data, both systems exhibited variations in electrical output throughout the testing period due to changes in solar irradiance and environmental conditions. The solar tracking system generally recorded slightly higher voltage, current, and power output values than the fixed solar panel system during most of the testing period. B. Voltage Analysis Fig. 1 illustrates the voltage comparison between the fixed solar panel and the solar tracking system throughout the day. The results indicate that the solar tracking system consistently produced slightly higher voltage values than the fixed solar panel system. During peak sunlight hours, the tracker maintained better alignment with the sun, allowing the Fig.3. Voltage comparison between fixed solar panel and solar tracking system. photovoltaic panel to receive greater solar irradiance. The highest voltage recorded by the solar tracking system was approximately 18.04 V at 3:00 PM, while the fixed solar panel recorded 17.42 V during the same period. This demonstrates the ability of the tracking mechanism to maintain improved solar exposure throughout the day. C. Current Analysis Fig. 2 presents the current comparison between both systems. Fig. 4. Current comparison between fixed solar panel and solar tracking system. The current output showed a similar trend to the voltage measurements. During the morning and midday periods, the solar tracking system generally generated higher current values than the fixed solar panel system. Higher current production indicates improved solar energy absorption due to better panel orientation. However, current values decreased significantly during the afternoon period and eventually reached zero after 2:00 PM. This behavior may be attributed to battery charging conditions and the operation of the solar charge controller, which could have limited the charging current once the battery approached full charge. D. Power Output Analysis Fig. 3 illustrates the power output comparison between the fixed solar panel and the solar tracking system.Time Fixed Voltage (V) Tracker Voltage (V) Fixed Current (A) Tracker Current (A) Fixed Power (W) Tracker Power (W) 8:00 AM 13.00 13.00 0.35 0.35 4.55 4.55 9:00 AM 13.40 13.41 0.33 0.32 4.42 4.29 10:00 AM 13.36 13.40 0.17 0.22 2.27 2.95 11:00 AM 13.47 13.50 0.25 0.28 3.37 3.78 12:00 PM 13.43 13.47 0.17 0.22 2.28 2.96 1:00 PM 13.63 13.71 0.29 0.34 3.95 4.66 2:00 PM 17.28 17.60 0.00 0.00 0.00 0.00 3:00 PM 17.42 18.04 0.00 0.00 0.00 0.00 4:00 PM 17.47 17.60 0.00 0.00 0.00 0.00 5:00 PM 17.23 17.37 0.00 0.00 0.00 0.00 6:00 PM 12.73 12.90 0.00 0.00 0.00 0.00Department of Electrical & Computer EngineeringFYP Proceedings81

Fig. 5. Power output comparison between fixed solar panel and solar tracking system. Power output is the most important performance indicator because it represents the actual electrical energy generated by the photovoltaic system. The results showed that the solar tracking system generally achieved higher power output compared to the fixed solar panel configuration. For example, at 1:00 PM, the solar tracking system generated approximately 4.66 W, whereas the fixed solar panel generated 3.95 W. This represents a noticeable improvement in energy generation due to the tracking mechanism. The improved performance can be attributed to the ability of the tracking system to continuously adjust the panel orientation according to sunlight intensity. By maintaining a more favorable angle of incidence between the solar panel surface and incoming sunlight, a greater amount of solar irradiance was converted into electrical energy. E. Comparison with Previous Studies The findings obtained in this study are consistent with previous research on solar tracking systems. Numerous studies have reported that single-axis solar trackers can improve photovoltaic energy generation compared to fixed solar panel systems. Previous studies summarized in Chapter II reported performance improvements ranging from approximately 20% to 40% depending on environmental conditions and tracker design. Similarly, the developed solar tracking system demonstrated improved electrical performance compared to the fixed solar panel configuration. Although the prototype utilized a simple Arduino-based control system and a small-scale photovoltaic panel, the results support the findings reported in previous studies regarding the effectiveness of solar tracking technology. F. Discussion The experimental results demonstrate that the proposed solar tracking system successfully achieved its intended function of automatically orienting the photovoltaic panel toward the direction of maximum sunlight intensity. The use of LDR sensors, Arduino Uno, and a servo motor enabled continuous tracking of the sun’s movement throughout the day. Compared to the conventional fixed solar panel system, the proposed tracker generally produced higher voltage, current, and power output values. These findings support the hypothesis that maintaining better alignment with the sun improves photovoltaic performance and increases solar energy collection. However, several limitations were observed during the experimental evaluation. The prototype utilized a small- scale solar panel and an SG90 servo motor, which may not accurately represent the performance of a full-scale residential installation. Furthermore, the testing period was limited to a single day of measurement, and weather conditions may have influenced the recorded results. Therefore, longer testing durations under varying environmental conditions are recommended for future studies. V. CONCLUSION This study successfully developed and evaluated a single- axis solar tracking system for residential photovoltaic applications. The proposed system utilized Light Dependent Resistor (LDR) sensors, an Arduino Uno microcontroller, and an SG90 servo motor to automatically adjust the orientation of a photovoltaic panel according to sunlight intensity. The developed prototype was capable of tracking the movement of the sun from east to west throughout the day, enabling improved exposure to solar irradiance compared to a conventional fixed solar panel system. Experimental testing was conducted under actual outdoor conditions, and the performance of the solar tracking system was compared with that of a fixed solar panel configuration. The evaluation was based on three electrical parameters, namely voltage, current, and power output. The experimental results demonstrated that the solar tracking system generally produced higher electrical output than the fixed solar panel system, indicating improved solar energy collection and photovoltaic performance. The findings of this study confirm that maintaining a more favorable alignment between the photovoltaic panel and the sun can enhance solar energy utilization. Furthermore, the developed system achieved all research objectives and demonstrated the feasibility of implementing a simple and low-cost solar tracking mechanism for residential photovoltaic applications. Future improvements may include the implementation of a higher torque actuator for larger photovoltaic panels, the development of a dual-axis tracking mechanism, and the integration of Internet of Things (IoT) technology for real- time monitoring and data analysis. These enhancements could further improve system efficiency and support the wider adoption of solar tracking technology in residential renewable energy systems. REFERENCES [1] M. von Rosing, L. Shepperson, and H. Czichos, “Beyond influence: Nurturing autonomy in organizational governance,” The Sustainability Handbook, Volume 2: The Body of Knowledge around Radical Sustaina

(IEEE), Jan. 2026, pp. 1–6. doi: 10.1109/isgtmiddleeast65737.2025.11314480. [3] M. K. Ab Rani, M. Z. Khalid, N. S. Salleh, R. M. F. T. Aziz, and M. F. Zahari, “Performance Analysis of Horizontal Single-Axis Tracker Versus Fixed Tilted Photovoltaic Systems,” in 2025 IEEE International Circuit and System Symposium, ICSyS 2025, Institute of Electrical and Electronics Engineers Inc., 2025, pp. 159–162. doi: 10.1109/ICSyS65149.2025.11379900. [4] M. H. de Sá Campos and C. Tiba, “Nptrack: A n- position single axis solar tracker model for optimized energy collection,” Energies (Basel)., vol. 14, no. 4, Feb. 2021, doi: 10.3390/en14040925. [5] Y. M. Manu, S. V. Shashikala, M. K. Hemchandru, P. Rachitha, B. N. Kavya, and D. Hongirana, “Single axis solar tracking system,” in 2nd IEEE International Conference on Advances in Information Technology, ICAIT 2024 - Proceedings, Institute of Electrical and Electronics Engineers Inc., 2024. doi: 10.1109/ICAIT61638.2024.10690503. [6] A. Khurwolah and V. Oree, “Investigating the Performance of a Dual-Axis Solar Tracking System in a Tropical Climate,” in Proceedings of the 3rd International Conference on Artificial Intelligence and Smart Energy, ICAIS 2023, Institute of Electrical and Electronics Engineers Inc., 2023, pp. 188–193. doi: 10.1109/ICAIS56108.2023.10073903. [7] K. Ananda-Rao, S. Taniselass, A. S. Rosmi, N. Thirumurugananthajiva, N. H. Baharudin, and I. Nisja, “Development of Dual-Axis Solar Tracking with IoT Monitoring System,” in Conference Proceedings -2025 9th International Conference on Man-Machine Systems, ICoMMS 2025, Institute of Electrical and Electronics Engineers Inc., 2025, pp. 268–272. doi: 10.1109/ICoMMS66553.2025.11200472. [8] Y. Liang et al., “Techno-Economic Feasibility of Off- Grid Renewable Energy Systems: A Comparative Case Study,” in IECON Proceedings (Industrial Electronics Conference), IEEE Computer Society, 2023. doi: 10.1109/IECON51785.2023.10312184. [9] M. Hammas, H. Fituri, A. Shour, A. A. Khan, U. A. Khan, and S. Ahmed, “A Hybrid Dual-Axis Solar Tracking System: Combining Light-Sensing and Time- Based GPS for Optimal Energy Efficiency,” Energies (Basel)., vol. 18, no. 1, Jan. 2025, doi: 10.3390/ en18010217. [10] A. Mohamad, H. Mhamdi, N. A. M. Amin, M. Izham, N. A. Aziz, and S. Y. Chionh, “A review of automatic solar tracking systems,” in Journal of Physics: Conference Series, IOP Publishing Ltd, Oct. 2021. doi: 10.1088/1742-6596/2051/1/012010. [11] A. Tripathi, A. K. Sree, B. Triveni, A. Rishitha, B. Malleswari, and S. Anjankar, “Implementation of IoT- Enabled Automatic Solar Power Tracking System,” in 2024 International Conference on Computing, Sciences and Communications, ICCSC 2024, Institute of Electrical and Electronics Engineers Inc., 2024. doi: 10.1109/ICCSC62048.2024.10830348. [12] C. N. Savithri, R. Nandhini, R. Harini, and C. Lekha, “Tilt Angle Determination for Solar Tracker,” in 2023 Intelligent Computing and Control for Engineering and Business Systems, ICCEBS 2023, Institute of Electrical and Electronics Engineers Inc., 2023. doi: 10.1109/ICCEBS58601.2023.10448776.Department of Electrical & Computer EngineeringFYP Proceedings83

Modelling and Analysis of Electric Vehicle Charging Infrastructure Impact on Distribution Network Mohamad Azim Shazmy Bin Mohamad Jasmy and S.M.A Motakabber Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract— The rapid growth of electric vehicle (EV) adoption causes significant operational challenges to medium-voltage distribution networks, including voltage deviation, increased power losses, and reduced hosting capacity. Uncoordinated EV charging, when combined with distributed generation such as photovoltaic (PV) systems, further complicates voltage regulation and power flow management. This study aims to analyze the impact of EV charging on a radial distribution network and to develop an effective mitigation framework using optimized reactive power compensation. The IEEE 69-bus test system is used to evaluate voltage profile and power losses under different EV penetration levels in line with the integration of PV generation. Manual capacitor placement is first assessed as a baseline solution, followed by optimization using the Lightning Search Algorithm (LSA) and its improved variant, Modified LSA (MLSA). The optimization objective reduces and minimizes total power loss and voltage deviation simultaneously. Simulation results demonstrate that the proposed MLSA-based PV and capacitor coordination significantly improves voltage stability and reduces system losses compared to conventional methods. The findings confirm that advanced metaheuristic optimization provide a robust and scalable approach for enhancing the performance of EV-integrated distribution networks. Keywords—Electric Vehicle, Distribution Networks, IEEE 69-Bus, Photovoltaic, Lightning Search Algorithm, Modified Lightning Search Algorithm I. INTRODUCTION he widespread adoption of electric vehicles (EVs) presents significant environmental benefits but creates significant technical challenges for infrastructure designed for predictable, linear loads. High-power charging networks increase the peak electricity demand and accelerate the degradation of primary substation transformers and cables. Bidirectional power flows are created by uncoordinated fast charging and highly intermittent renewable resources such as solar photovoltaic (PV) systems. Such dynamics affect the accuracy of traditional radial protection schemes and endanger the limits of the structure voltage. To mitigate these multivariable stress constraints without expensive physical line expansion, advanced utilities with a combination of active and reactive power mitigation strategies. In this paper, a coordinated optimization framework based on a Modified Lightning Search Algorithm (MLSA) is proposed for optimal capacitor allocation and sizing in an EV-PV-coupled medium- voltage grid. II. LITERATURE REVIEW A. Analysis of Existing Mitigation Approaches The technical challenges of integrating electric vehicle (EV) charging infrastructure have driven extensive research into software and hardware mitigation frameworks. Early studies focused primarily on quantifying grid bottlenecks. Research demonstrated uncoordinated residential EV charging in Malaysia rapidly accelerates transformer degradation and line overloads, though their scope was limited to low-voltage networks without proactive control [1]. There is also research that verified that high-power fast chargers inject substantial harmonic distortion and phase imbalances but offered no actionable mitigation [2]. WE establish network boundaries BY modelling A hosting capacity on 11 kV feeders [3]. But their framework relied on static rules without dynamic voltage or reactive power support. Fig.1 shows how PV penetration in the primary grid. were applied that contains EV using solar. Fig. 1. Diagram of EV + Solar + Battery Integration into the Primary Grid[4] Based on Figure 1, this study is using Simulink to model this exact grid-connected microgrid setup so I can analyze how heavy EV charging loads actually impact local distribution networks. The setup mimicking a typical urban distribution TDepartment of Electrical & Computer EngineeringFYP Proceedings84

