SEMG-based prediction of human lower extremity movements by using a dynamic recurrent neural network

Abstract

In this paper, a novel robust nonlinear model is proposed to predict human lower extremity motion based on the multi-channel surface electromyography (sEMG) signals. The prediction model is established by a data-driven dynamic recurrent neural network. The sEMG signals acquired from human lower extremity muscles are used as the inputs of the prediction model. The outputs of the model are joint angles of hip, knee and ankle. Different from the traditional feedforward network structure, this model has several feedback loops, thus it can take advantage of the output feedback information. To validate the effectiveness of the proposed method, five able-bodied people participated in the cycling exercises and relevant data were recorded in real time. The performance of the proposed prediction model is compared to those of the feedforward neural network with augmented inputs (FFNNAI) for the motion prediction accuracy and robustness. The results show that the proposed method provides acceptable performance which is clearly better than the FFNNAI-based approach under different experimental schemes.