Computing and Computational Intelligence (CCI) 1
Track B4F1 CCI 4.1: Computing & Computational Intelligence (CCI) 4.1
Room: F1. Sipadan I (Level 4)
Chair: Rosli Nurfatihah Syalwiah (Universiti Malaysia Sabah, Malaysia)
8:00 Domain-Specific Health Text Generation Through Low-Rank Adaptation of a Transformer Architecture
Kamalesh Debnath and Shrinjoy Das (Assam University, Silchar, India); Mousum Handique (Assam University Silchar, India & Assam University, India); Arnab Paul and Lalzo S. Thangjom (Assam University, Silchar, India); Megha Arakeri (Manipal Institute of Technology Bengaluru, India)
The growing demand for accessible and reliable health information has motivated the adaptation of domain-specific large language models (LLMs). LLMs perform well on general natural language processing (NLP) tasks but require fine-tuning for healthcare applications. In this work, Mistral-7B, a 7.3B parameter Transformer model, is fine-tuned for health text generation and noncritical symptom understanding using three parameter efficient methods-Low-Rank Adaptation (LoRA), Quantized Low-Rank Adaptation (QLoRA), and Rank-Optimized Reliable Adaptation (RoRA). A synthetic dataset comprising medical question answering, symptom descriptions, and home remedies was curated from public sources. Experimental results demonstrate that RoRA achieved the highest BLEU-4 (0.52), ROUGE-L (0.65), and F1-score (0.84), outperforming baselines such as BERT, RoBERTa, and LLaMA- 7B while maintaining low GPU memory usage. This work supports the use of fine-tuned LLMs for safe and efficient health communication, especially in low-resource settings. It also demonstrates that lightweight adaptation using Parameter Efficient Fine-Tuning (PEFT) can deliver high-quality outputs while minimizing computational demands.
8:15 A Region-Specific Nutritional Model Using LSTM Encoder and Attention-Enhanced Decoder
Kamalesh Debnath and Shrinjoy Das (Assam University, Silchar, India); Mousum Handique (Assam University Silchar, India & Assam University, India); Arnab Paul and Lalzo S. Thangjom (Assam University, Silchar, India); Megha Arakeri (Manipal Institute of Technology Bengaluru, India)
Malnutrition is a persistent challenge in rural India, where generic diet recommendations rarely reflect local food habits or economic realities. Many rural families in India rely on regionally available foods, and generic nutrition models typically overlook these dietary patterns. This paper presents an attention-augmented LSTM encoder-decoder model designed to generate affordable, region-specific meal plans for rural communities in Assam, India. A curated dataset was built using regional recipes, local expert input, and automated web scraping of authentic Bengali and Indian sources. By explicitly modeling cultural and seasonal food patterns, our approach ensures meal suggestions are realistic and easy to adopt in daily village life. The lightweight architecture also enables practical deployment in low-resource healthcare settings without sacrificing performance. On a held-out test set, the proposed model achieved strong results with a BLEU-4 score of 0.42, ROUGE-L of 0.54, F1-score of 0.81, and a BERTScore F1 of 0.581, outperforming both transformer and standard LSTM baselines. These results indicate that compact, locally adapted neural models can offer practical nutrition guidance in underserved settings.
8:30 Breast Cancer Detection: a Comprehensive Review of Multimodal ML Datasets
Jayendra Kumar (Vellore Institute of Technology, India); Priyanka Singh (VIT-AP University, India & Victorian Institute of Technology (VIT), Australia); Samineni Peddakrishna (National Institute of Technology Silchar, India); Banee Bandana Das (SRM University Andhra Pradesh, India); Saswat Kumar Ram (SRM University, Amaravati, Andhra Pradesh, India)
Breast cancer continues to be one of the most prevalent and life-threatening diseases affecting women worldwide. Early and accurate diagnosis significantly improves survival rates, and in recent years, machine learning (ML) has emerged as a transformative tool in enhancing diagnostic precision. This paper presents a comprehensive state-of-the-art review of publicly available machine learning datasets specifically designed for breast cancer detection. The review categorizes and analyzes a wide range of datasets including tabular, imaging, and genomic types such as the Wisconsin Breast Cancer Dataset (WBCD/WDBC), BreakHis, Digital Database for Screening Mammography (DDSM), MIAS, CBIS-DDSM, and TCGA-BRCA. Each dataset is evaluated based on key parameters such as data type, feature richness, class balance, sample size, and its suitability for various ML tasks like classification, segmentation, and multi-modal learning. Additionally, the paper outlines the strengths, limitations, and real-world applicability of each dataset, providing critical insights for researchers in selecting appropriate benchmarks for model development. The study also highlights current challenges and suggests future directions for constructing more diverse, annotated, and standardized datasets to support robust and generalizable breast cancer detection systems.
8:45 Opinion Polarization on Social Networks Based on Political Discourse
Susmita Das, Sounak Sadhukhan and Arunabha Tarafdar (Bennett University, India)
Social media platforms have become the nerve center of campaign discussion in regards to global political landscape. Political candidates are wielding social media as an influential instrument to advance their own election campaigns. In this paper, we have perused the online exchanges and conversations among supporters of different political parties in the intense environment of US Presidential Election 2024. Sentiment scores, abusive speech and stance detection of the tweets have been considered for understanding the perspective of voters on social media platforms. Each post has been considered whether it is a hate speech which contributes in polarization in the various conversation obtained from mainstream social media platform X/Twitter. A novel model has been proposed that factors in sentiment, stance and and hate speech for opinion dynamics estimation. Our method has been evaluated on various publicly available datasets and satisfactory results have been obtained. It is observed that there is higher level of opinion polarization when hate speech is involved.
9:00 HyLapDFN: a Hybrid Approach for Infrared-Visible Image Fusion Using Laplacian Pyramid and Decoder Fusion Network
Nithin Eswarappa and Jeevan K. M. (Gandhi Institute of Technology and Management (GITAM), India); Shefali Waldekar (Nirma University, India); Bikram Kumar Vivek (Gandhi Institute of Technology and Management (GITAM), India); Koshy George (GITAM University, India)
Deep-learning (DL) methods are popular in image fusion because of their generalisation abilities and resulting performances. However, these large fusion models restrict their usage in resource-constrained applications such as remote surveillance using UAVs and handheld devices. In contrast, the Laplacian pyramid (LP) method for feature extraction has a much smaller footprint. Accordingly, in this paper, we propose a hybrid Laplacian pyramid and decoder Fusion Network (HyLapDFN) for fusing infrared and visible images. While LP extracts features, the decoder network constructs the fused image. The network is trained unsupervised as a ground truth fused image is unavailable. The performance of HyLapDFN is compared with five state-of-the-art DL-based image fusion methods using eleven image quality metrics. Qualitative analysis of the fused images from all the methods are performed on two sets of images. Additionally the correlation of colour fusion metric (CFM) is evaluated. Additionally, we scrutinise the computational complexities of these fusion methods.
9:15 Particle Swarm Optimization - Artificial Neural Network Model for Predicting Rebar Corrosion in Fiber-Reinforced Concrete
Bon Ryan Aniban and Kevin Lawrence De Jesus (FEU Institute of Technology, Philippines); Dante L Silva (Mapua University, Philippines); Jimmy G. Catanes (Commission on Higher Education, Philippines); Sheina Pallega (National University, Philippines & Mapúa University, Philippines); Donna Ville Gante (FEU Institute of Technology, Philippines); Meriam Leopoldo (Mapúa Malayan Colleges Mindanao, Philippines)
Chloride-induced corrosion (CIC) is a primary reason of deterioration in reinforced concrete (RC), particularly in marine structures which causes cracking, degradation, and decreased service life. Advances in the 4th Industrial Revolution have enabled utilization of machine learning techniques in different fields of civil engineering. This study develops an Artificial Neural Network (ANN) enhanced by Particle Swarm Optimization (PSO) to predict rebar corrosion in polypropylene fiber reinforced concrete (PFRC). Accelerated corrosion tests were performed using the impressed current method on samples with varying polypropylene fiber content, concrete cover (CC), and bar diameter (BD). Experimental results showed that the 3-7-1 network structure (NS) (3 input neurons (IN), 7 hidden neurons (HN), 1 output neuron (ON)) achieved the highest accuracy with correlation coefficient (R) of 0.98969, mean squared error (MSE) of 0.18846, and mean absolute percentage error (MAPE) of 7.832%. Employing the generated connection weights (CW) from the governing model (GM), through Olden's connection weights approach, observed that the concrete cover had the most significant influence on corrosion (-43.231%), followed by bar diameter (33.717%) and fiber content (-23.052%). It highlights that increasing concrete cover and fiber content significantly reduces corrosion in PFRC, which may be used by civil engineering professionals as it offers insights for enhancing the durability of reinforced concrete structures. This approach supports SDG 9 (Sustainable Development Goal 9: Industry, Innovation, and Infrastructure) by promoting resilient, innovative construction methods and contributes to SDG 11 (Sustainable Development Goal 11: Sustainable Cities and Communities) by enhancing the longevity and sustainability of urban infrastructure.