system interfaces a 15 kV medium-voltage grid with a 0.4 kV low-voltage AC bus through a standard step-down transformer. This AC bus handles local AC loads and connects to a 480 kVA bi-directional AC-DC converter, which keeps a 653.2 V DC link stable. To help take some of the stress off the grid when multiple EVs plug in, the system integrates a 70 kW solar PV array directly into the DC bus via a DC-DC converter. This DC bus then feeds a 26.7 kW general DC load and five 40 kW Level 3 off-board fast chargers, which power up the vehicles at a maximum of 400 V and 100 A each. By building this framework in Simulink, this study focuses on capturing the real-time effects of bi-directional power flow and solar intermittency on the network's voltage profile and overall stability[4]. In Figure 2, There are 69 nodes and 68 branches in the IEEE- 69 bus distribution network. The system is intended to run at 12.66 kV of voltage and 100 MVA of capacity. The overall reactive power load is 2693.6 kVAR, and the total actual power load is 3801.4 kW. The line and load data have been sourced, with the substation identified as node one. In this investigation, an examination is conducted on a total of 100 EVs incorporating a maximum of five EVCSs equipped with numerous charging outlets. Additionally, up to five SDG units have been considered to offset power losses [5]. Fig.2. Integration of electric vehicle charging stations and solar distributed generation (SDG) with the IEEE 69 bus system [5]. To address reactive constraints directly, hardware-based Optimal Capacitor Placement and Sizing (OCPS) was explored. Using metaheuristic solvers for capacitor placement in radial systems, significantly reducing current magnitudes and ohmic line losses [6]. Nevertheless, their optimization models assumed static, conventional loads and failed to account for highly dynamic, unpredictable EV load variations. B. Research gaps A review of state-of-the-art literature reveals several important research gaps. First, software scheduling strategies and hardware-based reactive power compensation are almost universally studied in isolation, ignoring the combined impact of both on a medium-voltage feeder. Second, the optimization methodologies rarely consider the coordination of solar PV and capacitor assets under multiple different EV penetration scenarios (low, medium, and high) in a single multi-objective framework. Finally, in the case of the non-convex parameters of the OCPS problem in modern active grids, conventional metaheuristics such as GA and PSO suffer from slow convergence speeds and are prone to being trapped in local minima[7]. Physics-inspired algorithms such as Lightning Search Algorithm (LSA) and its adaptive variants are still under-explored in this domain[8]. Therefore, my research is needed to fill these important knowledge gaps. In this study, the multi-variable problems faced by the grid operators are addressed and an integrated multi-objective framework is proposed to minimize the total active power losses and mitigate the unified Voltage Deviation Index (VDI) simultaneously. This project extends the existing literature by deploying a highly efficient Modified Lightning Search Algorithm (MLSA) which includes an adaptive step- size mechanism for the optimal active solar PV injection and reactive capacitor sizes simultaneously. Simulation results on the IEEE 69-bus system validate the hybrid infrastructure under progressive penetration scenarios, offering a robust and computationally efficient solution independent of communication to maximize grid hosting capacity with voltage stability. III. METHODOLOGY Figure 3 represents the project workflow that begins by modeling the IEEE 69-bus network to analyze baseline power grid operations. Next, extra electrical loads are added to specific buses to simulate low, medium, and high EV charging station demands, which inherently increases system power losses and degrades the voltage profile. To mitigate these issues, solar PV systems are integrated to supply active power, and capacitor banks are introduced to provide reactive power support. The primary objective is to optimize the placement and sizing of these capacitor banks by comparing two metaheuristic techniques which is the standard Lightning Search Algorithm (LSA) and the Modified Lightning Search Algorithm (MLSA) . A key decision gateway evaluates whether MLSA achieves faster convergence velocity than LSA. Finally, the superior optimization algorithm is combined with solar PV integration to run a comprehensive performance analysis, validating that co-optimizing active and reactive power resources minimizes line losses and stabilizes grid voltage. Fig. 3. Mitigations WorkflowDepartment of Electrical & Computer EngineeringFYP Proceedings85

A. Distribution Network Topology and Power Flow Index Formulation Figure 4 shows the image representation of IEEE-69 bus system. The IEEE 69-bus radial distribution network benchmark, which is simulated in MATLAB environment, is the main validation model used in this study. This radial line is a typical utility network with 69 buses and 68 branches in the medium voltage level, with an uneven, non-uniform baseline load distribution, which is highly sensitive to changes in current and variations in voltage. Power flow simulations are performed iteratively (backward/forward sweep technique) to quantify network operational behavior for different scenario configurations. The system efficiency and the voltage stability are measured with the use of three performance indices: Fig. 4. IEEE-69 Bus System 1. Bus Voltage Magnitude: Checked at each separate node to verify that network voltage profiles remain within acceptable boundaries. 2. Total Active Power Loss: Calculated by summing real ohmic line losses across all distribution network branches (br): 𝑓ଵ௝ = ෍(𝑅௜ ⋅ 𝐼௜ ଶ + 𝑋௜ ⋅ 𝐼௜ ଶ) ௕௥ ௜ୀଵ (1) Where:  𝑉௝ is bus voltage magnitude.  𝑓ଵ௝ is total active power losses.  𝑅௜ is branch resistance.  𝑋௜ is branch resistance.  𝐼௜ is the corresponding branch current. 3. Average Voltage Deviation Index: Formulated to quantify the distance between the real voltage magnitude at bus j and the nominal reference value (𝑉௥௘௙ = 1.0 p.u) across all network nodes: 𝑓ଶ௝ = 1 𝑁𝑏 ෍ห1 − 𝑉௝ଶห ே௕ ௝ୀଵ (2) Where:  𝑓ଶ௝ is the average voltage deviation index  𝑉௝ is the real voltage magnitude at bus j.  𝑉௥௘௙ is the nominal reference value.  𝑁௕ is the nodes across all networks. To ensure optimal structural planning, these conflicting indices are unified into a single, weighted multi-objective fitness function (𝐹௝) to be minimized by the optimization solvers: 𝐹௝ = min[𝑓ଵ(𝑗) + 𝑓ଶ(𝑗)] (3) Where:  𝐹௝ is the fitness function.  𝑓ଵ is the total active power losses.  𝑓ଶ is the voltage deviation index across all buses. B. EV Charging Load Characterization Electric vehicle charging infrastructure is integrated into designated network buses as additional real (𝑃ா௏ ) and reactive (𝑄ா௏ ) lumped power sinks superimposed onto existing base load conditions: 𝑃௡௘௪ = 𝑃௟௢௔ௗ + 𝑃ா௏ (4) Where:  𝑃ா௏ is additional real power.  𝑃௟௢௔ௗ is original active power demand  𝑃௡௘௪ is total active power demand after EV integration 𝑄௡௘௪ = 𝑄௟௢௔ௗ + 𝑄ா௏ (5) Where:  𝑄௡௘௪ is new total reactive power.  𝑄ா௏ is reactive power demand introduced by EV charging stations.  𝑄௟௢௔ௗ is total reactive power demand of the system. To model different grid adoption phases and assess maximum asset stress during evening peak hours, three distinct penetration levels are implemented: 1. Level 1 (Low Penetration): Early-stage integration, representing an EV load addition of 2% to 20%. 2. Level 2 (Medium Penetration): Moderate grid expansion, representing a load addition of 30% to 50%. 3. Level 3 (High Penetration): Extreme saturation conditions, representing an EV peak load addition of 70% to 100%. C. Active and Reactive Compensation Infrastructure Hardware mitigation combines active power support from Distributed Generation (DG) and reactive power support from shunt capacitors. Solar PV systems are integrated at selected high-load buses and modeled at unity power factor, directly offsetting local real power demand:Department of Electrical & Computer EngineeringFYP Proceedings86

𝑃net = 𝑃load − 𝑃PV (6) Where:  𝑃net is remaining active power that must be supplied by main grid.  𝑃load is total active power demand from consumers.  𝑃PV is active power generated by photovoltaic. This local active injection reduces the current drawn from the upstream substation, lowering technical 𝐼ଶ𝑅 grid losses. To complement this, shunt capacitor banks supply local reactive power to compensate for the inductive demand of charger power electronics, rewriting the net node reactive balance as: 𝑄net = 𝑄load − 𝑄cap (7) Where:  𝑄net is the remaining reactive power that must be supplied by the grid.  𝑄௟௢௔ௗ is total reactive power demand of the system.  𝑄cap is the reactive power supplied by capacitors (capacitors banks). The baseline mitigation scenario uses manually placed capacitors at arbitrary nodes near the EV charging stations to serve as a comparative reference. D. LSA and MLSA Optimization Solvers The standard Lightning Search Algorithm (LSA) is a metaheuristic search algorithm based on the physical phenomenon of lightning step-leader propagation. Step leaders search for the path of least resistance to create a discharge channel, which corresponds mathematically to a trade-off between exploration and exploitation on a global search landscape. In this paper, a Modified Lightning Search Algorithm (MLSA) is applied to improve the search speed and solution accuracy of the optimal capacitor placement problem. Then, we propose the MLSA algorithm with an adaptive step-size parameter (𝛼) that decreases linearly over iterations to improve local optimization as time goes by. 𝛼 = 𝛼max − (𝛼max − 𝛼min) ⋅ iter max_iter (8) Where:  𝛼௠௔௫ = 2.0, Initial (maximum) value of 𝛼.  𝛼௠௜௡ = 0.1, Final (minimum) value of 𝛼.  iter is current iteration number.  max_iter is total number of iterations. Additionally, a 10% random mutation jump rate is integrated into the population loop to prevent the algorithm from getting stuck in local minima. The optimization engine controls a 6- variable space representing optimal bus coordinates and capacitor sizes (𝑄௠௜௡ = 50𝑘𝑉𝐴𝑟, 𝑄௠௔௫ = 300 𝑘𝑉𝐴𝑟) to minimize the global fitness function (𝐹௝). IV. RESULT AND DISCUSSION A. Operational Impact of Charging Infrastructure The load flow simulations showed that increasing and uncoordinated EV penetration severely degrades the stability of the radial grid. The base case scenario (no EVs, PV or capacitors) showed a stable Voltage Deviation Index (VDI) of 1.8177 and total active loss of 146.87 kW. Metrics dropped off a cliff when EV charging demand was added without coordination. For High EV Penetration (Level 3), total network power losses increased to 200.45 kW compared to the base case and the VDI increased to 2.1439, which indicates large voltage drops at the ends of long radial lines. B. Comparative Evaluation of Mitigation Frameworks Several mitigation approaches were implemented andcompared: 1. High EV + Opt. Caps: When using the highest efficiency EV along with optimal capacitors, the result is the correct global optimum as identified by conducting an exhaustive search of the line network and is also represent as the benchmark against other optimization approaches and compare their results. By using the complete search results, the location and the sizes of the capacitor banks are at buses 15, 25 and 65, while the sizes are 50kVAr, 200kVAr and 100kVAr. Internationally, this method yields the lowest possible line loss which is 176.15 kW for a system consisting only of capacitor banks. Thus, the Voltage Deviation Index has decreased from its original level 2.1439 to a value of 2.0338, obtaining an even more efficient network than would otherwise be possible. 2. High EV + PV Integration: Distributed solar PV units were connected to strategic high-load buses using busses at 15, 34, and 41 with 80 kW real power baseline injection per unit. The local active generation has successfully reduced the active power losses to 164.72 kW and enhanced the VDI to 1.9198. PV units effectively managed real power demand but did not solve the reactive power constraints of the grid. 3. High EV + Coordinated PV and MLSA Capacitors: The proposed hybrid approach combines local solar PV generation with MLSA-optimized shunt capacitors. The optimization engine successfully identified the global optimum coordinates for the capacitor banks at buses 15, 65 and 25 with optimal sizing configurations of 200 kVAr, 150 kVAr and 200 kVAr, respectively.Department of Electrical & Computer EngineeringFYP Proceedings87

C. System Performance Dataset Table I shows the summary results of the effect of adding EV charging stations on the power grid, and how different solutions help to mitigate it. In the Base Case, the grid has low power losses, 146.87 kW and stable voltage, 1.8177 under normal conditions. Adding "High EV" charging without a plan causes power waste to jump 36.5% to 200.45 kW and voltage to drop to an unstable 2.1439. Standard fixes, such as "PV Only" or simple manual capacitors, help a bit, but don't solve the problem. In contrast, the LSA and MLSA smart formulae both provide the best mathematical solution, reducing the power losses to 176.15 kW and fixing the voltage at 2.0338. Crucially, MLSA is much faster, finding the best solution in just 3 iterations, compared to 7 for LSA. Therefore, MLSA was chosen to optimize the capacitor location combined with PV to get the best power losses. TABLE I SUMMARY OF NETWORK PARAMETERS UNDER DIFFERENT MITIGATIONS. Scenario Voltage Deviation Index (VDI) Total Power Loss (kW) Best Iteration Base Case (No EV) 1.8177 146.87 - Low EV 1.8823 156.71 - Medium EV 1.9798 172.29 - High EV 2.1439 200.45 - High EV + Opt. Capacitors 2.0338 176.15 - High EV + PV 1.9198 164.72 - High EV + LSA 2.0338 176.15 7 High EV+ MLSA 2.0338 176.15 3 PV + MLSA Optimized Capacitors 1.8109 141.16 - Figure. 5 presents the total voltage stability of the grid showing the system capability to operate within the safe electrical limits of 69 buses. The chart shows a large spike upward in the unmitigated High EV charging case, which shows that the voltage drops are severe and the voltage is unstable near the heavy charging loads. Manual capacitors and standard solar panels help push this index down by boosting voltage locally, but do not fix the system. The smart search loops, LSA and MLSA, flatten out these voltage fluctuations perfectly down to a much safer, uniform level of 2.0338. The absolute best point on the chart is the deep valley at 1.8109 during the combined PV and MLSA capacitor stage, visually proving that mixing solar active power with optimized capacitor reactive power keeps the grid's voltage perfectly steady and safe from blackouts. Fig. 5. Result for VDI. The active power losses are presented in Figure 6 in kW. This directly means how much electricity is wasted in the power grid. As we can see from the chart, in the case of high uncoordinated EV charging there is a huge mountain peak at 200.45 kW. The heavy power waste can be explained by the overloaded lines with vehicles drawing high currents. From the graph, it is clear that both LSA and MLSA optimization formulas are aggressively shaving off this peak, flattening system losses to the absolute global floor of 176.15 kW by clearing out line bottlenecks. Finally, the trendline drops all the way down to an optimal low of 141.16 kW for the hybrid PV and MLSA case, Fig. 6. Result for Power Losses,kW. 1.6 1.7 1.8 1.9 2 2.1 2.2 Voltage Deviation Index (VDI) 0 50 100 150 200 250 Power Losses (kW)Department of Electrical & Computer EngineeringFYP Proceedings88

