Novel domain-adaptive Wasserstein generative adversarial networks for early bearing fault diagnosis under various conditions

Due to the scarcity of labeled data, early fault diagnosis under various conditions faces significant challenges. In this paper, a novel data augmentation method is proposed, called as Domain-Adaptive Wasserstein Conditional Generative Adversarial Network (DA-WGAN), to acquire a significant quantity of labeled samples for early fault of bearing in dynamic scenarios. DA-WGAN is characterized by its inclusion of a domain adaptation module, which allows for the incorporation of features from unlabeled samples in various operating conditions during training. This mechanism promotes DA-WGAN to generate a significant amount of labeled samples that closely resembles the features in the target domain's operational scenarios. In addition, a multi-scale transfer learning model with an attention mechanism is proposed to address the issue of the generated data not fully replicating the feature distribution of the target domain. This enhances the alignment of the feature distribution in the generated data with that of the target domain data. Experimental studies on early fault diagnosis of bearing demonstrate that the proposed method generates high-quality labeled samples for various conditions, which can significantly improve the classification accuracy of early fault of bearing under various operational conditions.