A Robust Sensor Fault Diagnosis System Using Neural Networks and a Dynamic Process Model

Abstract

Back propagation (BP) neural networks are applied for reactor fault diagnosis. Analyzed is the output prediction error between a neural network model and an independent dynamic process model, which serves as a residual for diagnosing actuator/component/sensor faults. It is found that the neural network without the process model is less sensitive to sensor faults than actuator or component faults. A scheme is developed utilizing a second neural model to analyze the difference between the output prediction of the process model and the neural network for diagnosing sensor faults in a simulated reactor. Results from the reactor fault diagnosis system are presented to demonstrate the satisfactory detection and isolation of sensor faults using this approach.