Computing and Computational Intelligence (CCI) 1
Track A1F1 CCI 1.1: Computing and Computational Intelligence (CCI) 1.1
Room: F1. Sipadan I (Level 4)
Chair: Khairil Anuar (Multimedia Universiti, Malaysia)
8:00 Benign Versus Malignant Classification of Lesions on Intraoral Fluorescent Images Obtained Using Blue Light Irradiation
Asahi Takahashi, Keishu Sano, Shigeo Ishikawa and Tadanori Fukami (Yamagata University, Japan)
In recent years, a technique for detecting diseases by irradiating the oral mucosa with blue light has been attracting attention. Under blue light irradiation, diseased tissues can be observed as a dark color because of a decrease in flavin adenine dinucleotide and destruction of collagen cross-linking structure in the diseased tissues, while blue-green fluorescence is observed in healthy tissues. However, the determination of tumor as benign or malignant (cancer) is still dependent on the knowledge and experience of the oncologist. In this study, we attempted to globally detect lesion regions using Mask R-CNN, an instance segmentation method, and to then locally classify the region into three classes (healthy tissue, benign tumor, cancer) using SVM (support vector machine) and Resnet (residual neural network). The results showed an accuracy of 70.7%. There were no cases in which cancer was classified as healthy tissue, but 27.0% of cases were misclassified as benign tumors.
8:15 Nrithya: a Curated Image Dataset of Indian Folk and South Indian Classical Dance Forms
Malavika V, Karthika Subbaraj and Srinivasan R (Sri Sivasubramaniya Nadar College of Engineering, India)
This study presents two carefully curated image datasets capturing the rich visual diversity of Indian folk and classical dance forms. The first dataset focuses on four South Indian classical dance styles - Bharatanatyam, Kathakali, Kuchipudi and Mohiniyattam - initially comprising 410 images, expanded to 1,621 through image augmentation. The second dataset encompasses six Indian folk dances - Bhavai, Garba, Karakattam, Malwai Giddha, Toda and Yakshagana - growing from 600 to 2,368 images after augmentation. All images were sourced from publicly available online platforms and YouTube videos, ensuring diversity in performers, regional styles, costumes and poses. Preprocessing techniques such as background removal and Canny edge detection were applied for costume-based feature extraction, and their effectiveness was systematically evaluated. This work addresses a critical gap in the availability of ethnically representative dance datasets by providing curated collections to support a range of computer vision tasks, including posture recognition, dance form classification, and cultural analytics - thereby contributing to both cultural heritage preservation and AI-driven dance research.
8:30 Comparative Analysis of Transformer-Based Models for Bail Prediction Using HLDC Dataset
Karthika Subbaraj, Aishwarya R and Nirranjana R (Sri Sivasubramaniya Nadar College of Engineering, India); Vishranth Msk (Vellore Institute of Technology, India)
This research presents a transformer-based approach to predict bail outcome using the Hindi Legal Documents Corpus (HLDC), a large-scale dataset consisting of legal texts from Indian district courts. The work employs a TFIDF-based extractive summarization technique to extract key factual information, enhancing relevance of input and shortening sequence length. Four transformer models-MuRIL, DeBERTa, DistilBERT, and mBERT-are fine-tuned and evaluated within a rigorously optimized experimental framework. Each model was trained for a binary classification task, to predict whether bail should be 'granted' or 'denied'. Experimental results show that MuRIL exhibits the highest performance with 84% accuracy, an AUC of 0.89, and an average precision of 0.94, which attributes to MuRIL's strong alignment with Indian language semantics and its synergy with the TF-IDF-based sentence ranking strategy. These findings validate the importance of linguistic alignment and input relevance in high-stakes legal NLP tasks. The proposed approach offers a scalable and interpretable framework for decision support in India's judicial context.
8:45 Scalable Face Recognition with Multiple Deep Learning Model Integration
Pawan Kamalkishor Ajmera, Ramakant Soni, Anmol Mathur, Jayesh Harlalka and Vipin Saraogi (BITS Pilani, India)
This paper presents the design of a scalable face recognition system to address real-world challenges including illumination-variation, occlusions and motion blur. This proposed method integrates multiple state-of-the-art models into a unified pipeline: YOLOv5-Face is used for preparation of the dataset, SCRFD for face detection with high accuracy and face alignment, ArcFace for discriminative feature embedding generation and BYTETracker for continuous multi-face tracking. Experimentations conducted on a locally generated dataset demonstrate strong performance by achieving the training accuracy of 90.13%, validation accuracy of 89.02%, and a ROC AUC score of 0.9592. The system is able to sustain real-time performance with 25-30 FPS and low latency of 33-40 ms per frame. These results make this system suitable for large-scale surveillance applications. The results also highlight the effectiveness of multi-model integration in improving the accuracy and adaptability when compared to traditional approaches. This paper contributes a practical solution for deployment in the domains such as campus surveillance, access control and crowd monitoring. It also provides foundation for future work which include liveness detection, super-resolution for low resolution recognition and lightweight edge-compatible model implementation.
9:00 Comparing Deep Learning Approaches for Low Resolution Face Recognition Challenges in Security and Surveillance Applications
Ramakant Soni and Pawan Kamalkishor Ajmera (BITS Pilani, India)
Facial recognition in the low-resolution environments remains as a critical challenge for security and surveillance systems. This paper comprehensively evaluates two advanced deep learning models: MobileNetV2 and EfficientNet-B0 across four facial datasets: LFW, LFW Funneled, TinyFace and a locally collected low-resolution dataset (Local LR). Using PyTorch with extensive data augmentation, model's performance was assessed on metrics: Accuracy, Loss, F1-score and ROC AUC. The experimental results shows that EfficientNet-B0 achieves higher accuracy on challenging low-resolution datasets like TinyFace (69.64% compared to MobileNetV2's 61.86%). EfficientNet-B0 shows 29.3% reduction in test loss and 6.4% higher F1-score maintaining comparable performance on higher-resolution datasets (97.81% on Local LR). MobileNetV2 shows more computational advantages with fewer parameters and lower FLOPs in comparison to EfficientNet-B0. This makes it more suitable for edge deployments despite slightly lower accuracy (96.43%) on LFW compared to EfficientNet-B0 (96.97%). These findings provide quantitative benchmarks for model selection. Results are highlighting EfficientNet-B0's strength in feature extraction from low-quality images and reflects the efficiency benefits of MobileNetV2 which is offering practical guidance for implementing facial recognition systems in real-world scenarios where image resolution and hardware constraints vary significantly.
9:15 Data-Driven Approach for Dynamic Comparator in Analog-to-Digital Conversion
Buddhi Prakash Sharma, Anu Gupta and Chandra Shekhar (BITS Pilani, India)
The performance of dynamic comparators significantly influences the efficiency, speed, and accuracy of Analog-to-Digital Converters (ADCs), especially in modern low-power and high-speed applications. Traditional design methodologies relying on exhaustive simulations and manual tuning are inefficient and fail to adapt to process, voltage, and temperature (PVT) variations. This paper proposes a comprehensive data-driven optimization framework that leverages machine learning (ML) and artificial intelligence (AI) to enhance dynamic comparator performance across key metrics. The framework introduces a four-stage pipeline comprising feature extraction from simulation and measurement data, supervised learning for offset voltage prediction, reinforcement learning (RL) for adaptive calibration, and Bayesian optimization for automated sizing of circuit components. These methods collectively reduce design complexity, improve accuracy and power efficiency, and enable real-time adaptability. While integration with Electronic Design Automation (EDA) tools remains a future direction, the proposed framework lays the foundation for scalable, intelligent comparator design targeting applications such as IoT, biomedical systems, and high-speed communications.
