A Study Of Hybrid Modeling Technique Based Fault Detection And Diagnosis And Its Application

Modern control systems are increasingly complex, which are usually characterized bynonlinearity, strong coupling and uncertainty, high-dimensionality, etc. Because of the complexity andaging problems, the systems are prone to various faults or failures. How to detect faults quickly,diagnose faults accurately and recover the system efficiently are important subjects in both academiccircles and engineering fields.The performance of fault detection and diagnosis highly depends on the quality of system model.Traditional modeling methods can be roughly divided into first-principle based methods, knowledge-basedmethods and data-driven methods. Parameters of first-principle models usually have straightforwardphysical meanings. The models are easy to understand and interpret. However, it is difficult andtime-consuming to develop accurate first-principles models. Knowledge-based models are suitable forthose systems that have difficulties for quantitative analysis and modeling. But knowledge-based modelshave portability issues. For those pure data-based modeling techniques, they usually require no priorprocess knowledge and are easy to be implemented; however, the performances of data-based models arestrongly reliant on the modeling data. For modern complex control systems, it is desirable to combine theadvantages of different modeling techniques. A hybrid modeling strategy can be a good solution to obtainaccurate and reliable system models for fault detection and diagnosis.This thesis attempts to develop a cascaded hybrid modelling technique based fault detection anddiagnosis framework, by making full use of the advantages of the first-principle modeling and data-basedmodeling techniques. The proposed hybrid modeling, fault detection and diagnosis methods are verified ona three-tank system and applied to a Quanser3-DOF Hover system.The key contributions are summarized as follow:1. For the systems with partial prior knowledge, a data-based non-parameter model is integratedinto the first principle model in a serial configuration, where BP neural network and T-S fuzzy modelare used and compared to identify the unknown parameters of the first-principle model.2. In practical application, a pure data-driven fault detection method using principle componentanalysis (PCA) or multiway PCA may be insensitive to the incipient faults. To overcome this problem,a fault detection method based on the combination of the aformentioned hybrid model and PCA isdeveloped, where PCA is conducted on the residual signals of the hybrid model. The proposed faultdetection method outperforms PCA or multiway PCA based methods with respect to the quickness ofii fault detection.3. Faults occurred in control systems can be divided into actuator fault and sensor fault, whichoften correspond to the variations in the state equation and the output equation of a system’sstate-space model, respectively. Based on this, a hybrid model based fault diagnosis method isproposed, where input reconstruction technique is used to determine the cause to fault. The proposed faultdiagnosis method outperforms the traditional model-based methods with respect to the accuracy of faultdiagnosis.4. The above proposed methods are verified on a three-tank system and applied to a Quanser3-DOFHover. The results can show their feasiblity and superority.