Computing and Computational Intelligence (CCI) 1
Track A2F1 CCI 2.1: Computing & Computational Intelligence (CCI) 2.1
Room: F1. Sipadan I (Level 4)
Chair: Helen Sin Ee Chuo (Universiti Malaysia Sabah, Malaysia)
2:30 DGE2I-Net: Exploring Depth Gait Energy and Entropy-Based Image Features for Human Gait Classification Using Deep Neural Networks
Mainak Ghosh, Sourav Biswas and Anup Nandy (National Institute of Technology Rourkela, India)
Image-based gait analysis has become an important research area for extracting comprehensive features for classification of gait abnormality. However, traditional Image-based models mostly focus on the pattern analysis from the image sequences, which affects the efficiency of the classification model. To overcome this problem, Gait Energy Image and Gait Entropy Image are used as silhouette-based feature extraction methods. But these silhouette-based methods frequently face challenges due to variations in lighting, background, and clothing conditions. To mitigate these challenges, depth information is captured to represent of gait features using Depth Gait Energy Image (DGEI) and Depth Gait Entropy Image (DGEnI). Due to 3D nature of the depth images, these are not affected by the environmental obstacles. We create a depth gait dataset of 10 subjects to evaluate these features with two Convolutional Neural Network based models. We achieve 96.30% and 98.62% accuracy for classification of human gait with DGEI and DGEnI-features respectively, which significantly outperforms many traditional methods.
2:45 Steel Classifications Based on Sparks Patterns Using Convolutional Neural Network ResNet Pre-Trained Models
Ali M. A. Daban (Universiti Technologi Malaysia, Malaysia & Business Media International, Malaysia); Siti Armiza Mohd Aris (Universiti Teknologi Malaysia, Malaysia); Syahid Anuar (Universiti Teknologi Malaysia, Kuala Lumpur, Malaysia)
Steel plays a vital role in numerous industries due to its strength, versatility, and wide-ranging applications. However, its production is highly energy-intensive and a significant contributor to global CO₂ emissions, raising environmental concerns. Recycling steel offers a more sustainable alternative, yet accurately classifying different steel types remains a persistent challenge, especially in resource-limited settings. This study introduces an automated method for steel classification based on the traditional spark test, utilizing pre-trained Convolutional Neural Networks (CNNs). Specifically, ResNet34 and ResNet101 models were trained on spark test images to evaluate their classification performance. ResNet34 achieved a notable accuracy of 95.0%, outperforming the previously used ResNet50 model, which recorded 92.6%. In contrast, ResNet101 significantly underperformed with an accuracy of just 20.6%. These results demonstrate the practical potential of deep learning models, particularly ResNet34, to enhance the efficiency of steel recycling processes, ensure higher material quality, and contribute to environmental sustainability through automated, precise classification.
3:00 Deep Representation Learning with Deep Belief Network for Petrophysical Properties Prediction
Pallabi Saikia (Rajiv Gandhi Institute of Petroleum Technology, India)
Deep neural networks are becoming predominant for the task of discrimination due to its capability to learn complex features. But its advantages get constrained in the scarcity of labelled data, as accomplishing proper representation by the network can be difficult in such scenarios. In this paper, we investigated deep belief network to learn deep representations of regression data, which ultimately behave as the feature detectors in deep neural network model. The model have been explored on a real-world regression problem of petro-physical properties prediction, in the domain of reservoir characterisation. The model leverages abundantly available unlabelled seismic data to learn better representation and such representations are applied in the deep neural network model for guiding the training of regression model in limited labelled seismic. We analyse the performance of the prediction capability of neural network using the feature detectors obtained from deep belief network. The results demonstrate that the model outperforms conventional neural network model in terms of generalised error and Computation units required.
3:15 Domain Knowledge Leveraged Inductive Transfer Learning on Solving a Regression Task
Pallabi Saikia (Rajiv Gandhi Institute of Petroleum Technology, India)
Deep neural networks (DNNs) are widely used for classification tasks due to their ability to learn complex features. However, their effectiveness is constrained in data-scarce scenarios. Transfer learning (TL) has proven useful in handling such limitations by leveraging knowledge from related tasks. While TL is extensively applied in classification, its use in regression remains relatively unexplored. This paper presents a TL-based approach for regression when labeled data is limited. The study focuses on predicting petrophysical properties in reservoir characterization, a real-world regression problem. A ResNet model, pre-trained on a related classification task within the domain, is adapted to regression by transferring learned parameters. This imposes an inductive bias that enhances model generalization. The experimental results demonstrate that the proposed method significantly improves regression performance compared to conventional approaches, particularly in low-data conditions. The findings highlight the potential of TL in improving regression models when data availability is a major constraint.
3:30 Object Detection for Intelligent Vehicles
Gonugunta Sai Prakash (SRM University-AP, India); Rituparna Choudhury (International Institute of Information Technology Bangalore, India)
Intelligent Vehicle System (IVS) is a very lucrative application of artificial intelligence or AI. These intelligent vehicles need to identify or classify objects present in a captured image. The detection latency is a very important factor for these high-speed applications. In real-time scenarios, this object classification or labeling latency should be minimal. In these applications, the accuracy of detection also plays a vital role. So, this paper proposes a hybrid algorithm to reduce detection latency while improving the accuracy of detection. The proposed approach uses a hybrid of feature selection, clustering, and classifiers to achieve the highest accuracy and lowest latency. The results show that the proposed method achieves better performance as compared to other algorithms. It can complete the classification in 2 msec while the detection latency of other deep learning or machine learning techniques is in the range of seconds. This algorithm also has the highest accuracy among all the existing classification models.
3:45 Multidomain HRV Features for the Smart Diagnosis of ADHD Using ML Models
Shafna V and S D Madhu Kumar (National Institute of Technology Calicut, India)
Attention-Deficit/Hyperactivity Disorder (ADHD) is a common neurodevelopmental condition that persists into adulthood and is often challenging to diagnose due to the subjectivity of traditional assessment methods. While neuroimaging and EEG-based approaches have gained popularity in ADHD research, they are often resource-intensive and less accessible. In contrast, physiological signals such as Heart Rate Variability (HRV) offer a non-invasive and cost-effective alternative that remains underexplored, particularly in adult populations. This study investigates the potential of HRV-derived features for distinguishing adults with ADHD from healthy controls using Machine Learning (ML) techniques. Time-domain, frequency-domain, and wavelet-domain features were extracted from RR interval data. To identify the most informative features, we employed Recursive Feature Elimination (RFE) using a Random Forest Classifier, selecting the top five features contributing most to classification performance. Among various models evaluated, the XGBoost classifier achieved the best performance with 91% recall 83% accuracy and f1-score. The results highlight that HRV-based metrics can provide meaningful insights into autonomic nervous system irregularities associated with ADHD in adults. The proposed approach demonstrates a viable path for scalable and accessible ADHD screening using physiological biomarkers. Future work will focus on expanding the dataset and incorporating advanced feature engineering and deep learning for improved accuracy.
Control Systems & Robotics (CSR)
Track A2F2 CSR 2: Control Systems & Robotics (CSR) 2
Room: F2. 501 Kadamaian (Level 5)
Chair: Fatanah Mohamad Suhaimi (Universiti Sains Malaysia, Malaysia)
2:30 Data-Driven Controllers for Nonlinear Discrete-Time Systems
Koshy George (GITAM University, India)
The emphasis on data-driven control has increased in recent years, even though the concept is over 100 years old. In one class of such controllers, the objective is to identify a model of the controlled plant and then control using any model-based control technique. Alternatively, methods have been introduced to tune the controller's parameters using the available data directly. Apart from these two, methods have evolved recently based on using a simplified mathematical representation of the plant, typically called model-free control (MFC). MFC is attractive as the objective is to control a plant without explicitly using a mathematical model. It is advantageous if the complex process can be controlled using a simpler model that may not wholly represent the entire process. Many successful applications have been reported. In contrast, model-based control based on the mathematical description of the behaviour has been studied and extensively used. This paper addresses the question of which data-driven controller to use in the context of a class of nonlinear discrete-time systems. It compares a method based on partial linearisation with a technique based on feedforward neural networks. The former is a model-free control and requires no prior information about the plant's dynamics. The latter, a model-based control technique with minimal prior information, can provide significantly better tracking performance.
