Application of Domain Invariant Features in Condition Monitoring of Rolling Element Bearings Under Changing Operating Conditions

Abstract

Data driven condition based monitoring of bearings has gained a lot popularity in recent times especially due to the fact that physics based models of equipment are difficult to formulate fully and accurately because of the complexity of machines. The low cost of sensors that has availed a large amount of data from operating machinery has further propelled the use of data-driven condition monitoring. Data driven models are heavily reliant on the domain of data they are trained on (source data). This means that they suffer in performance when applied to test data (target data) from a different domain. The most widely used techniques to counter this drop in performance are domain adaptation methods which seek to reduce the discrepancy between the two datasets. A key challenge with domain adaptation methods is the requirement for target data during training as a reference for the amount of discrepancy that exists. The other associated challenge is that the adaptation method has to be reconfigured or completely overhauled for each new test data because adaptation methods have varying capacity depending on the magnitude of the domain shift. This in turn means that models have to be retrained each time new test data are acquired. The goal of this work was to find and apply domain invariant features to the development of models so as to remove their dependence on target data but still be able to perform condition monitoring of REBs across domains. Publicly available datasets were used in the study: the Case Western Reserve and Ottawa Universities bearing datasets were used for diagnosis while the FEMTO-ST bearing dataset was used for prognosis. The Refined Composite Multi-scale Fuzzy Entropy Feature (RCMFE) was found to be a domain invariant feature for diagnosis. RCMFE had excellent fault detection ability, correctly detecting fault 100% of the time in different operating conditions. With the training data prepared such that each class of fault had a mixture of fault diameter sizes, RCMFE could easily differentiate inner race fault from ball and outer race faults with an average accuracy above 95%. However, the average accuracy for differentiating between ball and outer race fault fell to about 80%. With the training data arranged such that each fault type and size constituted a single class, RCMFE could isolate the three types of fault with an average accuracy of about 97%. Thus, this feature was able to achieve cross-domain diagnosis without domain adaptation. The hazard functions of kurtosis and shape factor were found to be trendable domain-invariant features for use in prognosis. Because the training data are full lifetime data, the challenge of low performance on test bearings with longer remaining useful life was overcome by supplementing the training data with its truncated versions. The combination of hazard functions and augmented training data enabled successful prognosis in changing operating conditions.