D. Algorithmic Convergence Characteristics The algorithms in figure 7 shows very different tracking speed. The standard LSA engine took 7 full iterations to find the global minimum. On the other hand, the adaptive step-size parameters and mutation mechanisms of the MLSA converged at the 3rd iteration. This reduction of computational burden highlights the high efficiency of the MLSA which is suitable for larger and more complex smart grid applications requiring fast and automatic control adjustments. Fig. 7. Convergence Curve of LSA and MLSA V. CONCLUSION The paper successfully modeled and analyzed the technical impacts of electric vehicle (EV) charging infrastructure on the IEEE 69-bus radial distribution system. The simulation results show that uncoordinated charging of EVs imposes heavy stress on the grid, leading to deep voltage drops and a significant increase in active power losses. While stand-alone solutions such as manual capacitor placement or solar PV integration offer some improvement, they do not take on the multi-variable voltage and load constraints of the network. The proposed hybrid framework that coordinates distributed solar PV generation with MLSA-optimized shunt capacitors provides the most effective mitigation by reducing the total losses to 141.16 kW and minimizing the voltage deviation index to 1.8109. Moreover, the modified optimization algorithm is also computationally more efficient than the traditional LSA in terms of the number of iterations for convergence. The work presented here will be extended in future work to include time-varying stochastic charging profiles, vehicle-to-grid (V2G) power flows, and localized energy storage systems to improve grid flexibility and reliability. REFERENCES [1] C. H. Tie, C. K. Gan, and K. A. Ibrahim, “The impact of electric vehicle charging on a residential low voltage distribution network in Malaysia,” 2014 IEEE Innovative Smart Grid Technologies - Asia, ISGT ASIA 2014, pp. 272–277, 2014, doi: 10.1109/ISGT- ASIA.2014.6873802. [2] M. Kezunovic, P. Pinson, Z. Obradovic, S. Grijalva, T. Hong, and R. Bessa, “Big data analytics for future electricity grids,” Electric Power Systems Research, vol. 189, p. 106788, Dec. 2020, doi: 10.1016/J.EPSR.2020.106788. [3] S. Y. Saçık, N. Yokuş, M. Alagöz, and T. Yokuş, “Optimum Renewable Energy Investment Planning in Terms of Current Deficit: Turkey Model,” Energies 2020, Vol. 13, Page 1509, vol. 13, no. 6, p. 1509, Mar. 2020, doi: 10.3390/EN13061509. [4] “pv research paper”. [5] M. Bilal, S. Gupta, P. N. Bokoro, and G. Sharma, “A probabilistic approach for optimal integration of EVs and RES using artificial hummingbird algorithm in distribution network,” IET Renewable Power Generation, vol. 18, no. 14, pp. 2305–2325, Oct. 2024, doi: 10.1049/rpg2.13074. [6] M. H. Aziz et al., “NiZrSe3/rGO modulated porous architecture for hybrid featured asymmetric supercapacitors,” J. Energy Storage, vol. 63, p. 106982, Jul. 2023, doi: 10.1016/J.EST.2023.106982. [7] M. H. Moradi and M. Abedini, “A combination of genetic algorithm and particle swarm optimization for optimal DG location and sizing in distribution systems,” International Journal of Electrical Power & Energy Systems, vol. 34, no. 1, pp. 66–74, Jan. 2012, doi: 10.1016/J.IJEPES.2011.08.023. [8] J. Chen, “Optimal Allocation of Distributed Generation Using Lightning Search Algorithm for Profit Maximization in a Distribution System,” Process Integration and Optimization for Sustainability 2024 8:5, vol. 8, no. 5, pp. 1545–1567, Aug. 2024, doi: 10.1007/S41660-024-00445-Z.Department of Electrical & Computer EngineeringFYP Proceedings89

Enhance PV Monitoring with Power Quality and Harmonic Detection Zaim Lukman Misran and Mohd Shahrin Abu Hanifah Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract— As Solar Photovoltaic (PV) integration grows, managing power quality is becoming as critical as tracking energy production. However, traditional monitoring solutions often fail to detect Harmonic Distortion and Power Quality (PQ) disturbances caused by high-frequency inverter switching, leading to harmonic distortion and potential equipment failure. This project aims to develop an Enhanced PV Monitoring System using an ESP32 microcontroller and isolated sensors (ZMPT101B and SCT-013) to provide high-resolution waveform analysis. By implementing Fast Fourier Transform (FFT) algorithms, the system calculates Total Harmonic Distortion (THD) and classifies system health status based on IEEE 519 standards. The result is expected to be a cost-effective, real-time diagnostic tool that ensures the reliability and compliance of PV- integrated power infrastructures. Index Terms—ESP32; Fast Fourier Transform; harmonic distortion, IEEE 519, IoT monitoring photovoltaic, power quality, Total Harmonic Distortion. I. INTRODUCTION The rapid integration of grid-tied Photovoltaic (PV) systems into modern power distribution networks has driven a fundamental shift from basic energy yield monitoring to comprehensive power quality (PQ) compliance. While commercial solar inverters convert DC energy from solar arrays into synchronised AC electricity, their internal high-frequency Pulse Width Modulation (PWM) switching networks inject non-linear current waveforms into the grid. When compounded by non-linear household loads such as variable speed drives, computers, and LED lighting matrices, this switching behaviour induces steady-state harmonic distortion characterized by unwanted integer multiples of the fundamental 50 Hz frequency. Left unmonitored, severe harmonic pollution accelerates equipment insulation degradation, causes distribution transformer thermal stresses, and induces localised grid instability [1], [2]. To mitigate these hazards, IEEE Std 519-2022 imposes a maximum voltage THD limit of 5.0 % at the point of common coupling (PCC) [4]. However, conventional residential smart meters and commercial IoT solar data loggers track only RMS voltage, current, and daily energy yield, rendering local harmonic pollution invisible to the end- user. Industrial-grade Power Quality Analysers (PQAs) capable of detailed harmonic profiling typically high in cost, making them economically inaccessible at residential scale [5]. Existing open-source telemetry designs frequently rely on low-tier 8-bit microcontrollers that lack the ADC resolution and processing throughput required to execute FFT algorithms directly at the edge. A notable prior work by Purnomo et al. (2021) demonstrated a low-cost IoT harmonic monitor using a ZMPT101B sensor and an Arduino Due platform; however, that system depended on an external MATLAB cloud server for FFT computation, introducing latency and proprietary software dependency, and its 10-bit ADC limited resolution for low-magnitude harmonics [6]. This paper addresses these gaps by presenting an edge-computing monitoring architecture that: (1) executes complete FFT-based harmonic analysis locally on the 32-bit dual-core ESP32; (2) classifies IEEE 519-2022 compliance in real time; (3) transmits results to a cloud dashboard without proprietary software; and (4) validates the system across three distinct supply topologies to characterise real-world PV inverter harmonic signatures. The scope is bounded to single-phase steady-state PQ monitoring, explicitly omitting transient analysis and multi-phase industrial modelling. II. BACKGROUND AND RELATED WORK A. Power Quality in PV Systems Power Quality (PQ) is defined by the degree to which electrical power deviates from an ideal sinusoidal waveform of fixed frequency and magnitude [7]. In grid-connected PV systems, two categories of PQ disturbance are recognised: transient events such as voltage sags and swells caused by irradiance fluctuations, and steady-state disturbances such as harmonic distortion generated by inverter switching [3], [7]. While transients are addressed by anti-islanding protection mechanisms, harmonic distortion represents persistent waveform pollution that degrades connected infrastructure over time. Three primary harmonic sources are documented in the literature. First, the DC-AC inverter’s SPWM switching creates high-frequency ripples and voltage notches in the output waveform. Second, environmental intermittency forces partial-load inverter operation at low irradiance, elevating the proportion of switching-frequency harmonics and increasing current THD significantly [8]. Third, the interaction of PV inverters with non-linear loads at the PCC can trigger harmonic resonance, amplifying specific harmonic orde

C. Research Gap The critical gap in the literature is the absence of a unified, edge- computing monitoring framework for small-scale PV installations that performs local spectral analysis, classifies IEEE 519 compliance in real time, and provides cloud monitoring without proprietary software. Existing low-cost IoT solutions neglect waveform health diagnostics, while high-accuracy methods are restricted to laboratory computing environments. Additionally, most prototypes validate against a single supply source, failing to characterise diverse harmonic signatures from different inverter topologies encountered in practice. III. SYSTEM DESIGN AND METHODOLOGY A. System Architecture The proposed system is structured as a three-stage pipeline. The Signal Acquisition layer consists of the ZMPT101B voltage sensor and SCT-013-010 current transformer connected to the ESP32’s 12-bit ADC channels. The Edge Processing layer executes RMS calculation, Hann windowing, 512-point FFT decomposition, THD computation, and IEEE 519-2022 compliance classification entirely on-chip. The Cloud Logging layer transmits a structured JSON payload to a Google Sheets dashboard via Wi-Fi every 10s, providing both real-time and historical power quality records. Figure 1: System Architecture of Proposed System B. Hardware Sensing Layer Voltage measurement is performed by the ZMPT101B module, featuring a 1000:1000 turns ratio that provides galvanic isolation, safely stepping down 240 V AC to a signal compatible with the ESP32’s 0–3.3 V ADC input range. A voltage scale factor of 967.2019 was determined through calibration against a reference meter. Current measurement employs the SCT-013-010 non-invasive current transformer (rated 0–10 A, sensitivity 0.1 V/A), which inductively couples to the conductor without circuit interruption. A DC zero-offset of 1935 ADC counts was established during calibration. Signal conditioning uses two 10 kΩ resistors to bias the sensor output to the ADC midpoint, and a 10 μF capacitor to stabilise the DC bias reference. Figure 2: Hardware Setup of Proposed System C. FFT Implementation and THD Computation A fixed sampling rate of 3200 Hz via hardware-driven timer interrupts collects N = 512 samples per measurement window, yielding a frequency resolution of: fres = fs / N = 3200 / 512 = 6.25 Hz per bin (1) A Hann window function is applied prior to FFT execution to mitigate spectral leakage, followed by DC offset removal on both channels. The arduinoFFT library processes the windowed buffer on the ESP32’s dual-core processor. THD for voltage and current is computed as: 𝑇𝐻𝐷𝑖= √∑ 𝐼𝑛 2∞ 𝑛=2 𝐼1 2 × 100 % (2) 𝑇𝐻𝐷𝑣= √∑ 𝑉𝑛 2∞ 𝑛=2 𝑣1 2 × 100 % (3) where V1 and I1 are the fundamental RMS magnitudes, and Vn and In are the nth harmonic RMS magnitudes. D. Classification and Cloud Logging If THDV < 5.0 % and THDI < 8.0 %, the cycle is classified as PASS; otherwise, FAIL is assigned. The complete dataset is encapsulated as a JSON payload and transmitted via HTTPS POST to a Google Apps Script web application, which appends a new timestamped row to the Google Sheets monitoring dashboard every 10 seconds. Figure 3: Figure of Monitoring Dashboard Figure 4: Figure of Real Time Data Monitoring E. Experimental Protocol Identical tests were performed under three distinct supply sources with a 33 W laptop as the non-linear load: (1) Main AC Utility Supply (240 Vᵣₘₛ, 50 Hz); (2) Pure Sine Wave (PSW) Inverter, high-frequency PWM and (3) Modified Sine Wave (MSW) Inverter. Five consecutive readings were captured per condition. TABLE I Bill of Materials for the Enhanced PV Monitoring Prototype No. Component Specification Qty 1 ESP32 Microcontroller 240 MHz dual-core, 12-bit ADC 1 2 ZMPT101B Voltage Sensor 1000:1000 turns, galvanic isolation 1 3 SCT-013-010 Current Sensor 0-10 A, 0.1 V/A, non-invasive CT 1Department of Electrical & Computer EngineeringFYP Proceedings91

4 Resistors 10 kOhm Voltage divider bias network 2 5 Capacitor 10 uF DC bias stabilisation 1 6 Pure Sine Wave Inverter Battery-powered, PWM+LC filter 1 7 Modified Sine Wave Inverter H-bridge switching topology 1 IV. RESULTS AND DISCUSSION A. Hardware Comparison Against Benchmark Table II summarises key hardware differences between the proposed system and the benchmark by Purnomo et al. (2021). The primary architectural improvements are the upgrade from an 8-bit Arduino Due (84 MHz, 10-bit ADC) to a 32-bit dual-core ESP32 (240 MHz, 12-bit ADC), and the migration of FFT computation from an external MATLAB cloud server to local edge processing. These changes eliminate processing latency, remove proprietary software dependency, and improve harmonic resolution for low-magnitude components. TABLE II Hardware Comparison: Proposed System vs. Benchmark (Purnomo et al., 2021) Component Benchmark System Proposed System MCU Architecture Arduino Due (8-bit) ESP32 (32-bit dual- core) CPU Clock 84 MHz 240 MHz ADC Resolution 10-bit 12-bit FFT Processing External MATLAB cloud Local edge on ESP32 Voltage Sensor ZMPT101B ZMPT101B (identical) Current Sensor SCT-013-000 SCT-013-010 Cloud Platform Google Sheet + MATLAB Google Sheets only Supply Sources Tested Single (mains) Three (mains, PSW, MSW) B. Voltage THD Benchmark Comparison Table III compares voltage and current THD readings from the Main AC supply against the Purnomo et al. (2021) benchmark. The proposed prototype recorded an average THDV of 5.47 %, compared to 7.20 % in the benchmark study, yielding a percentage error of 1.73 %. The higher benchmark THDV is attributable to testing at a campus electrical panel concurrently supplying multiple non-linear loads computers, variable speed drives, and LED matrices which collectively amplify voltage distortion at the PCC through harmonic resonance [3]. The prototype was tested in a controlled single-phase environment with a single 33 W laptop, yielding lower background distortion. Both systems detected THD levels at or above the IEEE 519-2022 limit of 5.0 %, confirming the prototype’s real-time compliance detection capability. TABLE III Voltage and Current THD: Proposed System vs. Benchmark Reading Benchmark THDV (%) Proposed THDV (%) Benchmark THDI (%) Proposed THDI (%) 1 5.00 5.32 24.00 131.70 2 10.00 5.98 24.00 102.00 3 9.00 6.08 23.00 157.67 4 6.00 5.35 29.00 114.52 5 6.00 4.62 27.00 133.28 Average 7.20 5.47 25.40 127.80 C. Current THD and Sensor Limitation A significant divergence was observed in current THD. While the benchmark reported an average THDI of 25.40 %, the prototype recorded 127.80 %. This discrepancy is not an algorithmic failure but results from two compounding factors. The 33 W laptop charger operates as a Switched-Mode Power Supply (SMPS) with a capacitive input rectifier, drawing current only during voltage peaks. This pulsed draw intrinsically produces high THDI percentages relative to the low fundamental current magnitude. More critically, the SCT-013-010 current transformer (rated 0– 10 A) was operated under a load current of only approximately 0.15 A, well below its optimal range. The resulting Signal-to-Noise Ratio (SNR) of approximately 1.9 (sensor noise floor ~0.08 A under inverter EMI) caused the FFT to process genuine harmonics superimposed on sensor noise, inflating THDI values. A minimum current load of approximately 0.8 A is required for reliable THDI measurement under the current hardware configuration. D. Multi-Source Comparative Analysis The proposed system was tested with three different types of sources which are main AC supply, pure sine wave inverter & modified sine wave inverter. Figure 5: Proposed System Tested using Main AC SupplyDepartment of Electrical & Computer EngineeringFYP Proceedings92

