Research On The Adaptability Of EMG Signal In Gesture Recognition

The EMG signal is a weak physiological electrical signal generated between flexion and extension by the muscle groups related to the movement of the human body.Surface EMG has been widely studied in recent years due to its safer and more convenient acquisition method.At the same time,the motion information contained in the EMG signal has attracted the attention of researchers as a control signal source.How to improve the accuracy of surface EMG gesture recognition has become a hot spot in the research field.Due to the instability and difference of surface EMG signals,there are many problems and challenges in the control system based on surface EMG signal gesture recognition.Given the current problems,the main research contents of this paper are as follows:(1)This paper firstly investigates the status of EMG signal acquisition and EMG signal gesture recognition at home and abroad and summarizes the current problems of EMG signal gesture recognition.Because of the existing problems,this paper designed a long-term EMG signal acquisition experiment,the time span was nearly 2 months,and the EMG signals of 10 movements of 8 subjects were collected.Next,the time domain and frequency domain characteristics of EMG signals at different times were explored.The experimental results show that EMG signals have different characteristics in time domain and frequency domain at different times,which affects the accuracy of EMG signal gesture recognition rate.(2)In this paper,after filtering the collected EMG signal with the Butterworth filter,the active segment in the EMG signal is extracted using the sliding window method.Next,this paper designs three different deep learning network structures to perform gesture recognition experiments on EMG signals.The experimental results show that the three networks with different structures can make the off-line recognition accuracy of EMG signals reach more than98%.Next,the influence of the time-varying EMG signal on the classification accuracy of EMG gesture recognition was explored.This paper designed a cross-experiment between EMG signals at different times,so that one day of EMG data was used as the training set,and the other day was used for testing.The results show that the performance of the EMG gesture recognition classifier drops by 80% due to the time-varying effect.At the same time,this paper proposes the use of transfer learning to reduce the impact of time-varying,and transfer learning can shorten the time for model retraining.It has been proved by experiments that the accuracy of the gesture recognition classifier can be restored to 90% by collecting the EMG signals of the action sequence three times again.Compared with this,the retraining time is reduced by50%.(3)Finally,this paper designs an EMG signal gesture recognition control system,which has the characteristics of high aggregation and low coupling.This system can switch different classifier algorithms and different controlled terminals.After training the basic model using the basic training set,the user’s 3 action sequence data is collected again as an additional data set to train the basic model,and finally the EMG signal gesture recognition control system can be used.In this paper,the Bluetooth car is used as the controlled terminal,and four gesture actions are used to control the Bluetooth car.The final experimental results show that the accuracy of gesture recognition reaches 90%,which verifies the availability of the EMG signal gesture recognition control system.