Many patients who have leg injuries from accidents or who are unable to walk independently are afraid of falling when participating in rehabilitation, which prevents these patients from participating in daily life. To overcome these limitations, the study designed a multidirectional weightlifting support system that can be trained in a safe environment. In addition to walking on the ground, the system can help patients train for more everyday walking tasks in a very realistic environment. Necessary assistance can be provided through task-related support design during user movement. One of the remaining challenges is how to manually switch between AIDS for different tasks. This approach is error-prone and cumbersome, distracting both the therapist and the patient, and interfering with the training workflow. Therefore, in this paper, the information flow gain algorithm is used to optimize the original system of the rehabilitation robot to improve the recognition accuracy of the robot motion pattern. Finally, the experimental results showed that the motion recognition accuracy of 8 subjects wearing rehabilitation equipment reached more than 80%, and the average accuracy of each muscle movement recognition was more than 85%.