Automatic Searching of Lightweight and High-Performing CNN Architectures for EEG-based Driving Fatigue Detection

The increasing number of vehicles has led to a rise in traffic accidents, with fatigued driving being a major contributing factor. Bio-electrical signals, particularly electroencephalograms (EEG), have emerged as a promising avenue for detecting driving fatigue. EEG signals can provide valuable insights into a person's brain activity and state of alertness. However, the complexity of EEG signals and the need for real-time detection pose significant challenges for traditional machine learning algorithms, leading to the growing popularity of deep learning in this domain. The objective of this paper is to design lightweight and high-performing convolutional neural network (CNN) models for detecting driving fatigue using multi-channel EEG signals. These models are intended to be deployed on resource-limited devices in intelligent vehicles, enabling timely alerts for fatigued driving. Rather than manually designing the deep neural network (DNN) architecture, we adopted the neural architecture search (NAS) approach to automate the architecture design process, considering both detection performance and computational cost. To evaluate the effectiveness of our approach, we conducted experiments using two publicly available EEG datasets widely used in driving fatigue detection studies. The performance of our NAS-derived model, named FD-LiteNet, was compared with a set of state-of-the-art baseline CNN models manually designed for EEG signal analysis. The results demonstrate that FD-LiteNet achieves significantly higher detection accuracy than all baseline models with a lower computational cost. Furthermore, our findings highlight the exceptional generalization capability of FD-LiteNet, as it can be fine-tuned with a small number of new samples to adapt to new datasets.