Neural networks identification of muscular response using extended Hammerstein models

Many researchers are exploring nonlinear approaches, among them the use of neural networks, which can be used to approximate a general nonlinear system, to find a suitable control approach for Functional Electric Stimulation (FES). This paper proposes to use a particular neural network structure and to perform only an off-line identification, allowing a linear adaptive controller to cope with the time-variances. This goal is reached by taking a Hammerstein model and extending it to allow an input hysteresis. The whole time-variance is lumped in the linear part of the model. A neural network with the corresponding structure can then be trained off-line to produce the inverse of the nonlinear part, while a standard adaptive self-tuning controller can cope with the time-variance. Simulation results on actual measurements are shown to prove both the suitability of the approach as well as the need of an adaptive linear model