Motor Imagery Recognition Based on Improved EEGConformer Algorithm
Menglong Li, Songlin Li, Yangqian Xu · 2024
Convolutional Neural Networks (CNNs) excel at extracting local temporal features from electroencephalograms (EEG); however, they encounter challenges with EEG data that contains complex temporal sequences. To overcome these limitations, the EEGConformer network has been developed, harnessing the strengths of both CNNs and Transformers. Despite its advances, the original EEGConformer network, which relies solely on two convolutional layers with large kernels during the feature extraction phase, may sacrifice finer local features. This paper introduces enhancements to the EEGConformer architecture aimed at improving feature extraction. Firstly, we replace the original one-dimensional spatiotemporal convolutional layers from the ShowllNet model, traditionally used for learning low-level local features, with those from the EEGNet model. The EEGNet model employs smaller convolutional kernels and a more extensive convolutional network to capture more nuanced local features. Secondly, to enhance training efficiency, we implement a distributed data parallel approach on a single machine equipped with multiple GPUs, configured for data parallelism. Finally, we conduct experiments using the BCI Competition IV 2a public dataset to assess the performance of our improved model in EEG-based motor imagery recognition. The results indicate that our proposed method not only achieves higher accuracy but also reduces the computation time significantly compared to the original network and single-machine training methodologies.