A Neural Network Approach to Electromyographic Signal Processing for a Motor Control Task

Abstract

A hybrid modeling structure composed of a one degree of freedom computational musculo-skeletal model and a multilayer perceptron neural network was used to effectively map electromyography (EMG) from a human exercise trial to muscle activations in a physiologically feasible and accurate fashion. Several configurations of the complete hybrid system were used to map four muscle surface EMGs from a ballistic elbow flexion to normalized muscle activations, estimated individual muscle forces and torque about the joint. The net joint torque was used to train the neural portion of the hybrid system to minimize kinematic error. The model allowed the estimation of the nonobservable parameters: normalized muscle activations and forces which was used to penalize the learning system. With these parameters in the learning equation, our system produced muscle activations consistent with the classic triphasic response present in ballistic movements.