Control Systems & Robotics (CSR)
Track B4F2 CSR 4: Control Systems & Robotics (CSR) 4
Room: F2. 501 Kadamaian
Chair: Rozita Jailani (University Teknologi MARA, Malaysia)
8:00 Real-Time Control of a Lab-Based Flexible Rotary Servo System Using Linear Matrix Inequalities
Mousumi Mukherjee (Indian Institute of Engineering Science and Technology, Shibpur, West Bengal, India)
We consider the problem of designing a linear quadratic regulator (LQR) for a lab scale flexible link rotary servo system using linear matrix inequalities (LMIs). First, data from the open loop system is collected in real-time. In particular, data corresponding to the two measured outputs are recorded in response to an input. This input-output data is used for identifying a model of the system. The identified model is compared with the given model for correctness. Once verified, the identified model is used to synthesize a linear quadratic regulator by solving linear matrix inequalities. It is known that the weighting matrices play a crucial role in the LQR design. The effect of the choice of the weighting matrices in the LQR design is examined, and accordingly a suitable pair of weighting matrices are chosen for the design. Thus, a state-feedback controller, based on measured input-output data is obtained. The design is experimentally validated by implementing the controller in real-time.
8:15 Preliminary Theoretical Study on Remote Outdoor Coordinate Measurement via Sensor Fusion
HyungTae Kim and Ohung Kwon (KITECH, Korea (South)); Sangwon Lee (Korea Institute of Industrial Technology, Korea (South))
This study presents a theoretical framework for estimating outdoor coordinates via using close-range laser sensing. The theoretical framework for estimating outdoor coordinates was formulated, incorporating a GPS receiver, a laser distance sensor, and an IMU. The laser distance sensor targets a point on a specific structure and measures the distance from the framework to that point. Then, the targeted GPS coordinate was calculated using the geometric relationships among the measured distance, GPS coordinate, and the framework's tilt angles. The geometric relationships can be described using Euler-angle and geodetic models, and the inverse of the geodetic models was obtained using optimal methods such as the simplex algorithm, gradient descent, and equal search. The spherical and WGS84 models, the most popular in geodesy, were adopted to verify the computations using reliable GPS distance calculators. The proposed method can be applied to surveying construction sites, inspecting large buildings, measuring hazardous places and designing outdoor performances.
8:30 Design and Development of an Integrated Water Spray and Brush System for Solar Panel Cleaning
Maizatul Zolkapli (Universiti Teknologi MARA, Malaysia); Mustaqim Muhizam (UiTM, Malaysia); Azween Hadiera Hishamuddin (Universiti Teknologi MARA, Malaysia); Shahril Irwan Sulaiman (Universiti Teknologi MARA Shah Alam, Selangor, Malaysia); Ahmad Sabirin Zoolfakar, Rozina Abdul Rani and Marianah Masrie (Universiti Teknologi MARA, Malaysia)
Solar energy facilitates the global transition to renewables; however, dirt accumulation can reduce efficiency by 30 percent, and existing cleaning technologies remain ineffective, expensive, and water-intensive. This project aims to develop an innovative solar panel cleaning system that integrates a water spray mechanism and a mechanical brush, all controlled by an Arduino-based system, to overcome these restrictions. The objective was to create an environmentally sustainable, cost-effective, and energy-efficient solution that consumes minimal water while assuring adequate cleaning. The system could autonomously traverse the panel, directed by limit switches that regulate the directional trajectory. The device maneuvered the cleaning brush along the panel via two motors, while a third motor regulated the rotation. The system's performance was assessed by comprehensive testing, emphasizing cleaning efficiency, water usage, and adaptation to various environmental circumstances. The system improves solar panel energy efficiency by decreasing soiling, while conserving water and reducing maintenance expenses, hence providing dependable performance under diverse environmental conditions. This project aims to enhance the efficacy, sustainability, and scalability of solar panel cleaning, thereby significantly benefiting the solar energy sector and promoting the broader adoption of renewable energy technologies.
8:45 Dehazing Algorithm Selection System with Feature Extraction for Single Image
Keizo Miyahara (Kwansei Gakuin University, Japan)
This paper proposes a selection system of dehaze-algorithm to apply an appropriate one for an input single image by means of feature extraction. It is known that the haze is one of the most critical challenges for camera image processing. In order to cope with the negative effect, a number of dehazing algorithms have been researched. Previous studies on both dehazing algorithms and feature descriptors were examined with the aim to select the most suitable algorithm for the input image based on the explicit indices of the image itself. The examined dehaze algorithms can be categorized in two groups: Image restoration based on physical model (GMRF, Optical filtering, DCP, Bayesian net, Pre-record data set) and Image improvement with image parameters (Wavelet, Retinex, High frequency enhancement, Histogram equalizing). The implemented system was validated through a series of object detection experiments using real-world hazy images. The results of the experiments confirmed the improvement of the evaluation metric, mAP, that depicts the effectiveness of the proposed system.
9:00 Cooperative Mapping Method with Distributed Autonomous Mobile Robots in Unknown Environments
Yuma Ikeda and Keizo Miyahara (Kwansei Gakuin University, Japan)
This paper describes a method for collecting on-site information in unknown environments, such as the size and location of obstacles, using a decentralized mapping algorithm with multiple mobile robots. Aiming at initial response to emergency disaster situations, each one of the homogeneous autonomous distributed robots determine own exploration direction by combining data clustering techniques based on only its local map at the time. The mobile robots communicate locally, when they encounter each other in the exploring field, to share the map information to enhance the efficiency of the entire mapping system. Even if some unexplored areas remained, another shape-fitting technique will be applied to the incomplete information obtained, and it enables the system to output the most-likely environment map for the subsequent principal task, such as intended rescue mission. A series of simulations were conducted to verify the proposed cooperative mapping method, and the results of the examination demonstrated its consistency and efficiency.
9:15 Modelling and Control of a 4-DOF Moving Base Manipulator
Aksa Santhosh (APJ Abdul Kalam Technological University, India); Lal Priya P S (College of Engineering Trivandrum, India); Sneha Gajbhiye (IIT Palakkad, India)
Moving-base robotic manipulators are increasingly utilized in various industries due to their flexibility and adaptability in diverse dynamic environments. This work focuses on the modeling and control of a 4-DOF robotic manipulator composed of a planar two-link manipulator mounted on a mobile base capable of translating along the dual axes. The dynamic behavior of the manipulator is derived using the Euler-Lagrange equation, which effectively captures the dynamic coupling between the manipulator and the base. Additionally, kinematic analysis is conducted to accurately determine the end-effector's position during operation. One of the main challenges in this setup lies in dealing with the system's nonlinear characteristics, external disturbances, and the need to manage the dynamic interactions between the base and the manipulator components. To address these challenges, a Sliding Mode Control (SMC) strategy is adopted, chosen for its well-known resilience to system uncertainties and external perturbations. The proposed control framework is implemented within the MATLAB Simulink and Simscape environments, where the entire 4-DOF system is modeled to reflect realistic dynamics and base-manipulator interactions. Through simulation, the effectiveness of the controller in achieving robust and precise motion control is demonstrated.