2:45 Evaluating Sensor Fusion Methods for Precise Differential Drive Robot Positioning
Luqmanul Haqeem Saiful Rizal and Saiful Rizal Abdullah (Malaysia); Khairulmizam Samsudin (Universiti Putra Malaysia, Malaysia); Nur Alifah Ilyana Binti Mohd Sharihan (IEEE Region 10 Conference, Malaysia)
This paper presents a comparative study on the odometry and positioning accuracy of a differential drive mobile robot equipped with motor encoders and an onboard Inertial Measurement Unit (IMU). The primary objective is to evaluate and improve localization performance by comparing three methods: basic dead reckoning, a Complementary Filter (CF), and an Extended Kalman Filter (EKF) for sensor fusion. Dead reckoning, while simple and commonly used, is prone to cumulative errors over time, particularly in dynamic or uneven environments. To address this limitation, the CF and EKF are employed to fuse data from the encoders and the IMU, aiming to enhance position and orientation estimation. Experimental evaluations demonstrate that both sensor fusion techniques significantly outperform basic dead reckoning, with accuracy improvements of up to 80%. Among the methods tested, the EKF-based approach delivers the most consistent and precise localization results, making it the most suitable option for applications requiring high-accuracy autonomous navigation.
3:00 Optimal Variable Gain Sliding Mode Controller Using Grey Wolf Optimization for Pitch Stabilization of Twin-Rotor MIMO System
Koteswara Rao Palepogu and Subhasish Mahapatra (VIT-AP University, India); Atanu Panda (Sister Nivedita University, India)
Helicopters play a pivotal role in military and rescue operations, necessitating precise control during critical maneuvers such as hovering, take-off, and landing. Stabilizing multi-rotor systems becomes increasingly challenging in the presence of external disturbances. This work has been implemented in a twin rotor MIMO (TRMS) system that mimics a helicopter model. This study introduces an innovative method for regulating the pitch angle of a twin-rotor MIMO system through a variable-gain sliding-mode controller. The proposed controller integrates a twisting algorithm with error-dependent variable gains, offering improved control efficiency compared to fixed-gain methods. In this work, the proposed controller parameters are optimized using the Grey Wolf optimization algorithm, ensuring accurate tracking and robust performance. Simulation result outcomes validate the capability of the controller to achieve precise pitch tracking under system uncertainties and actuator faults, highlighting its potential for real-world applications. Besides, the robust behaviour is analyzed under various uncertain scenarios to highlight the effectiveness of the proposed control.
3:15 Design and Development of an Indigenous Modular 5-DOF Robotic Arm for Multidisciplinary Tasks
Ashab Farhan Anon (Aviation And Aerospace University Bangladesh, Bangladesh); Md Samiullah Prodhan (Aviation and Aerospace University Bangladesh, Bangladesh); Holyjith Paul Himel (Aviation and Aerospace University Bangladesh & N/a, Bangladesh); Md Tariqul Islam (Aviation And Aerospace University Bangladesh, Bangladesh); Md. Ridoan Hasan (Aviation and Aerospace University Bangladesh, Bangladesh & N/a, Bangladesh); Khandokar Ahosanul Islam Jisan (Aviation And Aerospace University Bangladesh, Bangladesh); Tyseer Ninad, Md. Ehsanur Rahman and Afzal Hossain (Aviation and Aerospace University Bangladesh, Bangladesh)
This paper presents the design and development of an indigenous modular 5-degree-of-freedom (5-DOF) robotic manipulator aimed at diverse applications. The arm is built from locally sourced materials and cost-effective components to ensure affordability without compromising functionality. This research work describes materials selection, mechanical CAD modeling (SolidWorks), and the joint/actuator configuration. The control system uses open-source microcontrollers (e.g., ESP32, Arduino) and is integrated with sensors. Comprehensive simulations (static stress, dynamic motion, thermal and vibration analyses) were performed to validate the design. A prototype was fabricated and tested, demonstrating the expected reach, payload capacity, and positioning precision. Potential applications in industry, aerospace/space missions, healthcare, defense, and education are discussed. The results indicate that the proposed design meets the objectives of affordability and versatility for multidisciplinary use. Its modular architecture and low-cost parts make it suitable for a range of multidisciplinary applications. Overall, the research and innovative work demonstrates the feasibility and benefits of developing locally engineered robotic manipulators for diverse industries.
3:30 Digital Twin-Based Industry-Informed Environmental Quality Management (DTI2EQM)
Brijith Jacob and Santhosh Kumar G (Cochin University of Science and Technology, India)
The global struggle for clean air and water is a fight to secure access to these essential resources for everyone. This battle against pollution, environmental degradation, and the impacts of climate change engages individuals, communities, organizations, and governments worldwide. Urban centers, in particular, face the challenge of pollution, exacerbated by rapid urbanization. To combat this, cities are implementing various strategies, including air quality monitoring, stringent emissions targets, clean air zones, a transition to renewable energy sources, and expanded public transportation. These and other innovative solutions are crucial in the ongoing pursuit of clean air, clean water, and a sustainable environment. DTI2EQM represents a cutting-edge approach to addressing environmental challenges by merging the power of digital twin technology with industry- specific knowledge and data. We propose the design and implementation strategies to develop a smart city digital twin component leveraging existing (for example, LoRaWAN) infrastructure to interface with pollution sensors. The Digital Twin architecture will integrate multiple technologies, including inputs from the sensors, computational components such as machine learning models, ontology-based knowledge graphs, persistence and data repositories, and a comprehensive dashboard as the user interface.
3:45 Wiper Motor Control for Oscillation Suppression of Windshield Wiper Arms Using H-Infinity Synthesis
Tsutomu Tashiro and Sodai Kato (Osaka Sangyo University, Japan)
In this paper, a windshield wiper control and its design method to achieve both basic function of wiper control and oscillation suppression of wiper arms. A wiper model is described as a three-inertia resonance system to realize second mode oscillation characteristics. The control is designed based on H-infinity control theory to track the target motor angle and suppress the oscillation of the angular speeds of the driver's side arm and passenger's side arm. Verification is carried out with a wiper test rig using five types of rubber blades with different deterioration levels under a wide range of rainfall conditions. The effectiveness of the proposed control is demonstrated by comparing with the results of the conventional control designed based on optimal regulator theory. The experimental results are investigated from the viewpoints of suppressing the second mode oscillation and the accuracy of the reversal motor angle which determines the wiping range of the arms.
Power, Energy & Electrical Systems (PES)
Track A2F3 PES 2: Power, Energy & Electrical Systems (PES) 2
Room: F3. 502 Mesilau (Level 5)
Chair: Shamsul Aizam Zulkifli (UTHM, Malaysia & Universiti Tun Hussein Onn Malaysia, Malaysia)
2:30 Enhancing WLS State Estimation Using Future Load Profile Nomination (FLPN)
Earl Humprey M Bantug and Abdul Aziz G. Mabaning (Mindanao State University - Iligan Institute of Technology, Philippines)
Traditional pseudo-measurements often rely on operator forecasts, which lack consumer intent and adaptability. This paper proposes an enhanced Weighted Least Squares (WLS) State Estimation framework that integrates Future Load Profile Nominations (FLPNs), forward-declared load expectations with consumer-declared confidence levels, to improve estimation accuracy in low-SCADA or distribution-level networks. Unlike static forecast-based methods, FLPNs are modeled as inverse-variance-weighted pseudo-measurements, enabling direct consumer participation in grid monitoring.
The framework is evaluated on the IEEE 14-bus test system in two configurations: (1) FLPN-only, where SCADA data at participating buses is fully replaced, and (2) FLPN-augmented, where both SCADA and FLPN data are fused. A 4 by 4 sensitivity analysis across FLPN participation (25-100%) and accuracy levels (25-100%) shows that the FLPN-augmented mode consistently achieves the lowest voltage RMSE across most conditions, achieving up to 19.3% lower voltage RMSE than FLPN-only.
