A Multi-scale Residual Network Based on the Multi-head Attention Mechanism for Motor Imagery EEG Decoding

Recent advancements in deep learning have led to increased application in brain-computer interfaces (BCIs) for decoding motor imagery electroencephalogram (EEG) signals. Nevertheless, current methods still present considerable challenges in effectively decoding these signals. Convolutional neural networks (CNNs), most used in this domain, are constrained by limitations associated with their kernel size and inherently limited receptive fields. This restricts their capacity to fully capture the global spatial and temporal dependencies present in EEG signals, which is crucial for accurate decoding. In response to these limitations, we propose a novel approach: a multi-scale residual network based on the multi-head attention mechanism. This model is designed to overcome the constraints of traditional CNNs by incorporating a multi-head attention layer, four convolutional layers, and a pooling layer, which enable the extraction of global and local features. To verify the effectiveness of our model, we conducted an experiment involving left- and right-hand fist imagery. We evaluate the performance of our model against several state-of-the-art models currently used in the field. The experimental results indicate that our model performs significantly better than the baseline model, with superior decoding accuracy, and our model is faster and more stable in the convergence of the loss function. It achieves an optimal balance between model complexity and accuracy, which makes it more advantageous in practical applications.