Feasibility of EMG-based ANN controller for a real-time virtual reality simulation

Estimation of the joint angle from the surface electromyography (sEMG) is a quite complex task due to the complicated relationship between the kinematical variables and the raw sEMG. In this paper, we build a sEMG-to-motion model with the artificial neural network (ANN). EMG features, including Integral of absolute value (IAV), Zero crossing (ZC), Auto-regression coefficients (ARC), Median frequency (MDF), are extracted as the input of the ANN, and the output of the ANN is the operator´s elbow joint angle and the wrist motion. In addition, a 3D upper extremity model, which is built in SolidWorks and then transformed into MATLAB, will imitate the operator´s motion simultaneously with the estimations of the ANN. Thus, we accomplish a virtual reality system to realize the real-time simulation and validate the effectiveness of the sEMG-to-motion model. Experiment results show that the system achieves well in model accuracy, hardware compatibility and real time performance with a small mean square error of 1.921 degrees.