Moment-Wasserstein Distributionally Robust Virtual Domains for Reliable Single-Source Fault Diagnosis

Abstract

Single-source fault diagnosis trains a model on one labeled operating condition and must generalize to unseen speed, load, or sensor changes. Existing single-domain generalization methods either perturb only low-order statistics, which cannot fully describe the impulsive and heavy-tailed structure of rotating-machinery signals, or expand virtual domains without an explicit structural constraint, risking semantic drift. This paper proposes MWDG, a distributionally robust framework that models virtual-domain construction as optimization over a Wasserstein ball intersected with a moment-constrained ambiguity set. A learned transport potential generates moment-feasible virtual features, while a Cross-Layer Transport (CLT) constraint preserves dependency structure across network depth so that expansion remains label-consistent. The primary evaluation uses average accuracy and macro-F1 under source-to-unseen condition transfer. As complementary evidence, we include a lightweight reliability-oriented audit that reports weak transfer cases, class-level error patterns, and cross-rotation stability without using any target-domain supervision during training or model selection. Experiments on PU, HUST, and SDUST show that MWDG consistently improves source-only fault diagnosis over strong single-domain generalization baselines.