Power, Energy & Electrical Systems (PES)
Track B4F3 PES 4: Power, Energy & Electrical Systems (PES) 4
Room: F3. 502 Mesilau (Level 5)
Chair: Nur Atharah Kamarzaman (Universiti Teknologi MARA, Malaysia)
8:00 Reinforcement Learning for Smart Grid Stability Using Adaptive Control and State Abstraction
Sourav Datto, Mustakim Ahmed, Md Eaoumoon Haque, Kazi Redwan, Sajedul Islam and Nasif Hannan (American International University-Bangladesh, Bangladesh); Mohammad Shah Paran (Lamar University, USA); Abu Shufian (American International University-Bangladesh, Bangladesh)
Smart grids are under pressure due to rising energy use, more renewable sources, and unpredictable consumption. When control systems fail to respond in time, they can cause frequency changes, power imbalances, and reduced grid reliability. This paper introduces a reinforcement learning (RL) framework using Q-learning to handle these challenges through real-time, adaptive control. The approach uses Principal Component Analysis (PCA) to reduce data complexity and discretizes continuous variables to make learning more efficient. A custom Markov Decision Process (MDP) models the grid environment, where the agent chooses actions: Increase, Decrease, or Hold based on the current state. A tabular Q-learning algorithm helps the agent learn the best decisions by maximizing rewards over time. Results show that the RL agent improves power stability by 22% over baseline methods and reacts accurately to supply and demand shifts, with action preferences distributed as Increase (58%), Hold (31%), and Decrease (11%). Heatmaps and 3D plots reveal clear action patterns and strong confidence in decisions, with more than 85% of states showing a decisive optimal action. The model adapts well to changes, proving useful for intelligent and stable grid control. This work supports smarter energy systems.
8:15 Load Frequency Control of Multisource Interconnected System Using a Nature-Inspired Optimization Algorithm
Vishal Rathore (Maulana Azad National Institute of Technology Bhopal, India); Dhananjay Kumar (Govt. Engineering College Siwan, India); Anchal Raghuwanshi (Maulana Azad National Institute of Technology (MANIT) Bhopal, India); Sushma Gupta (Maulana Azad National Institute of Technology Bhopal, India)
Load frequency control (LFC) principally entails the appropriate design of controllers. For improved LFC, the controller parameters should be properly tuned. Therefore, this article presents a Honey Badger optimization algorithm (HBA) to guide the controller, which is designed and proposed for LFC of an interconnected system with multiple fuel inputs (ISMFI). The frequency and power deviations of the tie-line act as controller inputs. Synchro-phasor technology is used to measure them. The measured signals are transmitted via communication channels. In this work, the signal transmission time delay is compensated with the Padé approximation method. The sum of integral-time-absolute-error (ITAE) of deviation is set to zero for objective parameter optimization. The ITAE for an ac tie-line is calculated as 0.0594, and settling times (STs) of frequency deviation (FDs) with tie-line power deviation (TLPD) are found to be 6.3894 s, 5.3577 s, and 8.0262 s. While for ac-dc parallel tie-line ITAE is 0.0711, and the minimum values of STs of FDs with TLPD are 5.2501 s, 16.6955 s, and 13.1596 s. The simulation results of the proposed algorithm are then compared with the recently reported algorithms; for each case, the time domain simulations are demonstrated. Furthermore, the performance of the proposed HBA-based controller is tested for random step load changes.
8:30 Accurate Fault Classification and Location Identification Using Various Machine Learning Models for Self-Healing in Smart Grids
P. Swati Patro (Birla Institute of Technology and Science, Pilani, Hyderbad Campus, India); N Praneeth and STP Srinivas (Birla Institute of Technology and Science Pilani, Hyderabad Campus, India)
Parametric monitoring using machine learning techniques offers a promising solution for accurately detecting and classifying faults in electric power grids, providing greater precision and efficiency than traditional methods while supporting the development of resilient, self-healing systems in smart grid environments. This paper evaluates the applicability and performance of decision tree, random forest, XGBoost, feedforward neural networks, and long short-term memory models for real-time fault detection, classification, and localization in power system protection using the IEEE 9-bus test system. The test system is designed in DIgSILENT PowerFactory, and data is obtained from electromagnetic transient (EMT) simulations of three-phase voltage and current signals under both prefault and post-fault conditions. Various short circuit fault types and locations are considered to enhance model generalization. The machine learning models are trained on these EMT time series signals, and their effectiveness is analyzed in terms of accuracy, robustness, and computational efficiency. The results offer valuable insights into the capabilities and limitations of each model in supporting intelligent fault management in modern power systems.
8:45 Real Time Analysis of Solar Based Electric Vehicle Charging Station
Rakeshwri Agrawal (Lakshmi Narain College of Technology, Bhopal, India & Maulana Azad National Institute of Technology, Bhopal, India); Vishal Rathore and Sushma Gupta (Maulana Azad National Institute of Technology Bhopal, India); Mukesh Kirar (MANIT, Bhopal, India)
The continuous depleting petroleum resources and hike in their prices have diverted the attention towards Electric Vehicles (EV). Moreover, EVs are ecofriendly leaving minimal carbon footprint into the environment, and with the technological updates and government incentives it is most widely adopted in the present era. The growing number of EVs has put in the additional burden on the utility grid, potentially causing voltage fluctuations. Hence independent hybrid charging stations are being developed which can be energized via renewable resources. This paper presents a topology for solar based hybrid charging station for EV. The performance of the charging station has been analysed under variable irradiations and its effect is studied on the battery system of the EV and results are validated through OPALRT OP-5600. A dual-switch bidirectional DC-DC converter is designed to maintain the DC-bus voltage of the charging station, the performance of which is also analyzed under variable solar irradiances.
9:00 Optimized Sizing and Impact of BESS and PV on Grid Frequency Under Varying Droop Coefficients with Partial Load Shedding Mechanism
Irfan Ahmed (Bangladesh University of Engineering and Technology, Bangladesh); Ahmed Zubair (Bangladesh University of Engineering and Technology, Bangladesh & Rice University, USA)
Along with the increasing adoption of renewable energy sources, maintaining grid continuity has become a growing concern. As solar photovoltaic (PV) systems and battery storage systems (BESS) proliferate, their contribution toward increasing system resilience and offering frequency response is becoming critical. This research focuses on frequency regulation and assesses how BESS increases stability by modifying the droop coefficient in primary frequency control. We present a simulation-based analysis for different droop settings and construct a BESS sizing algorithm to mitigate a 0.5 Hz frequency dip and advocate a partial load-shedding approach with BESS aid. In addition, the integration of PV systems contributed to a significant reduction in the required size of the BESS, as the PV systems were able to directly supply a portion of the energy demand. The findings demonstrate that adaptive droop improves frequency response, optimal BESS size enhances energy efficiency, and load-shedding with BESS support reduces the burden of excessive load shedding to increase reliability and balance in the system.
9:15 Enhanced Load Frequency Control in Interconnected Power Systems Using Shuffled Shepherd Optimization Algorithm (SSOA) for PIDn Tuning
Jonathan David G Quiling (Mindanao State University - Iligan Institute of Technology, Philippines & GNPower Kauswagan Ltd. Co., Philippines); Rovick Tarife (Mindanao State University- Iligan Institute of Technology, Philippines)
This study presents an improved Load Frequency Control (LFC) strategy for a two-area thermal power system by employing a PID controller with derivative filtering (PIDn) optimized using the Shuffled Shepherd Optimization Algorithm (SSOA). The SSOA algorithm leverages a combination of adaptive step-size adjustment and multi-community guidance to achieve a balance between global exploration and local exploitation in the search space. The main objective is to reduce frequency fluctuations and tie-line power deviations arising from sudden load variations, particularly under nonlinear conditions such as Generator Rate Constraints (GRC). The effectiveness of the proposed SSOA-PIDn controller is validated through simulation studies under two different operational scenarios. Comparative analysis reveals that the SSOA-based tuning achieves superior performance over classical and several well-known metaheuristic optimization methods in terms of Integral of Time-weighted Absolute Error (ITAE) and system settling time. Furthermore, sensitivity analysis under parameter uncertainties confirms the robustness and reliability of the proposed method for practical deployment in modern interconnected power systems.
