Optimized Spatio-temporal Feature Representation Based on 3D-CNN Motor Image Decoding

The traditional convolutional neural network decoding model lacks the multi-domain information of electroencephalogram (EEG), which makes it difficult to decode motor imagery (MI) accurately. To solve this problem, this study proposes a 3D-CNN MI decoding method with optimized spatio temporal feature representation (OSTFR). Specifically, we utilize L21-norm optimization to generate a spatial filter set that captures highly distinguishable MI features in the multiband spatial domain. Secondly, with the sample covariance moments and spatial filter sets, the OSFR is generated by fusing the EEG time-domain and frequency-domain information. Finally, a new 3D-CNN model is designed to extract OSTFR features to complete MI decoding. Using the BCI Competition IV datasets 2a and 2b, which are available to the public, the value of the proposed method is verified, leading to decoding accuracies of 87.24% and 80.63%, respectively. The experimental results emphasize this method's potential for widespread use in BCI research by demonstrating its exceptional performance in MI decoding tasks.