Unsupervised Automatic Updating of Classification Models of Fault Diagnosis for Novelty Detection

Novelty Detection (ND) methods are mainly developed to detect novel samples without a clear and general strategy for updating the Fault Detection and Diagnosis (FDD) system. The present study addresses this problem through developing an automatic unsupervised data driven framework for updating FDD systems for the ND purpose. The proposed method is based on the combination of One Class Classifiers (OCCs) that is used to detect samples following novel patterns (i.e. faults), and automatic clustering technique to diagnose them according to novel clusters (i.e. figure out the specific novel fault type). The FDD updating is performed by modifying the existing clusters and detecting new clusters for assembling models in an unsupervised automatic mode. An observer is also incorporated for data processing and enhancing the FDD robustness. The proposed framework is studied and validated via its application to a simulated benchmark case study consisting of three coupled tanks. The results show that the framework can efficiently manage classification models, learn novel clusters, and update FDD system.