Research On Continuous Motion Estimation Of Elbow & Wrist Joint And Smooth Control Of Upper Limb Exoskeleton

Surface Electromyography(sEMG)is the superposition of skin muscle activity and nerve potential,which contains abundant human behavior information.Motion intention decoding based on sEMG provides an important information source for clinical medicine,rehabilitation medicine and other fields.In recent years,applying deep learning method to construct human motion intention recognition model based on sEMG has gradually become a research hotspot.However,most of the existing deep learning models are used to predict the continuous motion of a single joint.What’s more,multi-channel original sEMG or multi-dimensional feature matrix extracted from multi-channel sEMG are used as input.Which ignores the negative effects of muscle coupling on sEMG redundancy in the prediction process.Secondly,the fault tolerance rate of the point estimation method of parameters in the deep learning model is low,and the model lacks a reliability measure for the estimated results.Aiming at the above problems,this paper deeply studied the estimation method of synchronous kinematic and dynamic quantities of upper limb based on sEMG.The model was constructed based on multi-channel sEMG to estimate the continuous angle and torque of the elbow-wrist joint.The main research work includes:(1)Research on multi-channel sEMG redundancy removal algorithms for simultaneous angle estimation of double joints of upper limb.This paper analyzes the signal redundancy problem in predicting the continuous motion of double joints using deep learning models based on multi-channel sEMG.Based on the Deep Residual Shrinkage Network(DRSN),this paper proposes a Deep Residual Wavelet Shrinkage Network(DRWSN)model combining channel attention mechanism,spatial attention mechanism and Wavelet transform method respectively.Based on the proposed model,the adaptive redundancy removal of multi-channel sEMG was carried out for every subject,and the feasibility of the proposed method was verified by comparative experiments.(2)Research on continuous torque estimation method of double joint of upper limb against disturbance of uncertain factors.Firstly,reference torque is calculated by sensorless method.Then,the uncertainty of regression model based on deep learning is analyzed.Bayesian neural network model based on variational reasoning is constructed,and the weights and bias point estimates in the non-Bayesian neural network model are replaced by the updated probability distribution based on Bayesian back propagation.In this way,the problem of over-fitting in non-Bayesian neural network model can be eliminated.Furthermore,the influence of changing model input on the performance of Bayesian neural network model is studied.Comparative experiments show that the proposed model shows better robustness in joint torque estimation.(3)Smooth control application of rehabilitation exoskeleton based on joint continuous motion estimation.A phased rehabilitation training model has been designed to address the needs of upper limb rehabilitation at different rehabilitation period.Firstly,the position control based on the angle trajectory estimated by sEMG was carried out to realize the passive rehabilitation strategy.Then,based on the joint torque estimated by sEMG,the active rehabilitation strategy was realized by adjusting the relationship between position and interaction torque for soft control.Combined with the model-free adaptive control method,the controller is designed to realize the effective control of the upper limb exoskeleton.