Parameter identification of timevarying systems using neural networks
Abstract
A neural network solution is introduced for the problem of parameter identification of timevarying systems. An autoregressive (AR) model is used to identify the unknown system. The problem is formulated to minimize the least squares (LS) error between the output of the unknown system and the output of the AR model when they are driven by a white Gaussian noise. Because of the timevarying nature of the unknown system, an exponentially weighted LS scheme is employed to attach more weight to the current data. A neural network is designated to estimate the solution of the weighted LS problem that gives the parameter of the AR model. Since this is an unconstrained quadratic optimization problem and the AR parameters can assume either + or  values, a hyperbolic tangent function is used as the nonlinear activation function for the neurons. A space iterative search technique is incorporated into the computational procedure of the neural network such that it can find solution in an open space rather than within a hypercube. In order to identify the timevarying parameters efficiently, recursive formulations are derived so that the coefficient of the neural network can be updated whenever a new observation is available. Finally, computer simulation results are presented to demonstrate the performance of the scheme.
 Publication:

Electrical and Computer Engineering, Volumes 1 and 2 4 p (SEE N9330215 1131)
 Pub Date:
 1990
 Bibcode:
 1990ecev.confR....G
 Keywords:

 Autoregressive Processes;
 Neural Nets;
 Parameter Identification;
 Weighting Functions;
 Algorithms;
 Computerized Simulation;
 Least Squares Method;
 Optimization;
 Random Noise;
 Recursive Functions;
 Electronics and Electrical Engineering