Emotion Recognition Method Based on EEG in Few Channels

EEG acquired by wearable devices can effectively monitor human emotion. Some traditional methods based on some artificial designed features and deep learning technology have achieved good results in emotion recognition based on the EEG signals. However, in the actual experimental process, there are still some problems, such as the high number of channels, the feature redundancy and so on. These challenges hinder the application of the emotion recognition in portable wearable devices. This paper proposes a 4-channel method to achieve the classification accuracy in more channels (62 channels in SEED dataset and 14 channels in our dataset). In this paper, a parallel signal processing method, which is based on the intrinsic time scale decomposition (ITD), discrete wavelet transform (DWT), variational mode decomposition (VMD) and phase space reconstruction(PSR), is proposed to obtain more modes of signals. The differential entropy for each mode is extracted as the feature, and the linear dynamic system is used to smooth the feature. Based on the artificial features, this paper designs a simple CNN model with fewer parameters to complete the classification task. It also applies this method to SEED dataset and our own dataset, and carries out three types of experiments. Experimental results show that the average accuracy of the method in 4 channels can reach the performance of more channels. Due to the versatility of this method, it is expected to be more widely used in wearable devices with weak computing power under 4 channels.