Research On Fault Diagnosis Of Control System Based On Bayesian Network

Fault diagnosis technology has become an important research direction in the field of industrial process control in recent years. Although some achievements have been made on fault diagnosis of industrial processes, single fault diagnosis method still has some defects, and can not meet the diagnostic needs due to the increasing control accuracy requirements, the diversity of control algorithms and more complex external disturbances, etc. How to fully consider the system uncertainties, take full advantage of the diagnostic information to optimize fault diagnosis method is of practical significance.Bayesian network is a directed acyclic graph, which means a causal relationship between random variables to analyze and express uncertainty and probabilities. When applied in fault diagnosis, Bayesian network can conditionally depend on a variety of factors for decision analysis, and make inference from uncertain knowledge or information.This thesis studied fault diagnosis of control systems based on Bayesian network, the main work is as follows:First, signal processing-based Bayesian network was proposed for fault diagnosis. When an accurate mathematical model can hardly be identified for a complex control system, whereas the process data are obtained, the signal processing-based fault diagnosis methods would be integrated to build the Bayesian network. Also a parameter learning method was proposed for the diagnostic Bayesian network, in order to diagnose accurately the system faults that may occur. Tennessee Eastman benchmark simulation verified the effectiveness.Secondly, since some fault diagnosis methods are difficult to deal with the diagnosis problems as their models are often uncertain, the model-based fault diagnosis method was presented for system with model uncertainties. An Unknown Input Observer diagnostic subsystem is established to decouple the uncertain part of the model. The observer residual is integrated into the Bayesian network, and the posterior probability of the residual node is calculated. The failure probability from different resources is inferred. CSTR Simulation results illustrated the method.Finally, the model-based fault diagnosis scheme in Chapter4was verified in a pilot double-tank water level control system.