Graduate Projects (Electrical and Computer Engineering)
Permanent URI for this collection
Browse
Recent Submissions
Item Utilizing transformer for emotional understanding on Chinese mental-health dataset(2025) Du, Mingyu; Dong, XiaodaiThe rapid development of large language models has demonstrated successful performance in various areas. In terms of mental health, large language models exhibit the capability to understand emotional feeling to some extent. However, research in the mental health field requires a broad range of interdisciplinary knowledge and is often constrained by limited resources. This project focuses on the analysis of sentiment in conversational texts using large language models and investigating the model performances. By comparing 8 different open source models, the project demonstrates the outstanding performance of hfl/chinese-roberta-wwm-ext in emotional understanding using the mental health dataset released by Tongji University.Item GenomicRL: A DRL framework for cancer treatment recommendation using genomic and metastatic markers(2025) Beg, Heebatullah; Yang, Hong-ChuanEven with significant advances in cancer treatment, Gastroesophageal Junction (GEJ) cancers continue to present therapeutic challenges with a 5-year survival rate of just 21%. This work develops GenomicRL, a deep reinforcement learning (DRL) framework that integrates genomic, metastatic, and clinical markers to optimize treatment recommendations. Initially, a supervised learning (SL) baseline using ElasticNet achieves 64% exact match ratio (EMR) with clinician decisions. Augmenting training with synthetic data improves EMR to 70%, demonstrating generated data’s utility in mitigating limited real-world samples. However, SL’s reliance on historical decisions neglects post-treatment outcomes. To address this, a novel outcome-driven DRL agent is trained. Although the approach, incorporating survival, metastasis, and genomic stability into its reward function, reduces EMR from 99% (for clinician-mimicking reward function) to 73%, it achieves a higher average reward. Incorporating post-treatment signals, however, leads the agent to deviate from historical choices in ways that improve long-term outcome metrics—trading some immediate agreement for better anticipated patient benefit. This shift from pure imitation to outcome-oriented optimization highlights the promise of data-driven recommendation strategies that leverage diverse clinical and molecular information. Importantly, the proposed framework is not designed to supplant medical professionals but to assist them in refining treatment planning through personalized insights that account for individual patient variability.Item Enhancing security and safety in quadcopter drones through unsupervised machine learning(2024) Zhu, Lei; Traoré, Issa; Saad, SherifThe increasing deployment of drones in various sectors has raised significant security concerns regarding their vulnerability to cyber-attacks and sensor data manipulation. Our project presents a comprehensive anomaly detection system designed to enhance drone security and safety through unsupervised machine learning. The implementation utilized a custom drone platform based on open-source components, enabling complete access to system internals and sensor data for security monitoring. Through systematic vulnerability analysis and testing, the project collected an extensive dataset comprising approximately 11.5 million network traffic samples and 1.25 million MAVLink messages, representing both normal operations and various network-based attacks and sensor anomaly scenarios. The detection system employs machine learning algorithms to identify anomalies in both network communications and sensor data streams, with the Isolation Forest algorithm demonstrating superior performance in testing. The implemented system successfully detected various security threats, including network-based attacks such as man-in-the-middle attacks, denial of service, and port scanning, as well as sensor anomalies including GPS spoofing and rangefinder data manipulation. The deployment on a Raspberry Pi companion computer demonstrated the system's practical viability for real-world applications. This research contributes to the field of drone security by providing a systematic approach to anomaly detection and a comprehensive dataset for future research.Item Security Analysis for Vehicle Area Networks Protocol Using AVISPA(2024) Alahmar, Alaa; Gebali, Fayez; Altawy, RihamIn the era of smart transportation, Vehicle Area Networks (VANs) are critical in enabling secure communication between vehicles and infrastructure. This project examines the security robustness of the PUFGuard protocol, a physically unclonable function (PUF)-based authentication framework designed to protect Vehicle-to-Infrastructure (V2I) and Vehicle-to-Vehicle (V2V) communications in VANs. PUFGuard leverages the inherent uniqueness of