These findings demonstrate that integrating participatory declarations significantly enhances estimation accuracy and supports scalable, consumer-integrated grid operations. The study establishes a foundation for future extensions to dynamic and distributed estimation frameworks, advancing intelligent grid monitoring aligned with SDGs 7, 9, and 11.
2:45 Enhanced Bad Data Detection in Power System State Estimation Using Modified CUSUM with Future Load Profile Nomination (FLPN)
Franclein L. Francisco and Abdul Aziz G. Mabaning (Mindanao State University - Iligan Institute of Technology, Philippines)
Power system state estimation is highly sensitive to the quality of input data, particularly in distribution grids where consumer-driven load variability can introduce significant discrepancies between expected and actual power usage. Traditional residual-based detection methods, including classical cumulative sum (CUSUM), often misclassify these natural fluctuations as anomalies due to their reliance on static baselines and lack of normalization. This study proposes an enhanced anomaly detection framework that integrates Future Load Profile Nominations (FLPNs) into a modified two-sided CUSUM structure. The method evaluates normalized residuals relative to time-aware, user-declared load expectations, and incorporates drift compensation to suppress the accumulation of small benign deviations. Simulation results on the IEEE 14-bus system across 100 trials revealed that the Modified CUSUM with FLPN achieved up to 82% reduction in false alarm rate and up to 69% reduction in missed detection rate compared to traditional CUSUM (without FLPN) approach. This underscores the efficacy of integrating FLPNs and normalization in enhancing anomaly detection performance under dynamic consumer load behavior.
3:00 Future Load Profile Nomination (FLPN) in PV-BESS-EV Microgrids: A Philippine Use Case
Jeric Cesar Aquino Enriquez and Abdul Aziz G. Mabaning (Mindanao State University - Iligan Institute of Technology, Philippines)
Modern power systems face challenges amidst the increased integration of distributed energy resources such as photovoltaic (PV) systems, battery energy storage systems (BESS), and electric vehicles (EVs), which introduce complexity and uncertainty of demand profiles. Hence, this study proposes an optimization framework incorporating Future Load Profile Nominations (FLPN) as proactive, consumer-declared forecasts for energy dispatch planning, enabling a Proactive Energy Management (PEM) approach to grid operations. The IEEE 14-bus system is used as a testbed, with PV, BESS, and EV assets scaled proportionally to static loads at each bus. Dispatch optimization is performed under Philippine time-of-use pricing using a Monte Carlo approach to model FLPN uncertainty. Three scenarios are evaluated: (i) baseline dispatch without FLPN, (ii) FLPN-assisted dispatch with flexible EV charging, and (iii) FLPN-assisted dispatch with bidirectional Vehicle-to-Grid participation. Results show that while FLPN visibility alone yields limited operational benefits, its integration with flexible and controllable EVs achieves substantial improvements, reducing dispatch costs by up to 6.8% and lowering grid dependency up to 3%. This approach aims to contribute to achieving United Nation Sustainable Development Goals (SDGs) 7 (Affordable and Clean Energy) and 13 (Climate Action) by promoting proactive, more resilient, and consumer-integrated energy management strategies.
3:15 Privacy-Preserving Data Transformation Using Hybrid Signal Component Analysis for Forecasting Incorporating Future Load Profile Nomination
Perly Rica U. Flores and Abdul Aziz G. Mabaning (Mindanao State University - Iligan Institute of Technology, Philippines)
This study introduces a novel privacy-preserving data transformation framework for load forecasting that uses hybrid signal component analysis (SCA) techniques. Addressing the growing need for accurate energy predictions while protecting consumer privacy, the framework incorporates Future Load Profile Nomination (FLPN) with Load Profile Data (LPD) prior to transformation. A variety of basic and SCA-based methods including Principal Component Analysis (PCA), Independent Component Analysis (ICA), and Wavelet Transform (WT) were tested, both individually and in hybrid combinations, to address the challenge of balancing privacy and forecasting accuracy. Evaluation showed that WT-based combinations significantly boost privacy, achieving a privacy Mean Squared Error (MSE) as high as 4.680, but this severely decreased forecasting accuracy to an MSE of 4.786. These findings show that hybrid transformation methods can be systematically evaluated to optimize the trade-off between utility and privacy. This transformation-based approach offers an alternative to methods like Federated Learning and Differential Privacy, contributing to reliable, accurate, and privacy-respecting power systems aligned with SDG 7 and 11.
3:30 Proactive Energy Management Through Demand-Side Forecasting: A Future Load Profile Nomination Approach
Abdul Aziz G. Mabaning (Mindanao State University - Iligan Institute of Technology, Philippines)
The Future Load Profile Nomination (FLPN) framework is introduced as a novel approach to enhance power system load forecasting by enabling proactive participation from the demand side. As a specific application of the broader concept of Future Behavior Nomination (FBN) within power systems, FLPN addresses a key limitation of traditional forecasting methods, which depend primarily on historical and real-time data and often fail to anticipate deviations driven by future-oriented consumer behavior. FLPN empowers consumers to voluntarily nominate their anticipated electricity consumption in advance, effectively shifting the demand side from reactive to proactive. Mathematical models are developed at both the individual and system-aggregated levels, incorporating dynamic parameters such as participation rates, forecast blending, and time-varying accuracy. A simulation of a non-historical load anomaly validates the framework, demonstrating a significant reduction in forecast error even with imperfect user input. FLPN contributes to the broader objective of Proactive Energy Management (PEM) and directly supports the United Nations (UN) Sustainable Development Goals (SDG) by promoting efficient grid operations that reduce reliance on carbon-intensive spinning reserves (SDG 13), increase the hosting capacity for renewables (SDG 7), and empower consumers for responsible consumption (SDG 12).
3:45 Application of Combined GWO-HHO Techniques for Minimization of Energy Consumption Cost in Home Energy Management System
Avni Bagga and Shreya Ranjan (Vellore Institute of Technology, India); Prabhakar Karthikeyan Shanmugam (VIT University, India)
An increase in global energy requirements, coupled with growing environmental concerns, has led to a heightened focus on efficient energy management systems, specifically in residential settings. Home Energy Management Systems (HEMS) have emerged as a promising solution to optimize energy consumption, reduce costs, and minimize environmental impact. These systems utilize advanced technologies and algorithms to monitor, control, and manage energy usage within households. Recent advancements in optimization techniques have introduced novel approaches to enhance the performance of HEMS. This paper explores the application of two powerful metaheuristic algorithms, the Grey Wolf Optimizer (GWO) and the Harris Hawks Optimization (HHO) algorithm. The objective of this paper is to minimize the energy consumption cost and simultaneously satisfy the operational constraints of residential appliances, which include demand limits, minimum usage frequency, and energy requirements. To meet the constraints, penalty functions are inculcated. The GWO, inspired by the social hierarchy and hunting behavior of grey wolves, and the HHO, which mimics the cooperative hunting strategy of Harris hawks, have shown promising results in optimizing the energy consumption cost by scheduling the loads effectively.
Electronics, Circuits & Devices (ECD)
Track A2F4 ECD 2: Electronics, Circuits & Devices (ECD) 2
Room: F4. 503 Dinawan (Level 5)
Chair: Mazlina Mamat (Universiti Malaysia Sabah, Malaysia)
2:30 Frequency Dependence of Electron Emission Current from Crystal Defects in AlGaN/GaN HEMTs
Hideki Hoji, Soichi Sano, Junya Takeda and Hirohisa Taguchi (Chukyo University, Japan)
In this study, we investigated the drop in the drain current value of an AlGaN/GaN HEMT immediately after the collapse phenomenon occurred. Analysis of the IV characteristics clarified that this occurrence was owing to electrons being trapped by the current collapse phenomenon, causing a drop in the electron concentration. By applying a frequency that suppresses the current collapse phenomenon from the gate electrode side to the AlGaN-GaN crystal layer, we measured the process in which electrons were released from crystal defects in the AlGaN-GaN crystal layer. The frequency dependence of the carrier emission process was measured at 1 to 4 GHz, and we confirmed that the higher the frequency, the slower the carrier emission and the higher the saturated drain current value. This is because the time taken for electrons to be trapped in crystal defects and the process time for electrons to be emitted differ depending on the crystal defect. In crystal defects with long emission times, electrons are repeatedly emitted and recaptured even when the current collapse phenomenon is suppressed. Consequently, electrons cannot be recaptured at high frequencies, and the saturated drain current value increases. By comparing the carrier emission process at different frequencies, the components of crystal defects that cannot emit or capture electrons can be extracted. Crystal defects can be evaluated by comparing the carrier emission process at different frequencies.
