Single-domain incremental generation network for machinery intelligent fault diagnosis under unknown working speeds

Domain generalization based fault diagnosis methods are capable of maintaining good performance on unknown target domains. However, the need of multiple source domains for training limits the practical application of domain generalization fault diagnosis methods. Therefore, this study proposes a more practical single-domain generalization method, single-domain incremental generation network (SDIGN), for fault diagnosis under unknown speeds. First, an incremental domain augmentation strategy is proposed to address the changes in data sample distribution caused by machine speed changes. The purpose of incremental domain augmentation is to incrementally generate multiple augmented domains with different distributions but the same semantic information from a single speed source domain, thus simulating the data distribution changes due to speed changes. Inspired by adversarial learning, augmented domains with difference and completeness are generated by a generator during iterative domain augmentation. Then, the generalization performance of the model in unknown domains is improved by a contrastive learning to learn domain-invariant features in the source and augmented domains during each domain generation process. Finally, extensive experiments based on two datasets are conducted to validate the generalization of the proposed method in the unknown speed domains. Furthermore, the effectiveness of the proposed incremental domain augmentation strategy is verified by ablation experiments, showing the great potential of the proposed method in practical applications.