Control Systems & Robotics (CSR)
Track A1F2 CSR 1: Control Systems & Robotics (CSR) 1
Room: F2. 501 Kadamaian (Level 5)
Chair: Fatanah Mohamad Suhaimi (Universiti Sains Malaysia, Malaysia)
8:00 Data Driven Approach for Rapid Collision Checking Using Cross-Attention on Robot and Obstacle Sets
Chayan Maiti and Deep Patel (IIITDM Kancheepuram, India); Sreekumar Muthuswamy (IIITD&M Kancheepuram, India)
Collision checking is the process of determining whether a candidate robot configuration or trajectory intersects any obstacles in the workspace, and it underpins all safe motion-planning algorithms. Reliable real-time collision checking is critical for safe, autonomous robot manipulation in dynamic industrial settings. This paper presents a learning-based collision checking framework that leverages Set Transformers with cross-attention mechanisms to rapidly and accurately predict collisions between robotic manipulators and complex, variable obstacle environments. Unlike traditional geometric approaches, the proposed method efficiently processes a robot's configuration and a dynamic obstacle set to estimate collision likelihood in real-time. A synthetic dataset generated via CoppeliaSim for three industrial robot models (UR5, UR10, ABB IRB140) is used for training and evaluation. Experimental results demonstrate an average classification accuracy of 89.6%, with the model significantly outperforming established learning-based methods such as Fastron and CollisionGP, while achieving up to 7.6× faster inference than PyBullet-based checks. The proposed approach offers a scalable, generalizable, and computationally efficient solution for safe robot operation in dynamic and unstructured environments.
8:15 Modified INN Control Law for Efficient Flow Regulation with Pneumatic Control Valve
Manmahendra Singh Daksh and Puneet Mishra (Birla Institute of Technology and Science, Pilani, India)
In the industrial process plant, the control valve is the key component in regulating the flow of fluid that is done by the controller. Hence, the controller is crucial in improving the plant's overall efficiency. This paper proposes a new modified Inverse Neural Network (INN) control law that provides better setpoint tracking and disturbance rejection performance compared to the existing INN control law while being tested on a laboratory scale flow process to regulate fluid flow across the pneumatic control valve. In this proposed work, the data for process variables and controller outputs are taken for different set points, and then the identification of system parameters is done with the help of the Particle Swarm Optimization (PSO) technique. The identified parameters are used as initial parameters, while the updated INN control and existing INN control for the plant are tested for optimized control performance. In this process of hardware in the loop, the laboratory scale flow process is interfaced with the help of a DAQ USB device, NI USB-6001, and the flow across the valve gets controlled in real-time. In the case of the proposed modified INN controller (MOD-INN), the plant's actual output tracks the set point precisely with lower Integral Absolute Error (IAE) compared to existing INN control to regulate fluid flow across the control valve for different setpoints, as well as in the case of an introduced disturbance in the setpoint. Quantitatively, in terms of IAE, a 13.44 times improvement was achieved by MOD-INN over the existing INN controller under servo performance for a 1.2V value of setpoint, while a 10.94 times improvement was observed under regulatory performance.
8:30 Global Synchronization of Delayed Discrete-Time Networks Using Discrete Contraction Approach
Yashasvi Chauhan and Bharat Bhushan Sharma (National Institute of Technology Hamirpur, India)
This paper investigates the synchronization problem of delayed discrete-time systems interconnected in complex networks using contraction framework. The proposed approach provides a novel perspective on achieving global synchronization by using discrete contraction analysis, which ensures exponential convergence of the discrete networked system trajectories despite time-varying delays. The method systematically handles heterogeneous delays and nonlinear couplings while avoiding restrictive Lipschitz assumptions. This framework addresses multiple time-varying delays, including intrinsic system delays and coupling delays, to ensure robust stability in discrete-time networks. We establish sufficient synchronization conditions by constructing a contraction-based auxiliary system and deriving delay-dependent stability constraints in terms of matrix inequalities. Numerical simulations on discrete-time FitzHugh-Nagumo neural network demonstrate the effectiveness of the proposed methodology. Compared to existing Lyapunov-based techniques, proposed contraction approach offers improved convergence rates and reduced conservatism in stability conditions, making it more suitable to real-world applications involving secure communication, sensor networks, distributed control and large-scale interconnected systems with complex time-delay dynamics.
8:45 Advanced 3D Path Planning for Robotic Calligraphy Based on LLM-Driven Text Prompts
Cheuk Tung Shadow Yiu (The Hong Kong University of Science and Technology, Hong Kong); Dick Ho Cheung (Hong Kong University of Science and Technology, Hong Kong); Haolun Huang (HKUST, Hong Kong); Kam-Tim Woo (Hong Kong University of Science and Technology, Hong Kong)
Chinese calligraphy is a complicated art form that requires precise control of movement and brush dynamics, making it a challenging task for robotic systems. This paper presents a novel approach to enabling robotic manipulators to perform Chinese calligraphy using 3D path planning algorithms and large language model (LLM) as input generators. The proposed framework integrates an LLM to generate textual content prompts, which are then transformed into detailed 3D Bézier curve trajectories which represent the strokes of Chinese characters. These trajectories are optimized to ensure smooth and accurate motion of the robotic manipulator, taking into account the brush's pressure, orientation, and speed to emulate traditional calligraphy techniques. Experimental results demonstrate the effectiveness of the approach in producing visually authentic Chinese calligraphy with high precision and artistic quality. This work highlights the potential of combining LLM with advanced robotic path planning to bridge the gap between traditional art and robot world.
9:00 LiDAR-Based Cooperative Tracking Using IMM Extended Object Tracking and Information Filter
Yoshihiro Nakatani, Shuzo Uchiyama, Masafumi Hashimoto and Kazuhiko Takahashi (Doshisha University, Japan)
This paper presents a cooperative method for tracking objects, such as cars, two-wheelers, and pedestrians, using multiple light detection and ranging sensors (LiDARs) deployed in a road environment. Each LiDAR system independently estimates the motion and shape of objects, e.g., in terms of their positions and velocities, and sizes, within its field of view. A central server collects and fuses these estimates from LiDAR systems. The proposed approach uses an extended object-tracking method based on the multiplicative error model-extended Kalman filter (MEM-EKF) framework combined with an interacting multiple model estimator to accurately track the motions and shapes of objects exhibiting various movement patterns, such as stopping, constant velocity motion, and suddenly moving and stopping. In addition, information filter-based fusion is performed at the central server. Simulation results using two LiDAR systems in an intersection environment demonstrate the effectiveness of the proposed method compared to the conventional single-model MEM-EKF method and standalone tracking method with a single LiDAR system.
9:15 An Improved Auction Algorithm for Multi-Robot Task Allocation
Kuang Jui Lim, Shalini Darmaraju and Choon-Hian Goh (Universiti Tunku Abdul Rahman, Malaysia)
The problem of task allocation in multi-robot systems (MRTA) is critical for optimizing performance in various applications, such as warehouse automation and search and rescue operations. Current algorithms often fail to consider real-life complexities like robot dynamics and energy consumption, leading to inefficiencies. This study aims to address these gaps by proposing improved versions of the auction algorithm (A) and the enhanced auction algorithm (EA), focusing on optimizing task allocation by considering task complexity and robot capabilities. The methodology involves using auction and extended auction algorithms as the backbone, with two improved algorithms proposed to enhance fairness and efficiency. An energy handling function is also incorporated. Six scenarios were simulated, involving 10, 15, and 25 robots with 100 and 200 tasks, respectively. The mean and standard deviation of completion times were computed to evaluate performance. Results indicate that the improved auction (IA) algorithm consistently outperforms the auction (A) algorithm, while the improved enhanced auction (IEA) algorithm generally performs better than the enhanced auction (EA) algorithm. For instance, in scenarios with 25 robots and 200 tasks, the standard deviation of A, IA, EA, and IEA is 62.7s, 51.7s, 21.3s and 20.5s, respectively. A lower standard deviation indicates a better distribution. Additionally, the energy handling function significantly improved task allocation efficiency. In conclusion, our improved algorithms demonstrate better performance with the added energy handling function enhancing practical applicability. Future work should address real-life robot dynamics and communication delays to further improve the system's efficiency and effectiveness.
