A Motor Imagery EEG Classification Algorithm Based on ResCNN-BiGRU

In recent years, deep learning-based methods for motor imagery EEG classification have become increasingly popular in the field of brain-computer interfaces. However, most of the studies tend to use sequence-structured classification networks to extract spatial features when dealing with motor imagery EEG signal classification tasks, ignoring the fact that EEG signals as time-series signals contain rich temporal information and features between neural network layers, resulting in poor classification performance. Therefore, this paper proposes a feature fusion network called ResCNN-BiGRU, which consists of ResNet-based residual convolutional neural network (ResCNN) and bidirectional gated recurrent unit (BiGRU) connected in parallel. The two branches use different forms of EEG signal feature representation as input, the input to the ResCNN branch is a wavelet transformed time-frequency image, and the input to the BiGRU branch is EEG data in a two-dimensional matrix format. ResCNN extracts spatial features and utilizes interlayer features through residual connections. It also introduces convolutional block attention module (CBAM) to avoid introducing too much useless low-level feature information during interlayer feature fusion. BiGRU extracts temporal features. Finally, experiments are conducted on the authoritative four-category motor imagery dataset BCI Competition IV 2a to verify the performance of the proposed algorithm.