The aging trend of our population is becoming more and more obvious since entering the 21 st century.The elderly people have exceeded 200 million.Among them,the increase of patients with hemiplegia,paraplegia,cerebral palsy and limb dyskinesia is extremely obvious.They are increasing more than 5 million every year.At the same time,the western developed countries face the same problem.Those patients who want to regain their athletic abilities needing rehabilitation therapists to help them with certain rehabilitation treatments.Therefore,the developed countries such as Europe and the United States have developed rehabilitation exoskeleton robots for the medical market.It can not only help patients improve their autonomy in life,but also improve their quality of life.Gait phase control,which is the core in the control technology of rehabilitation exoskeleton robot,affects the safety and stability of rehabilitation exoskeleton robot.This paper presents several control models for the gait phase classification of rehabilitation lower limb exoskeleton robot.These models only require four joint angle signals of the lower extremity exoskeleton robot or the patients' four sEMG signals to accurately identify the four gait phases between the foot and ground during walking: Heel strike(HS),Foot Fat(FF),Heel off(HO)and Swing(SW)to help the hemiplegic walk.Compared with the long short-term memory(LSTM);the support vector network(SVM);and the back propagation neural network(BPNN)which are all rely on the Euclidean domain,the graph convolutional neural network model(GCNM)proposed in this paper is the first model for the gait phase classification problem of exoskeleton robots from non-Euclidean domain based on graph mechanism.The systems designed in this paper not only avoid complex sensor systems,but also ensure the accuracy of gait phase classification in contrast to the existing methods.The international industry standard Vicon system is used to verify the availability of real-time gait data acquisition system for lower limb exoskeleton robot.And the time delay of the system is also quantified by Vicon.The results show that the proposed GCNM is not only significantly higher on prediction accuracy but also has better robustness for gait phase classification to control the lower limb exoskeleton system to different people on level ground,uphill and downhill environments.Compared with the existing supervised learning algorithms where the label rates are usually 70% or 80%,the label rates of the semi-supervised learning model-GCNM are all under 10%.Finally,the maximum accuracy of GCNM on gait phase classification is 97.43% which could more accurately estimate the assist torque precision of exoskeleton robot in the future. |