2:45 Relationship Between Kink Phenomenon and Crystal Strain in AlGaN/GaN HEMT Under High Voltage
Soichi Sano, Junya Takeda, Hideki Hoji, Sho Nagai and Hirohisa Taguchi (Chukyo University, Japan)
In this study, we confirmed the dependence of voltage stress and temperature on the I-V characteristics of AlGaN/GaN high-electron mobility transistors (HEMTs) immediately after applying voltage stress. The high voltage stress and high-temperature environment intensified the kink occurrence, likely due to the expansion of the crystal strain. Results showed that, as the voltage stress increases, a crystal strain occurs in the GaN or AlGaN layer, which affects carrier transport. It has been reported that when the drain voltage is lower than 6 V, two dimensional electron gas (2DEG) is not stably formed; thus, carrier scattering occurs due to crystal distortion. Meanwhile, 2DEG stably forms above 6 V; hence, the influence of distortion is assumed to be minimal. The inflection point of the drain conductance is further calculated based on the I-V characteristics obtained from the kink phenomenon. This inflection point suggests a transition from 3D drift motion to 2DEG carrier transport. The impact of lattice scattering becomes more noticeable at drain voltages under 6 V in the I-V characteristics, where three-dimensional transport is dominant. The lattice scattering is confirmed to be relatively mitigated at drain voltages above 6 V in the I-V characteristics, where two-dimensional transport by the 2DEG is prevalent. It is reported that the lattice vibration of crystal strain due to temperature rise saturates above a certain temperature, and electron scattering also gets limited. When a certain temperature is exceeded under high electric field stress, the drain voltage, which is the inflection point, no longer changes. This is because the amount of crystal distortion due to temperature saturates and no longer affects the 2DEG.
3:00 Influence of Thermal Fluctuations on the Inverse Piezoelectric Effect in AlGaN/GaN HEMTs
Junya Takeda, Soichi Sano and Hirohisa Taguchi (Chukyo University, Japan)
With the growing demand for high-performance, high-frequency electronic devices, the development of devices based on gallium nitride (GaN) has garnered increasing interest. In this study, we examined the current-voltage characteristics of AlGaN/GaN high-electron-mobility transistors subjected to electrical stress. Stress application induced a kink effect, which became more pronounced with increasing gate reverse-bias stress and stress application duration. This observation indicated that local lattice strain, caused by the inverse piezoelectric effect, degraded the device performance. When combined with lattice vibrations owing to heat, this strain led to further degradation. However, the combined effect eventually reached saturation. Moreover, lattice vibrations likely suppressed electron trapping and strain during stress application, resulting in a nonlinear performance degradation trend. The inverse piezoelectric effect was sensitive to thermal fluctuations, resulting in instability. Furthermore, transient response measurements of the drain current revealed a long-term current recovery process, indicating a variation in the interaction between polarization and thermal fluctuations.
3:15 Design of Highly Stackable Charge Trap-Based 3D DRAM
Hyeongyu Kim (Jeonbuk National University, Korea (South)); Dabok Lee (Gyeonsang National University, Korea (South)); Hyun-Sik Choi (Kwangwoon University, Korea (South)); Yoojin Seol (Jeonbuk National University, Korea (South)); Jonghyeon Ha (Gyeongsang National University, Korea (South)); Kihyun Kim (Jeonbuk National University, Korea (South)); Jungsik Kim (Gyeongsang National University, Korea (South)); Won-Ju Cho (Kwangwoon University, Korea (South)); Zvi Or-Bach (MonolithIC 3D, USA); Sungil Chang (MonolithIC3D, Korea (South))
In this work, we propose a highly stackable Charge Trap-based 3D DRAM (CT 3D DRAM) structure that addresses key challenges in future memory scaling, including 3D integration, power consumption, and thermal management. Unlike conventional DRAM architectures that rely on complex capacitor structures, the proposed CT 3D DRAM utilizes a simple 1T memory cell with a poly-Si channel and Schottky barrier source/drain (S/D) contacts formed by metal silicide. Hot carrier injection (HCI) from the source side enables fast program operations through an ultrathin tunnel oxide. Key device parameters were optimized using 3D TCAD simulations, and planar CT DRAM devices were fabricated to validate the concept. The fabricated devices exhibited a program/erase window larger than 1 V under a 20 ns pulse, excellent retention characteristics exceeding 10 seconds at 85 °C, and endurance up to 10¹⁵ cycles with a remaining threshold voltage window of approximately 0.32 V. Moreover, the use of metal S/Ds significantly enhances heat dissipation and enables superior thermal management, critical for highly stacked 3D memories. The vertical integration of metal bit lines (BLs) and horizontal poly-Si channels results in lower RC delays, making the CT 3D DRAM scalable even beyond a thousand layers while maintaining effective cell area comparable to conventional 4F² DRAMs. Through the optimized design of the word line (WL) and bit line (BL) structures, as well as control of key materials such as the tunnel oxide and charge trap nitride, we demonstrate that CT 3D DRAM can achieve both high speed and reliability. This architecture offers a promising solution for next-generation 3D DRAM applications requiring high density, low power, and efficient thermal management, particularly in emerging memory platforms like Compute Express Link™ (CXL™) memory.
3:30 Bitstream-Level IO Tampering: Exploiting FPGA Bitstreams to Compromise ADC Integrity in Cyber-Physical Systems
Nur Alifah Ilyana Binti Mohd Sharihan (IEEE Region 10 Conference, Malaysia); Khairulmizam Samsudin (Universiti Putra Malaysia, Malaysia); Shaiful Hashim (UPM, Malaysia); Faisul Arif Ahmad (Universiti Putra Malaysia & Faculty of Engineering, Malaysia)
As Field-Programmable Gate Arrays (FPGAs) are increasingly deployed in mission-critical systems, they have become attractive targets for low-level hardware attacks. This paper presents a novel bitstream-level tampering methodology aimed at disrupting the functionality of FPGA-configured Analog-to-Digital Converter (ADC) applications by manipulating peripheral behavior. The proposed attack operates at post-bitstream generation, assuming adversarial access to the unencrypted bitstream file and the ability to reprogram the target device. Two distinct attack models are demonstrated: (i) Hardware-Level Denial-of-Service (HLDoS) attacks, which disable the ADC's conversion trigger by reverting its control port to a weak pull-up state, thereby halting successive data acquisition; and (ii) Digital Port Tampering (DPT) attacks, which corrupt data integrity by altering specific I/O bits associated with the ADC output interface. Experimental validation on an Intel Cyclone V FPGA confirms the viability and severity of these attacks, revealing both total functional loss and systematic bit-level corruption. The findings underscore a critical security gap in FPGA-based systems and motivate the need for robust bitstream authentication and runtime validation mechanisms.
3:45 Hardware Software Co-Design of 2D Modulation Schemes OTFS and OTSM on System-on-Chip
Sai Kumar Dora and Rakesh Kumar Yadav (IIT ISM Dhanbad, India); Himanshu Bhusan Mishra (IIT (ISM) Dhanbad, India); Amitav Panda (Nokia Solutions and Networks India Pvt Ltd, India)
In this paper, we design low-complexity hardware architectures for the basic modules of the two-dimensional (2D) modulation scheme, orthogonal time sequency multiplexing (OTSM). OTSM scheme works in the delay-sequence domain by using the basic modules inverse Walsh-Hadamard transform (IWHT) and Walsh-Hadamard transform (WHT) at the transmitter and receiver, respectively. We next compare the performance of the proposed architectures of the above-mentioned basic modules with that of its counterpart modules of the another 2D modulation technique Zak based orthogonal time frequency space (OTFS). Note that Zak-OTFS operates in the delay-Doppler domain, requiring the primary modules as 2D inverse Zak (IZak) and Zak transforms at the transmitter and receiver, respectively. This comprehensive comparative analysis is conducted on the computational complexity, timing performance, and power consumption of both schemes, evaluated on the ZCU706 Zynq SoC board. The results indicate that OTSM outperforms the Zak-OTFS in terms of area, power consumption, and latency. Zak-OTFS requires more programmable logic (PL) resources, utilizing 28,879 LUTs and 26,124 FFs, while OTSM uses significantly fewer resources, with 5,440 LUTs and 6,216 FFs.
