Neural networks performing system identification for control applications

In this paper the ability of a neural network to perform a multidimensional curve-fitting is used to let a multilayer perceptron system identify a nonlinear multivariable dynamic process. The identified model is the well-known innovation state space model (Kalmann predictor). The identification is based only on input/output measurements, so in fact the extended Kalmann filter problem is solved. The paper describes how the multilayer perceptron is structured, and two training methods for the recurrent network structure are mentioned; the recursive prediction error method using a Gauss-Newton search direction, known from linear system identification theory; and a modified backpropagation error algorithm allowing normal `static´ backpropagation in training recurrent networks. Finally, the model and training methods are tested on a noisy, strongly nonlinear, dynamic process, showing excellent results for the trained net to act as an actual system identifier, predictor and simulator (filter). The trained net allows actual online extraction of the parameter matrices of the model giving a basis for better control of the nonlinear process