Fast neural system identification: application to rehabilitation engineering

To develop a powered hybrid FES (functional electrical stimulation) orthosis for paraplegics, fast system identification techniques play a crucial role to obtain information about the neuromuscular system. However, the underlying systems have nonlinear and coupled dynamics with uncertainties in practice. Artificial neural networks provide such adaptive nonlinear models and can be well extended to system identification or prediction problems. In this research, fast and robust online identification using a multi-layer recurrent neural network is tackled in two steps: a multi-layer network architecture through a supervisory-signal and error recurrent neural network (SERNN); and robust training through a modified recursive least-squares algorithm, having a dynamic forgetting factor and an optimal teaching signal for the hidden layer.