Research On Prediction Of Lower Limb Joint Angle Based On Surface Electromyography Signals

Surface EMG signals contain a large amount of motion information and have a wide range of promising applications in human motion intention recognition.The purpose of this thesis is to continuously predict the joint angle of the lower limbs through surface elec-tromyography signals to provide multiple input information for man-machine interactive control of flexible exoskeleton assist suits.This thesis proposes a regression prediction method for continuous motion of human lower limb joints under different motion modes,which achieves synchronous acquisition of surface EMG signals and joint angle signals.After preprocessing,a mapping model of surface EMG signals and lower limb joint angles is established,thereby improving the accuracy and applicability of the regression predic-tion model.The main work of this thesis includes the following:Firstly,the synchronous acquisition scheme of lower limb motion information is de-signed.By analyzing the motion relationship between human lower limb muscles and the hip and knee joints,the experiment screened six characteristic muscles with a greater contribution rate,namely RF,VM,ST,TA,MG and SM.To determine the simultaneous acquisition scheme of motion information,a biomagnification signal acquisition system was used to provide surface EMG signals,and three angle sensors were utilized to provide the angle signals of hip and knee joints.An experimental scenario was set up to collect lower limb movement information from four movement models of human normal gait,sitting-standing,squatting-standing,and overstepping.Next,the surface EMG signals were pre-processed.The surface EMG signals are weak,low-frequency signals,and are easily affected by physiological and environmental factors.By analyzing the characteristics of surface EMG signals,the performance of three filtering and denoising methods,namely Butterworth band-pass filter,empirical modal decomposition denoising and wavelet threshold denoising,was compared.The results show that the surface EMG signals obtained by using wavelet threshold denoising exhibit periodic changes,possess the smoothness of the original signal,have a high signal-to-noise ratio,and have a large coefficient of determination.The sliding window method is used to extract temporal and frequency domain features from the surface EMG signal after denoising.Finally,a prediction model of hip and knee joint angles is constructed.A lower limb joint angle prediction model(PFA-LSTM-Attention)with optimized weight allocation is proposed.A two-layer attention mechanism is introduced into the LSTM long-term and short-term memory network architecture,which improves the input weight for surface EMG features with high prediction contribution rate and the output weight of the hidden layer at different times.The global optimal network parameters are obtained through the PFA pathfinder optimization algorithm.The experiment compares PFA-LSTM-Attention with random forest,LSTM,and LSTM-Attention models.The results show that the root mean square error of the PFA-LSTM-Attention prediction model proposed in this thesis is5.100,the average absolute value error is 3.5540,and the determination coefficient is 0.85 under multiple action modes,all of which are superior to the control group.This model can effectively improve the accuracy of angle prediction of the lower limb hip and knee joints,save the cost of artificial experience screening network parameters,and have good stability in various motion modes.