CWSTR-Net: A Channel-Weighted Spatial-Temporal Residual Network based on nonsmooth nonnegative matrix factorization for fatigue detection using EEG signals

Mental fatigue is the leading cause of accidents among industrial workers. Deep learning algorithms that harness multichannel electroencephalogram (EEG) data have proven to be the most effective means of detecting fatigue in industrial settings. However, the accuracy of current detection methods has not yet reached optimal levels. The inherent complexity and instability of EEG signals present challenges in feature extraction. Moreover, training models directly with all channel information risks capturing irrelevant data, which can lead to overfitting. To tackle these challenges, we have developed the Channel-Weighted Spatial-Temporal Residual Network (CWSTR-Net) for fatigue detection. This network combines Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) into a deep learning cascade residual network (STR-Net) to extract spatial-temporal features from EEG signals. Furthermore, we have incorporated an unsupervised channel weighting algorithm based on non-smooth nonnegative matrix factorization (nsNMF). The channel weighting is applied before the STR-Net processing. The nsNMF algorithm effectively identifies active brain channels. It derives weights for each lead by analyzing signals with complex internal correlations through matrix decomposition, thus providing crucial insights for decoding brain signals. Our network achieved an average accuracy of 97.23% on the SEED_VIG dataset, outperforming previous EEG fatigue detection models. This research introduces innovative perspectives to EEG-based fatigue detection and positively drives the development of the field. The code is available at https://github.com/ws101320/EEG_fatigue.git.