Figure 6: Proposed System using Pure Sine Wave Inverter Figure 7: Proposed System using Modified Sine Wave Inverter Table IV presents average THD results across the three supply sources. The dataset reveals a statistically consistent harmonic severity. The Main AC supply recorded the lowest average THDV of 5.47 %, marginally exceeding the IEEE 519-2022 limit. This is consistent with documented findings that urban residential low-voltage networks exhibit background THDV of 3 %–8 % from cumulative upstream non-linear load injection [5], underscoring the TABLE IV Comparative THD Results Across Three Supply Sources Parameter Main AC Supply PSW Inverter MSW Inverter VRMS (V) 240 230 226 Tested Load 33W Laptop THDV (%) 1st Reading 5.32 7.11 25.10 2nd Reading 5.98 5.52 24.32 3rd Reading 6.08 5.54 24.69 4th Reading 5.35 4.23 24.4 5th Reading 4.62 6.80 24.12 Avg 5.47 5.84 24.53 THDI (%) 1st Reading 131.70 321.86 228.24 2nd Reading 102.00 236.20 368.56 3rd Reading 157.67 195.39 479.84 4th Reading 114.52 195.19 219.92 5th Reading 133.28 291.86 113.41 Avg 127.80 227.72 281.99 importance of continuous monitoring even in environments assumed to be electrically clean. The PSW inverter recorded THDV of 5.84 %. Under ideal conditions a PSW inverter should produce THDV below 3 % due to its LC output filter. The elevated value reflects voltage degradation caused by the interaction between the lightweight filter stage and the capacitive pulsed current demand of the SMPS load, which introduces additional harmonic content via waveform clipping. The MSW inverter produced the highest THDV of 24.53 %, exceeding the IEEE limit by a factor of approximately 4.9. Across all three conditions, the prototype correctly classified as power quality severity MSW (most distorted), PSW (moderate), AC mains (least distorted) validating its core comparative classification capability without external instrumentation. Current THD showed a progressive upward trend: 127.80 % (AC mains), 227.72 % (PSW), and 281.99 % (MSW). While absolute magnitudes are inflated by the low-SNR sensor condition documented in Section IV-C, the relative severity ranking is qualitatively consistent with increasing EMI from each successive source topology, preserving diagnostic utility for comparative assessment. V. CONCLUSION This paper presented a low-cost, real-time Enhanced PV Monitoring System that executes FFT-based harmonic analysis entirely on the ESP32 microcontroller without dependency on external computational platforms or proprietary software. All three project objectives were fulfilled: (1) a high-resolution hardware sensing unit was integrated and calibrated; (2) a 512-point FFT with Hann windowing at 3200 Hz sampling correctly decomposes harmonic content from the 2nd to 11th order at 6.25 Hz/bin resolution; and (3) multi-source validation across utility mains, PSW, and MSW inverters demonstrated a clear harmonic severity progression, with benchmark comparison. The finding that all three supply sources approach or exceed the IEEE 519-2022Department of Electrical & Computer EngineeringFYP Proceedings93

THDV limit of 5.0 % including the utility mains highlights that harmonic pollution is a pervasive concern even in environments assumed to be electrically clean, reinforcing the need for continuous, affordable monitoring tools. The complete prototype was realised at a component cost below RM 200, establishing a validated, scalable, and remotely accessible alternative to industrial-grade PQAs. Future work includes: replacing the SCT-013-010 with a higher- sensitivity current transformer (e.g., ZMCT103C, 0–5 A) to improve current SNR under light loads; adding an RC low-pass filter (1 kΩ, 10 nF) on the current sensing channel to suppress inverter EMI; extending harmonic analysis to the 31st order as required by IEEE 519-2022; and integrating machine learning-based classification for predictive maintenance based on learned harmonic signatures across diverse operating conditions. REFERENCES [1] K. Osmani et al., “A critical review of PV systems’ faults with the relevant detection methods,” Energy Nexus, vol. 12, p. 100257, Dec. 2023, doi: 10.1016/j.nexus.2023.100257. [2] S. Ansari et al., “A Review of Monitoring Technologies for Solar PV Systems Using Data Processing Modules and Transmission Protocols,” Sustainability, vol. 13, no. 15, p. 8120, Jul. 2021, doi: 10.3390/su13158120. [3] S. M. Ahsan et al., “Harmonic Analysis of Grid-Connected Solar PV Systems with Nonlinear Household Loads in Low-Voltage Distribution Networks,” Sustainability, vol. 13, no. 7, p. 3709, Mar. 2021, doi: 10.3390/su13073709. [4] IEEE Standard for Harmonic Control in Electric Power Systems, IEEE Std 519-2022, 2022, doi: 10.1109/IEEESTD.2022.9848440. [5] Z. Sandi, “Examining Power Quality Challenges in Photovoltaic-Grid Integration: A Critical Review,” PEC, vol. 1, no. 1, pp. 12–23, Apr. 2024, doi: 10.62777/pec.v1i1.4. [6] P. Purnomo et al., “Realtime IoT based Harmonics Monitoring System Review with Potential Low-Cost Devices with Experimental Case Study,” JITEKI, vol. 7, no. 2, pp. 259–259, Aug. 2021, doi: 10.26555/jiteki.v7i2.21166. [7] T. Aditya et al., “Power Quality Analysis, Challenges and Feasible Solutions of 90 kWp Grid-Integrated PV System,” in Proc. ic-ETITE, Feb. 2020, doi: 10.1109/ic-etite47903.2020.452. [8] S. Adak et al., “Developed analytical expression for current harmonic distortion of the PV system’s inverter in relation to solar irradiance and temperature,” Elect. Eng., vol. 103, no. 1, pp. 697–704, 2020, doi: 10.1007/s00202-020-01110-7. [9] D. Rath et al., “Harmonic Distortion Assessment in the Single-Phase PV System Based on SPWM Technique,” Arabian J. Sci. Eng., vol. 46, no. 10, pp. 9601–9615, 2021, doi: 10.1007/s13369-021-05437-6. [10] K. Daniel et al., “A Review of Harmonic Detection, Suppression, Aggregation, and Estimation Techniques,” Appl. Sci., vol. 14, no. 23, p. 10966, Nov. 2024, doi: 10.3390/app142310966. [11] T. Tao and Y. Chu, “Harmonic Time–Frequency Analysis and Detection Method Based on Improved MSST,” Arabian J. Sci. Eng., vol. 49, no. 12, pp. 16421–16429, Apr. 2024, doi: 10.1007/s13369-024-Department of Electrical & Computer EngineeringFYP Proceedings94

Signal & Image Processing and Analysis T R A C K 6

Rooftop Photogrammetry for Solar Panel Dimensioning using Drone Amir Zuhni Abdul Rahman and Ahmad Fadzil Ismail Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract – This paper presents a method for measuring rooftops using a consumer-grade drone (DJI Lito X1) for solar panel installation planning. Aerial images were captured at 35 m altitude with 80% overlap and processed in Agisoft Metashape to generate 3D models and orthomosaics. Two different roof shapes (pentagon and triangle) were surveyed. Coordinates of corner points were exported and distances were calculated using the Euclidean distance formula in Microsoft Excel. The proposed method was validated against Agisoft's built-in ruler tool. Results showed a mean absolute error of 1.4 mm (0.03%) for the pentagonal roof and 2.6 mm (0.05%) for the triangular roof. The maximum error across both roofs was 0.07%. These findings demonstrate that the Excel- based coordinate method is accurate and suitable for pre- installation solar rooftop surveying. The DJI Lito X1 drone proved capable for this application, offering a safe and efficient alternative to manual measurement methods. Keywords – Earth Resistance Measurement, Fall of Potential method, Ground Rod Configuration, Multiple Rod Configuration, Single Rod Configuration. I.INTRODUCTION he adoption of solar photovoltaic (PV) systems has increased significantly worldwide. A critical step in solar system design is accurate measurement of the rooftop to determine available area, identify obstacles, and estimate panel capacity. Traditional manual measurement methods require physical roof access, which poses safety risks and is time-consuming. Drone photogrammetry has emerged as a promising alternative. By capturing overlapping aerial images, drones can generate 3D models and orthomosaics from which accurate measurements can be extracted [1]. However, most commercial photogrammetry software includes built-in measurement tools that may not be accessible to all users or may not allow batch processing of multiple measurements. This study proposes a simple, accessible method: exporting coordinates from Agisoft Metashape and calculating distances using the Euclidean distance formula in Microsoft Excel. The objectives of this study are: (1) to develop a coordinate-based measurement method for rooftop dimensioning, (2) to validate the method against Agisoft's built-in ruler tool, and (3) to assess the suitability of the DJI Lito X1 drone for this application. II. LITERATURE REVIEW A. Drone Photogrammetry for Rooftop Surveying Drone photogrammetry has been widely used for building inspection and rooftop mapping. Colomina and Molina (2014) reviewed the applications of UAVs in photogrammetry and remote sensing, highlighting their advantages in accessing difficult-to-reach areas. For rooftop surveying, drones offer a safer alternative to ladders and manual measurements. B. Solar Panel Installation Planning Accurate rooftop measurement is essential for solar panel installation. Freitas and Brito (2020) reviewed rooftop measurement techniques and found that errors in roof dimensioning can lead to incorrect panel count estimates and reduced energy yield. They recommended sub-centimeter accuracy for optimal system design. C. Comparison with Satellite Imagery Google Earth Pro is commonly used for preliminary roof assessment. However, its horizontal accuracy is approximately 15-50 cm (Google Earth Pro, 2024), which is insufficient for detailed solar panel layout planning. Small obstacles such as vents and chimneys are often invisible in satellite imagery. D. Coordinate-Based Distance Calculation The Euclidean distance formula is a standard method for calculating distances between points in 3D space. Exporting coordinates from photogrammetry software and performing external calculations allows users to customize their measurement workflow and batch process multiple distances. III. METHODS A. Drone and Flight Parameters A DJI Lito X1 drone was used for image capture. Flight parameters were as follows: Parameter Value Altitude 35 m Front & Side overlap 80% Camera mode 12 MP RAW Shooting interval 2 seconds Flight pattern Double Grid TDepartment of Electrical & Computer EngineeringFYP Proceedings95

B. Photogrammetric Processing Images were processed using Agisoft Metashape Pro (trial version). The workflow consisted of: 1. Align Photos: High accuracy, generic and reference preselection enabled. 2. Optimize Cameras: To refine camera alignment parameters. 3. Optimize Cameras: To refine camera alignment parameters. 4. Build Mesh: Height field surface type, source data from dense cloud. 5. Build Orthomosaic: Mosaic blending mode with hole filling enabled. C. Coordinate Export and Distance Calculation Five roof corner points (A, B, C, D, E) for the pentagonal roof and three points (A, B, C) for the triangular roof were marked on the orthomosaics. Their XYZ coordinates were exported as CSV files. Distances between adjacent points were calculated using the 3D Euclidean distance formula: 𝑑 = ඥ(𝑥ଶ − 𝑥ଵ)ଶ + (𝑦ଶ − 𝑦ଵ)ଶ + (𝑧ଶ − 𝑧ଵ)ଶ (1) Calculations were performed in Microsoft Excel. The same distances were measured using Agisoft's ruler tool for validation. D. Accuracy Metrics The following metrics were calculated:  Absolute error = |𝐸𝑥𝑐𝑒𝑙 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 − 𝐴𝑔𝑖𝑠𝑜𝑓𝑡 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒|  Relative error (%) = (𝐴𝑏𝑠𝑜𝑙𝑢𝑡𝑒 𝑒𝑟𝑟𝑜𝑟 / 𝐴𝑔𝑖𝑠𝑜𝑓𝑡 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒) × 100  Mean absolute error (MAE) = Average of absolute errors  Root mean square error (RMSE) = ටఀ (௘௥௥௢௥మ) ௡ IV. RESULTS AND DISCUSSION A. Pentagonal Roof Results Figure 1 shows the orthomosaic of the pentagonal roof with five marked corner points (A, B, C, D, E). Figure 1. Orthomosaic of pentagonal roof with labeled corner points A, B, C, D, E TABLE 1 EXPORTED COORDINATES FOR PENTAGONAL ROOF TABLE 2 DISTANCE MEASUREMENTS FOR PENTAGONAL ROOF Total Perimeter: 22.739 m B. Triangular Roof Results Figure 2 shows the orthomosaic of the triangular roof with three marked corner points (A, B, C). Figure 2. Orthomosaic of triangular roof with labeled corner points A, B, C TABLE 3 EXPORTED COORDINATES FOR TRIANGULAR ROOFDepartment of Electrical & Computer EngineeringFYP Proceedings96

TABLE 4 DISTANCE MEASUREMENTS FOR TRIANGULAR ROOF Total Perimeter: 16.375 m C. Accuracy Summary TABLE 5 ACCURACY SUMMARY FOR BOTH ROOF SECTIONS Metric Pentagonal Roof Triangular Roof Total perimeter 5 3 Number of edges 22.740 m 16.375 m Mean absolute error 0.0014 m (1.4 mm) 0.0026 m (2.6 mm) Mean percentage error 0.03% 0.05% Maximum error 0.05% 0.07% RMSE 0.0016 m 0.0029 m Figure 3. Comparison of Excel and Agisoft measurements for both roof sections D. Discussion The results demonstrate that the proposed Excel-based coordinate method achieved high accuracy on both roof shapes. For the pentagonal roof, the mean absolute error was 1.4 mm with an average percentage error of 0.03%. For the triangular roof, the mean absolute error was 2.6 mm with an average percentage error of 0.05%. The maximum error observed across both roofs was 0.07% (4 mm on a 6.7 m edge). These errors are well within acceptable tolerance for solar panel dimensioning, where typical allowable error is ±50 mm. The slight difference in accuracy between the two roofs can be attributed to image quality and marker placement precision. Compared to Google Earth Pro, which has a horizontal accuracy of approximately 15-50 cm, the proposed method offers significantly better precision (1.4-2.6 mm). Additionally, drone photogrammetry can measure roof pitch and detect small obstacles that are invisible in satellite imagery. The inclusion of both roof shapes strengthens the validation of the method, proving that it works consistently across different geometries. V. CONCLUSION This study successfully developed and validated a coordinate- based method for rooftop measurement using drone photogrammetry and Excel-based distance calculation. The key findings are: 1. The proposed method achieved a mean absolute error of 1.4 mm (0.03%) for the pentagonal roof and 2.6 mm (0.05%) for the triangular roof. 2. The maximum error across both roofs was 0.07%, well within acceptable limits for solar panel planning. 3. The total perimeters were measured as 22.74 m (pentagon) and 16.38 m (triangle). 4. The DJI Lito X1 drone is suitable for this application when flown at 35 m altitude with 80% overlap. The proposed method offers a simple, accessible alternative to built-in measurement tools, requiring only coordinate export from Agisoft and basic Excel calculations. This approach is particularly useful for users who prefer external data processing or need to batch multiple measurements. For future work, it is recommended to add Ground Control Points (GCPs) to improve absolute accuracy, validate with manual tape measurements, test on more roof shapes, automate calculations using Python, and measure roof pitch for energy yield calculations. REFERENCES [1] I. Colomina and P. Molina, "Unmanned aerial systems for photogrammetry and remote sensing: A review," ISPRS Journal of Photogrammetry and Remote Sensing, vol. 92, pp. 79-97, 2014. [2] S. Freitas and C. Brito, "Solar photovoltaic system design: A review of rooftop measurement techniques," Renewable and Sustainable Energy Reviews, vol. 120, 2020. [3] Google Earth Pro, "Measuring distances and areas," Google Support, 2024. [4]Agisoft LLC, "Agisoft Metashape User Manual: Professional Edition," 2024. [5] DJI, "DJI Lito X1 User Manual," 2024.Department of Electrical & Computer EngineeringFYP Proceedings97

