Conditional Adversarial Domain Adaptation With Discrimination Embedding for Locomotive Fault Diagnosis

Deep-learning-based methods have been widely applied to fault diagnosis for its capability of automatically extracting features from massive data. However, there are still two problems that limit the application of these methods to the locomotive fault diagnosis, which has practical significance for ensuring the safety and reliability of the locomotive. First, due to variation in the operating condition in locomotive operation, the distribution of the training data is different from the distribution of the test data, which degrades the performance of the deep-learning-based methods. Second, the existing domain-adaptation (DA)-based diagnosis models are not robust enough to diagnose locomotive with many disturbances in its vibration signals. To tackle these problems, conditional adversarial DA with discrimination embedding (CADA-DE) is proposed in this article. By adding the discriminative information conveyed in the label predictions to the domain classifier, the distribution discrepancy between the source and target domains is alleviated more sufficiently. Meanwhile, the center-based discriminative loss is introduced to enforce features more discriminative. The experimental results conducted on three different data sets demonstrate the effectiveness and superiority of the proposed CADA-DE.

Automated summary

This paper proposes a conditional adversarial DA with discrimination embedding (CADA-DE) method to address two problems in locomotive fault diagnosis. Experimental results demonstrate the effectiveness and superiority of the proposed CADA-DE.