Electronics, Circuits & Devices (ECD)
Track B4F4 ECD 4: Electronics, Circuits & Devices (ECD) 4
Room: F4. 503 Dinawan
Chair: P. Susthitha Menon (Universiti Kebangsaan Malaysia, Malaysia & Institute of Microengineering and Nanoelectronics (IMEN), Malaysia)
8:00 Energy Efficient Negative Capacitance L-Shaped Tunnel Field Effect Transistor
Alok Kumar Kamal (ABV-IIITM Gwalior, India); Neha Kamal (SIIC IIT Kanpur, India); Avinash Lahgere (Indian Institute of Technology Kanpur, India); Somesh Kumar (ABV IIITM Gwalior India, India)
In this paper, we have proposed a negative capacitance (NC) L shaped channel tunnel field-effect transistor (NC-LTFET), which consists of a hafnium oxide (HfO_2) based ferroelectric (FE) layer in the gate stack. The presence of polarization phenomena in the FE layer tends to enhance the internal voltage and electric field, which result higher ON-state current and steeper subthreshold swing (SS). From well calibrated 2-D TCAD simulation results, it is revealed that the NC-LTFET outperforms the conventional LTFET in terms of both static and dynamic energy dissipation. The NC-LTFET exhibits ∼ 430 × and ∼ 10^4 × higher ON-state current and I_ON /I_OFF ratio, respectively, as compared to the conventional LTFET. In addition, the proposed NC-LTFET shows ∼ 2.2 ×, ∼ 10^2 × and ∼ 10^2 × low SS, switching delay and energy delay product (EDP), respectively as compared to the conventional LTFET. As a result, NC-LTFET is 10 × higher energy efficient for both memory and logic designs switching at ultra-low supply voltage (< 0.2 V) when compared to the conventional MOSFET and LTFET.
8:15 Design, Optimization and Verification of the AMBA APB4 Protocol for Low-Power SoC Applications
Avneesh Singh and Alok Kumar Kamal (ABV-IIITM Gwalior, India); Somesh Kumar (ABV IIITM Gwalior India, India)
The Advanced Peripheral Bus (APB4) protocol, a component of the AMBA 4 specification, is designed to connect low-bandwidth peripherals to high-performance system-on-chip (SoC) designs. This paper presents the design, implementation, and verification of the APB4 protocol using Verilog HDL and SystemVerilog-based Universal Verification Methodology (UVM). We extend the standard APB4 protocol by incorporating power gating techniques and burst-mode support to enhance power efficiency and throughput. Simulation and synthesis were performed using Xilinx Vivado and Synopsys VCS, with verification data showing 100% functional coverage and zero protocol violations. Simulation results using ModelSim and Synopsys VCS confirm accurate protocol behavior, while synthesis using Xilinx Vivado and power analysis via XPower Analyzer demonstrate a 66% reduction in dynamic power and 7% area optimization compared to prior APB implementations. This work positions the APB4 protocol as a viable low-power solution for energy-constrained SoC designs, providing a reusable and modular verification infrastructure aligned with modern digital design flows.
8:30 Interactive Pattern Repetition Game Design Utilizing Real-Time Hardware Pseudo-Random Number Generator
Tee Hui Teo (Singapore University of Technology and Design, Singapore); Maoyang Xiang (8 Somapah Rd & Singapore University Technology and Desgin, Singapore); Kai Cao, Yiyang Fu, Si Yuan Lai, Qistina Binte Mohd Sahril, Samuel Lim and Wei Rui Ho (Singapore University of Technology and Design, Singapore)
This paper presents the design and implementation of a hardware-based memory game developed on an FPGA platform to promote cognitive stimulation among older adults. The game features a 3×3 LED and switch matrix that displays pseudo-random light patterns that users must recall and replicate. To enhance randomness, a seed is extracted from ambient electromagnetic noise using an Analog-to-Digital (ADC) and fed into a multi-Linear Feedback Shift Register (LFSR) pseudo-random number generator (PRNG). The PRNG employs XOR-combined LFSRs initialized with non-overlapping seed transformations to ensure statistical independence. A finite state machine (FSM) governs the game logic, translating 9-bit PRNG outputs into spatial LED patterns and evaluating user input for correctness. Statistical analyses of PRNG outputs, using normality tests and Central Limit Theorem-based smoothing, verify the Gaussian conformity and entropy quality of generated sequences. The fully integrated system includes custom PCBs for ADC input, a mechanical button interface, and an LED display, housed in a 3D-printed enclosure. Results from both simulation and hardware validation confirm the system's effectiveness in generating high-quality randomness and delivering a simple, tactile, and accessible cognitive exercise platform.
8:45 Incorporating FPGA-Driven Pseudo Number Generator into Python Tetris Game
Tee Hui Teo (Singapore University of Technology and Design, Singapore); Maoyang Xiang (8 Somapah Rd & Singapore University Technology and Desgin, Singapore); Yiyang Fu, Zhengyao He, Xavian B Muhammad Yunos, Wei En Phua, Sarah Cherian and Ahmad Naufal Bin Rozaini (Singapore University of Technology and Design, Singapore)
This project focused on the development of a Tetris-like game that incorporates a pseudo-random number generator (PRNG) implemented on a Field Programmable Gate Array (FPGA) to enhance the gameplay experience. By utilizing a hardware-generated seed instead of relying solely on software-based randomness, the project introduces a novel approach to determining the sequence in which Tetris blocks appear on the screen. This integration of hardware and software not only adds a layer of complexity to the game but also showcases the seamless real-time communication between a digital system powered by an FPGA and an interactive Python-based gaming application. Overall, this project represents a significant step forward in the realm of game development by showcasing the potential of integrating hardware and software technologies to enhance gameplay and create more engaging experiences for players. The fusion of FPGA-driven hardware components with a Python-based game exemplifies the innovative spirit driving advancements in the gaming industry, paving the way for future developments that blur the lines between virtual and physical gaming environments.
9:00 A Novel and Simple Measurement Technique for Two-Wire Resistive Sensors in Remote Applications
Elangovan K (IIITDM, Kurnool, India)
This work presents a simplified and accurate measurement technique for two-wire resistive sensors placed in remote locations, addressing the challenge of cable resistance. The proposed circuit utilizes bipolar current excitation and diode switching to generate a square-wave output voltage, from which the sensor resistance is extracted through voltage averaging. Simulation results confirm high precision with a nonlinearity of 0.0006 % and a relative error of 0.007 %, while cable resistance dependency shows a maximum error of only 0.004 %. Experimental validation using RTD-Pt100 characteristics over a temperature range of -100 °C to 850 °C demonstrates linear performance with nonlinearity and relative error within 0.44 % and 0.91 %, respectively. The circuit also shows limited sensitivity to cable resistance variations, maintaining output errors below 0.28 %. With its simple architecture, low component count, reduced power consumption, and cost efficiency, the proposed method offers an alternative for accurate remote measurement of resistive sensors.
9:15 A Power Efficient One-Shot Rectangular Pulse Generator for Temperature Sensor
Kuntal Chakraborty (National Institute of Technology Arunachal Pradesh, India); Subhajit Das (University of Engineering & Management, Kolkata, India); Abir Chatterjee (UEM, India); Abir J Mondal (National Institute of Technology Arunachal Pradesh, India)
This work presents a power efficient one-shot rectangular pulse circuit made up of a delayed step pulse generator (MDSPG) followed by a variable rectangle pulse width generator (VRPWG) with auto termination. The MDSPG generates a delayed feedback signal and the VRPWG generates variable rectangular pulse with four different pulse width options. At 90-nm CMOS technology, the design occupies 0.0058 mm2 and limits the power consumption between 5.78 μW to 5.81 μW at modes M0 and M3, respectively. Additionally, the proposed architecture ensures a large span of rectangular pulse width ranging from 155.66 μs to 1247.96 μs at M0 and M3, respectively. The proposed design is a suitable choice for a multi-mode resolution and multi conversion time-domain temperature sensor. A subthreshold time-to-digital converter is accommodated to realize the sensor in a 90-nm CMOS and the area is limited to 0.011 mm2. Lastly, the resolution varies between 0.29OC and 0.036OC at modes M0 and M3, respectively.
9:30 Smart Temperature Tracker for Thermo-Regulatory Disorders
Olanrewaju B Wojuola and Kutlwano Benedict Tigele (North-West University, South Africa)
Monitoring body temperature is crucial for detecting illness and evaluating the effectiveness of treatment. Particularly, temperature monitoring plays an important role for those with thermo-regulatory disorders, as accurate temperature monitoring is crucial for managing their conditions and for preventing serious complications. This paper proposes a temperature monitoring system that can be used for assisting individuals experiencing such a disorder. We also present test results for a prototype that we developed. The prototype monitors temperature in real time, records temperature data every five minutes and sends alerts to users and medical services during severe hyperthermia or hypothermia. The temperature monitoring system is designed to operate efficiently with low power consumption while providing real-time alerts for abnormal temperature conditions. Utilizing an Arduino board, a temperature sensor, and an OLED display, the system integrates advanced sleep modes and user interaction features. The designed prototype has demonstrated good accuracy and reliability in measuring body temperature when compared to both clinical digital thermometers and infrared gun thermometers.
