EEG signals classification plays a crucial role in motor imagery brain computer interface systems. Traditional convolutional neural networks tend to ignore temporal information when classifying motor imagery EEG signals, it uses a single-scale convolutional kernel, resulting in poor classification performance. In this paper, we propose a parallel fusion algorithm based on dual attentional multi-scale convolutional neural networks (DAMSCN) and long and short-term memory (LSTM). Firstly, DAMSCN uses convolutional kernels of different sizes at the same layer to extract time-frequency features of EEG signals at different scales, and introduces a dual attention mechanism. At the same time, LSTM extracts temporal features from the EEG signals. Then, the fusion and classification of all features is achieved with the help of fully connected layers and softmax layers. Finally, experiments are conducted on domain-specific public dataset to verify the performance of the algorithm.
|