An Integerated Estimation and Control Scheme for Piece-Wise Linear Systems using Neural Networks

Identification of the parameters and the structure of nonlinear discrete-time system models, in the joint frequency-position space, is investigated by using neural networks. A dynamical neuron model is first introduced. The neuron characteristic function NCF of the model, at higher values of signal to noise ratio SNR, resembles Gabor basis fuctions GBF. A simplification to the GBF´s is also presented, where the spatial Gaussian envelope of GBF´s is replaced with a triangular one. The neuron model is then encompassed in a three-layered neural net for parameter and structure identification. Simulation results, necessary modifications to improve convergence of the network, and possible extensions for this work are provided.