PUFs for secure key generation and authentication, aiming to establish trust and resilience against adversarial attacks in dynamic, multi-hop communication environments. To validate PUFGuard’s resilience, this research employs formal verification tools—AVISPA and SPAN—to simulate and analyze its effectiveness against common network threats, including replay attacks, manin-the-middle attacks, and impersonation attacks. The protocol is modelled in the High-Level Protocol Specification Language (HLPSL), where each component of the V2I and V2V authentication processes is systematically represented. Results from the AVISPA tests highlight the protocol’s strengths and potential vulnerabilities, providing insights into the adequacy of PUFGuard’s security measures in real-world VAN applications. The findings of this study suggest refinements to fortify PUFGuard further, offering a framework for secure, authenticated communication in modern vehicular networks.Item Modeling of Metal-Molecular Nanoelectronics Network: Analysis and Electrical Properties(2024) Ananna, Tabassum Perveen; Papadopoulos, ChristoComplementing electronic components with molecular counterparts offers a hopeful option for advancing beyond the current size limitations of traditional silicon electronic devices in the effort to create operational molecular nanoscale circuit components. Molecular modules have been extensively studied to evaluate their suitability for use in future nanoelectronic circuits. This study focuses on investigating the theoretical and experimental aspects of the electrical and electronic properties of metal-molecular networks bridged with dithiol molecules. The ratio of (di)thiol molecules and/or the type of molecules in the network can be adjusted to modify the electronic transport paths through the network. Furthermore, the electronic conductivity of small-scale networks made up of interconnected graphene clusters and thiolated molecules (benzene/alkanedithiol) in linear chains and extended networks is analyzed using simulations based on first-principles density functional theory. Geometry optimization and Energy Analysis using DMol3 that determines the electronic characteristics of molecules, surfaces, clusters, and crystalline solid materials through DFT were performed. The ability to adjust simulations by changing the molecule-to-nanoparticle ratio yields results that align well with the findings of the previously reported experiments. This offers valuable insights into manipulating network properties with various types of molecules. The analysis ended with VAMP Analysis on Carbon-based molecules such as benzene dithiol and graphene nanosheet. The outcome of experimental VAMP analysis presents a step-by-step process to work on carbon-based structures. The findings from these simulations are used to suggest molecular-level circuits for purposes like memory, switching, hardware security, and biosensors. The molecular electronic networks involving metal nanoparticles, as described in this study, offer a way to develop electronics at the molecular scale.Item EV charging station attack detection using machine learning(2024) Janwiri, Kamran Athar; Gebali, FayezElectric Vehicle (EV) charging stations are very important for supporting the adoption of EVs, but they are at risk of cyberattacks. This project looks at how Machine Learning (ML) can help to detect these attacks using the CICEVSE2024 dataset, which has data about normal operations and attack scenarios. ML models like Random Forest (RF), Gradient Boosting Machine (GBM), Support Vector Machines (SVM), Logistic Regression (LR), Naive Bayes (NB), and k-Nearest Neighbors (KNN) were tested. Principal Component Analysis (PCA) applied to simplify the dataset by minimizing the features, and SMOTE (Synthetic Minority Oversampling Technique) was used to balance the dataset. Models were evaluated with 21, 15, 10, and 5 features to find the best accuracy and speed. RF shows the best accuracy whereas KNN was the fastest.Item Audio analysis of customer calls for predicting purchase intentions: A novel approach to e-commerce insights(2024) Yu, Miao; Li, Kin FunClient audio recordings represent a valuable resource for many types of businesses. Utilizing these recordings to identify potential customers can help enhance purchase rates and reduce marketing costs, particularly with different kinds of machine learning methods that automatically label different groups, including positive, neutral, and negative buyers, instead of manual analysis. Though previous research has predominantly focused on text content analysis for this purpose, audio features, which effectively capture voice nuances such as tone, pitch, rhythm, and interaction patterns between interviewers and interviewees, may impact the model performance. This project explored an innovative method. It firstly investigates the effectiveness of emotion detection through audio features, leveraging