System Identification Using ARMA Modeling and Neural Networks

Abstract

Stochastic system identification is an important tool for control of discrete dynamic systems. Among the modeling strategies developed for this purpose, Auto Regressive Moving Average (ARMA for discrete systems) models offer an accurate identification technique. The disadvantage with these models are that they are extremely complicated to implement on-line, especially for nonlinear time-variant systems. This paper utilizes a Neural Network structure for identification of stochastic processes and tracks system dynamics by on-line adjustments of network parameters. Neural dynamics is based on impulse responses and an iterative learning algorithm is derived using conventional principles of gradient descent and backpropagation. The learning algorithm is analyzed and shown to be fast and accurate in the identification of parameters for stochastic processes in both time-invariant and time-variant cases.