Abstract—Speech signal processing has gained significant attention due to the increasing demand for clear and intelligible speech in applications such as telecommunications, hearing assistance devices, and voice- controlled systems. However, speech signals are often corrupted by background noise, resulting in degraded speech quality and reduced intelligibility. This project proposes a speech enhancement and noise classification framework based on the Discrete Wavelet Transform (DWT) and K-Nearest Neighbour (KNN) algorithm. The DWT is employed to decompose noisy speech signals into multiple frequency sub-bands, while soft thresholding is applied to suppress noise components and preserve important speech information. The enhanced speech signal is subsequently reconstructed using the inverse wavelet transform. In addition, statistical and spectral features extracted from the processed signals are utilized by the KNN classifier to identify different noise environments. The performance of the proposed system is evaluated using objective metrics, namely Signal-to- Noise Ratio (SNR) and Mean Squared Error (MSE). Experimental results demonstrate that the wavelet-based enhancement method effectively improves speech quality under various noise conditions, while the KNN classifier achieves an overall classification accuracy of 98.75% for four noise categories: white, babble, train, and traffic noise. The findings indicate that the proposed wavelet-KNN framework provides an effective and computationally efficient solution for speech enhancement and noise classification applications. Index Terms—Speech Enhancement, Noise Reduction, Discrete Wavelet Transform (DWT), K-Nearest Neighbour (KNN), Noise Classification. PEECH communication has become an essential component of modern technologies, including telecommunications, hearing assistance devices, voice- controlled systems, and speech recognition applications. The effectiveness of these systems depends heavily on the quality and intelligibility of speech signals[1], [2]. However, speech signals are often contaminated by various types of background noise during acquisition, transmission, or recording processes. The presence of noise degrades speech quality and reduces intelligibility, which can negatively affect both human communication and the performance of speech processing systems[3]. To address this issue, numerous speech enhancement and noise reduction techniques have been proposed. Traditional approaches such as Spectral Subtraction (SS), Wiener Filtering, Adaptive Noise Cancellation (ANC), and Kalman Filtering have been widely used to suppress unwanted noise components[3][4]. Although these methods can achieve satisfactory performance under certain conditions, they often struggle when dealing with non-stationary noise environments and may introduce speech distortion or residual noise artifacts[3], [5]. Wavelet-based speech enhancement has emerged as an effective alternative due to its ability to analyze signals in both time and frequency domains simultaneously. The Discrete Wavelet Transform (DWT) decomposes speech signals into multiple frequency sub-bands, allowing noise components to be separated from useful speech information more effectively[6]. By applying thresholding techniques to the wavelet coefficients, noise-dominant components can be suppressed while preserving important speech characteristics. This capability makes wavelet- based methods particularly suitable for speech enhancement applications involving complex noise environments[2]. In addition to speech enhancement, identifying the type of background noise can provide valuable information for adaptive speech processing systems. Machine learning techniques have demonstrated promising performance in classification tasks by recognizing patterns from extracted signal features. Therefore, this study proposes a speech enhancement and noise classification framework based on the Discrete Wavelet Transform (DWT) and the K-Nearest Neighbour (KNN) algorithm[7]. The proposed system performs wavelet-based denoising to improve speech quality and subsequently classifies noise types using statistical and spectral features. The performance of the framework is evaluated using Signal-to-Noise Ratio (SNR), Mean Squared Error (MSE), classification accuracy, and confusion matrix analysis[7]. Speech Enhancement and Noise Reduction via Machine Learning Muhammad Ferhad Eyuri Mohd Sukri, Ahmad Zamani Jusoh and Ani Liza Asnawi. Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia S I. INTRODUCTIONDepartment of Electrical & Computer EngineeringFYP Proceedings98

II. PROPOSED SPEECH ENHANCEMENT FLOW Fig. 1. Research flowchart of the proposed wavelet-based speech enhancement and noise classification framework. Figure 1 illustrates the overall research framework adopted in this study. The methodology is divided into three phases corresponding to the project objectives. The first phase involves a literature review of speech enhancement and noise reduction techniques, followed by the identification of research gaps and the selection of a wavelet-based approach. In the second phase, the clean speech signals were artificially corrupted with different background noise types to simulate realistic noisy communication environments and pre-processed before undergoing Discrete Wavelet Transform (DWT) decomposition and thresholding for noise suppression. This procedure enabled the evaluation of the proposed enhancement framework under multiple acoustic conditions. Finally, the enhanced speech signals are reconstructed using the inverse wavelet transform and evaluated using objective performance metrics. The results are then analysed to determine the effectiveness of the proposed speech enhancement system. A. Dataset Preparation The speech enhancement system was evaluated using clean speech signals and various background noise recordings. The speech signals were stored in WAV format with a sampling frequency of 16 kHz. Four types of noise were considered in this study, namely white noise, babble noise, train noise, and traffic noise. To simulate noisy environments, each clean speech signal was mixed with a corresponding noise signal at an input Signal-to-Noise Ratio (SNR) of 10 dB. A total of 400 samples were generated and used for both speech enhancement and noise classification experiments. B. Wavelet-Based Speech Enhancement The proposed speech enhancement method is based on the Discrete Wavelet Transform (DWT). The DWT decomposes the noisy speech signal into approximation and detail coefficients at multiple frequency levels. Noise suppression is performed by applying soft thresholding to the detail coefficients, where coefficients below a selected threshold are attenuated while significant speech components are preserved. After thresholding, the enhanced speech signal is reconstructed using the Inverse Discrete Wavelet Transform (IDWT). The wavelet decomposition process can be represented as(1): W(j,k) = ∫ x(t) ψj,k(t) dt (1) where x(t) represents the speech signal and ψj,k(t) denotes the wavelet basis function at scale j and position k. C. Feature Extraction and KNN Classification After enhancement, several statistical and spectral features were extracted from the speech signals. These features included Signal-to-Noise Ratio (SNR), Mean Squared Error (MSE), signal energy, standard deviation, kurtosis, skewness, and Fast Fourier Transform (FFT)-based spectral characteristics. The extracted features were used as input to the K-Nearest Neighbour (KNN) classifier for noise type identification. The Euclidean distance used in the KNN algorithm is given by(2): d(x,y) = √Σ(xi − yi)² (2) where xi and yi represent feature values of two samples. The class label of a test sample is determined through majority voting among its nearest neighbours. D. Experimental Setup All experiments were implemented in Python using NumPy, SciPy, PyWavelets, and SoundFile libraries. The dataset consisted of 400 samples, where 80% of the samples were used for training and 20% were used for testing. The performance of the speech enhancement system was evaluated using Signal-to- Noise Ratio (SNR) and Mean Squared Error (MSE), while the classification performance was assessed using accuracy and confusion matrix analysis.Department of Electrical & Computer EngineeringFYP Proceedings99

III. RESULTS AND DISCUSSION A. Waveform Analysis Fig. 2. Comparison of clean, noisy, and enhanced speech waveforms. Figure 2 shows the waveform comparison between the clean, noisy, and enhanced speech signals. The noisy speech signal exhibits larger fluctuations due to the presence of background noise. After applying the proposed wavelet-based enhancement method, the enhanced speech waveform more closely resembles the clean reference signal, indicating effective noise suppression while preserving the overall speech structure. The enhanced waveform more closely resembles the clean reference waveform, indicating that a substantial portion of the background noise has been suppressed. At the same time, the major speech patterns are preserved, suggesting that the wavelet thresholding process successfully reduced noise without introducing significant speech distortion. B. Speech Enhancement Performance Fig. 3. Comparison of clean, noisy, and enhanced speech spectrograms. Figure 3 illustrates the spectrogram comparison of the clean, noisy, and enhanced speech signals. The noisy speech spectrogram contains additional frequency components caused by background noise, particularly in the high-frequency regions. After enhancement, a significant reduction in unwanted spectral components can be observed while maintaining the important speech formants and time-frequency characteristics. This demonstrates the capability of the proposed wavelet-based method to suppress noise without severely distorting the speech signal. The enhanced spectrogram exhibits reduced spectral energy in regions previously dominated by noise. Important speech formants and time-frequency structures remain visible after enhancement, demonstrating the capability of the proposed method to preserve speech information while suppressing unwanted noise components. C. Speech Enhancement Performance The effectiveness of the proposed wavelet-based speech enhancement method was evaluated using Signal-to-Noise Ratio (SNR) and Mean Squared Error (MSE). Table I presents the average SNR values obtained before and after enhancement for different noise environments. TABLE I AVERAGE SNR RESULTS Noise Type SNR Before (db) SNR After (db) Improvement (db) White 10.22 12.91 2.69 Babble 10.17 10.59 0.42 Train 10.08 10.08 -0.01 Traffic 10.15 10.96 0.82 The results indicate that the proposed wavelet-based enhancement method improved speech quality under most noise conditions. White noise achieved the highest SNR improvement of 2.69 dB. This result indicates that the proposed wavelet thresholding technique is particularly effective in suppressing stationary noise components. Since white noise exhibits relatively uniform spectral characteristics, the wavelet decomposition process can more effectively separate noise from speech information, resulting in improved speech quality, demonstrating the effectiveness of the wavelet thresholding approach in suppressing stationary noise components. Traffic noise also showed a noticeable improvement of 0.82 dB, while babble noise achieved a moderate enhancement of 0.42 dB. In contrast, train noise exhibited minimal improvement due to its complex spectral characteristics, which are more difficult to separate from speech components.Department of Electrical & Computer EngineeringFYP Proceedings100

TABLE II AVERAGE MSE RESULTS Noise Type MSE Before MSE After Reduction White 0.002080 0.000611 0.001470 Babble 0.002104 0.001023 0.001082 Train 0.002152 0.002154 -0.000002 Traffic 0.002130 0.001543 0.000587 The MSE results further confirm the effectiveness of the proposed enhancement method. Significant reductions were observed for white, babble, and traffic noise, indicating that the enhanced speech signals more closely resembled the clean reference signals. The largest reduction was achieved under white noise conditions, while train noise remained the most challenging noise type for the proposed system. D. Noise Classification Performance The performance of the K-Nearest Neighbour (KNN) classifier was evaluated using the extracted statistical and spectral features. A total of 400 samples were used, with 320 samples allocated for training and 80 samples used for testing. TABLE III AVERAGE SNR RESULTS Metric Value Total Samples 400 Training Samples 320 Testing Samples 80 Classification Accuracy 98.75% Table III presents the performance of the K-Nearest Neighbour (KNN) classifier used for noise type identification. A total of 400 samples were utilized in the experiment, with 320 samples allocated for training and 80 samples used for testing. The classifier achieved an overall accuracy of 98.75%, indicating that the extracted statistical and spectral features were highly effective in distinguishing different noise environments. The high classification accuracy demonstrates that the KNN algorithm successfully identified the characteristics of white, babble, train, and traffic noise. These results suggest that the proposed feature extraction approach provides sufficient discriminative information for reliable noise classification. Fig. 2. Confusion matrix of the KNN noise classification results. Figure 2 illustrates the confusion matrix of the KNN classification results. Most samples were correctly classified, with white, babble, and train noise achieving near-perfect classification performance. Only one traffic noise sample was incorrectly classified as babble noise, which may be attributed to similarities in their spectral characteristics. Nevertheless, the overall classification performance remained highly reliable, further validating the effectiveness of the proposed noise classification framework. IV. CONCLUSION This paper presented a speech enhancement and noise classification framework based on the Discrete Wavelet Transform (DWT) and K-Nearest Neighbour (KNN) algorithm. The proposed system employed wavelet-based denoising to suppress background noise while preserving important speech characteristics. Experimental results demonstrated improvements in speech quality under various noise conditions, as indicated by increased Signal-to-Noise Ratio (SNR) values and reduced Mean Squared Error (MSE) values. In addition, the KNN classifier successfully identified four different noise categories, namely white, babble, train, and traffic noise, achieving an overall classification accuracy of 98.75%. The results confirm that the combination of wavelet- based speech enhancement and machine learning-based noise classification provides an effective and computationally efficient solution for speech processing applications. Future work may focus on adaptive thresholding techniques, larger datasets, and advanced machine learning models to further improve system performance.Department of Electrical & Computer EngineeringFYP Proceedings101

REFERENCES [1] D. Michelsanti et al., “An Overview of Deep-Learning-Based Audio-Visual Speech Enhancement and Separation,” IEEE/ACM Trans. Audio Speech Lang. Process., vol. 29, pp. 1368–1396, 2021, doi: 10.1109/TASLP.2021.3066303. [2] N. Das, S. Chakraborty, J. Chaki, N. Padhy, and N. Dey, “Fundamentals, present and future perspectives of speech enhancement,” Int. J. Speech Technol., vol. 24, no. 4, pp. 883–901, Dec. 2021, doi: 10.1007/s10772-020-09674-2. [3] S. Vihari, A. S. Murthy, P. Soni, and D. C. Naik, “Comparison of Speech Enhancement Algorithms,” in Procedia Computer Science, Elsevier B.V., 2016, pp. 666–676. doi: 10.1016/j.procs.2016.06.032. [4] S. M. Hattaraki, S. G. Kambalimath, L. Khedagi, K. Halemani, A. Bhairagond, and P. Hiremath, “Evaluation of Speech Enhancement Algorithms for Hearing Aids,” in Proceedings of NKCon 2024 - 3rd Edition of IEEE NKSS’s Flagship International Conference: Digital Transformation: Unleashing the Power of Information, Institute of Electrical and Electronics Engineers Inc., 2024. doi: 10.1109/NKCon62728.2024.10774613. [5] Q. Zhang, X. Qian, Z. Ni, A. Nicolson, E. Ambikairajah, and H. Li, “A Time-Frequency Attention Module for Neural Speech Enhancement,” IEEE/ACM Trans. Audio Speech Lang. Process., vol. 31, pp. 462–475, 2023, doi: 10.1109/TASLP.2022.3225649. [6] S. Pradeep Kumar, A. Daripelly, S. M. Rampelli, S. K. R. Nagireddy, A. Badishe, and A. Attanthi, “Noise Reduction Algorithm for Speech Enhancement,” in Proceedings of 2023 International Conference on Signal Processing, Computation, Electronics, Power and Telecommunication, IConSCEPT 2023, Institute of Electrical and Electronics Engineers Inc., 2023. doi: 10.1109/IConSCEPT57958.2023.10170204. [7] 2018 5th International Conference on Signal Processing and Integrated Networks (SPIN) : 22-23 February, 2018, Amity School of Engineering and Technology, Noida, India. IEEE, 2018.Department of Electrical & Computer EngineeringFYP Proceedings102

