Optimised decision model for vulnerable component classification in a vehicle health monitoring system using hybrid deep learning approach / Md Naeem Hossain

Faculty of Mechanical and Automotive Engineering Technology

Thesis (Master of Science) -- Universiti Malaysia Pahang – 2025

Includes bibliographical references

Automobiles play a crucial role in modern life, yet issues such as unexpected breakdowns, high maintenance costs, and environmental concerns persist due to overlooked irregularities in vulnerable components (VCs) and associated influential measures. Existing vehicle health monitoring systems (VHMS) lack the reliability required for real-time diagnostics, failing to comprehensively address component vulnerabilities and process vast data generated by modern vehicles. Traditional AI-based approaches struggle to capture spatial and temporal patterns, limiting predictive maintenance and necessitating more advanced solutions. This study proposes an AIdriven VHMS using a hybrid deep learning (DL) framework to improve fault detection and predictive maintenance of the vehicle. A vulnerability identification framework was devised and applied to determine VCs across five vehicle subsystems, and the Decision Making Trial and Evaluation Laboratory (DEMATEL) method identified the key influential measures affecting vehicle health. A decision model was developed using an infrastructure vulnerability assessment approach, which provides severity values to categorise vehicle health conditions into four distinct classes: Good, Minor, Moderate, and Critical. Convolutional neural network (CNN)- long short-term memory (LSTM) hybrid DL algorithm was employed to evaluate the performance of the VHMS decision model, and compared against multiple machine learning (ML) and DL techniques. Nine VCs and ten influential measures were identified, with damage influence (0.32537), maintenance influence (0.31081), and age influence (0.30566) emerging as the most critical due to their high weight values. Afterward, a numerical dataset was prepared using the decision model comprising 4870 severity values for each VC. The CNN-LSTM model achieved a 99.5% test accuracy, among other parameters using 80% training, 10% validation, and 10% test dataset, significantly outperforming ML models (85.5% average accuracy) and other DL techniques (94.8%). By ensuring timely maintenance and reducing vehicle breakdowns, the research supports the broader goals of smart cities by improving the safety, efficiency, and sustainability of urban transportation systems. This study presents a transformative approach to vehicle diagnostics, offering stakeholders including manufacturers, fleet operators, and individual vehicle owners - reliable, timely, and data-driven insights for proactive maintenance. The proposed model aims to enhance condition monitoring for autonomous, and electric vehicles, providing a user-friendly interface for vehicle owners and maintenance personnel to provide clear, actionable insights, and maintenance recommendations.