Research On Recognition Method Of Working Gesture Based On Arm Semg Signal

The surface electromyogram signal(sEMG)is generated by the movement and contraction of different muscle groups of the human body,and is a weak biological electric signal.Surface EMG signals are non-invasive,easy to collect,and to the human body without trauma.Therefore,they are widely used in many fields such as mechanical control,medical rehabilitation training,clinical medicine and sports science.This subject takes basic movements in mechanical assembly workings as the research object,collected and analyzed the characteristics of human arm surface EMG signals,and conducted theoretical and experimental research on working movement recognition methods,so as to effectively and quickly identify the working movements,lay the foundation for standardizing the working process and human-computer collaboration.Therefore,this thesis to carry out research in the following areas:(1)9 basic working gestures were extracted through research and analysis of the basic working movements involved in the assembly process of the reducer,and the goal was to standardize the assembly working movements.The thesis clarifies the problem and technical route of job action recognition based on sEMG signals,and analyzes the key problems that need to be solved,including surface EMG signal preprocessing,feature value extraction and selection,and the design of gesture recognition classifier.Aiming at these key issues,the basic process and solution of this subject research are proposed.(2)Aiming at the data collection and pre-processing the original sEMG signal of the working action,the 8-channel original sEMG data of 9 gestures were obtained by using the MYO sEMG sensor;the sixth-order Butterworth band-pass filter was used The original sEMG signal is filtered to reduce noise,reduce the interference between the sensors,and improve the recognition efficiency;the three detection methods of the sEMG signal active segment based on the improved short-term energy method,moving average absolute value method,and moving average energy method are studied.The detection method is analyzed and compared with the recognition accuracy.The recognition accuracy based on the moving average energy method is up to 98%.Therefore,this method is selected to perform preprocessing of the detection of the sEMG signal in the active segment,which laid the foundation for accurate recognition of subsequent working actions.(3)The surface EMG signal feature extraction and selection methods are studied,and the surface EMG signal characterization methods commonly used in time-domain features,frequency-domain features,and time-frequency domain features are analyzed;sliding windows are used to analyze the continuous signal in the active segment,the electrical signal is smoothed,and the processed data is averaged.A total of 15 feature variables and 120(8 channels x 15)feature values are extracted from the time domain,frequency domain,and time-frequency domain to characterize a certain gesture,and it is normalized;The XGBoost algorithm and the univariate feature selection algorithm are used to select and optimize the extracted feature values,and finally the intersection of the optimal feature sets obtained by the two methods is used as the feature variables for the job action recognition to achieve elimination The purpose of redundant features,improving recognition speed and accuracy.(4)The use of three supervised learning algorithms in machine learning,including K nearest neighbor algorithm,support vector machine and multi-layer perceptron,for the identification and classification of sEMG signals of work actions;comparative analysis of these three classifiers And optimize its parameters,and use cross-validation to evaluate the recognition accuracy of each algorithm.The final result shows that the recognition rate of the multilayer perceptron is higher than other algorithms,and its recognition rate is more than 98%;further on the algorithm and features The matching analysis in the time domain,frequency domain and time-frequency domain shows that MLP and time domain features have the best recognition effect.The results show that when the accuracy requirements are not particularly high,only the time domain features can fully meet the recognition needs.