In hospital, most stroke patients' rehabilitation training is assisted by rehabilitation physicians. However, this rehabilitation way is expensive and the number of physician cannot meet the demand. Therefore, surface electromyography (sEMG) and machine learning algorithms are introduced into the rehabilitation system, which can reduce the work of physicians and meet the needs of patients. We collect the sEMG signals of patients by picture guidance. Due to the mislabeling caused by the time mismatch, we re-calibrated the labels by means of peak detection. A model fusion algorithm, called stacking, is leveraged to improve the accuracy and robustness of hand action recognition. Before training, we will assess the patient's condition. Different rehabilitation training strategies will be adopted to patients under different conditions so as to each patient can receive effective training. According to rehabilitation psychology, virtual reality games were introduced to enhance the interest and pleasure of patients in the training process. At the same time, it can stimulate the development of nerve in the motor system and enhance the activity of the motor cortex in the cerebral cortex.