Power, Energy & Electrical Systems (PES)
Track A1F3 PES 1: Power, Energy & Electrical Systems (PES) 1
Room: F3. 502 Mesilau (Level 5)
Chair: Madihah Md Rasid (Universiti Teknologi Malaysia, Malaysia)
8:00 Optimal Integration of Solar-Powered EV Charging Stations in Radial Distribution System
Raj Chakraborty and Punam Das (National Institute of Technology Agartala, India); Sadhan Gope (NIT Agartala, India); Diptanu Das and Priyanath Das (National Institute of Technology Agartala, India)
Electric Vehicles (EVs) are critical to sustainable transportation systems owing to their energy efficiency and low environmental impact. However, their unpredictable charging patterns introduce stochastic loads into Radial Distribution System (RDS), leading to power quality degradation and elevated grid losses. This study presents a strategic methodology for deploying EV Charging Stations (EVCS) integrated with Photovoltaic (PV) units in RDS, aiming to minimize Active Power Loss (APL) and maintain a stable voltage profile. To determine the role of PV systems in enhancing network efficiency, the network parameters were analyzed under both solar-augmented and conventional charging configurations. Additionally, the PV system capacity was optimized to align with grid requirements. The proposed framework was tested on the IEEE 69-bus RDS, partitioned into four zones, each hosting one EVCS and one PV unit to balance charging accessibility with network efficiency. A Backward/Forward Sweep Power Flow (BFSPF) analysis was conducted to account for dynamic electrical parameters. The placement problem was formulated as an optimization problem and solved using Symbiotic Organisms Search (SOS) algorithm which demonstrated a reduction in APL of approximately 6.20% upon integrating PV units with EVCS.
8:15 Comparison of Deep Learning and Statistical Methods for EV Charging Demand Forecasting
Maizatul Shafiqah Sharul Anuar (Universiti Teknologi PETRONAS, Malaysia); Illani Mohd Nawi and Saeed S. Ba Hashwan (Universiti Teknologi Petronas, Malaysia)
The growing use of electric vehicles (EVs) poses challenges for maintaining power grid stability and managing energy effectively. Predicting EV charging demand accurately is essential for improving charging infrastructure and ensuring the reliability of the grid. This research evaluates two time-series forecasting models: Long Short-Term Memory (LSTM) networks and the Autoregressive Integrated Moving Average (ARIMA) model. The LSTM model, based on deep learning, captures complex time-dependent and nonlinear patterns in EV charging data, whereas ARIMA, a statistical approach, handles linear trends and seasonality. Both models were trained on 90% of the data and tested on the remaining 10%, with performance evaluated over a 90-day forecast horizon using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The LSTM model achieved significantly better forecasting accuracy compared to ARIMA. These results highlight the effectiveness of deep learning models in capturing complex, nonlinear patterns in EV charging behavior, offering a more reliable solution for demand forecasting in smart grid environments.
8:30 Minimizing Operation Cost of Fast Charging Station with Post-FIT PV by EV Charging Preference and Electric Market Price's Scenarios
Kento Imai and Shigeru Tamura (Meiji University, Japan)
Electric vehicles are rapidly penetrating and the number of their fast charging stations is also rising. Regarding the revenue of fast charging station, photovoltaic generation and battery energy storage supplying electricity, and utilizing power market effectively are necessary. There has been growing number of post-feed-in-tariff photovoltaic generation and it can be bought at a cheaper price in Japan. However, the revenue of fast charging station includes uncertainties such as electric vehicle's charging demand, photovoltaic generation, and power market price. The uncertain charging demand is acquired based on surveys of EV charging preference and traffic census data in this study. It also deals with the other uncertainties by using a clustering method, X-means, to generate the scenarios used for stochastic programming. The study aims to find the optimal capacity of battery and amount of post-feed-in-tariff photovoltaic generation to be bought under such uncertainties. The various prices of post-feed-in-tariff photovoltaic generation affecting the revenue are also studied.
8:45 Enhancing Thermal Imaging Accuracy for Power System Maintenance
Hasanka Gayan Somaweera, Gethwan Suriyawansa, Rusira Thispura Thilakarathna, Charithri Yapa and Nimantha Madhushan (University of Sri Jayewardenepura, Sri Lanka); Hasith Kapilapadma Liyanage (UUniversity of Peradeniya, Sri Lanka)
Accurate temperature measurement of electrical components is crucial for the health monitoring of electrical components for periodic maintenance. Traditional thermal cameras can apply a single emissivity value across the entire image. This gives an inaccurate reading due to different kinds of materials and factors like elevation from sea level, environmental temperature, and relative humidity. This makes analyzing a large number of thermal images ineffective. This research presents a novel method for machine learning-based object identification by practically applying formulas to increase accuracy considering all the factors. The proposed method can dynamically adjust the emissivity values based on the characteristics on the materials and the environmental conditions. The results show a significant improvement in the temperature of each component. This can improve better monitoring of components health as well as optimize the maintenance schedules. This leads to increase the reliability and efficiency of the electrical systems for periodic maintenance in the electrical industry.
9:00 Deep Learning-Based Resident Change Detection Using Electricity Metering Data
Jongin Kim, Dayeon Lee and Dong Sik Kim (Hankuk University of Foreign Studies, Korea (South)); Beom Jin Chung (Seoul National University of Science & Technology, Korea (South)); Young Mo Chung (Hansung University, Korea (South))
As smart meter technology advances, the advanced metering infrastructure (AMI) enables real-time monitoring of electricity usage at 15-minute intervals. Since prior studies mainly focused on industrial infrastructures, anomaly detection in residential electricity usage remains underexplored. In this paper, we propose a gated recurrent unit (GRU)-based deep learning model to detect resident changes in residential households using the electricity metering data through AMI. Monthly consumption patterns and energy consumption are predicted using actual electric energy metering data collected from 846 apartment households in Seoul, Korea. To detect the resident changes, combined losses are computed by integrating the consumption pattern and consumption energy prediction errors. The experimental results show an accuracy of 82.55%, specificity of 100%, and recall of 65.11% when allowing a +-1-month error margin. The proposed deep leaning model demonstrates the practical feasibility of detecting resident changes in residential households and is expected to contribute to the development of smart grid services.
9:15 Recent Evolution of Intelligent Approach in Electric Motor Drive System for EV Applications: a Concise Review
Syazmie Sepeeh (MAHSA University, Malaysia); Shamsul Aizam Zulkifli (UTHM, Malaysia & Universiti Tun Hussein Onn Malaysia, Malaysia); Huang-Jen Chiu (National Taiwan University of Science and Technology, Taiwan); Hanis Farhah Jamahori (Universiti Teknologi PETRONAS, Malaysia)
The fast development of electric vehicles (EVs) has motivated major studies on smart management techniques for electric motors, which are essential for guaranteeing optimal performance, efficiency, and dependability. This brief paper offers a summary of recent advances in artificial intelligence (AI)-based techniques used in electric motor management inside EV applications. The conversation covers several artificial intelligence models, including fuzzy logic (FL) systems, evolutionary algorithms (EA), machine learning (ML), and deep learning (DL). Every approach is quickly examined in terms of its working principles, benefits, drawbacks, and appropriateness for particular motor control duties, including speed control, torque optimisation, defect identification, and adaptive control. Particular focus is on how AI methods combine with motor types often used in EVs, such as permanent magnet synchronous motor (PMSM) and induction motor (IM). The paper highlights current trends and research opportunities to provide researchers and practitioners a modest but instructive reference for enhancing intelligent motor control in modern EV systems.
