Research On EMG Signal Gesture Recognition Method Based On Multi-channel Feature Fusion Network

In the field of human-computer interaction research,how to improve the accuracy of gesture recognition has always been a research hotspot.Among the existing research on gesture recognition based on Surface Electromyography(sEMG),one is based on deep learning,which mainly uses convolutional neural network to capture the spatial shape information of multi-channel sEMG,or uses long-short-term memory network to Extracts time-dependent information from single-channel sEMG without considering the correlation of electrode arrangement.The other is a method that combines traditional features with deep learning,mainly by stacking and replacing verified time-domain and frequency-domain features,but lacks the evaluation of feature importance.In response to the above problems,this thesis takes sEMG as the research object,aims to improve the performance of gesture recognition,and conducts research around the joint role of electrodes in different muscle regions and the importance of different traditional features.The main research contents are:(1)Using the SHAP(SHapley Additive explanations)interpretability method for feature screening,the sEMG feature set image was constructed.The SHAP method is used to score the importance of common features,and the main features with strong separability are screened out to construct feature set images,which provide more reliable data for model training.(2)A neural network framework for sEMG gesture recognition based on multi-stream feature fusion is proposed.The framework learns the relationship between different sEMG and specific gestures in stages,and extracts multidimensional spatial features from signal morphology,electrode space,and feature map space through multi-stream convolutional neural networks and residual convolutional attention modules.A recent view aggregation network fuses multistream features to improve the accuracy of gesture recognition.(3)A neural network framework for sEMG gesture recognition based on dualview fusion is proposed.The framework extracts the time-domain temporal and spatial features of the signal through the time-domain image sub-network,extracts the time-frequency domain statistical features of the signal through the feature-set image sub-network,and uses the time-domain image sub-network and the featureset image sub-network The dual-view feature layer and decision-making layer fusion strategy obtains the recognition results,which effectively improves the accuracy of gesture recognition.In this thesis,the proposed sEMG gesture recognition neural network framework is verified on Ninapro’s DB2 and DB4 sub-databases.The experimental results show that,compared with the existing EMG gesture recognition methods,the proposed method has higher recognition accuracy.The proposed method will provide a certain reference for promoting the academic research and application of gesture recognition based on sEMG signals.