A probabilistic framework for reference design for guaranteed fault diagnosis under closed-loop control

Active fault diagnosis (AFD) is crucial for safe, reliable, and high-performance operation of complex technical systems. For linear systems with uncertainties bounded within zonotopes, this paper addresses the AFD problem under closed-loop control. The AFD method reported in [17] is extended to design the reference signal of a feedback controller such that fault diagnosis is guaranteed within a prespecified time horizon. Designing the reference signal of a controller, instead of directly designing the system input, will enable application of the AFD method to systems with existing feedback controllers. The reference design problem seeks to separate the most likely model from the rest of the fault model candidates. The presented AFD method is implemented in a moving-horizon fashion through estimating the probability that each nominal/fault model can be active via sample-based Bayesian estimation. It is shown that the AFD method can enhance fault diagnosis in terms of reducing the number of time steps required for guaranteed model separation.