Communication Systems (CS)
Track A2F5 CS 2: Communication Systems (CS) 2
Room: F5. 504 Madai (Level 5)
Chair: Nur Idora Abdul Razak (Universiti Teknologi MARA, Malaysia)
2:30 Deep Reinforcement Learning-Based Dynamic Sharding for Blockchain IoT
Pooja Khobragade and Ashok Kumar Turuk (National Institute of Technology Rourkela, India)
Internet of Things (IoT) and Blockchain integration are having a huge impact on the future of technological progress. IoT has progressed from an emerging notion to a widely used technology, shaping the future of digital connectivity. With billions of networked IoT devices generating large amounts of data, efficient data management becomes critical. Blockchain has been explored extensively to enhance security in IoT networks however, its scalability limitations become evident when handling large-scale deployments. Sharding is recognized as a promising approach to improve blockchain scalability by partitioning the network into multiple independent groups. These groups, called shards, process transactions in parallel, increasing throughput while reducing communication, computation, and storage overhead. Despite its advantages, many existing blockchain sharding models rely on static algorithm, which fail to accommodate the dynamic nature of blockchain networks. Factors such as variable node involvement and possible security concerns present problems that static sharding cannot solve. To address these restrictions, deep learning provides a strong solution for dynamic and multidimensional sharding in blockchain-based IoT systems. Deep learning, with its capacity to understand complex patterns and adapt to changing network circumstances, can improve the efficiency, security, and scalability of blockchain-powered IoT networks. This article proposes a deep reinforcement learning-based dynamic shards in blockchain IoT applications to overcome scalability difficulties.
2:45 Fisheye Camera-Aided Standalone IRS Control for Transmitter Beamforming
Yoshihiko Tsuchiya (Tokyo University of Science, Japan); Norisato Suga (Shibaura Institute of Technology & ATR, Japan); Kazunori Uruma (Kogakuin University, Japan); Masaya Fujisawa (Tokyo University of Science, Japan)
Intelligent reflecting surface (IRS)-aided wireless communication is attracting interest as a technology that can improve communication quality and coverage in high-frequency bands such as millimeter waves. The proper control of the IRS requires channel estimation for each element, which reduces communication efficiency owing to a channel estimation overhead. Furthermore, when an IRS is used in multiple-input multiple-output (MIMO) systems, the overhead becomes larger than that of single-input single-output (SISO) systems. To control the IRS according to the beamforming of the transmitter, the IRS must cooperate with the transmitter. This requires a channel estimation of the number of elements for each transmitter antenna. Therefore, a standalone IRS that does not require a connection or channel estimation for cooperation between the IRS and transmitter is required. Recently, standalone IRS control in SISO systems using camera images has been proposed. In this paper, we propose a method to adapt this concept to transmitter beamforming. Our method achieves a standalone IRS by predicting the channel based on the 3D position estimated by detecting the user in the camera image and then applying the reflection coefficients corresponding to the estimated beamforming vectors. Numerical experiments confirm that the proposed method can control the IRS with only a slight degradation in communication quality.
3:00 Enhancing VLC Performance: an Experimental Study on System Parameter Tuning
Dilanka de Silva (University of Moratuwa, Sri Lanka & Sri Lanka Technology Campus (Pvt) Ltd, Sri Lanka); Ruwan Weerasuriya (Chalmers University of Technology, Sweden); Sumudu G. Edirisinghe (Sri Lanka Technological Campus, Sri Lanka); Madushanka Nishan Dharmaweera (University of Sri Jayewardenepura, Sri Lanka); Samiru Gayan (University of Moratuwa, Sri Lanka)
This research presents a comprehensive experimental and simulation-based investigation aimed at enhancing the link distance and performance of indoor Visible Light Communication (VLC) systems. A basic VLC setup was incrementally improved by optimizing key system parameters, including LED transmit power, transimpedance amplifier (TIA) feedback resistor, lens configuration, photodiode selection, and wall color. Each enhancement was experimentally validated, and corresponding simulations were conducted to analyze system behavior and verify measurement results. The study also examines the applicability of well-established VLC channel models, such as the Masao Nakagawa model and second-reflection models, in real-world scenarios. The findings confirm that incorporating optical and environmental optimizations can significantly extend link distance, up to 3 m with high accuracy, while aligning closely with simulation predictions. This work not only validates theoretical models but also provides practical design insights for reliable, long-range VLC deployment in indoor environments. The results demonstrate the importance of iterative experimental tuning in bridging the gap between theoretical assumptions and actual VLC system performance.
3:15 A Web-Based Testbed with Spatial Visualization Capabilities for IoT-Based Smart Systems
Josef Isaac Babaran, Ivan Blaise Gonzales, Julius Brian Ipac and Lord Peter Robin Dustin Suyat (University of the Philippines - Diliman Campus, Philippines); Paul John C Tiope, Adrian Cahlil Eiz Togonon and Jaybie A. de Guzman (University of the Philippines Diliman, Philippines)
In the world of the Internet of Things, testbeds offer the necessary infrastructure in which ideas, technologies, and practices are tested and validated. A smart classroom called Smart-iLAB has been previously developed to showcase advances in electronics laboratory instruction, and is equipped with IoT-based sensing and actuation features. However, the Smart-iLAB did not offer the infrastructure required for users to conduct experiments and tests using the available sensors and actuators. In this work, we established a REST-based API testbed for the Smart-iLAB's IoT platform leveraging on MQTT connections, python scripts, a SQL-based database, and a RESTful API. In addition, a digital twin has been developed to provide immediate visual feedback of the Smart I-LAB environment. The REST-based API testbed, with its accompanying digital twin, is hosted on a web-based platform that enables users to interact with the sensors and actuators of the Smart-iLAB externally. The REST-based API testbed can also be used to facilitate the development of scripts, from simple manipulation to complex automation. In addition, the visualization of the layout of the available sensors and actuators serves as a convenient way for users to maximize the capabilities of the smart system remotely. To illustrate its effectiveness, the REST-based API testbed was tested to assess its latency, scalability, and reliability.
3:30 Performance of Slotted ALOHA Systems with Successive Interference Cancellation and Feedback over Nakagami-m Fading Channels
Daiki Fukui, Yuhei Takahashi, Ryo Ozaki, Tomotaka Kimura and Jun Cheng (Doshisha University, Japan)
This paper explores the optimization of transmission probability and code rate in multi-device slotted ALOHA systems employing successive interference cancellation (SIC) and feedback over Nakagami-m fading channels. Although previous research has derived the optimal probability and the code rate to maximize the sum rate in a two-device scenario analytically using Markov process, its extension to configurations with more than two devices remains challenging. The complexity of Markov modeling arises from two main challenges: 1) an increase in the number of devices greatly expands the state space, precluding the analytical determination of system throughput; 2) deriving transition probabilities in the Nakagami-m channel model is difficult. In this study, we use computer simulations to evaluate the sum rate and to search for the optimal transmission probability and code rate to maximize the sum rate of the systems. Our simulation findings reveal that at an average SNR the optimal code rate is independent of the number of devices and transmission probability. In a 30-device slotted ALOHA system with an SNR of 15 dB, the maximum sum rate of 2.9235 is almost achieved at a transmission probability of 0.0665 and an optimal code rate of approximately 4.02, regardless of the number of devices.
