Driver fatigue classification based on electroencephalogram signal / Rafiuddin Bin Abdubrani

Faculty of Electrical & Electronics Engineering Technology

Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang - 2026

Include bibliographical reference

Driver fatigue is a critical risk factor for road accidents, underscoring the need for accurate and reliable detection systems. This research proposes an advanced EEG-based driver fatigue classification framework that integrates enhanced signal pre-processing with a hybrid deep learning architecture. The novelty of this research lies in the integration of an ICA–CWT pre-processing pipeline with enhanced modified z-score normalisation to produce high-fidelity EEG features, combined with a hybrid DBN–DQN architecture that significantly improves fatigue detection performance under crosssubject variability. The methodology employs independent component analysis combined with continuous wavelet transform (ICA-CWT) using the Morlet mother wavelet to improve artefact removal and time-frequency resolution. An enhanced modified z-score (EMZS) normalisation technique is introduced to stabilise feature distributions by reducing the influence of outliers. Extracted features from delta, theta, and alpha sub-bands are processed by a hybrid model combining a deep belief network (DBN) and a deep Q-network (DQN). The DBN performs hierarchical feature learning, while the DQN enhances decision-making through reinforcement learning principles, enabling improved classification accuracy. Experimental evaluation was conducted using stratified 5-fold cross-validation on a publicly available EEG dataset, with performance compared against traditional deep learning baselines (CNN and RNN) under identical conditions. The proposed method addresses key limitations of CNN/RNN models, such as poor frequency resolution and high sensitivity to noise, resulting in lower robustness in EEG-based fatigue studies. The proposed DBN-DQN framework achieved an average accuracy of 98.21%, precision of 98.15%, sensitivity (recall) of 98.24%, specificity of 98.18%, and an F1-score of 98.19%. These results significantly outperform CNN and RNN baselines, demonstrating the superiority of the proposed hybrid approach in EEGbased fatigue detection. The findings validate that integrating ICA-CWT, EMZS normalisation, and a hybrid DBN-DQN architecture substantially enhances fatigue classification performance. Future research should focus on real-time implementation, multimodal integration with additional physiological signals, and deployment of lightweight models for embedded systems. Overall, the study strengthens the link between methodology, results, and application by showing that ICA–CWT improves feature clarity, EMZS enhances data robustness, and the hybrid DBN–DQN architecture delivers highly reliable classification suitable for practical deployment in future intelligent transportation systems.