Research On SEMG Gesture Recognition Algorithm Based On Deep Learning

Surface electromyography(s EMG)signal,as one of the bioelectrical signals,is closely associated with the behavioral actions of the human body.By analyzing the EMG signal,researchers can identify human movements and behavioral intent.In recent years,gesture recognition based on s EMG signals has gradually become a frontier research direction in human-computer interaction(HCI)technology,which has been widely applied in industrial production,rehabilitation medicine,virtual reality and other fields.With the increasingly demanding requirements for recognition accuracy in these application scenarios,deep learning networks,represented by convolutional neural network(CNN),have attracted the attention of researchers.It has been shown that deep learning networks have good applicability in dealing with the problems of s EMG gesture recognition,but there are also some shortcomings that need to be solved urgently.To further improve the recognition accuracy,this thesis applies the deep learning network to solve the problem of gesture recognition based on s EMG signals.The main research contents are phrased as follows.(1)For the reason that existing s EMG gesture recognition algorithms based on deep learning uses a single network architecture,the model may ignore the strong correlation between gesture actions and muscle tissue during feature extraction or lose shallow feature information as the number of network layers deepens,which limits the accuracy of gesture recognition.Aiming at the problems mentioned above,this thesis proposes an s EMG gesture recognition algorithm based on dual-branch multi-stream network,which applies multi-stream fusion strategy and combines CNN and RNN architecture.The constructed network model consists of two branch network modules,which analyze the correlation between s EMG signals and corresponding gestures from the global and local perspectives respectively.On this basis,the algorithm combines the outputs of the two branch networks through a fusion network architecture to determine the gesture result.The experimental results on multiple NinaPro datasets demonstrate that the proposed algorithm achieves higher recognition accuracy than existing algorithms and achieves better recognition performance on different types of datasets.(2)For the Few-shot Learning(FSL)problem,the quantity of s EMG signal samples obtained from the acquisition are often too small due to the limitation of the acquisition conditions.Due to the lack of sufficient amount of supervised information,the recognition performance of the deep learning model which trained by using a limited number of samples directly will drop abruptly.To solve the problem mentioned above,this thesis introduces the idea of image data enhancement then proposes a data enhancement method for the FSL problem.The proposed method optimizes the s EMG gesture recognition algorithm,which expands the number of training samples through data cropping transformations and s EMG sample generation network based on variational auto-encoder(VAE),alleviating the problem of insufficient number of original samples.The experimental results on multiple independent subjects show that the optimized s EMG gesture recognition model can overcome overfitting problem significantly and achieve the improvement of recognition accuracy when dealing with FSL problem.