Latent subdomain assignment based on pseudo domain labels for fault diagnosis of unseen data

Intelligent fault diagnosis (IFD) is important for rotating machinery maintenance. Unfortunately, fault diagnosis training models often degenerate if unknown domain shifts exist between different working conditions when performing IFD. To deal with this problem, more generalized features related to rolling element bearing faults should be explored so that the generalized capacity of the training model is boosted for unseen target domain data. In this paper, a new algorithm using pseudo domain labels is proposed to explore subdomain distributions within each subdomain at the domain level. The idea behind the proposed method is that the domain shifts caused by variable working conditions, like varying speeds, should also be considered since the data may show a dynamic distribution of temporal features that are not limited to spatial distributions. That is, the original domain distribution could be further divided into several latent subdomains by introducing pseudo domain labels, which enables the proposed method to learn domain specific features. Furthermore, the diversity of learned features across subdomains ensures comprehensive feature coverage during model training, while the inherent similarities between these domains enhance the capacity of the model for domain generalization. To figure out how the domain label updates, a domain-class label is initially introduced to facilitate fine-grained feature learning, enabling the model to capture as many features as possible. Then an adversarial learning strategy is employed to separate the domain and class information. Specifically, pseudo domain labels are determined using class invariant features, while class labels are distinguished using features that are invariant across multiple latent subdomains. These two steps are equivalent to a min–max game, like adversarial learning. By exploring features from the class and domain levels, the domain generalization capabilities of the model can be improved, thereby further increasing the accuracy of results. Experiments on two public bearing datasets show that the proposed method outperforms state-of-the-art methods. Additionally, by limiting the number of accessible data from known source domains, the proposed method shows the potential to maintain satisfactory domain generalization capacities when combined with few-shot learning.