3:45 Enhancing Virtualization Security Through System Call-Based Anomaly Detection in Containers
Jie Zhang, Kuan-Chieh Wang, Po-Kai Hsu, Jhen-Jie Hsieh, Po-Shen Chen, Tze-Rong Jian, Kun-Hsiang Huang and Min-Te Sun (National Central University, Taiwan); Chun-Ying Huang (National Yang Ming Chiao Tung University, Taiwan)
In the current era of micro-services, containerized applications face unprecedented security challenges due to shared kernels and limited isolation. This research proposes a container security framework based on monitoring system call sequences to detect anomalies in containers providing various micro-services. We introduce a custom dataset named XXXX, which captures system call sequences behavior in containers running the micro-services and simulating attacks. The framework includes real-time system call monitors, parsers, dashboards, and an unsupervised anomaly detection model using unsupervised learning with autoencoders to enhance the detection capability of unknown vulnerabilities. It leverages containerization benefits - simplicity, scalability, and automation. Our evaluation emphasizes false alarm rate and average detection time. Results show that the attack detection performance of most containers meets expectations, though the detection time of one subset had slightly longer detection time due to the intrinsic complexity of vulnerabilities. This work offers valuable insights for improving container security in micro-service systems.
Computing & Computational Intelligence (CCI) 2
Track A2F6 CCI 2.2: Computing & Computational Intelligence (CCI) 2.2
Room: F6. 505 Sepilok (Level 5)
Chair: Siti Noraini Sulaiman (Universiti Teknologi MARA, Malaysia)
2:30 AI-Based Modeling of a Microwave Elliptical Sensor for Dielectric Measurement of Agricultural Samples
Kim Yee Lee, Yong-Hong Lee, Gobi Vetharatnam and Eng Hock Lim (Universiti Tunku Abdul Rahman, Malaysia); Cheng Ee Meng (Universiti Malaysia Perlis, Malaysia); Kok Yeow You (Universiti Teknologi Malaysia, Malaysia)
This study presents an AI-driven modeling approach for a microwave elliptical sensor used in dielectric measurements. Traditional analytical and numerical methods often face challenges such as high computational complexity, calibration difficulties, and reduced accuracy at higher frequencies, particularly with irregular sensor geometries like elliptical probes. To address these limitations, models were trained using reflection coefficient (S11) data from simulations across 0.5-8 GHz. Three techniques were evaluated: Multiple Linear Regression (MLR), Random Forest (RF), and artificial neural networks using Levenberg-Marquardt (ANN-LM). Performance analysis involved varying dataset sizes and input feature configurations (2-input and 3-input). The trained models were then tested on real agricultural samples and compared with results from a standard open-ended coaxial measurement. The findings show that the proposed AI models accurately predict the complex permittivity of samples under test, enhancing modeling efficiency and adaptability, especially for sensors with non-standard geometries. This capability supports real-time dielectric characterization in agricultural applications.
2:45 A Transfer Learning Based Decision Level Multimodal Framework for Continuous Sign Language Recognition
Navneet Nayan (Department of Computational Intelligence, SRM Institute of Science and Technology, KTR Campus, Chennai, Tamil Nadu, Indi); Debashis Ghosh (Indian Institute of Technology (IIT) Roorkee, India); Pyari Mohan Pradhan (IIT Roorkee, India)
Multimodal frameworks have appeared as a potential solution to achieve breakthrough results in the field of sign language and hand gesture recognition. In this paper, we propose a classifier combination based multimodal framework for continuous sign language recognition. In this work, we propose to use transfer learning to perform classification task on individual modalities. Further, we apply majority voting scheme to combine all the individual classification performances to obtain the final classification accuracy. For transfer learning, pre-trained deep neural networks like GoogleNet, MobileNet-v2 and EfficientNet-b0 are used independently from one another. We applied these networks independently on individual modalities of IPN Hand dataset and combined the classification results obtained on these modalities to obtain the final classification result. IPN Hand dataset is one of the most challenging and dynamic continuous hand gesture datasets. Among the used pre-trained networks, EfficientNet-b0 appeared as the best network in terms of accuracy whereas GoogleNet took the least computational time during classification. On this dataset, our proposed approach performs exceptionally well with individual modalities. After combining the classification results of individual modalities according to our employed algorithm, it is observed that our proposed approach performs better than the earlier reported results. Our method surpasses the reported benchmark performance as well as performs superior to many state-of-the-art results.
3:00 Multimodal Approach Based Sentence Level Sign Language Synthesis
Navneet Nayan (Department of Computational Intelligence, SRM Institute of Science and Technology, KTR Campus, Chennai, Tamil Nadu, Indi); Debashis Ghosh (Indian Institute of Technology (IIT) Roorkee, India); Pyari Mohan Pradhan (IIT Roorkee, India)
In this paper, we present a multimodal sentence level sign language synthesis system. Sentence level sign language synthesis requires synthesis of signs as well as synthesis of transition segments between the signs. In this paper, our focus is to develop. an efficient transition segment between two signs. For this, we propose to use edge features and trajectory features obtained from the transition segment of the query sign sentences. Edge features are extracted using the morphological operations on the hand images, whereas trajectory features are obtained from centroid detection of consecutive hand images. From trajectory, we obtain the direction of motion and shape and co-ordinates of the trajectory. Further, with the help of interpolation techniques we synthesize the hand gestures between the two signs. The correctness of the synthesis is analyzed and modified with the help of edge features and trajectory features. Our proposed approach is tested on some Indian Sign Language phrases and sentences. The proposed method is evaluated based on the mean opinion scores obtained from several users. Evaluation was based on three criteria, namely clarity in understanding, smoothness of the generated videos and similarity to the original sign videos. We obtained decent mean scores and encouraging feedbacks from the users.
3:15 Res-SH: Unbiased Residual Learning for Self-Healing Interface Toughness Prediction with Limited Data
Pei Sze Tan (Monash University, Malaysia Campus, Malaysia); Karen Koh (Monash University, Australia); Sailaja Rajanala, Arghya Pal and Raphael C.-W. Phan (Monash University, Malaysia); Ze Nan (Xidian University, China); Chee-Ming Ting (Monash University, Malaysia); Fuad Noman (Monash University Malaysia, Australia); Norfadilah Dolmat, Nik Nur Wahidah Nik Hashim and Afidalina Tumian (Petronas Research Sdn Bhd, Malaysia); Nik Nur Wahidah Nik Hashim (International Islamic University Malaysia, Malaysia)
The development of self-healing materials is often hindered by the high costs and material waste associated with traditional characterization methods. Current approaches to toughness prediction, primarily based on convolutional neural networks (CNNs), are limited by their tendency to capture only surface-level features, which can lead to biased predictions. Moreover, working with small datasets, which is common in materials science, further increases the risk of biased training due to overfitting, posing a critical challenge to the reliability and generalizability of predictive models. This study introduces an unbiased residual learning framework designed explicitly for predicting self-healing interface toughness under limited-data conditions. Our approach, ResNet-inspired approach for predicting self-healing material toughness, named Res-SH, used the power of residual networks to capture deeper, more complex patterns in the data, thereby addressing critical challenges in materials research. Res-SH minimises resource consumption and experimental overhead by focusing on unbiased learning, achieving accurate predictions with fewer training epochs and lower R^2 score and root mean square prediction errors compared to conventional CNN and lightweight model MobileNetv2. This novel framework provides a cost-effective and resource-efficient alternative to traditional material characterization methods, reducing material waste and accelerating the discovery and optimization of self-healing material systems.
