Multi-feature data fusion based on common space model and recurrent convolutional neural networks for EEG tristimania recognition used in upper limb rehabilitation exercises

Abstract

Traditional tristimania recognition methods cannot accurately recognize the mood of patients, which cannot provide effective adjuvant therapy for rehabilitation. Therefore, this paper proposes a new multi-feature data fusion method for Electroencephalography (EEG) tristimania recognition. It combines common space model and recurrent convolutional neural networks to classify the tristimania group and control group. According to the phase lock value, the phase
synchronization functional network between electrode channels is constructed, and the functional connection modes of two kinds under different frequency bands are analyzed. The Xception network and LSTM are used as two non-interfering parts to extract two feature matrices from EEG tristimania signals. They are fused into a single feature matrix by merge algorithm. The single feature matrix is input into the recurrent convolutional neural networks (RCNN) for feature extraction and pooling. L2 regularized Softmax function is used as the classifier to complete the training and testing of RCNN. Finally, combining the Fisher score feature selection method and the classifier dependency structure, a low dimensional and efficient feature subset is obtained. Experimental results on public tristimania data sets validate that the proposed method has better effect in terms of accuracy and PLV compared with other strategies.