Fault Detection and Isolation in Stochastic Nonlinear Systems using Unscented Particle Filter based Likelihood Ratio Approach

A novel unscented particle filtering based log-likelihood ratio (LLR) approach for Fault detection and isolation (FDI) in nonlinear stochastic systems is proposed. It is well-known that the particle filter (PF) is used for the state estimation of nonlinear and non-Gaussian system but the key step in the filter design is the selection of a suitable proposal density to represent the true posterior density. The traditional PF when used for the FDI problem does not always guarantee that all particles lie in the likelihood region as the proposal density for this filter is independent of the measurement data. The new approach utilizing PF with unscented Kalman filter (UKF) proposal to solve the FDI problem assures that the estimated states (particles) lie within the high likelihood region because the proposal density in the unscented particle filter (UPF) is dependent on the current measurement. The detection and isolation of faults are carried out through maximum likelihood estimation and hypothesis testing method. The efficacy of the proposed method is demonstrated through an implementation on two highly nonlinear systems- a chemical reactor system and a three phase induction machine. The simulation results obtained from this method are compared with that of FDI technique using the generic particle filtering algorithm as state estimator.