Research on EEG Emotional Recognition Based on LSTM

Abstract

Electroencephalogram (EEG) signals have been shown to provide insight into deeper emotional processes and responses. Emotion recognition based on EEG signals has been studied on a large scale. One of the research objectives is to find features suitable for EEG emotion recognition through various methods, then optimize model and improve accuracy of classification method. However, EEG signal is a random non-stationary time series signal, and the traditional classification method does not take into account its timing. The Long Short-Term Memory (LSTM) network in deep learning technology can solve this problem well due to its temporal recursive structure. However, a EEG sequence is generally long. If you use LSTM classification directly, it will lead to large computing resources and poor results. Therefore, in this paper, firstly the EEG signal is cut, and the differential entropy feature of each segment is extracted by wavelet transform. Finally, it is input into LSTM model. The experimental results show that the classification accuracy rate is about 0.89 when the 4 layers LSTM model is selected.