two datasets: the Toronto Emotional Speech Set (TESS) and the Surrey Audio-Visual Expressed Emotion Dataset (SAVEE). Furthermore, hierarchical clustering techniques are applied to explore the relationship between emotion-related audio features and customer categories using audio data provided by VINN Auto, an e-commerce firm. Next, Exploratory Data Analysis (EDA) is conducted to find the correlation between interaction-related audio features and customer categories, including positive, neutral, and negative buyers within the same dataset after labeling it. Using supervised learning, the results indicate that integrating audio features, including emotion-related and interaction pattern features, can affect the performance of models like Support Vector Machines (SVM), Decision Tree, and Extreme Gradient Boosting (XGBoosts), particularly when combined with traditional audio content-related features such as Term Frequency-Inverse Document Frequency (TF-IDF) scores while applying adjusted weight configuration for positive class. After these exploration, an ensemble method using a soft voting mechanism across these three models is developed to assess whether it can enhance the identification of potential purchasers. The approach of combining emotion-related audio features, interaction pattern features, and content-based features like TF-IDF scores with tailored weight configurations highlights the value of collaborating audio features in customer identification tasks compared with only using content-based features like TF-IDF scores. It could be a robust strategy for improving classification outcomes for the relevant analysis in the future.Item Emotion detection with data fusion(2024) Khuzhaniyazova, Maida; Li, Kin FunThis report explores the performance of three machine learning models — SVM with SGDClassifier, Gradient Boosting, and XGBoost — in detecting emotions using data fusion techniques. Early Fusion was chosen for integrating features due to its simplicity and reliable performance. The study employs the MELD dataset, which combines text, audio, and visual data from over 1,300 dialogues and 13,000 utterances in the “Friends” TV show. This dataset provides a unique multimodal approach to understanding emotions in conversational contexts, making it ideal for emotion recognition tasks. Evaluation metrics for the models included accuracy, F1-score, precision, recall, and AUCROC, calculated over multiple training iterations. By comparing the performance of these models on a comprehensive, multimodal dataset, this study meets the growing demand for accurate emotion detection in conversational AI. XGBoost demonstrated high and consistent performance on the MELD dataset; however, its effectiveness may vary under different conditions or datasets. SVM with SGDClassifier achieved the widest accuracy range, though less stable on nuanced emotions. Gradient Boosting delivered consistently strong AUC-ROC values but required full retraining with each data update, affecting its adaptability. Overall, while XGBoost and SVM delivered good performance, their accuracy was subject to fluctuations across iterations. Gradient Boosting consistently showed strong AUC-ROC values, but its disadvantage is the need to completely retrain the model when new data is added, which reduces efficiency.Item What determine soccer players' wages: The case of the English Premier League (EPL)(2024) Kouh Bor, Mohammadamin; Baniasadi, AmiraliThis study explores the determinants of wage structures in the English Premier League (EPL), one of the world’s most financially influential soccer leagues. Examining both traditional performance metrics—such as goals scored, assists, and minutes played—and emerging non-performance factors, including nationality, marketability, and club revenue, this study seeks to identify how these variables affect wage disparities among EPL players. After data cleaning, utilizing a dataset of 3,522 player-seasons across 31 teams- that attended at least on EPL season- spanning seven EPL seasons, this research employs machine learning techniques to predict wages based on a comprehensive range of performance and demographic factors. The findings suggest that while performance metrics remain crucial, non-performance factors, particularly nationality and Age play an increasingly prominent role in wage determination, contributing to sustained wage inequalities. Additionally, advanced models indicate that club financial resources, player positional roles, and club’s broadcast income significantly influence wage variability. This analysis provides an understanding Towards the economic and social forces shaping player wages and underscores the need for a balanced approach to wage determination to support both financial sustainability and equity in professional soccer.Item Joint adaptive transmission and numerology selection for 5G NR PDSCH with DQN-based reinforcement learning