CONSTRUCTION OF MODEL-BASED FAULT DIAGNOSIS SYSTEMS ROBUST TO MODELING ERROR

点击翻译按钮获取中文摘要

In this research project, for the purpose of construction of model-based fault diagnosis systems which are robust against to modeling error, we have investigated on the following subjects : 1). modeling and identification method for fault diagnosis of nonlinear dynamic systems, 2). evaluation method of model uncertainty(i.e. modeling error), 3). Fault detection scheme which is robust against to the modeling error, and 4). fault isolation scheme using knowledge information about the system to be diagnozed. The main results obtained for each subject arc as follows :1). We have proposed a Quasi-ARMAX model which is equiped with both of flexibility and linear structure, by imbedding nonlinear characteristics of the system into the ARMAX model parameters, and developed its identification method. It has been confirmed through simulation studies that the model is useful for the fault detection in wide class of nonlinear systems and for STR-based adaptive control as well.2). We have proposed a … More method for evaluating the modeling error due to linear approximation of nonlinear systems via the nonlinear terms in the Quasi-ARMAX modeling. Such the evaluation of modeling error is essential for the design of robust fault detection.3). We have analysed the relation between the modeling error and the Kullback Discrimination Information (KDI) which is used as the fault detection index, and developed a robust fault detection scheme based on the evaluation of modeling error. Another way of robust fault detection in nonlinear systems has been proposed based on multi-ARMAX models in the framework of the Quasi-ARMAX modeling and based on the analysis of the KDI index using the weighting coefficients of the multi-modeling. On the other hand, it is important issue to determine a resonable threshold value in the thrshold decision for fault detection. We have solved this problem by learning the probability density function of the KDI index using the data obtained from online identification of the system under the normal operation mode. As the result, a threshold value corresponding to a confidence revel of false alarm rate has been able to be detemined based on the learned probability density function of the KDI.Finally, as to the 4) research subject, it is still open problem due to the difficulty to estimate the physical parameters of the system through the Quasi-ARMAX modeling and identification procedures. Less