Electronics, Circuits & Devices (ECD)
Track A1F4 ECD 1: Electronics, Circuits & Devices (ECD) 1
Room: F4. 503 Dinawan (Level 5)
Chair: Maizatul Zolkapli (Universiti Teknologi MARA, Malaysia)
8:00 TCAD Study of SEE-Induced Reliability Analysis of PDSOI FinFET Based Capacitorless 1T DRAM
Mitali Rathi (NIT RAIPUR, India); Guru Prasad Mishra (NIT Raipur, India)
A comprehensive TCAD study of Single event effect on Capacitorless 1T-DRAM based on Partially depleted silicon on insulator (PDSOI) FinFET is demonstrated in this work. The device is well calibrated through 3D simulations using SILVACO ATLAS TCAD tool. First the memory operation and device characteristics is studied. The program operation is done by using Impact Ionization (II) method. Further the impact of high energy particle (HEP) strike on the Capacitorless 1T- DRAM operations such as Write-Hold-Read is carried out. The effect of various single event upset parameter such as HEP strike time, linear energy transfer (LET) value, HEP strike position are also evaluated. It is observed that as the device is scaled, it is highly susceptible to Single Event Effects (SEEs), even the drain and source regions are highly sensitive to SEEs. Hence, there is bit flip occurring in the device, and the read operation is corrupted, leading to malfunction of the DRAM memory.
8:15 Analysis of Dielectric Modulated Gate-All-Around (GAA) Junction-Less Transistor Based Biosensor
Achinta Baidya (Mizoram University, India); Vishal Kumar Pandey, Aditya Abhiraj and Kartik Swamy (Electronics and Communication Engineering, Mizoram University, India); Lalthanpuii Khiangte (National Institute of Technology Mizoram, India)
This work investigates a junction-less transistor (JLT) with gate all around (GAA) structure and dual cavity for the biomolecule sensing application. The GAA - JLT structure design with cavity exhibits good subthreshold swing and drain induced barrier lowering. The nanocavity is so placed below the gate that dielectric change due to the presence of biomolecule modulates the device's electrostatic characteristics. GAA-JLT have shown a significant variation in respect of threshold voltage, on-current, off current, ION/IOFF ratio due to the change in biomolecule in cavity. Due to immobilized higher dielectric biomolecules in cavity increase in gate capacitance occurs which results in increase in drain current, positive shift of threshold voltage, decrease of IOFF current. The device's sensitivity can be understood from the fluctuations of these parameters. The proposed structure with two 10 nm × 3 nm nanocavity has shown significant bio-sensing capability for detecting the biomolecules. All sensitivities of the biosensor are evaluated for biomolecules with dielectric constants (k) ranging from 1.6 to 12. Maximum ION/IOFF ratio achieved is 9.42× 10 9, Sensitivity (S(I on /I off )) ranges from 1.08 to 1357.6.
8:30 Design and Simulation of Multibeam Cylindrical EIK Cavities for a Ka-Band Klystron
Soumaya Mandal (NIT Mizoram, India); Santigopal Maity (NATIONAL INSTITUTE OF TECHNOLOGY MIZORAM, India); Chaitali Koley (National Institute of Technology, Mizoram, India)
This study reports on the computer-aided design and simulation of intermediate and input/output cavities configured in 2-pi Mode for a 4-beam Ka band klystron. Coaxial type extended interaction cavities have been investigated to achieve higher R/Q value. A 3-gap input/output cavity, in which, the central conductor was optimized for symmetric field distribution and uniform transmission coefficients, has been proposed for this multiple beam klystron (MBK). The coaxial cavity configuration with higher order mode allows usage of larger size cavities for the ease in fabrication complexity along with the stability compared to square shaped cavities. Here an R/Q has been achieved around 166 for intermediate cavity and 116 for I/O cavities. The Qext also has been analyzed with the variation of coupling slots. The proposed MBK has potential applications in high data rate millimeter-wave communication systems, satellite links, and 5G wireless backhaul networks, for which Ka-band frequencies are gaining increasing prominence.
8:45 An Advanced Electrothermal Small-Signal Compact Modeling of 250 nm GaN HEMTs Using ASM Methodology
Vijay Maitra, Ellapu Bhanu Prakash and Anil Prasad Dadi (National Institute of Technology Mizoram, India); Ashok Ray (North Eastern Regional Institute of Science and Technology, India); Sushanta Bordoloi (National Institute of Technology Mizoram, India)
Small-signal modeling of Gallium Nitride (GaN)-based High Electron Mobility Transistors (HEMTs) requires accurate self-heating and charge trapping representation. A small-signal model is proposed which simplifies yet accurately represents drain-lag behavior, improves the extraction flow and simulation precision of the physics-based compact model for the 250 nm device. To incorporate the thermal effects in the proposed small-signal model, the conventional ASM-HEMT model is referred and modified suitable. The model includes RSUB = 20 Ω and CSUB = 100 fF, along with a thermal network defined by RTH and CTH to capture transient thermal behavior. Pulsed I-V measurements and temperature-dependent characterization are used to validate the proposed model. Pulsed I-V measurements show peak IDS values of 230.45 mA at VDSQ = 8 V, 228.15 mA at 15 V, and 213.44 mA at 28 V as compared to the conventional ASM-HEMT reference model. The investigation is carried out in a calibrated Cadence Virtuoso simulation environment.
9:00 Simulation and Optimization of Top-Gated Graphene-Based Field Effect Transistor for Enhanced Biosensing Applications
Paramasudhen Ravendran (Asia Pacific University of Technology & Innovation (APU), Malaysia); Hemah Selvarajan (Universiti Putra Malaysia (UPM), Malaysia); P. Susthitha Menon (Universiti Kebangsaan Malaysia, Malaysia & Institute of Microengineering and Nanoelectronics (IMEN), Malaysia); Reena Sri Selvarajan (Asia Pacific University of Technology & Innovation (APU), Malaysia)
As nanotechnology and miniaturization lead the bioengineering domain, the quest to optimize the devices for biosensing applications remained as a gap to be fulfilled. Particularly, graphene-based field-effect transistors offer various advantages, and this has enticed researchers from all over the world to produce impactful research works for the past two decades. This work reports on the top-gated graphene-based field effect transistor (TG-GFET) device simulation and optimization that primarily focus on the key criteria such as electrode thickness, channel lengths, and electrode material to further enhance its practical implementation in biosensing applications. Optimized parameters obtained from the simulation work were verified with the theoretical concept prior to reporting in this work.
9:15 PWM and Differential PWM Generation Circuits Based on Conventional Schmitt Trigger Cooperating with Integrator: Analysis and Implementation
Sukkharak Saechia (Phranakhon Si Ayutthaya Rajabhat University, Thailand); Chanapat Kaew-in (School of Engineering, King Mongkut's Institute of Technology Ladkrabang, Thailand); Paramote Wardkein (King Mongkut's Institute of Technology Ladkrabang, Thailand); Nisukan Chatchaiphat (School of Engineering, King Mongkut's Institute of Technology Ladkrabang, Thailand)
This paper presents systematic approaches to generate the Pulse Width Modulation (PWM) signal by leveraging a simple configuration based on an operational amplifier (op-amp) comprising the Schmitt Trigger circuit and the integrator circuit. Two PWM modulation techniques are proposed according to the input position of the information signal. First, utilizing inverting terminal of the Schmitt Trigger circuit as an input terminal can function as the PWM modulator. The resulting PWM output exhibits its duty cycle proportional to the derivative of the input signal. Second, when input signal is applied to the non-inverting terminal of the integrator circuit, the generated PWM signal whose duty cycle is proportional to both input and its integral. To ensure that the generated PWM signal conveys the message correctly, PWM demodulator is used to validated and confirm the results. Therefore, both proposed configurations effectively function as PWM modulator for both information signal and its integral. Finally, the results of in-depth mathematical analysis are validated and align well with all experimental results.