solution(2024) San, Thet Naung; Yang, Hong-ChuanThe mission critical applications such as industrial automation, remote surgery and autonomous transportation systems demand low-latency, high-reliability communications service. As such, there is an urgent need to optimize transmission technologies in 5G New Radio (NR) to support Ultra-Reliable Low-Latency Communication (URLLC). This project introduces a joint adaptive transmission and numerology selection scheme for Physical Downlink Shared Channel (PDSCH) in 5G NR, targeting URLLC support. The transmission scheme selection problem is modeled as a Markov Decision Process (MDP). A Deep Q-Network (DQN) reinforcement learning agent is trained to dynamically adjust Modulation and Coding Scheme (MCS) and numerology based on real-time channel conditions and latency constraints. To evaluate the performance, we develop custom simulation environment by implementing PDSCH transmission model under frequency-selective fading channels, incorporating the Hybrid Automatic Repeat reQuest (HARQ) mechanism. The results demonstrate that the DQN agent effectively reduces transmission delays and improves reliability by optimizing transmission parameters. This approach enhances performance for 5G NR in URLLC support, achieving both higher reliability and lower latency than conventional adaptive transmission system.Item Implementing voice assistant for visually impaired using LLMs and Vision Language Models(2024) Jiang, Jinke; Yang, Hong-ChuanAs a result of population aging, the number of visually impaired people is growing. Unfortunately, there is limited accessibility measures to help improve the quality of life of these people. The recent technological development in Artificial Intelligence (AI), especially Large Language Models (LLMs), should offer effective and efficient solutions. Recognizing the limitation of existing products, we design and implement a user-friendly and privacy-safe voice assistant for visually impaired people. Using LLMs and Vision Language Models, the assistant can recognize and identify objects through low-latency speech-to-speech interactions. The assistant can be deployed on offline edge computing devices with camera/microphone/speaker, with easily extendable functionalities. In this report, we present the design, adopted technologies, and adjustment that we applied to arrive at the final implementation.Item A comparison of Long Short-Term Memory, Convolutional Neural Network, Transformer, and Mamba models for sentiment analysis(2024) Ruan, Hang; Gulliver, Thomas AaronSentiment analysis is a critical task in Natural Language Processing (NLP) that helps decode the emotions and opinions embedded in text. With applications spanning from market research and social media monitoring to political analysis and customer feedback evaluation, sentiment analysis provides invaluable insights into public opinion and consumer behavior. This project studies the evolution of sentiment analysis models, focusing on the advancements made by deep learning techniques such as Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNNs), and transformer-based models like Bidirectional Encoder Representations from Transformers (BERT) and Generative Pre-trained Transformer (GPT). These models have set new benchmarks for accuracy, efficiency, and versatility. Additionally, this explores Mamba, a recent State Space Model (SSM) designed to overcome the computational challenges of transformers in handling long sequences and demonstrates state-of-the-art performance on language modeling tasks comparable to transformers twice its size. This study examines the strengths and limitations of these models, comparing their performance on sentiment analysis datasets to provide a comprehensive understanding of their applicability and efficacy in various contexts.Item Design and development of a variable frequency drive test bench prototype and testing regime for repaired drives(2024) Kukreti, Akshat; Ilamparithi, T.Small and medium-sized industrial control panel makers, systems integrators, and repair shops frequently encounter issues with Variable Frequency Drive (VFD) testing. These companies frequently rely on third parties to evaluate their VFDs prior to field commissioning, increasing expenses. To address this issue, an in-house test bench was created for a control systems integrator's maintenance department (KJ Controls). The test technique includes static, functional, and operational testing of the drive under test. To completely verify the VFD's operational capability, it must be connected to a loaded motor. A load motor/dynamometer is required for this test. For this, a prototype was created in the repair department using readily available and stock components. This configuration facilitated the testing of the operation and functioning of a 2 HP drive. The tests run on the drive included a static test, a free run test, a motor stall test, and a speed