Vessel-Enhanced Deep Learning Approach for Intracranial Aneurysm Detection Ahmad Sayuti Hassan Shaari and Noreha Abd. Malik Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia Abstract—Intracranial aneurysms (IAs) are abnormal dilations of the cerebral arteries whose rupture can cause life-threatening subarachnoid hemorrhage. Their early detection on Time-of- Flight Magnetic Resonance Angiography (TOF-MRA) is clinically important, but manual assessment is slow and observer-dependent and is complicated by the very small size of aneurysms relative to the whole brain. This work proposes an automated two-stage 3D U-Net with a vessel enhancement subnet for intracranial aneurysm detection and segmentation. A first subnetwork generates a vessel enhancement map from the raw scan, which is concatenated with the scan and passed to a second, deeper subnetwork that produces the final aneurysm segmentation. The model was trained and evaluated on the publicly available ADAM dataset using N4 bias field correction, Z-score normalization, and balanced 3D patch cropping, with a combined Tversky-BCE loss to prioritize sensitivity. On the held-out test set, the model achieved a Dice coefficient of 0.4021, a sensitivity of 0.5630, and a specificity of 0.9997, and correctly located aneurysms in 14 of 17 patients, corresponding to a patient-level detection sensitivity of 82.35%. The proposed approach attains a patient-level detection performance competitive with a recent vessel-prior benchmark, although the voxel-level overlap remains modest, demonstrating the value of incorporating vascular information for aneurysm detection. Index Terms—ADAM dataset, deep learning, image segmentation, intracranial aneurysm, TOF-MRA, three-dimensional(3D) U-Net, vessel enhancement. I. INTRODUCTION ntracranial aneurysms (IAs) are pathological dilations of the cerebral arteries, typically arising at the bifurcations of the Circle of Willis, with an estimated prevalence of about 3–4% in the adult population [1], [2]. Their rupture leads to subarachnoid haemorrhage, a condition associated with high mortality and long-term disability, so the early detection of unruptured aneurysms is clinically important. Among the available imaging modalities, Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) is widely preferred for screening because it is non-invasive and does not require a contrast agent [3]. Manual identification of aneurysms on TOF-MRA is time- consuming and subject to inter-observer variability, particularly for small aneurysms. Automated detection is challenging for two reasons. First, aneurysms occupy only a tiny fraction of the whole brain volume, producing a severe class imbalance that biases a network towards the background class. Second, a standard 3D U-Net treats all voxels equally and does not explicitly exploit the fact that aneurysms arise from blood vessels. As reviewed in Section II, recent work has progressed towards two-stage, cascaded, and vessel-enhanced networks that incorporate vascular information. Motivated by these trends, this paper proposes a two-stage 3D U-Net with a vessel enhancement subnet, in which a learned vessel map guides a subsequent aneurysm detection subnet. The main contributions are: (i) a reproducible two-stage vessel-enhanced pipeline evaluated on the publicly available ADAM dataset; (ii) a sensitivity-oriented training scheme based on balanced patch sampling and a Tversky-BCE loss; and (iii) an evaluation using voxel-level metrics together with a patient with aneurysm detection analysis. II. LITERATURE REVIEW Among the available imaging modalities, TOF-MRA has become the most common input for deep learning–based aneurysm analysis because it visualises the intracranial arteries without a contrast agent [3]. Raw TOF-MRA volumes nevertheless suffer from scanner-induced intensity inhomogeneity and inter-scanner variability, so most pipelines apply a standardised preprocessing sequence. N4 bias field correction [4] is almost universally used to remove the low- frequency intensity bias, followed by intensity normalisation to place all subjects on a comparable scale [5]. To emphasise the vasculature in which aneurysms arise, Hessian-based vessel enhancement filters such as the Frangi filter [6] are widely employed, and the resulting vesselness maps are increasingly used as explicit vascular priors within deep networks. The U-Net [7] and its volumetric extension, the 3D U-Net [8], form the basis of most aneurysm segmentation methods. Because aneurysms are extremely small relative to the brain, single-stage networks struggle with class imbalance, which has motivated two-stage and cascaded designs that first localise the vasculature and then detect the aneurysm [9], [10]. A closely related direction incorporates explicit vascular information: Abbas et al. [11] applied Frangi enhancement before a deeply supervised U-Net, while Rainville et al.

III. METHODOLOGY A. Dataset The publicly available Aneurysm Detection And segMentation (ADAM) dataset [15] is used. It provides TOF-MRA scans from 113 subjects with voxel-wise annotations. Only untreated aneurysms (label 1) are retained as the positive class. The data is split at the patient level into training (70%), validation (15%), and test (15%) subsets to prevent leakage. B. Preprocessing Each volume undergoes N4 bias field correction [4] to remove scanner-induced intensity inhomogeneity, followed by Z-score intensity normalisation, 𝑥′ = 𝑥 − 𝜇 𝜎 (1) where μ and σ are the mean and standard deviation of the volume. Fixed-size 3D patches of 96 × 96 × 96 voxels are then cropped using a balanced strategy that centres a high proportion of patches on aneurysm voxels, and 3D spatial augmentation (random flips and 90° rotations) is applied to training patches only. C. Network Architecture The proposed network consists of two 3D U-Net subnetworks connected in series, as shown in Fig. 1. The first subnetwork, the vessel enhancement subnet, is a shallow 3D U- Net (channels 16→32→64) that produces a single-channel vessel enhancement map through a sigmoid activation. This map is concatenated with the raw patch to form a two-channel input. The second subnetwork, the aneurysm detection subnet, is a deeper 3D U-Net (channels 32→64→128→256) that outputs the voxel-wise aneurysm prediction. Each subnetwork is built from double-convolution blocks (Conv3D– BatchNorm–ReLU), max-pooling in the encoder, and trilinear upsampling with skip connections in the decoder. Fig. 1. Overview of the proposed two-stage 3D U-Net with vessel enhancement subnet pipeline D. Loss Function The model is trained with a combined Tversky-BCE loss [16] suited to severe class imbalance. The Tversky index is 𝑇𝐼 = 𝑇𝑃 𝑇𝑃 + 𝛼·𝐹𝑃 + 𝛽·𝐹𝑁 (2) with α = 0.3 and β = 0.7 so that false negatives are penalised more heavily, driving the model towards higher sensitivity. The Tversky loss is Lₜ = 1 − TI, and the total objective is 𝐿 = 0.3 · 𝐿𝐵𝐶𝐸 + 0.7 · 𝐿𝑇𝑣𝑒𝑟𝑠𝑘𝑦 (3) E. Training Configuration The network is optimised using AdamW (initial learning rate 1×10⁻³, weight decay 1×10⁻⁵) with a ReduceLROnPlateau scheduler and Automatic Mixed Precision. Training runs for 35 epochs with a batch size of 16 and a classification threshold of 0.2. The checkpoint with the highest validation Dice is retained. F. Evaluation Metrics Performance is measured by the Dice similarity coefficient (4), sensitivity (5), and specificity (6) [17], together with a patient with aneurysm detection analysis in which a patient is counted as detected if any predicted voxel overlaps the reference aneurysm. 𝐷𝑆𝐶 = 2 𝑇𝑃 2 𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁 (4) 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = 𝑇𝑃 𝑇𝑃 + 𝐹𝑁 (5)Department of Electrical & Computer EngineeringFYP Proceedings104

𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = 𝑇𝑁 𝑇𝑁 + 𝐹𝑃 (6) IV. RESULTS AND DISCUSSION A. Training and Validation Over 35 epochs the training loss decreased from 0.7757 to 0.3599 and the training Dice rose to approximately 0.43 (Fig. 2). The validation Dice improved overall and reached its highest value of 0.3322 at epoch 26, at which the best checkpoint was saved. The validation curve was unstable, occasionally collapsing to zero Dice; this is attributed to the small validation set and the sensitivity of voxel-level Dice on very small targets. The best checkpoint strategy was therefore essential. Fig. 2. Training loss curve (left) and training Dice coefficient (right) over 35 epochs B. Test Performance On the held-out test set the model achieved a Dice coefficient of 0.4021, a sensitivity of 0.5630, and a specificity of 0.9997 (Table I). The near-perfect specificity is expected under the extreme class imbalance and is therefore uninformative; Dice and sensitivity are the more meaningful measures. TABLE I. PERFORMANCE OF THE PROPOSED MODEL ON THE TEST SET Metric Value Dice coefficient 0.4021 Sensitivity 0.5630 Specificity 0.9997 C. Patient with Aneurysm Detection The model correctly located aneurysms in 14 of 17 patients, a patient with aneurysm detection sensitivity of 82.35% (Table II); three patients were missed. This is the most clinically meaningful result, as the goal of a screening tool is to flag the presence of an aneurysm for review. TABLE II. PATIENT WITH ANEURYSM DETECTION RESULTS Patient ID Outcome Patient ID Outcome 10027 Detected 10058F Detected 10030 Missed 10059B Missed 10031 Detected 10059F Detected 10032 Detected 10066F Detected 10043 Detected 10070F Detected 10044B Detected 10073B Missed 10048B Detected 10073F Detected 10053B Detected 10053F Detected 10054F Detected D. Qualitative Results Fig. 3 shows a representative test slice. The model produces a strong, high-confidence response co-located with the ground- truth aneurysm, confirming its detection ability. The predicted region is larger than the reference, indicating over- segmentation — a direct consequence of the false-negative- penalising Tversky loss and the low threshold of 0.2. The vessel enhancement map is diffuse rather than vessel-specific, reflecting that the vessel subnet is trained only implicitly, without direct supervision. Fig. 3. Qualitative result on a representative test slice (Slice 48): raw MRA, vessel enhancement map, ground-truth aneurysm, and predicted probability map E. Comparison with Existing Work The proposed model is compared against the Vessel-Prior UNet (VP UNet) [12], a recent method on the same dataset (Table III). The segmentation Dice (0.4021) is lower than VP UNet (0.614 internal, 0.49 external), but the patient-level detection sensitivity (82.35%) is close to its external-set value of 82.8%. The protocols are not strictly comparable, as the present evaluation is patch-based. The principal difference is that VP UNet uses an explicit, supervised Frangi vesselness prior [6], whereas the proposed model learns the vessel map implicitly; the stronger segmentation of the benchmark suggests that an explicit vessel prior is more effective. TABLE III. COMPARISON WITH THE VP UNET BENCHMARK ON ADAM Method Dice Detection sens. VP UNet [12] 0.614 (0.49 ext.) 0.929 / 0.828 ext. Proposed 0.4021 0.8235 V. CONCLUSION A two-stage 3D U-Net with a vessel enhancement subnet was proposed for intracranial aneurysm detection on TOF-MRA. On the ADAM dataset it achieved a test Dice of 0.4021 and a patient-level detection sensitivity of 82.35%, competitive with a recent vessel-prior method despite a modest voxel-level overlap. The results demonstrate the value of a sensitivity- oriented, vessel-aware design for aneurysm screening. Future work includes providing the vessel branch with an explicit supervised prior, adopting full-volume sliding-window inference with lesion-level metrics, applying brain-extraction preprocessing, and evaluating established frameworks such as nnU-Net. REFERENCES [1] N. Etminan, A. Dörfler, and H. Steinmetz, “Unruptured intracranial aneurysms,” Deutsches Ärzteblatt International, vol. 117, no. 14, pp. 235–242, Mar. 2020, doi: 10.3238/arztebl.2020.0235. [2] A. Keedy, “An overview of intracranial aneurysms,” Jul. 01, 2006. https://pmc.ncbi.nlm.nih.gov/articles/PMC2323531/Department of Electrical & Computer EngineeringFYP Proceedings105

[3] T. Sichtermann, A. Faron, R. Sijben, N. Teichert, J. Freiherr, and M. Wiesmann, “Deep Learning–Based Detection of Intracranial Aneurysms in 3D TOF-MRA,” American Journal of Neuroradiology, vol. 40, no. 1, pp. 25–32, Dec. 2018, doi: 10.3174/ajnr.a5911. [4] N. J. Tustison et al., “N4ITK: Improved N3 bias correction,” IEEE Transactions on Medical Imaging, vol. 29, no. 6, pp. 1310–1320, Apr. 2010, doi: 10.1109/tmi.2010.2046908. [5] S. Ham et al., “Automated detection of intracranial aneurysms using skeleton-based 3D patches, semantic segmentation, and auxiliary classification for overcoming data imbalance in brain TOF-MRA,” Scientific Reports, vol. 13, no. 1, p. 12018, Jul. 2023, doi: 10.1038/s41598-023-38586-9. [6] A. F. Frangi, W. J. Niessen, K. L. Vincken, and M. A. Viergever, “Multiscale vessel enhancement filtering,” in Lecture notes in computer science, 1998, pp. 130–137. doi: 10.1007/bfb0056195. [7] O. Ronneberger, P. Fischer, and T. Brox, “U-NET: Convolutional Networks for Biomedical Image Segmentation,” in Lecture notes in computer science, 2015, pp. 234–241. doi: 10.1007/978-3-319- 24574-4_28. [8] Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger, “3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation,” in Lecture notes in computer science, 2016, pp. 424–432. doi: 10.1007/978-3-319-46723-8_49. [9] T. Di Noto et al., “Towards Automated Brain Aneurysm Detection in TOF-MRA: open data, weak labels, and anatomical knowledge,” Neuroinformatics, vol. 21, no. 1, pp. 21–34, Aug. 2022, doi: 10.1007/s12021-022-09597-0. [10] F. Claux, M. Baudouin, C. Bogey, and A. Rouchaud, “Dense, deep learning-based intracranial aneurysm detection on TOF MRI using two-stage regularized U-Net,” Journal of Neuroradiology, vol. 50, no. 1, pp. 9–15, Mar. 2022, doi: 10.1016/j.neurad.2022.03.005. [11] S. F. Abbas, N. T. Duc, Y. Song, K. Kim, E. Srivastava, and B. Lee, “CV-Attention UNET: Attention-based UNET for 3D cerebrovascular segmentation of enhanced TOF-MRA images,” arXiv (Cornell University), Nov. 2023, doi: 10.48550/arxiv.2311.10224. [12] E. Rainville, A. Rasoulian, H. Rivaz, and Y. Xiao, “Weakly Supervised Intracranial Aneurysm Detection and Segmentation in MR angiography via Multi-task UNet with Vesselness Prior,” ArXiv.org, Aug. 2025, doi: 10.48550/arxiv.2508.00235. [13] N. Mu, Z. Lyu, M. Rezaeitaleshmahalleh, J. Tang, and J. Jiang, “An attention residual u-net with differential preprocessing and geometric postprocessing: Learning how to segment vasculature including intracranial aneurysms,” Medical Image Analysis, vol. 84, p. 102697, Nov. 2022, doi: 10.1016/j.media.2022.102697. [14] E. Aykaç, G. Göksungur, G. S. Albayrak, and M. E. Yüksel, “A Mask R-CNN-Based Approach for Brain Aneurysm Detection and Segmentation from TOF-MRA Data,” Brain Sciences, vol. 15, no. 12, p. 1295, Nov. 2025, doi: 10.3390/brainsci15121295. [15] K. M. Timmins et al., “Comparing methods of detecting and segmenting unruptured intracranial aneurysms on TOF-MRAS: The ADAM challenge,” NeuroImage, vol. 238, p. 118216, May 2021, doi: 10.1016/j.neuroimage.2021.118216. [16] S. S. M. Salehi, D. Erdogmus, and A. Gholipour, “Tversky Loss function for image segmentation using 3D fully convolutional deep networks,” Lecture Notes in Computer Science, pp. 379–387, Jan. 2017, doi: 10.1007/978-3-319-67389-9_44. [17] A. A. Taha and A. Hanbury, “Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool,” BMC Medical Imaging, vol. 15, no. 1, p. 29, Aug. 2015, doi: 10.1186/s12880-015- 0068-x.Department of Electrical & Computer EngineeringFYP Proceedings106

