SEMG-based continuous estimation of joint angles of human legs by using BP neural network

Abstract

In this paper, we propose an mth order nonlinear model to describe the relationship between the surface electromyography (sEMG) signals and the joint angles of human legs, in which a simple BP neural network is built for the model estimation. The inputs of the model are sEMG time series that have been processed, and the outputs of the model are the joint angles of hip, knee, and ankle. To validate the effectiveness of the BP neural network, six able-bodied people and four spinal cord injury (SCI) patients participated in the experiment. Two movement modes including the treadmill exercise and the leg extension exercise at different speeds and different loads were respectively conducted by the able-bodied individuals, and only the treadmill exercise was selected for the SCI patients. Seven channels of sEMG from seven human leg muscles were recorded and three joint angles including the hip joint, knee joint and the ankle joint were sampled simultaneously. The results present that this method has a good performance on joint angles estimation by using sEMG for both able-bodied subjects and SCI patients. The average angle estimation root-mean-square (rms) error for leg extension exercise is less than 9°, and the average rms error for treadmill exercise is less than 6° for all the able-bodied subjects. The average angle estimation rms error of the SCI patients is even smaller (less than 5°) than that of the able-bodied people because of a smaller movement range. This method would be used to rehabilitation robot or functional electrical stimulation (FES) for active rehabilitation of SCI patients or stroke patients based on sEMG signals.