Motion intention recognition from multi-channel surface electromyography through prediction of major muscle activation

Abstract

Recognizing human motion intentions from the human is very important in human-robot interaction applications to ensure stability. Electromyography (EMG), which measures a muscle’s electrical activity, has the advantage of high-speed synchronous control when applied to the wearable robot because it can speed up motion intention recognition. However, the sEMG signal characteristics are easily changed due to electrode location sensitivity and subject sensitivity. Changes in sEMG signal characteristics require a heuristic calibration for each wear, and it has been difficult to recognize human intention and generate a robot control signal. This thesis proposes a torque estimation method using multi-channel sEMG through the prediction of major muscle activation and aims to apply it to a wearable robot through torque-based motion intention recognition. The proposed method consists of pre-processing for predicting activation of the major muscle using multi-channel sEMG and post-processing for estimating joint torque using the major muscle signal. Prediction of the major muscle activation is a stochastic signal decomposition of the sEMG signal in static contraction motion of 80% MVC force, and the sEMG signal is orthogonally transformed in an uncorrelated state. This muscle signal decomposition model distinguished the major muscle signal for electrode location variation and inter-subject problems. The joint torque estimation model using the major muscle activation signal was made with a time-delay neural network (TDNN) that processes a dynamic time-series signal. To verify the proposed pre-processing model, the joint torque estimation model was used when the electrode location and the subject changed and compared with the basic method without applying the major muscle activation model. For the electrode location variation in an isometric contraction, the torque estimation accuracy was 87%, 55% larger than the basic method. For the inter-subject variation with isometric contraction, the torque estimation accuracy was 86%, 35% larger than the basic method. Verification by ten subjects in isokinetic elbow contraction, the proposed method had an accuracy of 81% when there was an electrode shift, and an error of 1.9% increased than before the shift. This was more robust to electrode shift than the existing multi-channel torque estimation method. The sEMG-based predicting of the major muscle activation showed that the torque estimation accuracy was improved for the electrode location variation and inter-subject variation. The proposed major muscle activation prediction and application overcome the limitations of sEMG-based motion intention recognition and can also be used in wearable robots and human-robot interfaces.