Balancing Discrepancy and Consistency: Adversarial Single Domain Generalization in Fault Diagnosis

Single-domain generalization (SDG) in fault diagnosis aims to train a model that generalizes to unseen conditions using data from a single operating condition. Despite its promise, SDG faces significant challenges. Recent methods focus on generating new domain features to expand data distribution and extract domain-invariant features, but they encounter three issues: 1) separating domain expansion from domain-invariant feature learning limits their synergy and generalization; 2) domain expansion often uses random or inconsistent enhancement directions, introducing noise or generating unreliable features; and 3) domain-invariant feature extraction methods emphasize global or class-level alignment, neglecting the optimization of their relationships. Therefore, an SDG fault diagnosis framework based on the adversarial interaction between discrepancy and consistency is proposed, where adversarial training is repeatedly applied to dynamically balance domain discrepancy and consistency to avoid single-direction bias. Specifically, a learnable statistical feature perturbation module for domain discrepancy perception is proposed, which purposefully guides the model to learn statistical perturbations that effectively characterize domain discrepancies, thereby increasing the reliability of domain extension. In addition, covariance alignment and semantic alignment strategies are proposed to maintain global and category consistency between original and perturbed features. Covariance alignment is employed to mitigate stylistic variations introduced by perturbed features while preserving mission-critical diagnostic information, and explicit semantic constraints are utilized to further promote consistency in the model’s prediction results. Finally, extensive experiments are carried out on five datasets to verify the effectiveness of the proposed method.