Decoupling Machine and Operational Variances: A Spectral Attention Framework for Robust Few‐Shot Cross‐Machine Fault Diagnosis

Deploying intelligent fault diagnosis models in real‐world industrial settings is severely hampered by a trio of challenges: data scarcity, cross‐machine heterogeneity, and time‐varying operating conditions. Existing domain adaptation methods, which primarily align statistical distributions, often fail because they are physics‐agnostic and implicitly assume data stationarity. To overcome these fundamental limitations, we propose a novel framework that learns representations invariant to both machine and operational variances. Our approach integrates a physical‐informed spectral attention (SA) mechanism with a dynamic spectral distribution alignment (DSDA) strategy. The SA mechanism adaptively identifies and focuses on the invariant harmonic structures of fault signals, making it robust to nonstationarity. Concurrently, the A‐distance‐guided DSDA dynamically balances physical constraints and statistical alignment to handle complex domain shifts. On 12 cross‐machine, constant‐speed tasks with only 10 labeled samples, our method achieves a state‐of‐the‐art accuracy of 97.22%. More critically, in ultimate stress tests under time‐varying speeds, it maintains an exceptional average accuracy of 93.55%, where traditional methods’ performance collapses. This work presents a paradigm shift toward building robust diagnostic systems by effectively decoupling physical and operational variances.