Electroencephalogram-Based Multi-Class Driver Fatigue Detection using Power Spectral Density and Lightweight Convolutional Neural Networks

Driver fatigue is the primary factor contributing to traffic accidents globally. To address this challenge, the electroencephalogram (EEG) has been proven reliable for assessing sleepiness, fatigue, and performance levels. Although alertness monitoring through EEG analysis has shown progress, its use is affected by complicated methods of collecting data and labelling more than two classes. Based on previous research, the original form of EEG signals or power spectral density (PSD) has been extensively applied to detect driver fatigue. This method needs a large, deep neural network to produce valuable features, requiring significant computational training resources. More observations regarding feature extraction and classification models are needed to reduce computational cost and optimize accuracy values. Therefore, this research aimed to propose a PSD-based feature optimization on a lightweight convolutional neural network (CNN) model. Five types of statistical functions and four types of signal power ratios were applied, and the best features were selected based on ranking algorithms. The results showed that feature optimization using the Relief Feature (ReliefF) algorithm had the highest accuracy. The proposed lightweight CNN model obtained an average intra-subject accuracy of 71.01%, while the cross-subject accuracy was 69.07%.