Smart Attendance System Using Face Recognition Identification Haikal Khusairi Ahmad and Khairayu Badron Department of Electrical and Computer Engineering, Kulliyyah of Engineering, International Islamic University Malaysia, Abstract—Traditional student attendance methods, such as manual roll calls and QR code scans, are time-consuming, prone to error, and lack automation capabilities. To address these challenges, this study develops an intelligent classroom management system that automates attendance tracking through facial recognition and real-time cloud data logging. The methodology utilizes a streaming camera module to capture facial images, which are processed using an optimized Python-based OpenCV pipeline. Attendance data is instantly synchronized and securely stored within a cloud-hosted spreadsheet for streamlined administrative access. To ensure algorithmic robustness, a multi- angle, variable-lighting dataset of student faces was used for evaluation. System performance was validated across key metrics, including recognition accuracy, processing latency, and power consumption across four hardware platforms. The experimental results demonstrate accurate real-time multi-face tracking and seamless attendance logging, significantly reducing manual workloads. Ultimately, this research offers critical insights into the optimization of practical, low-cost embedded systems for scalable classroom automation and data-driven administrative decision- making. Index Terms— Smart Attendance System, Facial Recognition, Raspberry Pi 5, OpenCV, Internet of Things (IoT), Classroom Automation. I. INTRODUCTION HE rapid growth of Internet of Things (IoT) technology and artificial intelligence has changed the way many systems operate today. These technologies are now widely used in areas such as education, healthcare, manufacturing, and security to improve automation and efficiency [1], [2]. In educational institutions and workplaces, attendance management is an important process because it helps organizations monitor participation, maintain accurate records, and support administrative activities [3]. However, traditional attendance methods such as signing attendance sheets manually, scanning RFID cards, using QR codes, or fingerprint verification still face several challenges. These methods can be time-consuming, prone to errors, and vulnerable to attendance fraud such as proxy attendance [7], [8], [12], [18]. To address these issues, facial recognition technology has gained attention as an alternative solution for attendance management systems [19], [21]. Facial recognition is a biometric technology that identifies individuals by analyzing unique facial features captured through a camera [22], [23]. Unlike conventional attendance methods, facial recognition allows attendance to be recorded automatically without requiring physical contact or manual input from users. This makes the process faster, more convenient, and more secure [20], [24], [25]. In addition, recent advancements in computer vision and machine learning have significantly improved the accuracy of facial recognition systems, allowing them to perform well under different lighting conditions, facial expressions, and viewing angles [5], [29]. Previous studies have shown that integrating facial recognition with IoT technology can create a more intelligent and efficient attendance management system [28], [30]. Through IoT connectivity, attendance information can be automatically transmitted, stored, and accessed in real time, reducing the need for manual record keeping [3], [28]. Despite these advantages, some existing systems require expensive hardware or complex system architectures, making them less practical for widespread implementation in schools and universities [21], [30]. Therefore, there is a growing need for a solution that is both cost-effective and easy to deploy while still providing reliable attendance monitoring. Another important consideration is the security and reliability of attendance records. Attendance systems should be able to prevent fraudulent activities while ensuring that records are stored accurately and can be accessed when needed [8]. By combining facial recognition with IoT technology, attendance data can be collected and managed automatically, reducing the possibility of human error and improving overall system efficiency [3], [28]. Furthermore, the use of Raspberry Pi as an embedded computing platform offers a low-cost and compact solution capable of performing real- time face detection and recognition tasks [30]. This study proposes an IoT-Based Facial Recognition Smart Attendance System using Raspberry Pi as the main processing unit. The system uses a camera module to capture facial images, while face detection and recognition algorithms are employed to identify registered users. Attendance records are then automatically stored and managed through IoT-based communication. The system is developed using Python and OpenCV to sup

system will provide a practical, secure, and efficient attendance management solution while overcoming the limitations of conventional attendance methods. The remainder of this paper is organized as follows. Section II presents the methodology, including the system design, hardware and software development, facial recognition process, and IoT integration. Section III discusses the experimental results and performance evaluation. Finally, Section IV concludes the study and suggests possible improvements for future work. II. METHODOLOGY This section outlines the methodologies, hardware, and software requirements implemented to develop the smart facial recognition attendance system. Following a comprehensive literature review of vision-based attendance framework, Haar Cascade and Local Binary Patterns Histograms (LBPH) were selected for their optimized performance on the Raspberry Pi 5 platform. The development pipeline initiates with data acquisition to establish a multi-angle, variable-lighting facial dataset of registered individuals. Subsequently, the system integration phase configures the Raspberry Pi 5 with a streaming camera module to validate real-time multi-face tracking and cloud-based data logging. Finally, the integrated architecture is evaluated under realistic operational conditions to benchmark recognition accuracy, processing latency, and operational constraints. A. Project Flowchart Figure 1 illustrates the methodology used to develop and evaluate the proposed Smart Attendance System using Face Recognition Identification. The process begins with a literature review, followed by system design and identification of functional requirements. Hardware development using Raspberry Pi and software configuration are then carried out before integrating all system components. The integrated system is tested to ensure it meets the project requirements. Performance evaluation is conducted based on face detection and recognition accuracy. Finally, the results are analyzed and discussed to assess the effectiveness of the proposed system and identify possible improvements. Fig. 1. Project Flowchart B. System Architecture Figure 2 shows the overall architecture of the proposed Facial Recognition Smart Attendance System. The system is divided into three main components: the input module, processing unit, and output module. The input module consists of a camera that captures real-time facial images of individuals for attendance monitoring. The captured images are transmitted to the processing unit, which is implemented using a Raspberry Pi. The Raspberry Pi performs face detection and recognition by comparing the detected facial features with the registered facial data stored in the system. Once a valid match is found, the system verifies the user's identity and automatically records the attendance information, including the name, date, and time. To prevent duplicate attendance records, each user is registered only once during a session. The output module stores the attendance records in an Excel (.xlsx) file, providing a simple and efficient method for data management and retrieval. Overall, the proposed architecture enables a contactless, automated, and reliable attendance management system through the integration of facial recognition technology and Raspberry Pi 5 processing. Fig. 2. System Architecture C. Dataset Collection The dataset used in this project consisted of facial images collected through real-time video recordings at the system deployment location, along with additional facial images obtained from publicly available open-source datasets. The use of multiple data sources helped improve the robustness and reliability of the facial recognition smart attendance system. The dataset included facial images captured from different angles and under varying environmental conditions. These variations allowed the system to learn and recognize facial features more effectively in real-world situations. Images with slight changes in facial orientation and expressions were included to enhance recognition accuracy. Lighting conditions were also considered during dataset preparation. Facial images captured in both well-lit environments and low-light conditions were included. This was important because facial recognition performance is generally affected under poor lighting conditions. By including low-light images, the system was better trained to handle challenging environments. Many facial images must be used as the datasetDepartment of Electrical & Computer EngineeringFYP Proceedings108

for this project. These images were utilized for facial feature extraction, training, and testing of the facial recognition attendance system. D. Face Detection using Haar Cascade Face detection in this project was performed using the Haar Cascade algorithm, which is a commonly used method for real- time face detection applications. The Haar Cascade algorithm scanned the input image at multiple scales to identify facial features such as the eyes, nose, and mouth. It utilized a set of pre-trained classifiers based on Haar-like features to differentiate facial regions from non-facial regions. During system operation as illustrates in figure 3, the images captured by the camera were first converted into grayscale to reduce computational complexity and improve processing speed. The Haar Cascade classifier was then applied to the grayscale image to detect the presence of faces. For each detected face, the algorithm generated a bounding box that defined the location and size of the facial region within the image frame. These detected facial regions were extracted and forwarded to the face recognition stage. Haar Cascade was selected for this project due to its fast execution speed and low computational requirements, making it suitable for real time implementation on the Raspberry Pi platform. Its efficiency ensured that face detection could be performed reliably without introducing significant processing delay. Fig. 3. Haar Cascade Algorithm workflow E. Face Recognition using Local Binary Pattern (LBP) Binary Pattern Histogram (LBPH) method as shown in figure 4. LBP is a texture-based facial recognition technique that represents facial features by analyzing the local texture patterns of the face. The method compares each pixel with its surrounding neighbours to generate a binary pattern that describes the facial texture. After a face was detected using the Haar Cascade algorithm, the extracted facial region was resized and normalized before being processed by the LBPH recognizer. During the training phase, facial images from the prepared dataset were used to generate LBP histograms for each registered individual. These histograms were stored as reference models in the system database. During the recognition phase, the LBP histogram of the detected face was compared with the stored histograms. A confidence value was calculated to determine the similarity between the input face and the stored facial data. If the confidence value was within the predefined threshold, the identity of the individual was successfully recognized and the attendance record was updated. Otherwise, the face was classified as unknown and no attendance entry was recorded. The LBPH method was chosen due to its robustness against lighting variations, low computational complexity, and effectiveness when working with a limited dataset. These characteristics made it suitable for real-time facial recognition on an embedded system such as the Raspberry Pi. Fig. 4. LBP Algorithm workflow III. RESULTS This chapter presents the results and discussion of the proposed Facial Recognition Smart Attendance System developed using Raspberry Pi 5. The chapter begins with the hardware setup and system configuration, followed by the dataset training process used to register and store facial information. The implementation of the system is then demonstrated to show its real-time operation. Next, the performance of the face detection and face recognition modules is evaluated to assess the accuracy and reliability of the system. The IoT-based attendance recording process is also presented, highlighting the automatic storage of attendance data. Finally, the overall system performance is analyzed and compared with findings from previous studies. The limitations of the system are discussed, followed by a summary of the results in relation to the project objectives. A. Hardware Setup and System Integration Fig. 5. Full Hardware Setup and System Integration PrototypeDepartment of Electrical & Computer EngineeringFYP Proceedings109

The integration of the whole system can be visualized from the prototype in Figure 5 above. Face detection and recognition tasks are fully processed locally by the Raspberry Pi 5 microcomputer using the live feed from the connected USB camera. The cloud- based Google Sheets database, on the other hand, is responsible for receiving and compiling the attendance records. Communication between the facial recognition system and the online database utilizes an internet connection, where Python network libraries are used to automatically push the recognition data and timestamps directly to the cloud. B. Dataset Training Fig. 6. Registered Facial Image Dataset Used for System Training Figure 6. illustrates the facial image dataset stored in the system database and used during the training phase of the Facial Recognition Smart Attendance System. Each image represents a registered user whose facial information has been collected and stored for future recognition. The images are organized within the Raspberry Pi file directory and labelled according to the corresponding individual's name. This dataset serves as the reference database that enables the system to identify and distinguish users during real-time attendance monitoring. Before the recognition process can be performed, the system must first learn the unique facial characteristics of each registered individual. The collected images are processed using OpenCV and the Local Binary Pattern Histogram (LBPH) algorithm to extract distinctive facial features. These features are then converted into numerical representations and stored as a trained model. During operation, newly captured facial images are compared against this trained database to determine the identity of the detected user. C. Face Recognition Result Figure 7 illustrates the face recognition process performed by the developed system Haar Cascade for face detection to locate the facial region and LBPH for face recognition to match the extracted features against the registered database. The system captures real-time video through a USB camera connected to the Raspberry Pi 5. Once a face is detected, a red rectangular bounding box is drawn around the detected facial region. The system then extracts facial features and compares them with the stored facial database to determine the identity of the individual. As shown in the figure 4.1, the detected user was successfully recognized and labeled with the corresponding registered name. The successful identification demonstrates that the LBPH algorithm is capable of distinguishing facial features and matching them with trained images stored in the database. The algorithm was selected because of its computational efficiency and suitability for embedded systems such as Raspberry Pi. The recognition process was completed within a short period, allowing the system to operate in real time without noticeable delay. Fig. 7. Face identification and the individual name is displayed D. Attendance Record Fig. 8. Attendance Records Stored in Google Sheets As shown in Figure 8, the database successfully captures and organizes the attendance information into two main columns: Timestamp and Staff. Every single time a user stands in front of the camera, the system creates a new row containing the exact date, hour, minute, and second of their arrival, followed immediately by their name. For example, the logs show Faizah checking in at 9:42:38 and Nabilah checking in just one second later at 9:42:39. This rapid, row-by-row logging proves that the backend pipeline is fast enough to handle continuous check-ins without dropping user data or delaying the list Additionally, Figure 7 shows how the system safely manages security concerns. On row 4, an unrecognized face was detected at 9:42:42 and automatically labeled as "Unknown". This ensures that unregistered people do not break the program or save corrupted files into the system. By instantly switching from image processing to live internet database updates, the project achieves its core objective of automatic, paperlessDepartment of Electrical & Computer EngineeringFYP Proceedings110