3:30 Enhanced CNN Models for Accurate Classification of Calcification Patches in Breast Cancer Detection Using DBT Images
Syafiqah Aqilah Saifudin (Universiti Teknologi MARA Cawangan Pulau Pinang, Malaysia); Siti Noraini Sulaiman (Universiti Teknologi MARA, Malaysia); Muhammad Khusairi Osman (Universiti Teknologi Mara (UiTM), Malaysia); Iza Sazanita Isa (Universiti Teknologi Mara, Malaysia); Mohd Firdaus Abdullah (Universiti Teknologi MARA, Cawangan Pulau Pinang, Malaysia); Noor Khairiah A. Karim and Nor Ashidi Mat Isa (Universiti Sains Malaysia, Malaysia); Slamet Riyadi and Yessi Jusman (Universitas Muhammadiyah Yogyakarta, Indonesia)
Breast cancer remains one of the most common and serious health concerns worldwide, particularly among women. Early detection substantially decreases death rates, which has fuelled the development of deep learning-based medical imaging analysis. This research focuses on recognizing calcification patches in Digital Breast Tomosynthesis (DBT) images using Convolutional Neural Network (CNN) architectures. To evaluate classification performance, several CNN architectures are compared, including ResNet-18, SqueezeNet, GoogleNet, as well as their modified versions. The enhanced models incorporate additional convolutional layers to improve feature extraction and classification of DBT patches. Specifically, the Modified ResNet-18, Modified SqueezeNet and Modified GoogleNet represent enhanced versions of the original architecture, with integrated designs to capture multi-scale features specific to DBT images and improve sensitivity to subtle calcifications. Experimental results show that the modified GoogleNet outperforms the conventional and other experimented networks, with accuracy, sensitivity, precision, F-Measure, and Jaccard Index values of 95.91%, 92.63%, 90.79%, 91.29%, and 0.844, respectively. Although numerically modest, these gains are clinically meaningful, as even small increases in sensitivity can translate into earlier detection of additional cases across large screening populations. This work highlights the potential of tailored CNN architectures to improve diagnostic accuracy in DBT imaging and support radiologists in early breast cancer detection.
3:45 Analysis of a Regression-Based Feature Set for Classifying Lung Lesion and Non-Lesion Regions in CT Scans
Nurul Najiha Jafery ('None', Malaysia); Siti Noraini Sulaiman (Universiti Teknologi MARA, Malaysia); Muhammad Khusairi Osman (Universiti Teknologi Mara (UiTM), Malaysia); Noor Khairiah A. Karim (Universiti Sains Malaysia, Malaysia); Zainal Hisham Che Soh, Ir (Universiti Teknologi MARA, Malaysia); Iza Sazanita Isa (Universiti Teknologi Mara, Malaysia)
Lung lesion classification in CT scans is critical for the early diagnosis of lung cancer and other pulmonary diseases. Accurate distinction between lung lesion and non-lesion regions can significantly support clinical decision-making. Existing feature extraction methods often rely on conventional image characteristics, which may not fully capture subtle shape variations or inter-slice dynamics. This study introduces a regression-based feature extraction approach designed to enhance lesion classification in axial CT images. Geometrical characteristics across consecutive slices were modelled as signal-like inputs for a hybrid deep learning framework, enabling the capture of both absolute values and inter-slice variations. Four regression feature configurations (RFE_1 to RFE_4) were systematically evaluated, incorporating combinations of roundness, diameter, centroid coordinates, area, and perimeter. Results showed that RFE_2, comprising roundness, diameter, and centroid coordinates, achieved the highest classification accuracy of 96%, outperforming other configurations. In contrast, including area and perimeter reduced accuracy, highlighting the negative impact of redundant features. These findings demonstrate that careful optimisation of regression-based features can improve the robustness and reliability of AI-driven lung lesion detection systems.
Engineering Technologies & Society (ETS)
Track A2F7 ETS 2: Engineering Technologies & Society (ETS) 2
Room: F7. 506 Selingan (Level 5)
Chair: Yan Yan Farm (Universiti Malaysia Sabah, Malaysia)
2:30 A Graph Based Attention Model and Calibrated Random Forest for Breast Cancer Classification Using Histopathology Images
Dipti Deb (National Institute of Technology Rourkela, India); Ratnakar Dash (NIT Rourkela, India); Durga Mohapatra (NIT, Rourkela, India)
Artificial intelligence and computer vision advancements have revolutionized computer-aided diagnosis (CAD) systems, enabling more accurate breast cancer (BrCan) detection using histopathology images. This study proposes a classification framework that integrates Vision Transformers (ViT), Graph Attention Networks (GAT), and Calibrated Random Forest (CRF) to enhance diagnostic accuracy. ViT effectively captures rich visual representations and helps to form a graph-like structure, while GAT models the structural relationships within histopathology images, providing a more comprehensive understanding of tissue morphology. Extensive experiments were conducted with various model combinations, demonstrating that the ViT + GAT + CRF architecture achieved the highest performance. The experiment is carried out on the BreakHis dataset, and the model acquires an accuracy of 97.33%. These results highlight the effectiveness of incorporating both visual and structural features to improve diagnostic reliability. Our proposed framework represents a significant advancement in digital histopathology-based (BrCan) diagnosis and holds promise for broader applications in medical imaging.
2:45 Development of a Smart Financial Tool for Computing High-Yield Savings in Digital Banks to Advance Financial Literacy Through a Blended Agile Methodology
John Heland Jasper Ortega (FEU Institute of Technology, Philippines)
This study developed the Digital Banks PH Notebook, a mobile application designed to support financial literacy among Filipinos by optimizing savings through high-yield digital banking platforms. The application featured a savings portfolio tracker, interest forecasting calculator, and savings goal management to address gaps in financial planning and savings behavior. Development followed a blended Agile methodology integrating Scrum, Extreme Programming, and Feature-Driven Development, ensuring iterative improvements aligned with user needs. Software quality was assessed using the ISO/IEC 25010 model, while qualitative feedback was analyzed through word cloud visualization to capture user sentiment and key focus areas. Findings indicated that the application effectively enhanced users' understanding of savings strategies and promoted responsible saving practices. The tool successfully connected the opportunities presented by digital banking with the practical requirements of financial education, providing users with actionable insights to manage their savings more strategically. By leveraging agile development practices and rigorous evaluation frameworks, the project demonstrated that technology-driven solutions can play a significant role in advancing financial literacy and supporting sustainable financial behaviors in an evolving digital economy.
3:00 Inculcating Soft Skills in Requirements Elicitation: a Dynamic Role-Play Approach
Chien Ching Lee (SIT, Singapore); Ryan Fraser Kirwan (Singapore Institute of Technology, Singapore)
Requirements elicitation (RE) is a critical yet often underemphasized component of software engineering education. This study investigates the effectiveness of a dedicated workshop designed to introduce undergraduate computing students to RE through experiential role-play with instructors simulating dynamic client interactions. A structured four-step elicitation framework, grounded in industry practices, guided students in preparation, question formulation, documentation, and confirmation. Post-workshop surveys and tutor feedback indicated that while students improved in areas such as question scoping and clarity, many continued to face challenges in rapport building and adaptive communication-aligning with prior research on the importance of interpersonal skills in RE. Students valued the workshop's practical focus but expressed a strong desire for additional practice, particularly in probing and clarifying requirements in dynamic client environments. Findings suggest that controlling the complexity of sample systems allowed students to concentrate more effectively on interview techniques. Only a minority preferred chatbot-based elicitation, citing concerns about authenticity and emotional nuance. Consolidated tutor feedback confirmed frequent student mistakes in sequencing and probing, echoing established literature. Although the small sample size limits generalizability, the study highlights the importance of structured practice, feedback, and reflection in developing RE competencies. Future research should explore targeted interventions to strengthen rapport-building and questioning strategies. Overall, the workshop effectively raised students' awareness of the dynamic, human-centered nature of requirements elicitation and underscored the need for continuous skill development in this area.
3:15 Honing Internship Students' Social-Emotional Competence: Unpacking and Mitigating Challenges
Chien Ching Lee (SIT, Singapore); Ryan Fraser Kirwan (Singapore Institute of Technology, Singapore)
Technological innovations do not just happen. They are spurred by teams with strong technical skills and effective social-emotional competence. This paper presents an interpersonal skills training workshop conducted three months into students' internships which focused on improving students' social-emotional competence specifically, using Goleman's' Emotional Intelligence model to regulate their emotions and that of others, The Johari Window to raise their awareness about their level of openness and the Thomas Kilmann conflict management model to help them read situations more accurately. Data was gathered via pre- and post-surveys and a Communication Challenges Scenario worksheet where they reflected on communication incidents during their internships. The findings from the pre-survey showed that the more severe communication issues were in adapting to professional communication styles, communicating with international colleagues, and understanding technical jargon. The improvements from the workshop as reflected in the post-survey were mainly in students adopting a more proactive communication stance in task execution. The top three improvements were in practicing active listening, navigating cultural and hierarchical dynamics and providing progress updates. The students mentioned that after the workshop, they were more confident in identifying and regulating their emotions and responding to others with empathy and clarifying tasks. They needed more support however, in managing conflicts.
