Using neural networks and statistical tests for detecting changes in the process dynamics

Abstract

In this paper, a real-time system for detecting changes in dynamic systems is designed. This study involves the determination and identification of black-box models, the design of a model-based residual generator and finally the evaluation of the residuals via two statistical information tests; the Page-Hinkley test and the Bayes classifier. It is intended to reveal any drift from the normal behaviour of the process. The process behaviour under its normal operating conditions is established by a reliable model. In order to obtain this reliable model for the process dynamics, the black-box identification by means of a NARX (Non-linear Auto-Regressive (AR) with exogenous input) model has been chosen in this study. It is based on the neural network approach. This paper shows also the choice and the performance of the neural network in the training and test phases. An analysis of the inputs number, hidden neurons and their influence on the behaviour of the neural predictor is carried out. Three statistical criteria, Aikeke's Information Criterion (AIC), Minimum Description Length (MDL) and Bayesian Information Criterion (BIC), are used for the validation of the experimental data. After describing the system architecture and the proposed methodology of the fault detection, we present a realistic application in order to show the technique's potential. The purpose is to detect the occurrence of change, and pinpoint the moment it occurred.