Communication Systems (CS)
Track A1F5 CS 1: Communication Systems (CS) 1
Room: F5. 504 Madai (Level 5)
Chair: Chee Yen Leow (Universiti Teknologi Malaysia, Malaysia)
8:00 Adaptive Deep Reinforcement Learning for Coordinated Autonomous Vehicle Platooning on Edge Networks
Nischal Maheshwari, Vaishvi Shah, Nitya Balar, Dhruv Luniya, Vasukumar Parsaniya, Bhabani Prasad Mishra and Nitin Singh Rajput (Pandit Deendayal Energy University, Gandhinagar, India); Santosh Kumar Satapathy (Pandit Deendayal Energy University, India & Assistant Professor, India)
Vehicle platooning is a promising application for achieving greener, faster, and more economical transportation. In a platoon, vehicles operate in a coordinated convoy, where the lead vehicle is responsible for navigation and driving instructions, which are followed by the trailing vehicles. However, the operation of such a system requires not only intra-platoon communication but also reliable connectivity with infrastructure and cloud services, which are often located at distant locations. Ensuring seamless and uninterrupted communication in this setting poses a significant challenge. To address this, we propose an adaptive multi-agent deep reinforcement learning framework for enhancing communication reliability in vehicle platoons, particularly within edge-assisted networks. In the proposed model, each vehicle functions as an intelligent agent that learns an optimal policy for selecting communication channels that ensure reliable data transmission, both within the platoon and between the platoon and infrastructure nodes. The framework is built on deep Q-learning with experience replay, enabling agents to learn robust strategies over time. The simulation results demonstrate that the proposed model effectively learns optimal policies and maintains high performance even under dynamic and challenging network conditions.
8:15 Exploration of Spatial Diversity in Wi-Fi CSI Crowd Counting for LOS and NLOS Environments
Juliane G Cudal, Charles S Dela Cruz, Nicolas John P Ellovido, Justin Andrei J Fajarda, Kyle Dominic R Reantoquio and Josyl Mariela Rocamora Reyes (University of Santo Tomas, Philippines)
Conventional crowd counting (CC) systems employ cameras, LIDAR, or environmental sensors to detect the number of people in a target environment. However, these systems require line-of-sight (LOS) to recognize and estimate the crowd effectively. An alternative sensing method uses Wi-Fi channel state information (CSI) derived from the received signals. Despite Wi-Fi's capability to work in non-line-of-sight (NLOS) conditions, recent Wi-Fi CSI CC systems focus mostly on LOS scenarios, with some systems utilizing multiple pairs of TX-RX devices to improve the classification performance. Therefore, this study investigates LOS and NLOS scenarios for Wi-Fi CSI CC systems utilizing single- or multi-pair devices. We gathered diverse datasets and applied two preprocessing techniques: discrete wavelet transform (DWT) for noise filtering and principal component analysis (PCA) for dimension reduction. These datasets were utilized in training machine learning models to recognize three classes: 0, 1, and 3 persons. Results show that adding another pair of TX-RX devices boosts the CC classification performance to reach 100% accuracy even in NLOS or obstructed conditions.
8:30 Sectorized Cooperative Relay Transmission Utilizing Surplus Energy for WPSNs
Shoma Taniguchi, Kosuke Sanada, Hiroyuki Hatano and Kazuo Mori (Mie University, Japan)
Wireless powered sensor networks (WPSNs) suffer from the double near-far problem, in which supplied and consumed energy depends on the location of sensor nodes (SNs). To mitigate the double near-far problem, cooperative relay transmission has been proposed, in which the packets from the SNs far from a power transmitter are relayed by the close SNs utilizing their surplus energy. In this scheme, however, all the SNs closer to the power transmitter are candidates for relaying, causing energy waste problem at candidate SNs. Moreover, the relay SNs selection problem in WPSNs is commonly solved by iterative algorithms with high computational complexity, which is impractical for practical networks. To address this problem and improve the data collection performance, this paper proposes surplus energy aided cooperative relay transmission employing sectoring technique for WPSNs. The proposed scheme divides the network area into sectors based on the number of SNs accommodated in each sector, and information transmissions are carried out sector by sector. This reduces energy waste in the relay candidate SN and does not require high computational complexity. Thus, this can also improve packet collection rates. The computer simulation results show that the proposed scheme provides better performance over the conventional one.
8:45 LTE Bands for Human Presence Detection Using Software Defined Radio
Allen Gabriel T. Estorque, Miguel Leo S. Malibiran, Kurt Louis A. Mariano, Krizelle Anne Lou M Recinto, Alyza Joyce D. Sones, Josyl Mariela Rocamora Reyes and Jehiel D. Santos (University of Santo Tomas, Philippines)
Human presence detection has various applications, with most existing methods relying on fixed-frequency RF signals like Wi-Fi, limiting transmitter-receiver distance. This study investigates the feasibility of using LTE bands such as bands 28, 3, 1, and 41 for human presence detection in indoor environments. LTE signal data were collected during video call sessions using a Software-Defined Radio (SDR). Support Vector Machine (SVM) models were trained and evaluated on 0 vs. 1 person, 0 vs. 3 person, multiple classifications per band. Accuracy results on 0 vs. 1 person, 0 vs. 3 person comparisons were: Band 28 - 70% & 85%, Band 3 - 50% & 55%, Band 1 - 75% & 80%, and Band 41 - 60% & 85%. Multi-class SVM training yielded accuracies of 60% (Bands 28 ), 50% (Band 3 & 1), and 62% (Band 41). This highlights the impact of LTE frequency variation on human presence and crowd density detection.
9:00 Blind ISI Channel Estimation, Symbol Detection, and Message Recovery Using Clustering
Wakana Yagyu, Tomotaka Kimura and Jun Cheng (Doshisha University, Japan)
In wireless communication systems, the presence of inter-symbol interference (ISI), resulting from multipath propagation and channel memory, poses a significant challenge to reliable symbol detection and message recovery. A blind estimation of the ISI channel with a memory length-(L) is proposed. The iterative (K)-means clustering algorithm is used in the receiver for this blind channel estimation process. The algorithm produces (K^L) potential sets of channel taps as a consequence of channel phase uncertainties caused by phase rotation. Subsequently, these are reduced to (K) sets using the squared error comparison between the received and the reconstructed symbols. The transmitted message is subsequently obtained using these (K) candidate channel tap sets through Viterbi demodulation, channel decoding, and verification of cyclic redundancy check (CRC). The block error rate of the proposed approach is close to that of a perfectly known channel, yet it nearly aligns with the performance of the conventional MMSE estimator utilizing a few pilot symbols.
Computing & Computational Intelligence (CCI) 2
Track A1F6 CCI 1.2: Computing & Computational Intelligence (CCI) 1.2
Room: F6. 505 Sepilok (Level 5)
Chair: Farashazillah Yahya (Universiti Malaysia Sabah, Malaysia)
8:00 Weigh&Pay: Semi-Automated Self-Checkout and Pricing for Non-Barcoded Items Using AI Vision
Maxine Van Caparas, Fernan Frans Pelobello, Jan Kevin Albior Galicia and Maria Leonora Guico (Ateneo de Manila University, Philippines)
Long queues in supermarkets are one of major problems for consumers and retail establishments around the globe. Thus, self-checkouts using barcode scanning have been developed in pursuit of shorter checkout times. However, some customers find these systems to be tedious to use. Furthermore, there is a gap in implementation when it comes to non-barcoded items which require weighing and manual code entry. Thus, deviating from the barcode system and eliminating the need for manual barcode scans may simplify the checkout process by eliminating the need of packaging, weighing, and labeling in advance. This research aims to develop a semi-automated self-checkout system for shopping item classification and price calculation of non-barcode items such as fresh proteins and produce. This was done by developing a computer vision module using YOLOv8 trained on a custom dataset, a time-based digital scale module for real-time weight reading, a price calculation module, and a hardware configuration.
