Class Imbalanced Domain Adaptation for Rolling Element Bearing Fault Diagnosis

Abstract

For fault diagnostics of rotating machines, it is commonly observed that fault data are scarce while the operating conditions vary. Existing fault diagnostic models account for only one of the two problems of either class imbalance or variable operating conditions. To address the problem, a new fault diagnostic method, namely class-imbalanced domain adaptation (CIDA), is proposed in this study. The time-shifting data augmentation strategy is employed with pseudo-labeling. Labeled source fault data and pseudo-labeled target fault data are augmented. Then, the deep convolutional domain adaptation networks extract features by minimizing two losses including classification loss and class-wise central moment discrepancy (CCMD) loss. The classification loss quantifies the disagreement between the predicted labels and target labels, while the CCMD loss evaluates the distance between source and target domain distributions attributed to different operating conditions. The validity of the proposed CIDA method was evaluated with two datasets, including artificially-seeded-fault and run-to-failure bearing data. It is shown that the proposed CIDA method is effective in addressing the class-imbalance problem and various operating conditions.