Communication Systems (CS)
Track B4F5 CS 4: Communication Systems (CS) 4
Room: F5. 504 Madai (Level 5)
Chair: Mohamad Yusoff Alias (Multimedia University, Malaysia)
8:00 Logarithmic Hyperbolic Cosine Adaptive Filter with Variable Center Based Channel Estimation Under Non-Zero-Mean Non-Gaussian Noise
Mahima Chouksey (Madhav Institute of Technology and Science Gwalior, India); Sandesh Jain (ABV-IIITM Gwalior, India)
The design of robust adaptive filters is critical for applications such as system identification, channel estimation, noise suppression, and channel equalization, especially under non-Gaussian/impulsive noise environments. The existing maximum correntropy criterion, hyperbolic cost adaptive filter (HCAF), and logarithmic HCAF (LHCAF) based adaptive algorithms deliver suboptimal performance for non-Gaussian distortions having non-zero mean due to the incorporation of order statistics about the origin. In this paper, we propose a novel adaptive filtering algorithm called LHCAF with variable center (LHCAF-VC), which is robust against impulsive noise with non-zero mean, as it considers central order moments of the error distribution rather than origin-based moments. To further enhance convergence performance and adaptability in dynamic environments, a variable step-size (VSS) mechanism is also integrated into LHCAF-VC, resulting in a new VSS-LHCAF-VC algorithm. Simulation results for the channel estimation task under Gaussian-mixture noise conditions demonstrate that the proposed LHCAF-VC and VSS-LHCAF-VC algorithms achieve superior performance compared to existing state-of-the-art methods.
8:15 Learning to Identify RF Devices from Few Pilots via Reptile
Longqi Shen, Tomotaka Kimura and Jun Cheng (Doshisha University, Japan)
We examine a neural network-based system designed to identify $K$ devices, each with device-specific I/Q imbalances in the RF modulator, as they transmit signals to an access point using time-division multiple access. Traditional methods begin training with randomly chosen initial parameters and need numerous pilot symbols, which can be impractical in scenarios such as when a unmanned aerial vehicle (UAV) serves as the access point, collecting data from massive device sensor networks. In this work, we focus on adaptive learning with minimal pilot symbols to train the neural network to classify devices by their I/Q imbalances. Reptile is an efficient meta-learning algorithm designed to help models rapidly adjust to new tasks with minimal training data. By employing Reptile, we pre-train neural network model parameters offline, facilitating rapid online adaptation and fine-tuning for effective identification of new devices. Simulations indicate that Reptile surpasses conventional methods in device identification with fewer pilot symbols.
8:30 Machine Learning Models Accuracy Study Using P4 Programmable Data Plane in SDN - IOT Networks
Yashwanth A Doddegowda and Vidya Sagar Thalapala (Indian Institute of Information Technology Kottayam, India); Preeth Raguraman (IIIITDM Kancheepuram, India); Koppala Guravaiah (Indian Institute of Information Technology Kottayam, India)
Programmable data plane is the current trending area of research in networking. P4 is a programming language is using to handle different issues in data plane such as flow management, traffic management, and firewall management. Distributed Denial of Service (DDoS) attack is one of the crucial attacks in the network. Handling DDoS attacks at router level give more security to the network. This paper explores how the attack detection can be handled at router level in SDN-IoT Network. The proposed paper analyzes three different machine learning algorithms such as Decision Tree, Logistic Regression, and Tsetline Machine models on programmable data plane. Each machine learning model is implemented in a P4 programmable data plane environment using the BMv2 behavioral model and tested on a Mininet test-bed. The models uses the standard IoT data set of ICMP-based DDOS traffic features. The analysis explores the accuracy, memory efficiency, and feasibility for real-time deployment. The study proves that, novel adaptation of the Tsetlin Machine model into P4 logic, demonstrating its effectiveness and practicality for next-generation programmable networks.
8:45 QUIC-AID: Adaptive Intrusion Detection for QUIC Traffic Using Online Learning with ADWIN-Based Drift Detection
Carlo M Alamani, Mary Crizelle Jamielane T Figueroa, John Raphael Mundo and Jaybie A. de Guzman (University of the Philippines Diliman, Philippines)
QUIC is a transport protocol that is gaining adoption among content providers, offering reliability and speed by building an encryption layer similar to TCP's atop a UDP framework. Due to the vulnerabilities of both protocols, the demand for a QUIC-based Network Intrusion Detection System (IDS) continues to grow. QUIC-AID, an adaptive IDS framework designed for QUIC traffic, is proposed. Labeled network traffic flows were collected through a simulated environment for benign traffic and two types of QUIC attacks: flooding and fuzzing. QUIC-AID incorporates (1) dynamic feature selection-OFS and FIRES, (2) an online learning classifier-Adaptive Random Forest (ARF), and (3) drift detection-Adaptive Windowing (ADWIN). Prequential evaluation on the collected QUIC dataset (390,000 flows) showed that QUIC-AID outperformed other benchmarks (KNN and OC-SVM) in terms of true positive detection with a false positive rate below 0.1%. It also yielded the best results for ARF, achieving 99.94% recall. With dynamic feature selection, the rate of byte flow during attacks consistently had the highest weighting feature through FIRES, whereas OFS shifted weights based on the attack type.
9:00 Design of Optimal Reflection Coefficients and Low-Complexity Equalizer for IRS-OTFS System
Rakesh Kumar Yadav and Sai Kumar Dora (IIT ISM Dhanbad, India); Himanshu Bhusan Mishra (IIT (ISM) Dhanbad, India); Samrat Mukhopadhyay (IIT ISM Dhanbad, India)
For high Doppler scenarios, intelligent reflecting surface (IRS)-aided orthogonal time frequency space (OTFS) systems exhibit enhanced performances in terms of bit-error rate (BER), achievable rate (AR), signal-to-noise ratio (SNR), etc. These performances can be achieved by optimally designing the IRS coefficients through solving proper optimization problems. Note that in the existing literature, for IRS-OTFS systems, the optimal reflection coefficients were designed by minimizing the BER, which may not achieve a highly spectral-efficient system. Therefore, in this work, we design reflection coefficients by developing an AR optimization framework. We propose a root-mean squared propagation (RMS-prop) approach to solve this optimization problem. On the other hand, OTFS system comprises a high dimension delay-Doppler matrix, which can increase the computational complexity of linear equalization techniques. Thus, in this work, we also design low-complexity linear-minimum-mean-squared-error (LMMSE) equalizers, for OTFS system (with and without IRS), which relies on the principle of Cholesky-based decomposition. Our simulation results demonstrate the efficacy of the proposed AR optimization framework and low-complexity equalizers in terms of AR, BER and computational complexity, compared to the existing state-of-the-art techniques.
9:15 Possibilities of a Novel DIO Duplication Attack in Multi-Instance RPL Networks: an Empirical Study
Renya Nath N (National Institute of Technology Calicut, India); Jaisooraj J (Amrita Vishwa Vidyapeetham, Coimbatore); Hiran V Nath (National Institute of Technology Calicut, India)
The Internet of Things (IoT) is no longer a figment of imagination. With more ``things" acquiring IP connectivity, the scope of IoT applications is also widening. However, with greater opportunities presented by the IoT applications come equally critical security concerns. Low-power and lossy networks (LLNs), being a key IoT enabling technology, have to be immune to attackers for the IoT applications to work without disruptions. The IPv6 routing protocol for LLNs must ensure secure communication among IoT devices, as it is considered the de facto IoT routing protocol. The current literature concerning RPL lacks the exploration of attack possibilities in a multi-instance RPL context. This paper presents a novel DIO duplication attack at the LLN Border Router (LBR) level in an LLN with multiple RPL instances. The paper analyses the attack on three attack scenarios based on the number and position of malicious nodes. The simulation results provided in this paper justify the effects of the proposed DIO duplication attack on packet delivery ratio, power consumption of nodes, end-to-end delay, and packet overhead for all three attack scenarios.
