A hybrid self-attention convolution network for MI signals decoding

Research on Motor imagery electroencephalogram (MI-EEG) can reveal the motor intentions that trigger electrical changes in the research of brain–computer interface (BCI), which enables assistance of rehabilitation treatment of patients with neurological and muscle dysfunction. Recent deep learning methods generally fail to obtain the most discriminative motor imagery features due to the uneven information distribution on various electrode channels. In addition, the lack of sufficient sample sizes, leading to the limited decoding performance of neural networks. To address these two problems, a hybrid self-attention convolution network (HSACNet) method combined with optimization method based on attention mechanism is proposed in this paper. HSACNet contains an improved optimization method which employs the Particle Swarm Optimization (PSO) to enhance the adaptivity ability of the model. The proposed method is evaluated on BCI Competition IV public dataset and achieves the highest average accuracy of 56.75 % in cross-subject experiments, which is significantly higher than that found for recently published algorithms. Competitive experimental results show that the proposed classification algorithm promotes the global connection and the performance of decoding motor intentions, which can provide innovative perspectives for BCI systems.