Research on driving fatigue detection based on multi-scale wavelet packet-inspired convolutional network under small sample conditions

This study proposes a Multi-scale Wavelet Packet Inspired Convolutional Network (MWPICNet) based on electroencephalogram (EEG) signals for driving fatigue detection. MWPICNet alternates between multi-scale wavelet packet convolution layers and soft threshold activation layers to extract time-frequency fatigue features. It incorporates a frequency band weighting layer and a global power pooling layer to enhance feature recognition. Tested on the SEED-VIG and SADT datasets, MWPICNet achieves accuracies of 98.67% and 99.39%, respectively, maintaining over 95% accuracy under noise interference. Compared to CNN and GCN, MWPICNet improves accuracy by 5.99% and 4.79%.