Computing & Computational Intelligence (CCI) 2
Track B4F6 CCI 4.2: Computing & Computational Intelligence (CCI) 4.2
Room: F6. 505 Sepilok (Level 5)
Chair: Lorita Angeline (Taylor's University, Malaysia)
8:00 Design and Implementation of an AI-Driven Academic Path Forecasting System Using Sequential and Classification Models
John Heland Jasper Ortega and Abigail Lopez Alix (FEU Institute of Technology, Philippines)
An AI-driven academic path forecasting system is proposed to support data-informed advising and early academic intervention in higher education. In the Philippine context, where delayed graduation, student dropouts and lack of personalized academic guidance persist, machine learning in education offers a scalable and intelligent solution. The system combines three educational data mining techniques: a Long Short-Term Memory (LSTM) network for course sequence prediction, a decision tree classifier for student progress classification as regular or irregular and a K-Means clustering algorithm for grouping students based on academic trajectories. These models are developed in TensorFlow and deployed on a web platform built with CodeIgniter, enabling functionalities such as academic path forecasting, curriculum tracking and real-time risk alerts. Evaluation shows that the LSTM model achieves strong precision and recall in predicting next-term courses, while the decision tree classifier accurately detects off-track students with interpretable decision rules. K-Means clustering reveals meaningful groupings aligned with academic outcomes, further supporting early identification of at-risk learners. Confusion matrix analysis confirms high model accuracy across tasks. By integrating AI into higher education through course prediction, student classification and cluster-based insights, the system offers a practical framework for enhancing student success through targeted academic support.
8:15 Flood Induced Economic Damage Assessment from Satellite Imagery Using Vision Transformers
Md. Ashrif Rahman Arian, Md. Mehedi Hasan Shishir, Sadman Islam Chowdhury Samin and Shahnewaz Siddique (North South University, Bangladesh)
Every year floods cause substantial threats to lives, livelihoods, agriculture and infrastructure. Rapid assessment of economic damage caused by floods is necessary for disaster management, resource allocation and policy making. In this study, we propose a novel method for calculating flood induced economic damage using before and after flood satellite imagery. By leveraging Vision Transformer techniques, we perform semantic segmentation to identify land cover changes after the disaster. By measuring the area loss per class and assigning economic value to each class, we provide a method to estimate the monetary damage due to the disaster. We used Segformer B3 for segmentation which is a model of the Vision Transformer and achieved much higher pixel accuracy (0.98) and mIoU (0.53) compared to state of the art segmentation models UNet and DeeplabV3+. Moreover, Segformer B3 demonstrated considerably higher computational efficiency compared to the two other models experimented. Our approach offers an innovative and automated solution for post disaster flood damage assessment.
8:30 Edge-Optimized Machine Learning Model for Real-Time Prediction of Feed and Water Intake Using Multimodal Sensor Data
Lalith Reddy Tekulapalli, Anshika Verma, Aditya Ray Baruah, Varshith Srinivasa Peddada, Diptanshu Malviya, Likhita Paul Indupalli and Anakhi Hazarika (BITS Pilani Hyderabad Campus, India)
The growing integration of digital technologies in agriculture is reshaping livestock farming by enabling automation, data-driven decision-making, and real-time monitoring. Among critical tasks, the timely and accurate prediction of feed and water intake is essential for ensuring animal health, early disease detection, and sustainable resource management. Traditional manual methods are labor-intensive, error-prone, and lack responsiveness to dynamic conditions, while existing smart solutions often rely heavily on cloud infrastructure that introduces latency, increasing operational costs, and limiting usability in rural or resource-constrained areas. This paper presents a lightweight, edge-compatible machine learning (ML) model for real-time prediction of livestock feed and water intake using multimodal sensor data. The proposed approach emphasizes optimizing the model size to minimize dependence on cloud infrastructure to make it suitable for deployment in connectivity-limited farm environments. The use of systematic feature selection and correlation-based modeling enhances the interpretability and accuracy of the ML models, including Random Forest, CatBoost, XGBoost, and Neural Networks. Experimental results validate the performance of the proposed solution for feed and water intake prediction that enables timely and autonomous interventions and promotes operational efficiency, sustainability, and scalability in modern farming practices.
8:45 Modified Viterbi Algorithm for Religious Text: A Part-of-Speech Tagging for Waray-Waray
Jeneffer A Sabonsolin (FEU Institute of Technology, Philippines); Robert R Roxas (University of the Philippines-Cebu, Philippines); Ace Lagman (FEU Institute of Technology, Philippines)
Part-of-speech tagging (POS) is a vital process in natural language processing, enabling the identification of grammatical categories within sentences. This research emphasizes the lack of attention given to POS tagging for Asian languages, particularly Waray-waray. Limited studies on Waray-waray religious texts have hindered linguistic documentation and the deeper understanding of its grammar and vocabulary. To address this gap, the study introduces a POS tagging system for Waray-waray utilizing a Modified Viterbi Algorithm, which also incorporates a strategy for handling unfamiliar words. Evaluated on a corpus of 50,000 religious text datasets, the algorithm demonstrates outstanding performance-achieving an accuracy of 93%, precision of 90%, recall of 90.52%, and an F1 score of 92%. These results underscore the algorithm's effectiveness in navigating linguistic challenges across specialized genres. Beyond technical contributions, the study promotes linguistic diversity and fosters inclusive language technologies, advancing the goals of the Sustainable Development Goals (SDGs). Specifically, it enhances language learning and literacy among Waray-waray speakers, supports inclusive education through computational tools for minority languages, and aligns with SDG 4 by providing foundational resources for mother-tongue instruction and educational content development. Additionally, it offers new insights into Waray-waray's grammatical structures, laying a robust groundwork for future linguistic and computational research. Beyond technical contributions, the study promotes linguistic diversity and fosters inclusive language technologies, advancing the goals of the Sustainable Development Goals (SDGs). Specifically, it enhances language learning and literacy among Waray-waray speakers, supports inclusive education through computational tools for minority languages, and aligns with SDG 4 by providing foundational resources for mother-tongue instruction and educational content development. Additionally, it offers new insights into Waray-waray's grammatical structures, laying a robust groundwork for future linguistic and computational research.
9:00 Design and Optimization of Graph Neural Networks for EEG-Driven Anxiety Classification at the Edge
Mugdha Gupta, Eshaa Aranggan, Kavya Ganatra, Abinav Venkatagiri, Chinmayee P, Sameera M Salam and Anakhi Hazarika (BITS Pilani Hyderabad Campus, India)
The growing burden of anxiety disorders highlights the urgent need for scalable and non-invasive systems for mental health monitoring. Electroencephalography (EEG)-based Brain-Computer Interfaces (BCIs) offer a promising solution by capturing neural oscillations linked to anxiety. However, conventional detection methods oversimplify the brain's graph-structured connectivity and are computationally intensive, limiting their feasibility for real-time, edge-based deployment. To address these limitations, we propose two edge-optimized Graph Convolutional Network frameworks (GatedGCN and GAT) for anxiety classification using EEG signals. By modeling multi-channel EEG data as dynamic graphs, the system captures spatial-temporal brain dynamics critical for detecting anxiety-related patterns. The architecture incorporates adaptive graph construction, hierarchical spatio-temporal convolutions, and quantization-aware training to enable a reduced-size inference model with minimal accuracy loss. Our approach achieves real-time, resource-efficient performance on low-power edge devices that enables continuous, private, and accessible anxiety monitoring to pave the way for practical mental health interventions in wearable and mobile healthcare settings.
9:15 SOC Estimation in Electric Vehicles: a Comparative Evaluation of Kalman Filter and Coulomb Counting Methods
Hasanur Zaman Anonto (American International University-Bangladesh, Bangladesh); Md Ismail Hossain (American International University Bangladesh, Bangladesh); Azad Shahriyar, Tanay Banik and Abu Shufian (American International University-Bangladesh, Bangladesh); Md Mukter Hossain Emon and Md Akteruzzaman (Lamar University, USA)
The accurate estimation of the state of charge (SOC) in batteries is a critical component of battery management systems (BMS), especially in electric vehicles (EVs), where it directly impacts the efficiency, longevity, and safety of the system. This paper investigates two widely used SOC estimation techniques: the Coulomb counting method and the Extended Kalman filter (EKF) algorithm. The Coulomb counting (CC) method estimates SOC by integrating the battery current over time, making it simple and computationally efficient. However, it suffers from errors due to inaccuracies in the initial SOC estimate and does not account for battery self-discharge. In contrast, the Kalman filter algorithm is a dynamic estimation technique that uses a probabilistic model to estimate SOC, providing more accurate results even in noisy measurements and initial errors. This study compares both techniques by evaluating their performance in terms of accuracy, adaptability, and computational complexity in various battery usage scenarios. The Coulomb counting method, starting with an initial SOC estimate of 80%, shows a maximum estimation error of 15% over a six-hour charge-discharge cycle. The Kalman filter, initialized with a SOC of 80%, converges to the real SOC value of 50% within 10 minutes, achieving an estimation error of less than 5%. The results show that while Coulomb counting is effective in short-term applications with accurate initial estimates, the Kalman filter excels in long-term SOC estimation, offering superior performance in dynamic and noisy conditions. The paper concludes by discussing the practical applications of both methods and providing recommendations for choosing the appropriate technique based on specific system requirements.
