A CNN-SVM Combined Regression Model for Continuous Knee Angle Estimation Using Mechanomyography Signals

Compared with the pattern recognition of discrete human motions, the continuous human motion estimation is more significant for the motion control of wearable power-assisted robots. In previous studies, researchers estimated human motion from the surface electromyography (sEMG) signal based on the Hill-type muscle model or using the traditional machine learning algorithms with the hand-crafted features. However, these methods generally require the domain knowledge about the muscle dynamics or complicated signal processing. In this study, 3-channel time series mechanomyography (MMG) signals were detected on clothes from the thigh muscles, and a CNN-SVM combined regression model was proposed to estimate the knee angle under different motion velocities. The convolutional neural network (CNN) model was used to automatically extract the features from the MMG signals, and the support vector machine (SVM) regression was used to process the features for angle estimation. The results showed that the CNN-SVM combined regression model obviously improved the estimation performances and avoided using the muscle model or hand-crafted features. The methods used in this paper would promote the development of motion control system for wearable power-assisted robots, and could be further extended to the fields of rehabilitation training, medical diagnosis, etc.