Wavelet-enhanced temporal convolution and Transformer encoder for motor imagery EEG signal classification

Convolutional Neural Networks (CNN) have become a research hotspot in motor imagery brain-computer interfaces, but their limited receptive field may be deficiencies in handling long-distance dependent information in EEG signals. In this study, we present EEGWaveformer, an innovative model based on deep learning, which incorporates the advantages of CNNs, Wavelet Convolution (WTConv), and Transformer encoders, aiming for efficient classification of EEG signals. The model extracts time-domain features through a multi-scale convolution module and further explores cross-channel information using a channel-wise convolution-based spatial feature extraction module. To capture long-range dependencies, the model uses Transformer encoders and attention mechanisms to enhance feature recognition. Finally, a convolutional encoder and a fully connected layer are used for feature integration and classification. Experimental results show that EEGWaveformer achieves average accuracies of 85.22%and 97.10%on the BCIC-IV-2a and HGD datasets, respectively, outperforming existing methods.