Abstract: Aiming at the problem of a low accuracy of motion intention recognition for stroke patients at different disease degrees, a motion recognition method based on surface electromyography (sEMG) for patients at different Brunnstrom stages is proposed. Firstly, sEMG data of patients at different Brunnstrom stages are fused, and tsfresh library is used for feature extraction. Then, the features are selected based on random forest (RF) model, and the selected features are used for training action classification model. Furthermore, the rehabilitation evaluation actions are determined by studying the relationships between actions and rehabilitation grades, and an automatic evaluation algorithm of rehabilitation grade is designed. In order to verify the effectiveness of the proposed method, the sEMG data of 24 patients are tested. The experimental results show that the proposed method can improve the average recognition accuracy of 9 kinds of actions and 6 types of rehabilitation stages to 89.81% and 94% respectively. The hand rehabilitation robot system based on the proposed method can realize the automatic evaluation of rehabilitation grade.