Multi Model Fusion Based Prediction of Human Joint Angles Using sEMG and Historical Angles

Active rehabilitation training for patients with lower limb motor dysfunction mainly depends on exoskeleton robots that can accurately and timely identify human motion intentions, which can be implemented by continues estimation of human joint angles. In this paper, a multi model fusion based ridge regression named as 'MMF-RR' for the prediction of human joint angles is proposed. Specifically, four selected basic learners are first end-to-end trained for the subsequent analysis from the data set composed of surface electromyography signals (sEMG) and historical angles. Then the results are spliced into ridge regression with penalty terms for joint angle prediction. The proposed MMF-RR was evaluated on publicly available dataset for 22 participants, where half of them were healthy participants, while the other half had various knee joint diseases. And the results show that the proposed MMF-RR can significantly improve the predicting performance. In particular, the average mean absolute error (MAE) and average mean square error (MSE) of knee joint angle prediction of healthy participants and participants with knee joint lesions were 6.73%, 2.36% and 8.97%, 3.26%, respectively, which indicate that MMF-RR can provide more precise human motion intention information for lower limb motor under rehabilitation training than five comparison methods.