Engineering Technologies & Society (ETS)
Track B4F7 ETS 4: Engineering Technologies & Society (ETS) 4
Room: F7. 506 Selingan (Level 5)
Chair: S.M. Anisuzzaman Status (Universiti Malaysia Sabah, Malaysia)
8:00 Capsule Neural Network for Inertial Sensor-Based Autism Spectrum Disorder Detection Through Multiple Gait Activities
Jayeeta Chakraborty (KIIT University, India); Anup Nandy (National Institute of Technology Rourkela, India)
The cerebellar deficit in children with Autism Spectrum Disorder (ASD) leads to motor deficits, resulting in gait pattern abnormalities. Wearable inertial measurement unit (IMU) sensors have emerged as an acceptable alternative to high-end motion sensors for cost-effective gait assessments using automatic feature extraction techniques. In this study, we develop a capsule network model to train on a dataset acquired from a small group of children with ASD and healthy children using wearable IMU sensors outside the lab environment. The gait data are collected from participants performing various gait activities- walking overground, ascending, and descending stairs. The capsule network model is enhanced with transfer learning to differentiate between autistic and healthy children's gait patterns. The trained models are evaluated using the precision-recall (PR) curve across varying thresholds to analyze the overfitting problem. Comparative analysis with the state-of-the-art automated feature learning methods shows that the proposed model outperforms other methods with an average F1-measure of 89.09% and an average area value under the PR curve (PR-AUC) of 0.919. The result analysis also indicates that considering different gait activities for the dataset improves the performance and applicability of the proposed models.
8:15 Transformative Pedagogy in Object-Oriented Programming: the COOP Model and AI-Enhanced Case-Based Learning
Wen Zhan Chee, Chin Ann Ong and Owen Newton Fernando (Nanyang Technological University, Singapore)
Advancements in generative artificial intelligence (AI) and large language models (LLMs) offer new opportunities to enhance object-oriented programming (OOP) education. Building on these developments, this study implements a case-based learning pedagogy within an AI-driven framework. Specifically, the ‘Case-Based Learning for Classroom Management Problem Solving (CBL-CMPS)' model proposed by Choi and Lee is adapted to the OOP context, resulting in the development of the ‘Cooperative Object-Oriented Programming (COOP)' model. A chatbot based on the COOP pedagogy was developed and deployed for testing with computer science students. This chatbot incorporated modern features of generative AI, such as prompt engineering and multi-agentic workflow, to enhance the effectiveness of delivering the pedagogy. Findings indicate that the COOP model effectively enhances student motivation and understanding of OOP. Furthermore, a phase-wise evaluation of the COOP model suggests that, although generally effective, targeted refinements could further improve its impact. This study concludes by discussing the potential, challenges, and implications of integrating LLM-assisted case-based learning into programming education.
8:30 Cardiac Care IoT and ML: Portable Home-Based Cardiovascular Monitoring for Early Risk Assessment
Fariah Mahzabeen, Khondoker Ahmed Zubaier, Fatiha Tultul and Intesar Hassan Bhuiyan (North South University, Bangladesh); Riasat Khan (North South University, Bangladesh & New Mexico State University, USA)
Sudden cardiac arrest is a significant global health concern; however, triage of cardiac events typically occurs only in clinical settings, often when the patient is already experiencing a cardiac episode. Effective and efficient at-home cardiovascular virtual health monitoring is essential for early intervention, potentially preventing sudden fatalities. A highly accurate machine learning model combined with a Veroboard integrated cardiac care kit has been developed in this work, which addresses the high mortality rate of sudden cardiac arrest through home-based monitoring. This device integrates three critical parameters ECG, blood pressure, and heart rate-into a machine-learning model to analyze real-time cardiovascular health data. A combined dataset of 1,690 cardiac patients from the UC Irvine Machine Learning Repository is used for model training and testing, encompassing 11 health features, including critical cardiovascular indicators such as fasting blood sugar, ECG results, exercise-induced angina, and ST slope. While the dataset includes individuals across different age groups, it primarily focuses on individuals aged 40 to 60. Min-max scaling for continuous features and one-hot encoding for categorical features have been applied in the dataset preprocessing stage. A Stacking classifier is implemented, using Decision Tree, Random Forest, and Gradient Boost classifiers as base estimators, with KNN as the final meta estimator. The applied Stacking ensemble model achieves an accuracy of 95.6% and an F1 score of 95.9%. This proposed device ensures a user-friendly interface and high accuracy, making it suitable as a household monitoring tool to reduce fatalities from unanticipated cardiac events.
8:45 Enhancing Cognitive Engagement and Inclusion for the Blind and Visually Impaired Through Sensor-Based Board Game
Shahrinaz Ismail (Goolee Sdn Bhd, Malaysia & IEEE Consultants Network Affinity Group Malaysia, Malaysia); Hayatul Nabihah Khairul Anwar and Kamazulzzaman Ahmad (Goolee Sdn Bhd, Malaysia)
Humans take for granted of board games without realizing that they could nurture higher order thinking skills (HOTs). Not being able to see, or being blind and visually impaired, would unintentionally discriminate the opportunity to cultivate HOTs through board games. The question in mind is how to give the same opportunity of developing HOTs to the blind and visually impaired through a game only the sighted can play. With the mission to make the board game "visible" to the blind people, B-Goolee is developed. This study went through multiple cycles of designing, developing, integrating, and testing all elements necessary to make B-Goolee an assistive educational game technology. The context consists of the board with sensors, the 3D printed game pieces, the mobile device with installed AppsGoolee connected to the board with Bluetooth technology, and the manual book printed in Braille. With this ecosystem, the blind and visually impaired individuals could enjoy the game while challenging their cognitive skills, and the tests results have proven it. The actual performance metrics from the user evaluation results quantifies the major findings from this board game usage and usability to the users. This contributes to the assistive technology that leads to the educational outcomes for the blind and visually impaired individuals.
9:00 Smart Campus Evolution via IoT: Technical Constraints, Stakeholder Analysis, and Future Opportunities
Bishwajit Banik Pathik (Military Institute of Science and Technology, Bangladesh & American International University-Bangladesh, Bangladesh); Abir Tirtha Das and Md. Kothan Miad (American International University-Bangladesh, Bangladesh)
This article aims to address the technical challenges and potentials associated with implementing smart campuses in conventional university frameworks. Integration of cost effective IoT sensors, improved wireless communication, different machine learning algorithms for data analysis, sustainable energy infrastructure are valuable tools for facilitating smart campus architecture. To gain insight into the feasibility and future prospects of smart campus adoption, a comprehensive survey was conducted among various campus stakeholders including students, faculty, and administrative staff. A total of 138 responses were received from the survey which had eighteen questions inquiring various aspects of any smart campus. The survey responses not only highlighted diverse opinions and priorities on the implementation of advanced technologies in smart university campuses but also indicated future aspects of the same in terms of educational and administrative growth. This study provides valuable recommendations for universities seeking to transition toward smart campus architectures while addressing the technical challenges inherent in such initiatives.
9:15 Decentralized Cross-Border Financial Aid Distribution Utilizing Blockchain-Based Tokenization
Md. Raisul Hasan Shahrukh (University of Malaya, Malaysia); Nafees Mansoor (University of Liberal Arts Bangladesh, Bangladesh)
Cross-border financial aid distribution encounters substantial obstacles, such as inefficiencies, elevated transaction costs, complications in regulatory compliance, and transparency issues inherent in traditional banking systems. This study presents a decentralized system for financial aid distribution that utilizes blockchain tokenization and smart contract automation to tackle these difficulties. The proposed approach specifically incorporates Hyperledger Besu, a permissioned blockchain, to enable traceable, transparent and efficient international financial transactions. The system manages digital tokens within a blockchain-based system that utilizes its smart contract, facilitating rapid and verifiable aid transfers while ensuring compliance with international regulatory guidelines. Performance evaluations demonstrate improved transaction throughput up to 4.87% higher than Hyperledger Fabric alongside reduced latency and enhanced operational efficiency compared to existing blockchain solutions. Furthermore, the modular consensus mechanism in Hyperledger Besu ensures Byzantine fault tolerance and maintains sub-second transaction finality under simulated loads of over 1,000 transactions per second. Moreover, automating compliance using smart contracts minimizes human mistake and potential fraud, thus considerably improving the traceability and dependability of cross-border financial aid disbursements.
