A hybrid inference approach for health diagnostics with unexampled faulty states

System health diagnostics provides diversified benefits such as improved safety, improved reliability and reduced costs for the operation and maintenance of engineered systems. Successful health diagnostics requires the knowledge of system failures. However, with an increasing complexity it is extraordinarily difficult to have a well-tested system so that all potential faulty states can be realized and studied at product testing stage. Thus, real time health diagnostics requires automatic detection of unexampled faulty states through the sensory signals to avoid sudden catastrophic system failures. This paper presents a hybrid inference approach (HIA) for structural health diagnosis with unexampled faulty states, which employs a two-fold inference process comprising of preliminary statistical learning based anomaly detection and artificial intelligence based health state classification for real time condition monitoring. The HIA is able to identify and isolate the unexampled faulty states through interactively detecting the deviation of sensory data from the known health states and forming new health states autonomously. The proposed approach takes the advantages of both statistical approaches and artificial intelligence based techniques and integrates them together in a unified diagnosis framework. The performance of proposed HIA is demonstrated with a power transformer health diagnosis case study, where Mahalanobis distance serves as a representative statistical inference approach.