Real-time Human Motion Recognition And Prediction Based On Autoregressive Learning Algorithm

Enhanced exoskeleton is a wearable mechanical device that provides assistance for human movement.It can enhance human body power and reduce metabolic consumption during long-term and heavy-load work.This article focuses on the individual-enhanced exoskeleton human-computer interaction system in the military field,and conducts innovative research on the perception and prediction methods of human motion intentions.First of all,the surface EMG signal is acquired by myoelectric sensor.Because the signals are mixed with a lot of interference noise,the filtering and regularization of the signal is completed through a certain preprocessing method.According to the possible rotation offset problem of myoelectric sensor,an offset correction method based on active polar angle in polar coordinate system was proposed.The purpose is to solve the problem of estimation accuracy degradation caused by the offset from the original position when the electrode patch is first positioned.Experiments show that using the corrected surface EMG signals to predict intent information has higher accuracy.Secondly,in terms of data preprocessing,due to the coupling relationship between the EMG signal data and the correlation between the channels,the principal component analysis method is used to reduce the dimension and convert the original variables with correlation into a set of linear Irrelevant variables effectively reduce the amount of input data to the network,reduce computational complexity,and improve response time.Then,a human motion intention estimation model derived based on the skeletal muscle model is proposed.The human intention information includes the joint angle and angular velocity of the three degrees of freedom of the upper limbs.The estimation function uses NARX neural network with autoregressive and time series characteristics.Due to the recursive structure of the NARX neural network,it may lead to the accumulation of errors in the prediction of the sequence,which will affect the prediction accuracy of the model.Therefore,combined with the unscented Kalman filter algorithm,a closed-loop structure is realized,and the state quantity is estimated in real time.Finally,experiments verify the feasibility and accuracy of the overall algorithm.The motion capture system Motion Analysis and MYO armband are used as experimental equipment.A set of pre-designed arm movements are as the basis for experimental data collection and verification.Compared with the estimation results of the back-propagation feed-forward neural network,the mean square error value of the angle estimation is less than 3.86 degrees,and the mean square error of the angular velocity estimation is less than4.02 degrees / sec.The results prove that the intent recognition algorithm proposed in this paper can effectively improve the recognition accuracy and reduce the real-time errors of joint angle and angular velocity estimation.