8:15 Improved Genetic Algorithm Based on Fractals for Packet Scheduling in MANETs
Mercy Sharon Devadas (Chennai Institute of Technology)
Mobile Ad-hoc Networks (MANETs) have become an integral component of modern wireless communication systems, especially in scenarios where infrastructure-based networks are either unavailable or impractical. These self-configuring, decentralized networks are used in various applications, such as disaster recovery, military operations, and mobile conferencing, where efficient data transmission is critical. Ensuring the effective operation of MANETs is therefore essential, particularly in terms of Quality of Service (QoS), which encompasses factors like delay, packet loss, and throughput. This study presents an innovative technique aimed at enhancing QoS in MANETs by combining the mathematical concept of fractals with the optimization power of genetic algorithms. Specifically, the research focuses on developing an improved packet scheduling algorithm-a core element in achieving better QoS. The proposed system classifies packets into service classes based on their priority levels and acceptable delay thresholds. Following classification, a modified genetic algorithm is applied to schedule packets efficiently. To assess the effectiveness of this approach, simulations were performed using the NS3 network simulator, and the results were thoroughly analyzed.
8:30 An Efficient Knowledge Graph Construction Using LLM-Based Retrieval Augmented Generation for Agricultural Data
Rohini Naik (COEP Technological University, India & VPKBIET Baramati, India); Sunil Mane (COEP Technological University, India)
The agricultural industry produces a large volume of unstructured data from various sources, including research papers, field reports, weather data, and satellite images. Extracting valuable insights from unstructured data is crucial for improving decision-making, boosting productivity, and driving innovation. Knowledge graphs (KGs) provide a structured method to represent and link information, yet constructing them from unstructured agricultural data poses a significant challenge due to diversity of data and complexity of the domain. In this study, we introduce a methodology for creating agricultural knowledge graph using the Large Language Model (LLM)-based Retrieval-Augmented Generation (RAG) technique. This approach combines 1) the generative capabilities of LLM with the automated extraction and organization of entities-relationships, and 2) the RAG framework, which enhances the precision of information retrieval pertinent to the agricultural field. We describe the design, implementation, and evaluation of our method, showcasing its effectiveness in generating accurate and comprehensive knowledge graph. Our experimental evaluation, demonstrated that proposed framework achieves average ROUGE performance of 29% higher with respect to RAG retrieval and 53% more compared to vector retrieval. Also, average BLEU metric achieves 20% and 68% higher respectively. This work establishes a foundation for scalable, intelligent knowledge system that can support various agricultural applications, such as advisory services, research, and policy development.
8:45 Unlocking Secure Clouds: a Modern Perspective on Access Control and Privacy
Prateek Sharma, Junaid Khan, Mohammed Rehan Deshnoor, Dhruti Dobariya, Darshit Verma and Abhishek Upadhyay (BITS Pilani, Hyderabad Campus, India); Jay Dave (BITS Pilani Hyderabad Campus, India)
As cloud computing continues to support diverse and critical applications, ensuring secure, efficient, and privacy-preserving data sharing across dynamic, multi-cloud environments remains a significant challenge. Traditional solutions often suffer from inflexible access policies, high computational overhead, and complex key management, making them unsuitable for scalable and real-time scenarios. In this paper, we propose a lightweight and practical Data Protection as a Service (DPaaS) framework that enables fine-grained access control, efficient key management, and secure deduplication. Our scheme integrates ciphertext-policy attribute-based encryption (CP-ABE) with policy attribute masking to protect user privacy and utilizes distributed key management to support dynamic role updates without re-encrypting the entire dataset. Additionally, the system incorporates client-side deduplication techniques to reduce storage overhead while maintaining data confidentiality. We implement a prototype and evaluate its performance and security in a real cloud environment. Experimental results demonstrate that our scheme achieves low computation time, supports secure access control, and minimizes resource consumption, making it suitable for deployment in resource-constrained and large-scale cloud-based systems.
9:00 TASNet: a Temporal Attention Network for Front-Wheel Axis Angle Prediction in Real-World Driving
Pritam Chakraborty and Anjan Bandyopadhyay (Kalinga Institute of Industrial Technology, India)
Autonomous navigation in unstructured environments such as off-road terrain remains a significant challenge due to dynamic obstacles and unpredictable surfaces. Most vision-based models depend on structured road features, limiting their applicability in such settings. We propose TASNet, a vision-only deep learning architecture that integrates a lightweight CNN backbone, bidirectional LSTM for temporal context, and spatial attention to prioritize obstacle-relevant features. The model is trained using CARLA-generated synthetic data with domain randomization (e.g., lighting, weather), improving generalization to real-world variability. TASNet reduces steering angle error by 27.4% compared to NVIDIA PilotNet and achieves a 92% collision-free success rate in simulation. Real-world tests on an all-terrain vehicle (ATV) dataset confirm its robustness across varied outdoor scenarios. Moreover, TASNet is optimized for deployment on resource-constrained hardware, delivering real-time inference at 45 FPS with 8.2W power on NVIDIA Jetson AGX. Ablation studies underscore the importance of temporal modeling and attention. TASNet offers a cost-effective alternative to LiDAR-based systems for field robotics and disaster response.
9:15 Real-Time Fabric Defect Detection System Using Yolov8
P Kaythry (Anna University, India & Sri Sivasubramaniya Nadar College of Engineering, India); Swetha A and Shree Varshini N (Sri Sivasubramaniya Nadar College of Engineering, India)
Fabric defect detection represents an essential and pivotal process within the textile industry, which is fundamental for guaranteeing both product quality and consistency across various manufacturing outputs. Nevertheless, the traditional methods of manual inspection that have been employed historically are not only labor-intensive and time-consuming but also inherently subjective, thus making them highly susceptible to human error and misjudgment. In response to these significant challenges, this study presents a machine learning approach that utilizes the You Look Only Once model (YOLO) to effectively tackle and ameliorate these pressing issues. The intricate nature and vast diversity of fabric defects, which can arise from a multitude of factors including machine malfunctions, inferior quality yarns, spoilage of materials, and excessive stretching during the manufacturing process, greatly exacerbate the challenges associated with accurate detection. To overcome these formidable obstacles, the proposed YOLO model serves as a more efficient and precise alternative to the traditional inspection.
Engineering Technologies & Society (ETS)
Track A1F7 ETS 1: Engineering Technologies & Society (ETS) 1
Room: F7. 506 Selingan (Level 5)
Chair: Dalila Mat Said (Centre of Electrical Energy Systems (CEES), Universiti Teknologi Malaysia, Malaysia)
8:00 Indoor Comfort Analysis of Integrated Radiant Ceiling Panels with Fan Coil Units in a Tropical Climate
Mohammad Hadi Ghasemi (Universiti Kuala Lumpur Malaysian France Institute, Malaysia); Kushsairy Kadir (Universiti Kuala Lumpur British Malaysian Institute, Malaysia); Mohammad Mehdi Salehi Dezfouli (Norwegian University of Science and Technology, Norway); Muhamad Fahezal Ismail (UniKL MFI, Malaysia); Alireza Dehghani-Sanij (University of Waterloo, Canada); Mohammad Miqdad Abdul Aziz (Universiti Kuala Lumpur British Malaysian Institute, Malaysia)
This study investigates the integration of Radiant Cooling Panels (RCPs) with Fan Coil Units (FCUs) in the Sustainable Energy Living Lab (SELL) at the University of Kuala Lumpur British Malaysian Institute (UniKL BMI) in a tropical climate. The thermal performance of the building is evaluated using real-world data and detailed simulations in TRNSYS, following ASHRAE standard 55 for thermal comfort. The results show that the combination of RCPs and FCUs significantly improves Thermal Comfort (TC), with more than a 50% reduction in the Predicted Mean Vote (PMV) and Predicted Percentage of Dissatisfied (PPD) indices compared to scenarios using only FCUs. Annual and monthly analyses reveal stabilized PMV values closer to neutral levels and substantial PPD reductions, indicating enhanced Indoor Environmental Quality (IEQ). These findings highlight the potential of Radiant Cooling Technology (RCT) to achieve both energy efficiency and occupant satisfaction, thereby supporting the advancement of sustainable building practices in tropical climates.
