Continuous Motion Estimation Of Multiple Joints Based On SEMG And Acceleration

With the increase of social aging and social pressure surge,more and more people fall into the problem of physical movement obstacle because of stroke.Therefore,people pay more and more attention to the field of rehabilitation training,and the research of limb rehabilitation training robots has become a hotspot.The purpose of our project is to study the intention of patients' muscle movement.And we want to use surface electromyography(SEMG)and acceleration(ACC)signals to predict the trend of patients' continuous joint motions and control the rehabilitation robot for rehabilitation training.In this paper,we choose 5 channels SEMG signals and 5 channels ACC signals as the input source to study the continuous motions of multiple joints of the lower limbs.The main work and innovations are as follows:(1)Aiming at the mainstream rehabilitation training mode,this paper put forward the corresponding standard experimental movements and collection rules.Trigno wireless acquisition system was used to collect 5 channels SEMG signals and 5 channels ACC signals.The SEMG signals were collected by the analysis of the contribution of the muscle.The experimental muscles were gluteus maximus,vastus medialis,rectus femoris,tibialis anterior and gastrocnemius.(2)For the collected signals,the subsampling processing was carried out to unify the frequency.At the same time,the noise reduction experiment based on empirical mode decomposition(EMD)was carried out for the SEMG signals.Aiming at the feature extraction methods of SEMG signal in time domain,frequency domain and timefrequency domain,this paper proposed a new eigenvector based on wavelet multi-level decomposition and correlation dimension coefficient as the input feature of SEMG signals.For the ACC extraction method,this paper selected the signal amplitude vector as the input feature of ACC signal.(3)Aiming at the nonlinear and non-stationary characteristics of SEMG signals,the Extreme Learning Machine(ELM)and GA-Elman network were used as joint angle prediction models.The ELM was an optimization of BP neural network.It can train faster than the traditional BP algorithm on the premise of ensuring the learning accuracy.GAElman was an Elman network optimized by genetic algorithm(GA).Traditional Elman network had been superior to other methods in predicting time series.(4)According to the experimental data,we used the above method to extract the features to form the input feature vector.We determined the network parameters of ELM and GA Elman prediction models.Firstly,we used the same prediction model to predict the joint angle,tested the performance of our feature extraction method and proved that this method is better than others.Then,in the case of extracting the same feature,comparing different prediction models,it was verified that the prediction error of GAElman network is significantly smaller,and the time-consuming was also less.Next,every volunteer was tested separately.It was found that the prediction error of each volunteer was about 4%.Our prediction method had good individual adaptability.Finally,we analyzed the synchronization of SEMG signals and prediction angles based on the detrended cross-correlation analysis algorithm.And the angle prediction results can be optimized according to the results.