Computing & Computational Intelligence (CCI) 3
Track B4F8 CCI 4.3: Computing & Computational Intelligence (CCI) 4.3
Room: F8. 507 Monsopiad (Level 5)
Chair: Wan Sieng Yeo (Universiti Malaysia Sabah, Malaysia)
8:00 Parallel Lightweight Hybrid Attention BiGRU Framework for Multi Resident Human Activity Recognition with Sparse Sensor Data
Abisek Dahal (National Institute of Technology Meghalaya, India); Kaushik Ray (North Eastern Regional Institute of Science and Technology, India); Soumen Moulik (National Institute of Technology, Meghalaya, India)
Recognizing human activity in multi-resident smart homes is complex due to sparse sensor activations and overlapping occupant actions. This paper presents a Parallel Lightweight Hybrid Attention BiGRU framework, integrating Multi-Head Attention (MHA) with Bidirectional GRUs (BiGRUs) to address these challenges effectively. The model processes global sensor relationships and local temporal dependencies in parallel, overcoming information loss typical in sequential processing. Experiments on real world smart home ARAS datasets show that the framework achieves 97.99% accuracy in Multi resident settings in House A and 99.49% accuracy in multi resident scenarios in House B. It also generalizes well across different households, demonstrating strong adaptability and robustness. By combining attention mechanisms with recurrent neural networks, the proposed architecture efficiently captures key patterns in sparse and concurrent sensor data. This work marks a significant advancement in multi resident human activity recognition, offering a scalable and reliable solution for real world smart-home applications and establishing a strong foundation for future research in pervasive computing and ambient intelligence systems.
8:15 Action-Aligned Video Pairing for Video Augmentation
Randy C Wihandika (Kumamoto University, Japan & Brawijaya University, Indonesia); Israel Mendonca and Masayoshi Aritsugi (Kumamoto University, Japan)
Video augmentation is an effective strategy for improving the performance of action recognition models. A recent video augmentation strategy addresses scene bias by mixing human regions from one video with the background from another. However, this often produces artifacts due to limitations in the video mixing process, which degrade training quality. This study proposes a video augmentation strategy that produces compatible action-scene video pairs rather than choosing them randomly, to improve the quality of mixed videos. To achieve this, two compatibility metrics are introduced to guide this selection to significantly reduce the occurrence of visual artifacts and generate higher-quality augmented videos. Our method improves alignment between actions which leads to more effective augmentation. The performances are further enhanced by applying a temporal morphological operation to improve object detection consistency. Experimental results on the UCF101, HMDB51, and Kinetics-100 datasets show that our approach improves classification performance. Code is available at https://github.com/rendicahya/video-action-alignment.
8:30 A Context-Aware PDF Query Chatbot
Ravi Kishore Kodali (National Institute of Technology, Warangal, India); Sai Veerendra prasad Kuruguti (National Institute of Technology Warangal, India); Varsha Sanga (CloudAngles, India); Lakshmi Boppana (National Institute of Technology Warangal, India)
Modern Retrieval Augmented Generation often lacks an inherent understanding of document-specific relationships and structured knowledge. By combining large language models and graph-based retrieval, the PDF Query Chatbot presented in this research fills this gap and provides more precise and contextually aware responses. In order to explicitly record entity relationships and structural dependencies, the system uses Neo4j to create a knowledge graph after extracting textual content from the uploaded documents. To facilitate a semantic similarity search, the text is simultaneously shredded and embedded in a vector store. A hybrid retrieval system that combines vector-based search for contextual relevance and graph traversal for relational comprehension is activated when a user submits a query. To produce grounded, document-specific responses, the results of the two retrieval pipelines were combined and sent to the LLM. By synergizing graph databases, semantic search, and LLMs, this architecture provides a context-aware solution for intelligent document interaction, addressing key limitations in traditional LLM-based question resolution systems.
8:45 Causal LIME: Enhancing Local Explanations with Causal Perturbations for Military Sensor Data
Trupthi Rao (IIIT Allahabad, India); Navjot Singh (Indian Institute of Information Technology Allahabad, India); Sonali Agarwal (Indian Institute of Information Technology, Allahabad, India)
Interpretability is vital in safety-critical domains such as defense, where understanding model behavior is crucial for building trust, ensuring accountability, and supporting decision-making. Traditional local explanation techniques, such as Local Interpretable Model-Agnostic Explanations (LIME), often neglect the causal relationships among input features. This oversight can result in misleading or spurious interpretations, particularly in complex, high-stakes environments. To address this limitation, we propose Causal LIME, an extension of LIME that incorporates causal graphs to guide the generation of perturbations in a manner consistent with the underlying data-generating mechanisms. This ensures that explanations respect the causal structure of the domain, leading to more trustworthy insights. We evaluate Causal LIME along three dimensions: (i) comparative analysis with traditional LIME, (ii) validation against permutation-based feature importance, and (iii) application to real-world military sensor data for vehicle classification. Experimental results demonstrate that Causal LIME produces more stable, causally grounded explanations, reinforcing its value in mission-critical AI applications where interpretability, reliability, and trust are paramount.
9:00 Improving Dynamic Time Warping in Gesture Recognition for Autonomous Vehicles
Yu Wu, Pin-Yu Lin, Yu-Chiu Lin and Min-Te Sun (National Central University, Taiwan)
Gesture recognition is a key component in developing intuitive and efficient human-computer interaction systems, enabling machines to interpret human intentions through body movement analysis. A common approach employs the Dynamic Time Warping (DTW) algorithm to compute similarity between keypoint sequences extracted via pose estimation from both reference and live video streams. While effective and flexible, DTW has limitations, including high space complexity and sensitivity to keypoint misidentification often seen in pose estimation. To address these issues, we propose two enhancements to the DTW-based gesture recognition pipeline. First, we apply a space compression technique to reduce memory usage without sacrificing performance. Second, we introduce a frame-skip mechanism to mitigate the impact of incorrectly detected keypoints on recognition results. To evaluate our method, we construct a dataset of traffic gestures commonly used in Taiwan. Experimental results show that the proposed enhancements improve the efficiency, scalability, and accuracy of gesture recognition, making the approach more suitable for real-time use in constrained environments.
9:15 Efficient Task Scheduling Algorithms for Decentralized Large Language Model Serving
Sanjaya Kumar Panda (National Institute of Technology Warangal & NITW Techsammelan Private Limited, India); Sankalp Dubey (NIT Warangal, India); Siba Mishra (C. V. Raman Global University, Bhubaneswar, India)
Large language models (LLMs) have gained enormous popularity for processing and generating text. They are a subset of generative artificial intelligence (GenAI) that require higher availability of graphical processing unit (GPU) resources for inference services. However, making GPU resources available in a centralized infrastructure is quite challenging. Therefore, recent works have focused on decentralized physical infrastructure networks (DePIN) to utilize idle GPU resources, enabling scalable LLM inference services across the decentralized network. These inference services may experience inherent latency (i.e., measured in time per output token (TPOT)) due to communication overhead or time between GPU resources responsible for generating consecutive tokens. The task scheduling algorithm is crucial in decentralized LLM inference services to minimize TPOT and maximize GPU resource utilization, particularly when GPU resources are constrained by computational capacity. This paper introduces two task scheduling algorithms, the improved greedy heuristic shortest path algorithm (IGHSPA) and the dynamic programming-based task scheduling algorithm (DPTSA), for decentralized LLM serving to achieve these objectives. Each task involves assigning a layer to a GPU resource, which IGHSPA and DPTSA accomplish using greedy heuristic and dynamic programming. Both algorithms are extensively simulated and compared with one of the recent algorithms, namely the greedy heuristic shortest path algorithm (GHSPA), in terms of TPOT and execution time (ET). Our simulation results demonstrate that DPTSA improves TPOT up to 47.50% and 35.50% and ET up to 99.95% and 35.00%, compared to GHSPA and IGHSPA.