PREDICTION OF LOWER EXTREMITY JOINT ANGLES USING NEURAL NETWORKS FOR EXOSKELETON ROBOTIC LEG

摘要

Joint angles are one of the fundamental parameters to control the exoskeleton robotic leg. This research examines the performance of neural networks for the prediction of joint angles in various walking processes consisting of walking on the ground, walking on the treadmill, ascending and descending the stairs. A gait monitoring system was designed to collect gait kinematics and kinetics. The system consists of magnetic rotary encoders and force-sensitive resistors. Using these sensors, joint angles and foot contact states were obtained from a total of 40 healthy subjects. Moreover, subjects' demographic information such as age, sex, weight and height were recorded. Multilayer perceptron neural networks (MLPNNs) were used to predict future states of a leg movement by processing either only joint angles or both joint angles and foot contact states of the other leg. In addition to these two networks, a third MLPNN was designed with inputs from joint angles, foot contact states and demographic information of subjects. The results demonstrate that the overall prediction accuracy of 96% is achieved for the input data set consisting of joint angles and foot contact states.