Research Of Fault Diagnosis Method Based On Bayesian Inference

With the development of computer science and control engineering, the scale and complexity of the systems have increased rapidly. The issue, how to improve the basic performances such as efficiency, reliability, stability and safety is increasingly significant. This is the main reason for the speedy development of fault diagnosis technology. To avoid the production safety accidents, the faults are supposed to be detected and isolated in time. Recently, fault diagnosis combining a quantity of advanced algorithm and technology has become more efficient theoretically and practically. The sorts of fault diagnosis methods are becoming significantly diverse. It is still a hot topic that to built up a robust analytic scheme to diagnosis the faults, especially for the nonlinear, non-Gaussian systems.The method of state estimation based on filtering can estimates the state of the system in noise environment. So it is fit to fault diagnosis of nonlinear systems particularly. More attention on filtering technology based on Bayesian inference has been raised recently. Bayesian inference synthesizes the information of prior probabilities and observers to diagnose the system. It provides an easy but instructional technology for fault diagnosis. Particle filter is an effective application based on Bayesian inference which combines recursive Bayesian estimation and Monte Carlo method. In it, a group of random sample particles with corresponding weights to propagate forward in time sequence through the system model, to get all kinds of the system states and to seek for the system state of the posterior probability density function finally. In this thesis, the fault diagnosis methods based on particle filter is proposed for the tank control system.The primary task of fault diagnosis is to establish a robust residual generator which can generate the residual correctly without the effects of disturbance. The key tasks of fault diagnosis are:fault detecting, fault isolating and fault identification. Particle filter is used for the fault detection in nonlinear, non-Gaussian systems in this thesis. It utilized two methods both based on particle filter:likelihood probabilities function and smoothed residual method to detect the fault. Then compared the results of simulating based on both two methods, we can find the advantages and disadvantages of the two methods. After the fault has been detected, a method combining particle filter and multiple-models is proposed for the fault isolating. The method of multiple-models is an adaptive one for solving the problems in uncertain systems effectively. Particle filter is designed for each of the fault models, and then we can isolate the position of the fault by comparing the deviation to match the fault model. The effects of fault diagnosis can be improved through this method. In this thesis, fault set was established according the model of the system. Then comparing the state values with the fault set, the fault can be isolated when they are matching with each other.The results of simulations based on tank control system and one-dimension nonlinear system manifest that the method of likelihood probabilities function can detect the fault immediately. The method of multiple-models particle filter can isolate the fault correctly. The research results can be used for diagnosing faults rapidly and correctly in nonlinear, non-Gaussian systems.