A cross-domain fault diagnosis method for mixed-fusion samples based on data generation and class-level domain adversary

Abstract With the widespread application of rotating machinery in intelligent manufacturing, aerospace, and other industrial fields, accurate and reliable fault diagnosis and maintenance have become increasingly critical for ensuring system safety and operational efficiency. However, existing domain-adaptation-based cross-domain intelligent fault diagnosis methods primarily focus on achieving feature transfer at the global domain level, often overlooking the complexity, imbalance, and significant class-level variability arising from the simultaneous distribution of samples across the source and target domains. This oversight can lead to inaccurate recognition of fine-grained class-level features, thereby limiting diagnostic accuracy. To address these challenges, this paper presents a class-level domain alignment method (CDD_DANN) that combines classifier deterministic difference (CDD) loss with a dual-classifier structured domain-adversarial neural network (DANN), effectively improving class-level feature alignment and transfer in cross-domain fault diagnosis. Additionally, to effectively address the challenge of sparse marginal samples at deeper levels, we propose the PMCDAN (proxy-based metric learning and classifier deterministic difference adversarial network) method, which replaces CDD with a proxy-based metric learning approach, proxy neighborhood component analysis (ProxyNCA), to capture deeply shared features between the source and target domains more robustly. This enables global domain and class alignment under challenging conditions. Furthermore, to tackle the data imbalance, this paper incorporates a diffusion-generative adversarial networks-based fault sample augmentation method, which facilitates both domain and class-level alignment when data are scarce, thus enabling more accurate fault diagnosis. The effectiveness and superiority of the proposed approach are validated through experimental evaluations against existing methods using the Paderborn University bearing data set and a self-collected gear fault data set. The proposed method provides valuable insights and practical guidance for fault diagnosis in complex real-world industrial scenarios.