Research On Myoelectric Signal Generation Model Based On Adversarial Neural Network

In recent years,the accuracy of pattern recognition has been continuously improved with the development of deep learning,but deep learning requires the support of large data volume,and the larger the amount of data,the higher the stability and accuracy of the model.Therefore,the data generation technology represented by generative adversarial network(GAN)has attracted the attention of researchers.Surface electromyoelectric signal(sEMG)is a weak electrical signal caused by human muscle activity,which has broad application prospects in human-computer interaction,prosthetic control,motor function rehabilitation and other fields.However,the sEMG data acquisition process is complex and requires ethical approval,resulting in pattern recognition accuracy based on EMG signals limited by the size of the data and not significantly improved.Therefore,it is of practical significance to generate EMG simulation data by technical means.In this thesis,sEMG simulation sample generation is realized based on GAN,and the accuracy of gesture recognition based on sEMG signals is improved by adding simulation samples to the training dataset.The main innovations in this thesis are as follows.1.A method for generating sEMG image samples based on deep convolution generative adversarial networks(DCGAN)is proposed.This method first converts the original multi-channel sEMG time series into image data,which is convenient for DCGAN processing.Secondly,all image data is improved by histogram equalization.Finally,EMG image samples representing each type of action are fed into DCGAN for individual training to obtain EMG image samples that can convert random noise into EMG images with specific categories.Experiments show that the EMG pictures generated by DCGAN are similar to the real EMG pictures,and the dataset mixed with real samples and simulated samples can obtain higher classification accuracy.2.A sEMG feature sample generation model based on wasserstein generative adversarial networks(TWGAN-GP)is proposed.The model is improved on the basis of WGAN-GP architecture,and the two-dimensional convolutional neural network(2D-CNN)is replaced with the bi-directional long short-term memory(Bi LSTM)network and one-dimensional convolutional neural network(1DCNN).The model fully considers the correlation between time series and pays attention to local features,which can significantly improve the ability of the framework to model time series.On this basis,the wasserstein distance loss function is used instead of the cross-entropy loss function of the original GAN,and the gradient penalty is used to optimize the loss.Experimental results show that the model has stability and robustness in network training,and can generate high-quality sEMG feature samples.3.A data distribution shift test method to evaluate the performance of the time series generation model is proposed.Data distribution skew testing refers to the correlation analysis of the spatial distribution of real data sets and synthetic data sets,including statistical evaluation and machine learning evaluation.Statistical evaluation uses spearman’s coefficient to correlate the mean and standard deviation of the real and synthetic data sets.Machine learning evaluation measures whether the classification model is compatible on the real and synthetic datasets by comparing the classification measures of the classification model on the real and synthetic datasets,reflecting the correlation between the two datasets.Experiments show that this method can achieve good results in the EMG signal generation model proposed in this thesis,and can be further extended to the performance test of other timing signal generation models.