Conditional Adversarial Domain Generalization With a Single Discriminator for Bearing Fault Diagnosis

Abstract

Traditional deep learning assumes that the probability distributions of the test set and the training set are exactly the same. When the test samples and the training samples are from different working conditions, the distribution gap between different conditions might dramatically lower the accuracy. To address this problem, domain adaptation (DA) via joint using training and test sets is adopted to increase the final accuracy. The disadvantage of DA is that it requires test samples to participate in the training phase. However, sometimes it is impossible to collect the test samples beforehand. To solve this problem, this article proposes a conditional adversarial domain generalization with a single discriminator, which aims to extract domain-invariant features from different working conditions and generalize the features to unseen working conditions. To realize conditional adversarial training, a new conditional adversarial strategy that the feature extractor can allow the discriminator to distinguish the fault classes, but cannot distinguish the domains, is designed to better confuse the discriminator and generalize the features. Meanwhile, three loss functions, including mean square error (MSE), cross-entropy coupled with entropy, and cross-entropy combined with traditional adversarial methods, are proposed to achieve the new conditional adversarial strategy. The effectiveness of the conditional adversarial strategy is verified by training conditional generative adversarial networks (CGANs) on two image data sets, and the performance of the proposed domain generalization method for bearing fault diagnosis is also verified by bearing data sets. Compared with traditional conditional adversarial domain generalization, the proposed method can save computing resource consumption and improve accuracy that does not appear in the training phase.

Topics & Concepts