Fault Detection and Diagnosis of Multiple-Model Systems With Mismodeled Transition Probabilities

This paper proposes an improved interacting multiple-model (I²MM) algorithm with inaccurate transition probabilities (TPs) for fault detection and diagnosis (FDD). We first study the influence of inaccurate TPs by inspecting the expectation and covariances of residual error vectors in the traditional IMM method. It shows that Kalman filters can be retained as subfilters in the presence of mismodeled TPs, and the effect of TPs can be removed naturally if all of the probabilities of true modes are equal to one. In view of this, a modification operator is proposed to make the real mode probabilities heuristically approach one. The modification degree is governed by a parameter determined by the online measurements. When the modification parameter calculated is identical to one, the I²MM method reduces to the conventional IMM algorithm. An experiment designed through a ball-and-tube testbed is presented to demonstrate that the I²MM-based FDD method can provide more reliable FDD results and reduce the possibility of false alarms.