8:15 Evaluation of a Neural Network for Personal Identification from the Frequency Spectrum of Ballistocardiographic Signals
Hitoshi Ueno (Tokyo Information Dsign Professional University, Japan); Karin Takahashi (Tokyo Information Design Professional University, Japan)
Ballistocardiogram signals are vibration signals on the surface of the body caused by changes in blood pressure, and can be obtained by pressure sensors. Since all that is required is to detect pressure fluctuations, there is no need to place the sensor in close contact with the skin of the human body, and biosignals can be obtained even through clothing, making them widely used. Conventionally, the vital signs obtained by pressure sensors and used as biosignals are heart rate and respiratory rate, but ballistocardiogram signals also contain other information about the living body. We have been focusing on the fact that the waveform of ballistocardiogram signals contains personal characteristics, and have been researching a personal identification algorithm that can identify individuals from ballistocardiogram signals. In this study, we analyzed ballistocardiogram data from 10 different individuals by inputting the frequency spectrum data of the ballistocardiogram signals into a three-layer fully connected neural network. As a result, we were able to achieve an F1 score of 0.7 or higher, which is an evaluation method of the classification system. In addition, in the analysis of nine individuals excluding subjects with a lot of body movement, we achieved a high classification accuracy of 0.87.
8:30 The Portable Fingerprint Attendance Using NB-IoT
Thanakrit Tawatthanadech, Thitirath Chaewsuwan and Thana Udomsripaiboon (University of Phayao, Thailand)
This paper presents a portable fingerprint scanner for recording the attendance of students using Internet of Things (IoT) technology through the Narrowband Internet of Things (NB-IoT) communication system, which is a low-power long-distance communication. This device is compact and energy-saving. It can be carried to various locations through the NB-IoT communication system to connect to the Internet. Moreover, the proposed fingerprint scanner system is able to monitor the attendance data of students by using the R307 fingerprint sensor module to receive information from each user's fingerprints and then send it to store and process on the Google service platform, which is the Google Sheet, via the Internet system based on cloud technology. The data on the cloud technology can be processed, report and summarize the attendance results. It also has the ability to store and forward all information to those involved in developing the quality of students for further proceedings.
8:45 Building Students' Industry Workplace Skills Using Artificial Intelligence
Chien Ching Lee (SIT, Singapore); Ryan Fraser Kirwan (Singapore Institute of Technology, Singapore)
Effective written communication is a crucial skill for students transitioning to industry, where accurate documentation of project activities and decisions are paramount. This paper evaluates the impact of a structured series of embedded workshop designed to enhance logbook writing skills among Applied Computing students to prepare them for their eight-month internships under an Integrated Work Study Programme (IWSP). The workshop was held over a three-week period, two hours per week during scheduled online classes. Generative AI (GenAI) performed different roles for the workshop: as a chatbot taking on the role of a group of clients in a Requirements Engineering (RE) exercise (lesson 1), as a coach to identify weaknesses in their writing (lesson 2) and as an editor to help them revise their RE documentation drafts (lesson 3). Students performed RE documentation through iterative logbook entries. Data was collected via a comparison of students' logbook entries in lesson three compared to lesson one and an end-of-workshop survey. The primary improvements observed between the students' logbook submissions were in critical analysis, content relevance and depth, and the integration of supporting ideas. The survey findings revealed that GenAI, when deployed as a user, helped generate ideas during brainstorming sessions and provided a basic and structured skeleton to kickstart the writing process. GenAI, as a coach, helped students be more objective by considering multiple perspectives. As an editor, it offered clarity and formatting suggestions, which helped students write more clearly and coherently. The findings underscore the effectiveness of AI-assisted workshops in preparing students for industry by developing essential written communication skills, thereby fostering better project documentation and reporting practices.
9:00 Deep Learning-Based Detection and Auditory Summation of the Philippine New Generation Series Peso Coins for Visually Impaired Individuals
Robert G. de Luna (Polytechnic University of the Philippines, Philippines & University of Sto. Tomas, Philippines); Jana Elyssa O. Bandong (Polytechnic University of the Philippines - Sto. Tomas Campus, Philippines); Selwin Adam R. Garcia (Polytechnic University of the Philippines - Santo Tomas Campus, Philippines); Trishia L Mirabel, Maria Carmela L. Panaligan and Jullianne Christille Narito Sunga (Polytechnic University of the Philippines - Sto. Tomas Campus, Philippines)
This research is a deep learning-driven assistive device specifically engineered to support visually impaired individuals in identifying Philippine New Generation Series (NGS) peso coins. The primary goal of this study is to facilitate independent and accurate coin recognition through an accessible and user-friendly auditory feedback system. The system leverages computer vision and artificial intelligence to distinguish between various denominations based on their visual features. To train deep learning models, a custom dataset consisting of 1,419-coin images was manually collected, ensuring diversity in coin orientation, lighting conditions, and background environments to reflect real-world usage. Three deep learning architectures-Convolutional Neural Network (CNN), ResNet50, and VGG16-were selected for evaluation, given their proven efficacy in image classification tasks. These models were trained and validated on the prepared dataset, and their performance was compared in terms of accuracy. Among the three, the CNN model outperformed the others, attaining an accuracy of 99.89%, demonstrating its superior capability in recognizing coin types with high precision and minimal error. The finalized CNN model was then integrated into a functional hardware prototype which combines the trained model with a camera module and a microcontroller-based processing unit. Upon capturing an image of the coin, the system classifies its denomination and provides real-time auditory feedback, thereby enabling blind or visually impaired users to accurately identify coins without external assistance.
9:15 Integrating Generative AI into Creative Workflows: a TAM-Based Investigation of AI Adoption in Video Production and Editing
John Heland Jasper Ortega (FEU Institute of Technology, Philippines)
This study investigates the integration of generative artificial intelligence (AI) into video production and editing workflows using the Technology Acceptance Model (TAM) as the theoretical framework. Given the creative industry's substantial economic contribution globally and in the Philippines, the research focuses on how video professionals perceive the usefulness and ease of use of AI tools such as Adobe Premiere Pro, After Effects and OpenAI's Sora. A purposive sample of 60 professionals, including editors, motion graphics artists and content creators, was surveyed to assess behavioral intention, user attitude and the influence of external factors on AI adoption. Findings indicate a generally positive attitude toward AI, with many respondents highlighting increased efficiency and automation of repetitive tasks. However, the study also identifies key barriers such as skill gaps, tool complexity and limited training opportunities, emphasizing the need for competency-based training and institutional support. Correlation and word cloud analysis reinforce the importance of enhancing education quality and multimedia learning to support effective adoption. The results confirm TAM's applicability in this context and suggest that with improved accessibility, policy alignment and targeted capacity building, generative AI can significantly enhance creative workflows while supporting the development of employability skills in the digital content sector.
