Equipment operating conditions, referred to as domains, can induce domain drift in monitoring data, affecting data-driven fault diagnosis. Researchers have explored multi-domain generalization methods to tackle this issue. However, in actual industrial scenarios, the availability of fault data may be limited to a specific condition due to the cost or feasibility constraints associated with collecting extensive monitoring data. This limitation hampers the generalization ability of these methods, posing a major challenge for robust fault diagnosis under variable operating conditions. To address this challenge, we proposed a gradient-based domain-augmented meta-learning (GDM) single-domain generalization method. We analyze the restrictions of generating fake domains and construct a domain-augmented loss by evaluating diagnostic tasks minimization, semantic consistency, and distribution diversity for fake samples. Using a gradient-based technique, fake domains are generated iteratively, providing diverse fault knowledge for improved generalization. Instead of using time-consuming ensemble methods, we develop a novel meta-learning method to train a highly efficient and generalizable model, relaxing the requirement for auxiliary datasets in existing meta-learning methods. Two case studies consistently demonstrate the effectiveness and superiority of the proposed GDM method. Our findings suggest that this study offers a promising and competitive solution for single-domain generalization in fault diagnosis within real industrial scenarios.
As the operating conditions (also known as domains) of rotating machinery become increasingly diverse, fault diagnosis has garnered growing attention. However, fault diagnosis frequently encounters challenges such as long-tailed data distributions, domain shifts in monitoring data, and the unavailability of target-domain data. Existing approaches can only address some of these challenges, limiting their applications. To address these challenges concurrently, we introduce a novel learning paradigm called long-tailed multi-domain generalized fault diagnosis (LMGFD) and propose a two-stage learning framework for LMGFD, comprising domain-invariant feature learning and balanced classifier learning. In the first stage, we leverage a balanced multi-order moment matching (BMMM) module to align subdomains with long-tailed distributions. Additionally, a balanced prototypical supervised contrastive (BPSC) module is developed to effectively alleviate the contrastive imbalance. The combination of BMMM and BPSC enables the effective learning of long-tailed domain-invariant features. In the second stage, we extend the focal loss to a multi-class version and re-weight it using effective sample numbers to strengthen tailed-class loss, thereby mitigating the overfitting problem. Experimental results on both a public dataset and a private dataset support the competitiveness and effectiveness of the proposed method. The findings suggest that we present a promising solution for fault diagnosis of rotating machinery under variable operating conditions.
Data-driven fault diagnosis approaches have been extensively investigated for rotating machinery. However, real-world industrial environments present significant challenges: shifting operating conditions lead to distribution differences in monitoring data, and unknown target operating conditions result in a lack of target domain data for training. Additionally, monitoring data often exhibit class imbalance, with common healthy states dominating the dataset, while rare but critical fault types are underrepresented. These issues severely impair the generalization capability of diagnostic models and remain largely unaddressed. To tackle these challenges, we propose a novel learning framework, Prototype-guided Supervised Contrastive Learning with Dynamic Temperature Modulation (DTM-PSCL), designed to effectively capture imbalanced domain-invariant features. The proposed approach incorporates class prototypes to address contrastive imbalance and introduces a regularization technique to approximate cross-domain prototypes. Furthermore, we develop an exponential dynamic temperature modulation strategy in contrastive learning, which enhances training stability and enables the model to adaptively learn imbalanced domain-invariant features. Additionally, a class-balanced loss is used to prioritize underrepresented classes, reducing the model’s bias toward dominant classes. Extensive experiments on four well-known datasets consistently demonstrate the effectiveness and superiority of the proposed framework, suggesting a promising methodology for fault diagnosis of rotating machinery under unknown operating conditions.