| With the progress of society,the incidence rate of stroke is rising.After stroke,there will be a great chance to leave limb movement dysfunction.Rehabilitation medicine shows that rehabilitation exercise can help patients recover their limbs’motor function.Traditional exercise rehabilitation training is mainly completed by doctors.Patients are usually treated passively when they are treated by traditional methods,which can not give full play to the wishes of patients.Using intelligent rehabilitation robot to carry out rehabilitation training for patients has become a new rehabilitation method.An angle prediction method based on human biological signal is proposed.This angle prediction method can give full play to the subjective will of patients and reduce the workload of doctors by analyzing the movement intention of human body and making real-time prediction of the angle,and then controlling the rehabilitation robot to assist the movement of patients.The main work is as follows(1)Based on the analysis of upper limb rehabilitation equipment,combined with the process of rehabilitation training,the requirements of angle prediction are put forward:MSE between the predicted angle curve and human joint angle data curve is not more than 0.1,and the goodness of fit R~2 is not less than 0.6;the time of the whole experimental scheme is not more than 0.1s;MSE fluctuation range of the same subject’s data is not more than 0.02,The fluctuation range is not more than 0.1;surface electromyography(sEMG)is selected for angle prediction,and the angle prediction scheme is formulated.(2)Using g.hiamp equipment,sEMG signals needed for human movement are collected from four upper limb muscles,and wavelet threshold denoising algorithm is used to remove the noise.(3)For the denoised sEMG signal,the feature extraction algorithm based on principal component analysis(PCA)and the feature selection algorithm based on genetic algorithm(GA)are used respectively.The experimental results show that the angle prediction error MSE of feature extraction is 0.1087,R~2 is 0.7472;the angle prediction error MSE of feature selection is 0.0674,R~2 is 0.7771,so the angle prediction algorithm of feature selection is selected.(4)Three different angle prediction algorithms are constructed,including BP neural network(BPNN),convolutional neural networks(CNN)and classification and regression tree(classification and regression)Tree(CART)algorithm is used to predict the joint angle,and BPNN algorithm is selected according to the angle prediction error.The experimental results show that:the error MSE of BPNN algorithm is 0.1087,R~2is 0.7472;the error MSE of CNN algorithm is 0.1365,R~2 is 0.6386;the error MSE of cart algorithm is 0.1231,R~2 is 0.6565,so BPNN algorithm is selected.To solve the problem of insufficient data of sEMG signal,transfer learning is used.The experimental results show that after using transfer learning,the generalization of the algorithm is better,the number of training is less,and the prediction error is smaller.After using transfer learning,the prediction MSE is 0.0651 and R~2 is 0.7736.(5)The sEMG signals of eight subjects were collected,and the four degree of freedom rehabilitation robot was used to verify the angle prediction algorithm.The off-line experiment mainly includes the verification of feature processing method,the verification of angle prediction accuracy and repeatability experiment.The experimental results show that the average prediction error MSE of the algorithm constructed by the feature selection method is 0.0532,R~2 is 0.7905,the average prediction error MSE of the algorithm constructed by the feature extraction method is0.0270,R~2 is 8508,which verifies that the prediction error of the feature selection algorithm is smaller,the average prediction error MSE of BPNN algorithm is 0.0571,R~2 is 0.8231,the average prediction error MSE of CNN algorithm is 0.0975,R~2 is0.8231 6878,the average MSE and R~2 of cart algorithm are 0.1069 and 0.6661respectively,which verify that BPNN algorithm has smaller prediction error.The average error fluctuation MSE is 0.0012 and R~2 is 0.6661 through five repeated experiments The experimental results show that the average motion lag of the rehabilitation robot is 28.49ms,which meets the real-time requirements,and the average acceleration of the rehabilitation robot is 1.66°/s~2,which meets the stability requirements.That is,the proposed scheme can meet the requirements... |
PDF Full Text Request