regulation test. These were then suitably recorded on a factory acceptance test sheet. It was eventually calculated that having an in-house testing system would save the company at least 10% on the total cost of testing repaired VFDs. It was predicted that the payback period for such a test bench arrangement would be close to three years.Item Sentiment Analysis on IMDB Reviews leveraging Transfer Learning and Neural Network Models via Keras API(2024-09-13) Kharmandar, Armita; Amirali BaniasadiThis project focuses on sentiment analysis using a pre-labeled IMDB dataset containing 50,000 movie reviews. Our approach leverages transfer learning and a pre-trained neural network from TensorFlow Hub, specifically the NNLM model, to streamline the preprocessing stage. The sentiment classification model is built using the Keras API with a Sequential architecture. To design the model, several attempts have been made to improve the performance and overall accuracy and the finalized model includes first layer of an embedding layer that utilizes the pre-trained NNLM embeddings, followed by a three-layer network consisting of dense layer, dropout layer and another dense layer for classification respectively. To compile the model, Adam optimizer and binary cross-entropy loss are defined while achieving an accuracy of almost 88%. In addition to this deep learning model, traditional sentiment analysis algorithms such as logistic regression, random forest, and SVM were also trained for comparison. Each attempt’s output is visualized and finally in conclusion part it is discussed which of the models output the desired performance considering two factors of efficiency and accuracy. The results highlight that by incorporating transfer learning and Keras API, the overall model complexity and computational cost were significantly reduced while maintaining competitive accuracy.Item A simulation platform for connected autonomous vehicles incorporating physical and communication simulators(2024) Chen, Yuhao; Cai, LinThis project report provides a holistic record of the development of a connected autonomous vehicle simulation framework incorporating a physics simulator and a communication simulator. The development of this tool aims to help researchers in vehicle communication protocols to evaluate the simulated performance of their solutions in the physical world. By using this tool, communication researchers can observe the impact of their communication protocols on the actual connected autonomous vehicle operation process without the need to delve into the underlying logic of vehicle kinematic simulation. They only need to configure simple parameters and deploy their own protocols on the communication simulator and see the effect. This project report will start by introducing the components and operating principles of the entire system, and then demonstrate its usage through a simple simulation example.Item Dynamic object detection using moving cameras for UAV collision prevention(2024) Dadrass, Arman; Baniasadi, AmiraliIn this project, the focus is on object and motion detection in video streams from a camera mounted on a moving platform, such as a drone. Motion segmentation and moving object detection are fundamental for object tracking and collision avoidance in autonomous vehicles and flying drones. Moving object detection and tracking with a moving camera is challenging due to the combined effects of the camera’s motion relative to its mounting base and the movement of the platform within the environment. The limited field of view of the camera results in frequent object discontinuity and severe background variations, which conventional background subtraction approaches for fixed cameras cannot handle. To address these challenges, dense optical flow clustering is used to detect moving objects with a moving camera. The clusters correspond one-to-one with moving objects and background motion due to camera's motion; however, objects frequently enter or leave the scene, necessitating frequent redefinition and recalculation of clusters. Additionally, since the number of objects in the scene is unknown and can vary over time, the clustering algorithm must adapt quickly to changing scenarios. Therefore, the Adaptive Resonance Theory-2 (ART2) network was adapted to eliminate the need for pre-tuning the number of clusters as a hyper-parameter, automating the process similarly to human perception and enabling rapid redefinition and recalculation of varying cluster numbers. The performance of the proposed approach was evaluated in terms of execution time and accuracy using the VisDrone and KITTI datasets, and the results are discussed.Item IoT-enabled smart shopping carts: Automating checkout processes for frictionless shopping(2024) Chander, Kiran; Thirumarai Chelvan, IlamparithiThe growing crowds at shopping destinations like malls, grocery stores, supermarkets, and discount stores highlight