3:30 Study on Quality Evaluation Models for Collaborative Architectures
Yo Suzuki, Kei Sugawara and Sumie Morita (Akita Prefectural University, Japan)
This paper proposes a hierarchical decision-making-based architecture evaluation model (H-DRC) designed to quantitatively assess the quality of collaborative architectures in complex systems. The model introduces a structured methodology for evaluating architectural characteristics by applying weighted metrics derived from prior studies, which are mapped to ISO/IEC 25010 quality attributes such as reliability, maintainability, and performance efficiency. By integrating these metrics, the H-DRC model enables a comprehensive visualization of architectural suitability and facilitates effective evaluation of system design decisions. To validate the proposed model, we applied it to an in-house system, the BabaCAFE System, which incorporates a Context Broker also implemented in FIWARE, an open-source platform framework widely recognized as an "urban operating system" for smart cities and industrial IoT solutions. The evaluation demonstrates that the H-DRC model can accurately capture architectural strengths and weaknesses, even in small-scale systems, and confirms its potential applicability to large-scale industrial domains such as supply chains, smart cities, and healthcare systems. This work contributes a practical and generalizable evaluation framework that bridges academic research and real-world implementation scenarios.
3:45 Automated KOJI AWARENESS Physical Function Assessment Using Pose Estimation and Web-Based Real-Time Evaluation
Shura Osako (Fukuoka Institute of Technology, Japan & FIT, Japan); Hiroyuki Fujioka (Fukuoka Institute of Technology, Japan)
This paper presents a web-based system for automating a subset of the KOJI AWARENESS physical function test using pose estimation. The KOJI AWARENESS test consists of 50 items designed to evaluate mobility, flexibility, and posture. In this study, five items focusing on neck and shoulder mobility were selected for implementation. The system uses YOLOv8 and MediaPipe to detect body and hand keypoints, providing real-time evaluation through a browser-based interface. Built with React and Flask, the system requires no special software installation, enabling deployment in both clinical and non-clinical environments. Initial testing with six participants confirmed that the system could accurately detect postures and evaluate movements in real time. These results demonstrate the feasibility of using pose estimation techniques for automated physical assessments. Furthermore, this study suggests that such systems may serve as a foundation for accessible and scalable approaches to physical function evaluation, with potential applications in preventive healthcare, remote rehabilitation, and physical fitness monitoring.
Computing & Computational Intelligence (CCI) 3
Track A2F8 CCI 2.3: Computing & Computational Intelligence (CCI) 2.3
Room: F8. 507 Monsopiad (Level 5)
Chair: Jay Dave (BITS Pilani Hyderabad Campus, India)
2:30 CardioScan: a Multimodal Approach for Congenital Heart Disease Diagnosis Using PCG Signals
Aditya Svs (IIIT Naya Raipur, India); Sai Sriram Gonthina (International Institute of Information Technology, Naya Raipur, India); Debanjan Das (IIT Kharagpur, India); Rajarshi Mahapatra (IIIT Naya Raipur, India)
Congenital heart disease (CHD) is a leading cause of morbidity and mortality in infants and early diagnosis with access to specific diagnostic tools is often limited and the interpretation of heart sounds is subjective. Traditional auscultation relies heavily on clinician skill and is highly variable, while advanced imaging techniques can be expensive, or even impossible in low-resource settings. In this study, we present a multimodal machine-learning framework that combines demographic data and 66 hand-crafted features from denoised and resampled phonocardiogram (PCG) recordings. Signal preprocessing included a third-order Butterworth bandpass filter set between 65-1000 Hz and resampling to uniform time series recordings of 2000Hz to preserve signal integrity and reduce noise. For classification we utilize LightGBM achieving binary task accuracy of 94% and multi-class murmur classification accuracy of 87%. The entire system can be deployed to a Raspberry Pi 4 connected to a digital stethoscope and able to conduct real-time inference as an embedded system without cloud or internet access. Experimental results verified high accuracy and low latency operational characteristics making it well suited for an embedded deployment with such limited computation resources. The proposed end-to-end framework provides an affordable, portable, and clinically useful tool to assist in the early detection of abnormal heart sounds and has the potential to revolutionize CHD screening in resource challenged communities.
2:45 Dual-Stage Feature Refinement and Wavelet Denoising for Enhanced VIX Prediction Using Residual BiLSTM
Akanksha Sharma (Maulana Azad National Institute of Technology, Bhopal, India); Priya Singh (Vellore Institute of Technology, Tamil Nadu, India); Chandan Kumar Verma (Maulana Azad National Institute of Technology, Bhopal, India)
Derivative pricing and financial risk management rely heavily on volatility predictions. Given the dynamic and nonlinear nature of financial markets, accurately predicting volatility remains a persistent challenge. Traditional econometric models often struggle to capture the complex patterns and time-dependent behaviors present in market data. This research presents a Residual Bidirectional Long Short-Term Memory (ResBiLSTM) model-based deep learning framework for VIX price prediction. By integrating residual connections and bidirectional temporal processing, the model is able to successfully capture intricate patterns found in financial time series data. A complete array of 64 designed features, comprising technical indicators and wavelet-denoised inputs, was employed to train and assess the model. The suggested ResBiLSTM surpasses conventional models, including LSTM, GRU, CNN, BiLSTM, and Residual LSTM, across multiple criteria. The performance was additionally confirmed using 5-fold cross-validation and statistical significance assessment via paired t-tests across many experimental iterations. The findings illustrate the model's resilience, precision, and applicability for implementation in practical volatility forecasting scenarios.
3:00 A Predictive Approach to Energy Loss in Solar Panels Affected by Soiling
Yi Feng Law (Universiti Tenaga Nasional (UNITEN), Malaysia); Faridah Hani Mohamed Salleh (University of Tenaga Nasional, Malaysia); Murthy Parventanis (Universiti Tenaga Nasional, Malaysia)
The use of solar energy continues to grow, tackling challenges such as soiling is vital to sustain efficiency. Soiling, the accumulation of dirt, dust, and pollutants on solar panels, significantly reduces the sunlight reaching photovoltaic cells, thereby diminishing energy output. This study examines the key variables influencing energy loss from soiling in Sepang, Selangor, Malaysia, with a focus on irradiance and rainfall in a tropical climate. By using linear regression, decision tree, and the random forest models, we predict energy loss from soiling and validate the model accuracy with R-Squared, Mean Squared Error, and Root Mean Squared Error. The results show that the random forest model, driven solely by irradiance data, provides the most accurate predictions. Including rainfall did not enhance predictive accuracy, underscoring irradiance as the primary factor. This research highlights the impact of soiling in tropical climates, offering a valuable insight to improve solar panel maintenance and to optimize energy production.
3:15 Sensor-Based Gait Recognition Using Ensemble Network Unified with Independent Subnetworks
Sonia Das (National Institute of Technology, Rourkela, India); Pradosh Ranjan Sahoo (VIT-AP University, India)
Smartphone-based gait recognition is increasingly important for surveillance, security, and health monitoring. However, traditional deep learning methods often struggle to model long-term dependencies in gait sequences and suffer from the computational demands of large convolutional kernels. Although multi-kernel approaches attempt to address these challenges, their fixed sizes may fail to capture relevant variations in dynamic signals and are often difficult to train efficiently. This paper proposes a multi-scale deep ensemble network that overcomes these limitations by using independent subnetworks, each processing different temporal resolutions of down-sampled input signals. These subnetworks extract diverse and complementary gait features. A unified ensemble training strategy integrates the outputs using multiple loss functions, enhancing the model's robustness and generalization. Extensive experiments on two benchmark datasets demonstrate that our method captures both spatial and temporal complexities more effectively than existing approaches. The complementary learning achieved through multi-scale ensemble modeling leads to superior performance, setting a new standard in smartphone-based gait recognition.