Research On Technology Of Human Lower Limb Motion Recognition Based On Multi-Source Information Fusion

EMG(Electromyogram)signal,a kind of bioelectric signal,is generated ahead of limb movement and contains numerous motion characteristic information.It not only can accurately feedback human motion,but also has advantages of easy integration and less interference with human body,making it has broad prospects for application in the field of human motion recognition and wearable exoskeleton.As a weak bioelectric signal,EMG signal has many problems,such as complex noise,immature decoding methods,and low utilization of effective information,resulting in significant limitations in its application.Based on this,this dissertation takes exoskeleton robot for astronauts as application object,deeply excavates motion information in EMG signal,and explores the multi-source information fusion method with the goal of improving accuracy of human motion intention recognition.The specific researches are summarized as follows:Firstly,research on the basis of human biomechanics and astronaut movement patterns.The generating mechanism of EMG signals and the internal relationship among the nervous system,EMG signals,muscle contraction,joint changes,and macroscopic motion were explored;A mapping model of human lower limb joint movement and muscle changes was established based on the principles of human lower limb anatomy and Open Sim;The factors influencing muscle force were studied,and the intrinsic relationship between EMG signals and joint motion was discussed based on classical Hill muscle model.And a detailed analysis was conducted on the motion patterns of astronauts in and out of spacecraft,as well as on planet surface.This research provides a foundation for the subsequent research on multi-source information fusion algorithms.Secondly,in terms of muscle selection methods and EMG signal preprocessing for human motion intention recognition,a muscle screening method based on Open Sim model and correlation analysis was proposed to address the issue of irregular selection of lower limb EMG signals leading to incomplete motion information feedback and affecting motion recognition rate.Specifically,this method selected muscles based on the correlation of lower limb muscles during gait,which were most relevant to motion and have low correlation with each other.So that there was no excessive interference among the measured EMG signals,simultaneously,it can more sufficiently reflect human motion information.In addition,considering the weak characteristics of EMG signals and the susceptibility to noise during detection,noise processing was carried out for EMG signal,and outlier and trend items in the signal were eliminated.In terms of human motion recognition at the data level of EMG signal,a multi-source information data layer fusion method for human lower limb motion and motion recognition was proposed to achieve accurate recognition of astronaut motion patterns in ground environments.Specifically,using a single hidden layer BP neural network to achieve pattern classification for 7 road slopes(-15～15)and 5 categories of gait(flat walking,up/down slopes,up/down steps).To study the discrimination of EMG signals between continuous gait and movements in lower limbs,two squat movements(squat and side squat)were added to the classification category and the outputs under different input conditions were compared.The obtained results validate the mapping relationship between EMG signals and lower limb movements at the data level.The linear/nonlinear feature fusion of EMG signals were research on motion recognition of lower limb joints.In view of the incomplete feedback of motion information on the single eigenvalues of EMG signals,series splicing method for EMG features was proposed.This method constructed linear and nonlinear features into a higher dimensional feature matrix,which improved the accuracy of identifying lower limb movements using EMG signals.Compared with the recognition effect of a single feature,series splicing matrix improved the average recognition rate to 92.32%for 3gaits+squat+sitting positions.A complex vector method based on matrix fusion of EMG signal eigenvalues was proposed to address the problem of low correlation between the eigenvalues of EMG signals,which integrates linear and nonlinear features of the signal into a complex matrix.Specifically,the linear features of EMG signals(Time domain RMS and VAR,Frequency domain MPF and MF)were taken as the real part of the complex matrix,the nonlinear features(Intrinsic Mode Function,IMF)were taken as the imaginary part of the complex matrix,and the Matrix norm was taken as the feature input to identify lower limb movements.This method fully integrates the linear and nonlinear features of electromyography signals,and improves the recognition rate of BP neural network for lower limb movements to 96.1%.Finally,in terms of improving the recognition rate of knee joint movements and predicting angles,a multi-source information fusion method with multi weak classifier based on a parallel structure was proposed to address the problem of low recognition rate of lower limb gait by individual classifiers.Specifically,the multi weak classifier signal information fusion classification system was designed according to the different number of weak classifiers and input data.3 information fusion methods(equal weighting method,logarithmic equal weighting method,and autonomous competition method)were used to fuse the weak classifier results.To further increase the connection between weak classifiers in parallel classification system,this dissertation combined the nonlinear mapping ability of BP neural network with the characteristics of AdaBoost algorithm to dynamically adjust the weight distribution of the training set,and proposed BP＿AdaBoost algorithm with BP neural network as the weak classifier.This method increased the average recognition of knee joint movements by electromyography signals from 78.52%to 93.52%;The BP＿AdaBoost strong predictor was used to predict the angle of knee joint.Comparing the effects of several predictors,it was found that the designed BP＿AdaBoost predictor controlled the error within ±13~°.