Research On Hand Gesture Recognition Based On SEMG Signal Under Fatigue State

The surface electromyogram(sEMG)is a weak and complex biological signal generated by the human brain’s motor commands transmitted through the nervous system,causing muscle contraction and relaxation that ultimately result in limb movement and detected on the skin surface.sEMG has advantages such as non-invasiveness,simple collection methods,and accurate reflection of limb movement characteristics.In recent years,sEMG-driven bionic prosthetics technology has been extensively studied.In the field of bionic prosthetic control,the technology of gesture recognition based on sEMG has been a research hotspot,and numerous researchers have made achievements in this area.However,most researchers have only focused on sEMG-based gesture recognition research in non-fatigue,static states,while neglecting the influence of muscle fatigue on the accuracy of gesture recognition.Muscle fatigue is a common physiological phenomenon,often characterized by a temporary decrease in the ability of muscles to produce maximum force or output power after sustained or intermittent exercise.This decrease can result in a noticeable weakening of muscle strength,which inevitably affects the sEMG signals generated during hand gestures.This paper explores the impact of muscle fatigue on the accuracy of sEMG-based hand gesture recognition,and proposes a data augmentation method based on Generative Adversarial Networks(GAN)to enhance sEMG signals in a fatigued state and mitigate the negative effects of muscle fatigue on gesture recognition.This study investigates sEMG-based hand gesture recognition under fatigue conditions as follows:(1)Fatigue and non-fatigue sEMG acquisition.The sEMG signals of four muscle groups,including the brachioradialis,extensor digitorum,extensor carpi radialis,and flexor carpi radialis,were collected from 8 participants performing 7 common hand gestures using the MP160 multi-channel physiological signal recorder from Biopac Inc.Two datasets were acquired from non-fatigued and fatigued states through the implementation of fatigue-inducing experiments.(2)Active segment detection and sEMG denoising.Two methods,namely the sliding window absolute mean method and the improved short-time energy multi-threshold method,were employed for the detection of activity segments.Results indicated that the latter method exhibited stronger anti-noise capability and higher accuracy.Through noise analysis of the collected sEMG signals,it was discovered that there was powerline interference at the 50 Hz frequency and its harmonics,as well as equipment noise at 59 Hz.A comparative study suggested that the adaptive filter based on normalized least mean square(NLMS)algorithm was more effective at filtering the 50 Hz powerline interference.Therefore,a hybrid filter consisting of the NLMS filter and the Butterworth filter was utilized to eliminate the powerline interference at 50 Hz and the equipment noise at 59 Hz,as well as the unwanted frequency components below 20 Hz and above 500 Hz.(3)Explore the influence of muscle fatigue on gesture recognition based on sEMG.Four distinct methods have been presented for selecting training and testing sets,and seven hand gesture actions have been classified and recognized using three conventional machine learning algorithms K-nearest Neighbor(K-NN),Support Vector Machine(SVM),Decision Tree(DT)and a deep learning algorithm,Deep Residual Network(Res Net).These methods have demonstrated that muscle fatigue diminishes the accuracy of gesture recognition.Specifically,the sEMG testing accuracies under non-fatigue conditions were 96.7%,89.0%,87.3%,and 97.5%for the K-NN,SVM,DT,and Res Net classifiers trained under non-fatigue conditions,respectively.However,under fatigue conditions,the testing accuracies decreased to 53.3%,55.4%,45.8%,and 64.8%,respectively.(4)Overcoming the effects of muscle fatigue.A method for suppressing muscle fatigue based on GAN is proposed,which involves transforming sEMG signals under fatigue conditions into those under non-fatigue conditions by means of data augmentation,and subsequently classifying them using a trained classification model based on non-fatigue sEMG signals.Experimental results show that the testing accuracy in fatigue conditions is improved by over 20%,with recognition rates of 72.3%,80.9%,69.9%,and 92.1% achieved by four classifiers,indicating the effectiveness of the proposed method in overcoming the negative impact of muscle fatigue on gesture recognition accuracy and improving the robustness of the sEMG-based gesture recognition model.