Using neural network with virtual sensors to generate optimum FES gait controllers

Abstract

In Functional Electrical Stimulation (FES) control systems, artificial intelligence has been employed for feedback or adaptive control to assist paraplegic walking. Neural networks with a three-layer structure can be used to generate control replacing the manual control to deliver the stimulation during walking. Sensors which have been used to provide information for the controller range in complexity from simple heel or hand switches to accelerometers. There are three basic problems connected with the selection of sensors: sensor types, number of sensors and the optimum location on the limb. Kinematic signals can be simulated from 3D data collected from a motion analysis system. These 'virtual' sensors (goniometers, gyroscopes, inclinometers and accelerometers) showed a good correlation with their physical counterparts. The aim of this study was to use neural network to generate optimum FES controllers. 32 sensors ('virtual' kinematic sensors and physical sensors recording crutch forces and foot floor contacts) were used to find an optimum sensor set. The results have shown that neural networks with a small optimum sensor set could produce a robust controller with a higher degree of accuracy than a traditional heel switch controller. After few months, the controller still maintained a high accuracy.