Gait control system for functional electrical stimulation using neural networks

In functional electrical stimulation (FES) systems for restoring walking in spinal cord injured (SCI) individuals, hand switches are the preferred method for controlling stimulation timing. Through practice the user becomes an 'expert' in determining when stimulation should be applied. Neural networks have been used to 'clone' this expertise but these applications have used small numbers of sensors, and their structure has used a binary output, giving rise to possible controller oscillations. It was proposed that a three-layer structure neural network with continuous function, using a larger number of sensors, including 'virtual' sensors, can be used to 'clone' this expertise to produce good controllers. Using a sensor set of ten force sensors and another of 13 'virtual' kinematic sensors, a good FES control system was constructed using a three-layer neural network with five hidden nodes. The sensor set comprising three sensors showed the best performance. The accuracy of the optimum three-sensor set for the force sensors and the virtual kinematic sensors was 90% and 93%, respectively, compared with 81% and 77% for a heel switch. With 32 synchronised sensors, binary neural networks and continuous neural networks were constructed and compared. The networks using continuous function had significantly fewer oscillations. Continuous neural networks offer the ability to generate good FES controllers.