Integrating Expert Knowledge With Domain Adaptation for Unsupervised Fault Diagnosis

Data-driven fault diagnosis methods often require abundant labeled examples\nfor each fault type. On the contrary, real-world data is often unlabeled and\nconsists of mostly healthy observations and only few samples of faulty\nconditions. The lack of labels and fault samples imposes a significant\nchallenge for existing data-driven fault diagnosis methods. In this paper, we\naim to overcome this limitation by integrating expert knowledge with domain\nadaptation in a synthetic-to-real framework for unsupervised fault diagnosis.\nMotivated by the fact that domain experts often have a relatively good\nunderstanding on how different fault types affect healthy signals, in the first\nstep of the proposed framework, a synthetic fault dataset is generated by\naugmenting real vibration samples of healthy bearings. This synthetic dataset\nintegrates expert knowledge and encodes class information about the fault\ntypes. However, models trained solely based on the synthetic data often do not\nperform well because of the distinct distribution difference between the\nsynthetically generated and real faults. To overcome this domain gap between\nthe synthetic and real data, in the second step of the proposed framework, an\nimbalance-robust domain adaptation~(DA) approach is proposed to adapt the model\nfrom synthetic faults~(source) to the unlabeled real faults~(target) which\nsuffer from severe class imbalance. The framework is evaluated on two\nunsupervised fault diagnosis cases for bearings, the CWRU laboratory dataset\nand a real-world wind-turbine dataset. Experimental results demonstrate that\nthe generated faults are effective for encoding fault type information and the\ndomain adaptation is robust against the different levels of class imbalance\nbetween faults.\n