A Framework on Optimization Strategy for EEG Motor Imagery Recognition.

Abstract

At present, in the process of encephalogram motor imagery decoding, facing the background of big data analysis, it has the necessity to design an effective system which is subject-independent. Pre-training is common to carry out before each experiment, which affects the practicability of the EEG system. In order to solve this problem, the most feasible method is to design a unified framework for deep learning optimization, which could capture the spatial and spectral dependence of original motor imagery EEG signals according to the features extracted by CNN and the temporal dependence extracted by RNN-LSTM. The framework is superimposed from both end-to-end and time-frequency domains so as to retain and learn interpretable motor imagery features. In addition, artificial EEG signals can be automatically generated by training the generated adversarial network, which can generate the feature distribution similar to the original EEG signals, increase the capacity of EEG samples, and ultimately improve the classification performance and robustness of EEG motor imagery recognition. This deep learning framework can improve the classification accuracy of motor imagery for different subjects. In addition, the network can learn from the original data with the least amount of preprocessing, thus eliminating the time-consuming data preparation process.