Classification Of SEMG Signals Based On Deep Learning

Surface electromyogram(sEMG)signal is a summation of motor unit action potential trains(MUAPTs),which can be readily measured on the skin during the muscle contraction.Through the analysis of sEMG signal,we can accurately know the motional intention of human body.At the same time,it is widely used in prosthetic limb control,human-computer interaction and other fields because of its advantages of easy collection and non-invasive.The current research on using sEMG for gestures recognition mainly focuses on designing EMG signal features,decent designs can significantly improve the result.Nevertheless,the process of designing and selecting features can be complicated,on the other hand,deep neural networks can automatically extract features.So,based on the characteristics of sEMG signals and deep learning,we propose several models for classifying sEMG signals.Firstly,the sEMG signal belongs to the time series signal.By combining the advantages of the LSTM and the convolutional neural network(CNN),we propose a LCNN network and CNN_LSTM network that can infer human motion intention directly from the original sEMG signal.The LSTM module is used to extract timing information in the signal,and the CNN model is used to extract freatures from the signal.Secondly,the CNN network has strong abstract feature extraction ability for images.After contimuous wavelet transform,the sEMG signal can not only retain the timing information in the signal,but also the transformed parameters can be regarded as the spectrum images.We use the characteristics of wavelet transform to design the EMGNet network.The EMGNet network consists only of convolutional layers.Finally,we validated LCNN and CNN-LSTM on our own MyoDataset dataset which containing 4 volunteers and 5 gestures.The accuracy of the two models was 98.14% and 98.03%,respectively.At the same time,we compare the above three models on the online MyoArmbandDataset dataset and NinaPro DB5 dataset with the classic machine learning method.The experimental results show that the EMGNet,LCNN and CNN-LSTM networks have achieved a certain improvement compared to the classic machine learning method.At the end of this paper,we also explored the impact of the number of gestures categories on the classifier and the impact of the amount of gesture data on the classifier.