Conditional disentangled representation augmentation network for imbalanced domain generalization fault diagnosis of rotating machinery

Abstract

Domain generalization has been widely applied to cross-operating conditions and cross-machine fault diagnosis of rotating machinery, especially when operating conditions or machines are unseen. However, existing domain generalization-based diagnosis methods are typically constructed under the assumption of class balance in source domains, and thus imbalanced data in real industrial scenarios will greatly impair their fault diagnosis performance. To tackle this problem, we present a Conditional Disentangled Representation Augmentation network (CDRAnet) for rotating machinery fault diagnosis under imbalanced domain generalization scenarios. In CDRAnet, a parameter-sharing 1D convolutional neural network is first constructed as the feature extractor to capture the high-level abstract fault features from multiple source domains. Then, a disentangled representation augmentation module is designed to augment the minority fault class’s features, thereby balancing the inter-class data distribution. Meanwhile, a class-conditional triplet-center loss is derived to tighten the intra-class features of the same class across domains and thus improve the model’s generalization to unseen domains. Lastly, comprehensive cross-operating condition and cross-machine experiments reveal that the fault diagnosis performance of CDRAnet significantly surpasses other advanced methods under imbalanced domain generalization scenarios.