Research On SEMG Feature Generation And Classification Performance Based On Generative Adversarial Network

At present,domestic and foreign scholars have successively conducted research on surface electromyography(s EMG)signals in the fields of electromyography control and human-computer interaction,and have achieved gratifying research results.However,due to the large number of samples required for surface electromyography signal recognition technology to ensure the accuracy of training results,the common problems of low data volume,difficult and complex collection,and environmental impact of electromyography signals hinder the improvement of the accuracy of electromyography classification.In response to the above issues,this article proposes an electromyographic feature generation method based on energy-based generative adversarial network(EBGAN).Focusing on s EMG signal acquisition,preprocessing and feature extraction,generative model construction,generation data classification performance verification and other aspects,the research aims to expand s EMG data to improve the accuracy of s EMG classification.The work of this article mainly includes the following points:(1)s EMG signal acquisition,preprocessing,and feature extraction.First of all,the signal acquisition scheme is determined.Five action modes are selected,namely,walking on the ground,going upstairs,going downstairs,standing up,and sitting down.Surface EMG signals at the lower biceps femoris muscle,lateral femoris,and medial femoris muscles are collected;Further using bandpass filtering and wavelet threshold denoising for preprocessing,and evaluate the denoising effect using mean square error(MSE)and signal-to-noise ratio(SNR);Secondly,time window segmentation is performed on the preprocessed signal.Finally,three time-domain features of electromyography signals under different actions are selected for feature extraction.The mean absolute value(MAV),root mean square(RMS),and variance(VAR)are extracted to form a feature dataset,providing data support for subsequent generation of electromyography features and validation of data classification performance under five action modes.(2)Research on the method of generating electromyographic features.This article Proposes a feature generation method for electromyography signals based on energybased generative adversarial network(EBGAN).This article sets the generator and discriminator architecture according to data distribution,and designs the hidden layer of the generator to incorporate a batch normalized one-dimensional fully connected layer architecture,so that the data in each layer maintains its original distribution characteristics.The discriminator is designed as an automatic encoder structure,with two parts: encoder and decoder,and outputs an energy function to minimize reconstruction errors.Carry out a comparative study on the data generation quality between this method and the traditional GAN,WGAN,and DCGAN models,and finally judge the s EMG feature generation performance of the generated countermeasure network model by observing the generated data pattern,the maximum mean difference(MMD),the dynamic time warping index(DTW),and the t-SNE visualization method.The experimental results show that compared to the other three GAN networks,the EBGAN model proposed in this paper generates electromyographic features similar to real data,achieves smaller MMD and DTW,and has the best performance.(3)Research on validation of generated data classification performance.To verify the effectiveness of the electromyographic feature generation method,this article applies the EBGAN synthetic dataset,raw dataset,and other GAN synthetic datasets to some typical classification models: K-nearest neighbor(KNN),support vector machine(SVM),linear discriminant analysis(LDA),and long-term and long short-term memory(LSTM),and conducts sample quality and universality generation Research on performance validation of sample generation scale and EBAGN compared to other models in terms of superiority.The results show that the classification results of each classification method are better than those of the original training set and other generative model through training on the EBGAN synthetic data set,and the accuracy of each classification model is improved by 1%～5%,which provides a new way of thinking and method for machine learning in EMG recognition research.