Rolling Bearing Fault Diagnosis Based on Deep Adversarial Networks with Convolutional Layer and Wasserstein Distance

Intelligent bearing fault diagnosis techniques have been well developed to meet the economy and safety criteria. Machine learning and deep learning schemes have shown to be promising tools for rolling bearing defect diagnosis. They require multitudinous labelled data in the training phase and assume that the training and testing samples abide by the same data distribution. However, in real-world industrial contexts, these two preconditions are almost impossible to be satisfied. Conversely, approaches based on transfer learning are potent instruments for proactively reacting to the above two challenges. Consequently, this paper presents an unsupervised method for diagnosing rolling bearing defects based on transfer learning. Convolutional neural networks, adversarial networks, and Wasserstein distance are adopted to extract domain invariant features, narrow the discrepancy between the source domain and target domain, and precisely categorize the faulty samples. A series of experiments corroborate that the proposed model can effectively facilitate the overall performance and outperform several traditional approaches under six measurement metrics.