attendance recording, greatly improving day to day management efficiency E. Performance Analysis The data in Table I shows that the system is both accurate and fast. In real-world field tests, it achieved a 96.67% accuracy rate, successfully recognizing 29 out of 30 students despite variations in lighting and head positions. The average system latency of 1.4 seconds is highly acceptable for a small edge computing network. This speed is achieved because the system shrinks the video frames to a smaller size before processing them, reducing the workload on the Raspberry Pi 5 processor. Under perfect conditions with bright, even lighting and users looking straight at the camera the accuracy increases to over 99.0%. This difference shows that while the underlying code is highly precise, real-world environments introduce variables like shadows that can slightly affect performance. TABLE I SYSTEM PERFORMANCE AND EFFICIENCY ANALYSIS Metric Result Findings Recognition Accuracy 96.67% Successfully recognised 29 out of 30 students in the test set. Average Latency 1.4 seconds Measured from the moment of frame capture to the finalised attendance log in the database. Accuracy (Ideal Conditions) >99% Achieved in optimal lighting with frontal face capture. IV. CONCLUSION Overall, the proposed Facial Recognition Smart Attendance System was successfully developed and implemented using Raspberry Pi 5. The system was able to perform real-time face detection and recognition, enabling attendance to be recorded automatically without manual intervention. The integration of facial recognition technology with IoT-based data storage provided a contactless and efficient attendance management solution. The experimental results demonstrated that the system could accurately identify registered users and record attendance information, including the user's name, date, and time. In addition, the automated attendance logging process reduced the possibility of human error and eliminated issues associated with conventional attendance methods such as manual registration and proxy attendance. The findings indicate that the proposed system is capable of providing reliable and efficient attendance monitoring while maintaining a low-cost and compact implementation. Therefore, the developed system is suitable for use in educational institutions and organizations seeking a secure, automated, and user-friendly attendance management solution.. REFERENCES [1] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, and M. Ayyash, “Internet of Things: A survey on enabling technologies, protocols and applications,” IEEE Communications Surveys & Tutorials, 2015. [2] S. H. Alsamhi, O. Ma, and M. S. Ansari, “Survey on collaborative smart IoT in education,” IEEE Access, 2021. [3] P. S. Rokade and S. N. Kharche, “Smart attendance tracking system using IoT,” International Journal of Engineering Research, 2022. [4] M. A. Khan and K. Salah, “IoT security: Review, blockchain solutions and open challenges,” Future Generation Computer Systems, 2018. [5] Q. X. Ying, “Haar-VGG: Face Attendance System,” Journal of [6] Computer Vision and Pattern Recognition, 2025. [7] S. B. Reddy and P. Rao, “Smart Attendance System using Face Recognition and RFID Technology,” EAI Endorsed Transactions on Smart Cities, 2021. [8] K. Okereafor, “Public Safety Risks of Touch-Based Fingerprint [9] Biometric Systems During Infectious Disease [10] Outbreaks,” JMIR Public Health and Surveillance, 2020. [11] A. Nwabuwe et al., “Fraud Mitigation in Attendance Monitoring Systems: A Review of Biometric and Non- Biometric Approaches,” International Journal of Information Security Research, 2023. [12] L. S. Ezema, C. K. A. Joe-Uzuegbu, J. N. Eneh, and I. Amanze, “Fingerprint Based Attendance Management System,” International Journal of Scientific and Engineering Research, vol. 6, no. 7, pp. 1623– ?, Jul. 2019. [13] “Understanding choosing RFID tag based on the tag frequency,” [14] RFID4U. https://rfid4u.com/rfid-frequency/ (accessed Dec. 15, 2022). [15] J. J. Soriano, R. Cabatuan, E. Dadios, A. Calilung, and R. A. Medina, “Fingerprint Based Attendance Management System,” Proc. IEEE, 2019. [16] S. Nainan, R. Parekh, and T. Shah, “RFID Technology Based Attendance Management System,” CoRR, vol. abs/1306.5381, 2013, doi: 10.48550/arXiv.1306.5381. [17] Application Notes, “Introduction to RFID Technology,” [18] CAENRFID: The Art of Identification, 2008. [19] L. Srivastava, RFID: Technology, Applications and Policy Implications, Presentation, International Telecommunication Union, Kenya, 2005. [20] R. Rouillard, “Contextual QR Codes,” Proc. IEEE, 2009. [21] B. A. Talip and M. Z. Zulkifli, “Mobile Attendance System Using QR Codes Technology,” Journal of Computing Technologies and Creative Content, vol. 3, no. 1, pp. 1–3, 2019. [22] R. N. Khairnar, S. G. Ingale, A. S. Nichit, R. S. Barse, and P. U. Mandlik, “QR Code Based Student Attendance System,” International Journal of Advanced Research in Science, Communi

[30] Smita, P. S. Hegde, and A. Afshin, “Face Recognition Based Attendance Management System,” International Journal of Engineering Research & Technology (IJERT), vol. 9, no. 2, pp. 120–124, 2020. [31] S. Sawhney, K. Kacker, S. Jain, S. N. Singh, and R. Garg, “Real-Time Smart Attendance System Using Face Recognition Techniques,” International Journal of Computer Sciences and Engineering (IJCSE), vol. 7, no. 5, pp. 777–784, 2019. [32] M. S. Akbar, A. M. Al Ashray, P. Sarker, J. Uddin, and A. T. Mansoor, “Face Recognition and RFID Verified Attendance System,” International Conference on Computer and Information Technology (ICCIT), vol. 21, no. 4, [33] pp. 42–47, 2018. [34] S. Sawhney, K. Kacker, S. Jain, S. N. Singh, and R. Garg, “Real Time Smart Attendance System Using Face Recognition Techniques,” in Proc. 9th International Conference on Cloud Computing, Data Science & Engineering (Confluence), 2019, pp. xxx–xxx, doi: 10.1109/CONFLUENCE.2019.8776934. [35] D. D. Nguyen, X. H. Nguyen, T. T. Than, and M. S. Nguyen, “Automated attendance system in the classroom using artificial intelligence and Internet of Things technology,” in 2021 8th NAFOSTED Conference on Information and Computer Science (NICS), IEEE, 2021, pp. 531–536. [36] P. D. Mishra, “A comparative study of face recognition models for smart attendance,” IJRASET, vol. 11, no. 5, pp. 6666–6670, May 2023. [37] A. Kumar, V. Indragandhi, A. Chitra, R. Paul, and S. Banerjee, “Smart attendance management system using Raspberry Pi and deep learning technique,” in Proc. First International Conference on Combinatorial and Optimization (ICCAP), Dec. 7–8, 2021, Chennai, India, 2021.Department of Electrical & Computer EngineeringFYP Proceedings112

Titles Published in Other Publications 2 0 2 5 / 2 0 2 6

Department of Electrical & Computer Engineering FYP Proceedings 113 NO TITLE AND AUTHORS 1 A Personalized Digital Reading Tool for Dyslexic individuals Qistina Faqihah Binti Abdul Samat, Huda Adibah Mohd Ramli 2 AI Model Protection Against Model Extraction in Cyber-Physical Security Fatimah Azzahrah Binti Razali, Mohamed Hadi Habaebi `3 AI-Assisted Bounding Box Generation for YOLO Training in Object Detection Systems Muhammad Haiqal Danial Bin Mohamad Rasid, Amelia Wong Azman 4 Air-Writing Recognition System for Dyslexia-Friendly Learning Nik Nurasilah Binti Nik Aziz, Nur Shahida Binti Midi 5 Artificial Neural Network-Based Optimization of Free Space Optical Communication System for High Speed Data Transmission Insyirah Afiqah Binti Huzalene Deliza, Norun Farihah Abdul Malek 6 Attendance Monitoring System using IOT and PowerBI Muhammad Fahmi Bin Muhamad Fadzil, Siti Noorjannah Ibrahim 7 Classification of Meat Spoilage using Electronic Nose Data and Machine Learning Wan Ahzha Aiman Bin Wan Zamri Afendi, Faridah Abd Rahman, Aliza Aini Md Ralib 8 Computational Modelling of Earthing Resistance Based on the Fall-of-Potential Method Farris Asmawi Bin Amran, Mohd Fahmi Bin Hussin 9 Deep Learning-Based Political Fake News Detection Nursyazleen Syazrina Binti Mohammad Jasni, Noreha Abd. Malik 10 Design and Development of a Generative AI Chatbot using Retrieval-Augmented Generation (RAG) Frameworks for Academic Management and Admission Services Muhammad Syakir Aiman Bin Shahrol, Hasmah Mansor, Teddy Surya Gunawan 11 Design and Fabrication of a Portable Bioassay Imaging Unit for Brine Shrimp Toxicity Assessment Nur Irdina Fatini Binti Mohd Saupi, Ahmad Fairuzabadi Bin Mohd Mansor 12 Development of a DC Series Arc Fault Detection System for Photovoltaic Installations using Discrete Wavelet Transform Khairun Nazmi B Razian, Mohd Shahrin Bin Abu Hanifah 13 Development of a Rubber-Based Antenna at 2.4 GHz Mohamad Ammar Bin Azmil, Sarah Yasmin Mohamad 14 Development of a Weight Sensor using Macrobending based Single Mode Optical Fiber Nik Arief Irfan Bin Nik Ab Rahman, Norazlina Bt Saidin 15 Development of an AI-Based Bioassay Analysis Module for Brine Shrimp Toxicity Assessment Nuraina Faqihah Binti Norhisham, Ahmad Fairuzabadi Bin Mohd Mansor

Department of Electrical & Computer Engineering FYP Proceedings 114 NO TITLE AND AUTHORS 16 Development of an Arabic–Malay Online Dictionary with for Language Learners using 3-Letter Root Word System Nur Syuhada Binti Mohd Ridzuan Lean, Siti Noorjannah Ibrahim 17 Development of Finger-Actuated Microfluidic Pump for Sample Transport in Pathogen Detection Application Muhammad Imran Bin Affendi, Rosminazuin Ab. Rahim 18 Development of Piezoelectric Nanogenerator using ZnO Nanomaterial for Wearable Applications Muhammad Nabil Bin Bakar, Aliza Aini Md Ralib 20 Exclude Analysis of a Variable Frequency Drive (VFD) for Electric Vehicles (EV) Mohamad Izzat Za'Im Bin Abdul Khalid, A.H.M. Zahirul Alam 21 Finite Element Analysis on Thermal Stress and Warpage Bifurcation of Thin Wafer Packaging Muhammad Azam Bin Rahmat, Nurul Arfah Binti Che Mustapha, Suazlan Bin Mt Aznam 22 Graphite-on-Paper Temperature Sensors using Direct Pencil Drawing Technique Rosnah Binti Abdullah, Norazlina Bt Saidin 23 Intelligent Optimization for Improved Battery Recycling Efficiency Nur Dania Batrisyia Binti Amirudin, Siti Hajar Binti Yusoff 24 Investigation of Earth Resistance Measurement using the Fall of Potential Method Muhammad Aiman Bin Fazil, Mohd Fahmi Bin Hussin 25 Investigation of Thermal–Mechanical Effects in Soft-Walled Microfluidic Devices for LAMP- Based Virus Sensor Khairul Harisin Bin Ahmad Shafri, Rosminazuin Ab. Rahim 26 IOT-Based Battery Management and Active Load Shedding Strategy for Solar-Powered Fertigation Harmi Hazwan Arifi Bin Harmi Thamri, Mohd Shahrin Bin Abu Hanifah 27 Machine Learning Based Energy Management System for Electric Vehicle Charging Network Saidatina A'Ishah Binti Abd Kadir Jailani, Siti Hajar Binti Yusoff 28 Modelling of Partial Discharge in Different Void Geometry under DC Stress Muhamad Haiqal Bin Mohamad Rasidi, Nur Shahida Binti Midi 29 N8N GenAI for Computer Networks Monitoring and Management Alif Najwan Bin Hishamuddin, Mohamed Hadi Habaebi 30 Optimization of DC-DC Buck Converter Based on Conducted Emission using Machine Learning Amin Ikhwan Bin Ahmad Nizam, Siti Noorjannah Ibrahim 31 PPE Detection using Computer Vision for Workplace Safety Muhammad Faris Bin Mazlan, Nurul Fariza Zulkurnain 32 Predictive Modelling of Levelized Cost of Energy for Solar Photovoltaic Systems in Malaysia Ain Syazwani Binti Shahril Azman, Siti Hajar Binti Yusoff

Department of Electrical & Computer Engineering FYP Proceedings 115 NO TITLE AND AUTHORS 33 Random Forest-Based Predictive Modeling for Student Specialization Stream Recommendation Muhammad Izzamie Bin Mohd Yusop, Nor Farahidah Za'Bah 34 Reduced - Switch Multilevel inverters (RS-MLI) for Electric Vehicle Applications Muhammad Aidil Syazril Bin Khairul Anuar, Nurul Arfah Bt. Che Mustapha 35 Speech-Based Stress Detection for Children and Adolescents via Convolutional Neural Networks Nur Dayana Binti Juanidi, Ani Liza Asnawi 36 Star Sensor Star Pattern Recognition in Lost in Space Mode for Space Application Muhammad Hafiz Shauqi Bin Mohd Yusri, Yasser Asrul 37 Starlink Latency Degradation Analysis and Classification using LSTM in a Tropical Region Ruqayyah Nur Balqis Binti Mohamad Habibullah, Yasser Asrul 38 Tropospheric Scintillation Prediction Model for Ku-Band Satellite System in Kuala Lumpur using Curve Fitting Method Qaisara Batrisyia Binti Mohamad Sazaly, Nadirah Abdul Rahim 39 Ultra-Low Voltage Hybrid-GDI Full Adder for IOT Edge Devices with Improved Energy Efficiency at near Threshold Operation Muhammad Syazwan Bin Mat Saat, Nurul Arfah Bt. Che Mustapha 40 Wearable Health Monitoring System for the Elderly using IOT Sensors Ahmad Fakhri Hirzan Bin Murad, Hasmah Mansor 41 Development of an Image Processing Framework to Measure Green Coverage and Predict Oxygen Levels in City Environment Muhammad Al-Kasysyaf Bin Mohamad Nor, Khairayu Badron 42 AI-Based Dynamic Spectrum Allocation Management System in 5G Networks Using Reinforcement Learning Muhammad Haziq Bin Halim, Farah Diyana Abdul Rahman 43 AI-based Color Detection and Quality Inspection System for Automotive Painting in Car Manufacturing Ibad Harzan Bin Hasri, Farah Diyana Abdul Rahman 44 Thermal simulation for optimized heat sink and cooling fin for airborne electronic package Alhaziq Bin Azhar, Anis Nurashikin Nordin 45 Development of Eco-Friendly Laser-Induced Graphene Electrodes using Natural-Wax Paper Substrates Iman Haniff Bin Jaffrey, Anis Nurashikin Nordin 46 High Precision General Purpose Comparator with Reduced Offset Voltage Muhammad Aqeef Fahmi Bin Mohd Fariq, Anis Nurashikin Nordin 47 Detecting Deepfake Videos Through Temporal Analysis of Facial Landmarks and Audio Synchronization Using Machine Learning Nik Nurul Zulaikha Binti Nik Robidin, Othman O-Khalifa

Department of Electrical & Computer Engineering FYP Proceedings 116 NO TITLE AND AUTHORS 48 Deep Learning for Spectrum Sensing in Cognitive Radio Networks Muhammad Fahmi Bin Azaha, Khaizuran Abdullah

Department of Electrical and Computer Engineering Kulliyyah of Engineering International Islamic University Malaysia P.O.Box 10 50728 Kuala Lumpur Malaysia https://engine.iium.edu.my/ece/