Computing & Computational Intelligence (CCI) 3
Track A1F8 CCI 1.3: Computing & Computational Intelligence (CCI) 1.3
Room: F8. 507 Monsopiad (Level 5)
Chair: Anup Nandy (National Institute of Technology Rourkela, India)
8:00 Disentangle Class Imbalance in Micro-Expression Recognition with Causal Structure Learning
Pei Sze Tan (Monash University, Malaysia Campus, Malaysia); Sailaja Rajanala and Raphael C.-W. Phan (Monash University, Malaysia)
Micro-expression recognition is a critical task in affective computing, yet it is often hindered by biases inherent in datasets, leading to skewed and unreliable outcomes. This work introduces a novel approach to disentangling bias in micro-expression recognition using causal structure learning. By modeling causal relationships within the data, we identify and mitigate sources of bias that traditional machine learning models often overlook. Our framework integrates causal discovery techniques to uncover biased patterns in widely used micro-expression datasets. We employ debiasing strategies to enhance the fairness and accuracy of recognition models and generate counterfactual examples to address sensitive attributes such as gender and age, allowing us to observe the effects of imbalanced classes on classification results. Experiments conducted on baseline micro-expression recognition models demonstrate comparable results after undersampling to create emotion class balance, revealing label bias in current training datasets including CASME2, SAMM, and SMIC. Further evaluation on balanced gender and age classes using generated counterfactual data as additional training instances showed performance improvements for the 4DME dataset.
8:15 Unsupervised Machine Learning of Historical Flood Data for Flood Events Clustering
Jackel Vui Lung Chew and Sean Woon (Universiti Malaysia Sabah, Malaysia)
Floods remain one of the most frequent and destructive natural disasters in Malaysia. To understand flood characteristics, which aid the development of early warning systems, we investigate the effectiveness of unsupervised machine learning techniques to identify clusters of flood events based on historical flood data provided by the Department of Irrigation and Drainage Sabah. The dataset used in this study consists of hydrological features such as average recurrence interval, flood duration, flood depth, and affected areas. Due to the nature of raw dataset such as missing data, varying scales, and outliers, we generated six different preprocessed datasets using an extensive data preprocessing which includes handling of missing values, feature normalization, and outlier removal. The main reason for doing this task was to ensure the clustering model robustness. We then applied five clustering algorithms, which are K-means, K-medoids, hierarchical clustering, density-based spatial clustering of applications with noise (DBSCAN) and its extended version, HDBSCAN, to derive natural clustering in the flood events. The clustering quality was carefully assessed using internal validation metrics such as silhouette score, Davies-Bouldin index, and two-dimensional visualization by principal component analysis. In our study, K-means, K-medoids, and hierarchical clustering outperformed DBSCAN and HDBSCAN when a predefined number of clusters was required. The best clustering quality was achieved using datasets with imputation and normalization. The major contribution of our study is to provide meaningful insights into the characteristics of flood events, offering potential applications in early warning systems.
8:30 DR-VQA: Large Language Model Based Vision Question Answering System on Diabetic Retinopathy
Saleh Musleh (Hamad Bin Khalifa University & Aspire Academy for Sports, Qatar); Hamada Al Absi and Md Rizwan Parvez (Hamad Bin Khalifa University, Qatar); Anant Pai and Ghassan Salih (Hamad Medical Corporation, Qatar); Tanvir Alam (Hamad Bin Khalifa University, Qatar)
Medical Visual Question Answering (VQA) presents challenges due to the complexity of imaging data and the need for precise, context-aware responses. Traditional VQA models often struggle in clinical settings, limiting their utility in decision-making. This study fine-tunes the BLIP (Bootstrapped Language-Image Pretraining) model for medical VQA, leveraging both visual and textual data to generate accurate diagnostic answers. To enhance semantic relevance, BERT-based similarity evaluation is integrated. Using a dia- betic retinopathy dataset, the model achieves a validation BERT similarity (BERTsim) score of 0.94 and a test score of 0.95 on 450 samples, demonstrating strong alignment with expert an- notations. These results highlight the model's potential to assist clinicians by improving diagnostic accuracy and efficiency. The proposed approach can streamline medical workflows, reduce clinician workload, and enhance patient outcomes. Future work will focus on expanding datasets and refining the model for broader medical applications. We believe our approach will support to enhance the patient care as well democratization on AI technology for community.
8:45 Machine Learning-Based Prediction of Fuel Consumption for Diesel and Gasoline Vehicles Considering Terrain Variations
Shaun Patrick B Calumba, Curt Johann M Maracha, John Darren E Casidlac and Cedrick Johnlery Sayson Judilla (National University, Philippines)
This study presents a machine learning-based approach to real-time fuel consumption monitoring for diesel and gasoline vehicles considering terrain variations. The study leverages machine learning algorithms to develop a predictive model that accounts for the impact of terrain variations on fuel consumption. The proposed model integrates terrain data acquisition and various vehicular parameters, including speed, RPM, throttle position, elevation, road grade, tire circumference, engine torque, piston area, and cylinder capacity. To enhance accuracy, the study utilizes Support Vector Machine, Random Forest Regression, and Gradient Boosting Regressor models. The results are validated using K-Fold cross validation and the Hold Out method, ensuring the robustness of the predictive model. Through this model, Comparative analyses including correlation matrix evaluation, scatter plots, Feature Importance, boxplots, and 3D visualizations demonstrate the relationships between key variables and fuel consumption patterns. The results highlight the impact of terrain and vehicle dynamics on fuel efficiency, offering valuable insights for optimizing fuel consumption strategies.
9:00 ChatGPT in the Classroom: A Qualitative Study of Educators' Pedagogical and Ethical Perspectives
Sareen Kaur Bhar, Nurul Sakinah Aziz, Munirah Munawar Ali and Aimi Hazwani Abdullah (Multimedia University, Malaysia)
This qualitative study investigates educators' perceptions of ChatGPT integration in higher education through 12 semi-structured interviews. The objective of this study is to examine Malaysian higher-education educators' perceptions of integrating ChatGPT into teaching, focusing on pedagogical benefits, ethical concerns, and institutional readiness. Findings indicate that educators view ChatGPT as a useful tool for enhancing teaching strategies, improving engagement, and supporting professional tasks. Reported benefits include idea generation, efficiency, and personalized learning. ChatGPT was also noted for helping overcome language barriers and resource limitations. However, challenges such as limited institutional support, ethical concerns, reduced critical thinking, and unequal access remain significant barriers. By centering educators' experiences often overlooked in AI-in-education research, this study offers valuable, practical, and context-specific insights for responsible and effective ChatGPT adoption in diverse contemporary higher education settings. The findings are based on 12 participants from a single Malaysian institution and may not be generalizable across all contexts.
9:15 Personalising AI-Generated Jokes: an Empirical Study
Ally Tze-Rou Teh (Australia); Toh Xiao Ying (Monash University, Australia); Chai Xuen Seow (Malaysia); Tazeek Bin Abdur Rakib and Lay-Ki Soon (Monash University Malaysia, Australia)
This paper presents our investigation of the potential of generative AI in crafting personalised humour, focusing on how user-specified keywords and demographic information can guide joke generation. We introduce a two-part system: (1) a personalised joke generator built using GPT-3 that tailors humour based on keywords and user demographics (age, gender, continent), and (2) an automated joke ranking pipeline powered by fine-tuned BERT models to filter and prioritise high-quality, humorous content, trained on Reddit-sourced joke data. To evaluate the effectiveness and reception of the generated jokes, we conducted a human-centered evaluation involving 32 participants with diverse personality profiles. Evaluators rated the jokes across four humour dimensions: Funniness, Offensiveness, Surprise, and Reality Representation. Our findings reveal notable variations in humour perception across personality types based on MBTI. For instance, Sentinels responded more positively to the generated jokes, while Analysts tended to rate them lower across most categories. Additionally, a strong correlation between perceived funniness and surprise highlights the role of unexpectedness in humour appreciation. These results underscore the importance of personality-aware humour generation and suggest that current AI models require further refinement to effectively cater to diverse audiences.