Detecting Sensor Faults for a Chemical Reactor Rig via Adaptive Neural Network Model

Abstract

An adaptive neural network model based approach to sensor fault detection is proposed for multivariable chemical processes. The neural model is used to predict process output for multi-step ahead with the prediction error used as the residual, while the model is on-line updated to capture dynamics change. The recursive orthogonal least squires algorithm (ROLS) is used to adapt a radial basis function (RBF) model to reduce the effects of data ill conditioning. Two error indices are developed to stop the on-line updating of the neural model and its corresponding threshold is used to distinguish the fault effect from model uncertainty. The proposed approach is evaluated in a three-input three-output chemical reactor rig with three simulated sensor faults. The effectiveness of the method and the applicability of the method to real industrial processes are demonstrated.