A Multiview-Clustering-Based Fault Diagnosis Architecture for Complex Systems Using Probabilistic Multi-Signal-Flow Graph Model

Under industrial circumstances, the insufficiency of anomaly samples makes data-driven approach limited within the level of anomaly detection, which could hardly support the really-wanted LRU-level fault diagnosis tasks. However, different with anomaly samples, the empirical knowledge on the qualitative signal-fault correlation is usually available on the complex systems and could well complete the fault-relevant information, which is deficient inside the unsupervised data-driven anomaly detection mechanism. Therefore, we propose a signal-orienting strategy in order to combine the signal-level data-driven anomaly detection mechanism with the signal-fault knowledge, which realize eventually an unsupervised fault diagnosis architecture via an inductive Multiview-Clustering-based anomaly detection and probabilistic Multi-Signal-Flow-Graph-based fault attribution in stack. Experiments on a public UAV dataset could demonstrate the state-of-the-arts performance of the proposed method on the complex-system fault diagnosis.