Abstract

In this thesis, a framework for multi-resolution fault diagnosis in discrete-event systems (DES) is introduced. Here a sequence of plant models, with increasing resolution, are used in fault diagnosis and the range of possible diagnosis is narrowed down step by step, until the failure node is isolated. In this way, the original problem of fault diagnosis is replaced by a sequence of smaller problems. The plant models used at each step of diagnosis are abstractions of the original plant model. We propose to use model reduction through the solutions of the Relational Coarsest Partition problem to obtain these abstractions. For each diagnosis step, minimal sensor sets are chosen to have a coarser output map, and hence, to improve the efficiency of model reduction. In this thesis, a polynomial algorithm is proposed that verifies failure diagnosability by examining the distinguishability of two plant (normal/faulty) conditions at a time. A procedure is presented that finds minimal sensor sets, referred to as minimal distinguishes for distinguishability of one condition from another. A polynomial procedure is introduced that combines minimal distinguishers to obtain a minimal sensor set for fault diagnosis. The proposed method reduces the computational complexity of sensor selection. A benefit of using minimal distinguishers is that their computation maybe speeded up using expert knowledge. The proposed method for sensor selection is particularly suitable for multi-resolution diagnosis since it permits some of the results of computations, performed for sensor selection at the lowest (finest) level of multi-resolution diagnosis to be reduced at higher levels. This feature is particularly useful in reducing the computations necessary for online reconfiguration of the multi-resolution diagnosis system. An important procedure used in sensor selection is testing diagnosability. In this thesis, a new procedure for testing diagnosability in timed DES is introduced based on the relatively timing of plant output sequence. It is shown through example that the proposed test maybe executed with significantly fewer computations compared to tests developed for untimed models and adapted for timed systems. Furthermore, two new sets of sufficient conditions are provided under which diagnoser design and diagnosability tests based on relative timing of output sequence can be performed efficiently