the necessity of refining billing protocols to handle the increased number of shoppers. Moreover, considering the decline in sales experienced by physical stores like grocery stores in the face of growing online shopping trends, it becomes imperative to address this challenge and devise viable solutions. The Smart Shopping Cart project aims to enhance the retail shopping experience by leveraging the Internet of Things (IoT) technology to create a smart device that can be attached to standard shopping carts. By incorporating features such as RFID capabilities, MySQL database connectivity, and the Raspberry Pi Pico W microcontroller, the system provides seamless integration with user-friendly interfaces, allowing customers to interact with the shopping cart and scan products themselves effortlessly. This approach expedites the checkout process, reduces wait times, and improves operational efficiency for stores, while also offering valuable insights into consumer purchase patterns for inventory control and targeted marketing. By eliminating the need for manual scanning at checkout, customers save time and enjoy added convenience, thereby enhancing the overall shopping experience. Ultimately, Smart Carts offer an innovative application of IoT technology to streamline and modernize retail shopping, making it more efficient, convenient, and enjoyable for both customers and businesses alike.Item Automating warehouse inventory management(2024) Hajibabaei, Neda; Baniasadi, AmiraliEfficient inventory management is crucial for the smooth operation of warehouses in large retail and chain stores, where traditional methods of manual barcode scanning are often labor-intensive and prone to errors. This project addresses these challenges by developing an automated system that utilizes QR codes and computer vision techniques for inventory tracking. By implementing OpenCV to detect QR codes from images captured within Walmart’s warehouse environment, the system processes and categorizes the data to identify warehouse sections and product details. Comprehensive reports are generated to facilitate accurate inventory management, and visualizations are created to provide clear insights into inventory distribution. This approach significantly enhances the accuracy and efficiency of inventory processes, reducing labor costs and improving decision-making and resource allocation.Item Enhancing image classification accuracy using convolutional neural network on CIFAR-10 dataset(2024) Ghafouri, Sara; Baniasadi, AmiraliThis project explores the application of Convolutional Neural Networks (CNNs) for image classification on the CIFAR-10 dataset, a widely recognized benchmark in computer vision. The CIFAR-10 dataset comprises 60,000 32x32 color images across 10 distinct classes. This study aims to build and optimize a deep learning model to achieve high classification accuracy on this dataset. A CNN with multiple convolutional, pooling, and dropout layers was implemented, enhanced with batch normalization to prevent overfitting. Data augmentation techniques were applied to improve the model's generalization capabilities. The model was trained using the Adam optimizer, with callbacks for learning rate reduction and early stopping to fine-tune the training process.[1] The results demonstrate the effectiveness of deep learning techniques in achieving substantial performance on the CIFAR-10 dataset, with detailed analysis of training and validation metrics. The model was evaluated on selected data from the CIFAR-10 dataset, showcasing its predictive capabilities. The findings provide insights into the design and optimization of CNNs for practical image classification tasks.Item Automated title generation for research papers: A transformer-based approach(2024) Hosseinzadeh, Maryam; Baniasadi, AmiraliThe task of generating concise, informative, and relevant titles for research papers is a critical but challenging aspect of academic publishing. This project report explores the application of Transformer-based Large Language Models (LLMs) to automatically generate and suggest research paper titles from their abstracts. Building on the foundational Transformer architecture introduced by Vaswani et al., we evaluate the performance of different models, including a custom-built LLM, a pre-trained GPT-2 model, and a fine-tuned version of GPT-2. Through qualitative analysis, we demonstrate that fine-tuning GPT-2 on a specific dataset of research paper abstracts and titles significantly enhances the coherence, relevance, and contextual accuracy of the generated titles. We address challenges such as hallucinations in LLM-generated text and discuss the importance of high-quality datasets and task-specific fine-tuning. This work contributes to the broader understanding of the capabilities and limitations of LLMs in specialized NLP tasks, offering insights for future research and applications in academic publishing.