Research On SEMG Gesture Recognition Algorithm Based On Multi Feature Fusion

Electromyogram(EMG)is an important bioelectrical signal generated by muscle activity.The detection and analysis of EMG signal is of great significance to clinical diagnosis,human-computer interface and rehabilitation exercise.At present,in many related research fields,the EMG signal gesture recognition has been the focus of research.Traditional researches in this field mostly use machine learning methods to classify gesture recognition.In recent years,with the advent of the application boom of deep learning,many researchers began to apply deep learning algorithm in this field and achieved better results than the traditional methods.However,there are still some limitations in the current research,such as the loss of original data information in the process of deep learning training,and the training data sample size of real scene is difficult to meet the training requirements.In view of the above problems,this thesis has carried on the research and has done the following work.(1)When using traditional deep learning methods for EMG gesture recognition,there is a problem that when using Convolutional Neural Network(CNN)training,the extracted features are continuously abstracted and feature information is lost as the number of layers deepens,resulting in insufficient recognition accuracy.Aiming at this problem,this thesis optimizes the classic convolutional neural network,and combines the inherent characteristics of the EMG signal,and proposes an EMG gesture recognition model based on multi-feature fusion CNN.The model uses a parallel network architecture to process the split EMG signal data.In the iterative process of the network,the shallow features in the initial training stage,the intermediate features in the middle of the network,and the deep features after the convolution training are fused and extracted;finally,the fused feature values are input into the traditional machine learning classifier to replace the Soft Max layer in the neural network for classification,and the final EMG gesture recognition result is obtained.The simulation results on the Nina Pro EMG database show that the algorithm has better robustness and can also obtain better classification accuracy.(2)In a small-sample training scenario,when the target data set has too few samples,the direct use of deep learning methods may lead to overfitting and greatly reduce the recognition rate.To solve this problem,this thesis uses the triple siamese neural network method in meta-learning,combined with the feature fusion module,and optimizes the surface EMG signal features at the same time,selects an appropriate convolution kernel to extract more accurate signal features,and the final recognition result is obtained by calculating the distance relationship between samples.The simulation experiment results on multiple small sample training data segmented based on the Nina Pro data set show that compared with the ordinary neural network model,the use of this model can effectively improve the accuracy of gesture recognition in small sample scenarios.