Integrating Expert Knowledge With Domain Adaptation for Unsupervised Fault Diagnosis

Data-driven fault diagnosis methods often require abundant labeled examples for each type. On the contrary, real-world data is unlabeled and consists of mostly healthy observations only few samples faulty conditions. The lack labels imposes a significant challenge existing data-driven methods. In this paper, we aim to overcome limitation by integrating expert knowledge with domain adaptation in synthetic-to-real framework unsupervised diagnosis. Motivated fact that experts have relatively good understanding on how different types affect signals, first step proposed framework, synthetic dataset generated augmenting real vibration bearings. This integrates encodes class information about types. However, models trained solely based do not perform well because distinct distribution difference between synthetically faults. To gap data, second an imbalance-robust adaptation~(DA) approach adapt model from faults~(source) faults~(target) which suffer severe imbalance. evaluated two cases bearings, CWRU laboratory wind-turbine dataset. Experimental results demonstrate faults are effective encoding type robust against levels imbalance