Dual‐Domain Fusion Graph Convolutional Network for EEG‐Based Driving Fatigue Detection

Driving fatigue detection technology plays a crucial role in preventing traffic accidents and ensuring road safety. In recent years, fatigue detection techniques based on electroencephalography (EEG) have received widespread attention for their objectivity and accuracy. However, existing studies are deficient in exploiting the brain topology and inter-electrode dependencies, resulting in under-exploitation and under-utilisation of much information, insufficient diversity in feature extraction and over-smoothing. For this purpose, a Dual-Domain Fusion Graph Convolutional Network (DDFGCN) model is proposed. The model combines neuroscience theory and analyses the electrode connectivity relationship from both local and global dimensions to achieve multi-level feature aggregation. First, multi-scale temporal convolution is used to extract the temporal dynamic features of EEG as input. Second, two different brain map construction methods were developed to capture the local channel dependencies and the global topology of EEG signals, respectively. Finally, accurate prediction of fatigue state is achieved by feature integration and classification modules. The model achieved 94.67% and 95.6% accuracy on the SADT and SEED-VIG datasets, respectively, under the nested cross-validation setting. The proposed DDFGCN model has excellent classification performance compared to other methods. The experimental results show that DDFGCN fully considers the local activity information within the functional regions of the brain while integrating the remote dependencies of the brain regions, enhances the relational modelling of the whole scalp, improves the classification performance and provides a promising new method for the application of fatigue driving detection technology.