Fault Diagnosis Research Based On The Expectation Propagation Algorithm

ABSTRACT:With the development of science and technology, the processing ability of actual control system and the level of modernization have gradually improved, the scale and investment of overall system and complexity are higher and higher. Once Large-Scale Complex System fails, it will cause huge property losses and casualties directly. So in the past few decades, fault diagnosis problem got wide attention of scholars both domestic and abroad. In2001, Thomas p Minka presents a new family of algorithms for approximate Bayesian inference, he pointed out that the primary computing tasks in Bayesian inference is numerical integration, but the biggest obstacles to use Bayesian methods has been its computational expense. As a result, he put forward a new family of algorithms for approximate Bayesian inference---Expectation Propagation, compared to before inference algorithms faster and more accurately with less amount of calculation. Using Expectation propagation combined with Bayesian theory to perform fault diagnosis can improve the efficiency of fault diagnosis and have important theoretical and practical significance for improving the existing fault diagnosis methods. According to Expectation Propagation algorithm, this thesis puts forward a new method for fault diagnosis named Extended Kalman Smoothing with Expectation Propagation (EP-EKS) and does some research for the fault diagnosis method based on EP-EKS.The main work of this thesis is:1. We present a new method of fault diagnosis, namely Expectation Propagation Extended Kalman Smoothing fault diagnosis method, according to the Expectation Propagation algorithm (EP) and related theory in this paper, then we elaborate the principle of fault diagnosis method of EP-EKS theoretically.2. We use the traditional extended kalman filter (EKF), unscented kalman filter (UKF), and the new extended kalman smoothing with expectation propagation (EP-EKS) in this thesis combined with the residual vector method for fault diagnosis and use a univariate nonstationary growth model (UNGM) as the research object, by setting the system parameters changed in the process to simulate the failure. The simulation results show that EKF and UKF can’t achieve the goal of fault diagnosis when the parameters changed in small scope, but EP-EKS can accurate diagnosis; and for the same fault, the diagnosis results of EP-EKS are better than EKF and UKF in veracity (using missalarm and faultalarm as index), accuracy (using RMSE and NMAD as index) and efficiency (using simulation time as index), so we consider that EP-EKS can improve the efficiency of fault diagnosis.3. In order to verify the theoretical feasibility of the method, we use the water level/temperature control system as the research object, by setting the switch-parameters changed to simulate failure. Because the application principle of EKF is local linearization and use the first derivative as the approximation at the nonlinear point, but the observation equation of water level/temperature control system exists non-differentiable points, so we use particle filter instead of EKF. We also use the UKF, PF and EP-EKS combined with the residual vector method for fault diagnosis. Simulation results show that the UKF, PF and EP-EKS all can achieve the goal of fault diagnosis, considering the precision, accuracy and timeliness comprehensively, EP EKS is better than UKF and PF, so we consider it can improve the efficiency of fault diagnosis.The results show that the method based on Expectation Propagation for fault diagnosis is feasible, and by introducing Expectation Propagation algorithm we can reach the purpose to improve the efficiency and these works provide reference on the Expectation Propagation Research, especially on the fault diagnosis research.