Randomized algorithms aided analysis and design of model-based fault detection systems

In this thesis a probabilistic performance assessment framework of FD systems is developed, where the FD system performance indicators are given in the probabilistic framework. Furthermore, their estimation schemes are studied. In order to incorporate the probabilistic properties of industrial systems, the so-called randomized algorithm (RA) is applied for the estimation of the proposed performance indicators, which guarantees the accuracy of the proposed estimation schemes. The further study of this thesis is focused on the design of FD systems for uncertain LTI systems. The design methods studied in the literature are in general limited to special structures of uncertainties, which basically have two major drawbacks. First, the structured uncertainties are rather unrealistic in practical cases. The true uncertainties in the plant description, however, often only form a small subset of the structured uncertainties. Secondly, the distributions of the u ncertainties are often not considered during the design procedure. Therefore, the FD systems are always designed for the „worst-case“ scenario, which leads to rather conservative design results. Motivated by the above observations, in this thesis, RA-aided FD system design methods are proposed where the probabilistic properties of uncertain systems are incorporated into the design procedure. Furthermore, the probabilistic performance of the achieved FD system is guaranteed due to the application of RAs. As an essential element of FD system design, the threshold setting strategies are investigated in this thesis for both single feature and multi-feature FD configurations. Finally, the proposed design methods are demonstrated on the three-tank laboratory benchmark process.