Physics-guided adaptive spectral transformation for unsupervised cross-domain vibration-based compound fault diagnosis

Abstract

Diagnosing multiple faults that are simultaneously occurring in a rotating machinery is theoretically a domain adaptation problem where different defect sizes cause systematic spectral changes which make the standard classifiers assumption of identically distributed data invalid. Current unsupervised domain adaptation methods make use of marginal alignment through multi-kernel Maximum Mean Discrepancy (MK-MMD) or adversarial training. They are not explicitly imposing a physical constraint on the variability of defect scale which suggests that they are likely to fail in negative transfer and harmonic decoupling. This work introduces a physics-informed adaptive spectral transformation (PGAST) method that consists of three main elements. A severity-aware spectral transformation unit that first separates the signal into frequency bands and then a signal-dependent warping coefficient is applied with the help of time-domain and spectral regularization. A Lipschitz-stabilized 1D-CNN encoder that obtains domain-invariant features at a controlled sensitivity and a spectral-aware class-conditional MMD objective that aligns feature distributions separately across low, mid, and high-frequency regimes which help avoid inter-class mode mixing. When tested on the publicly available bearing datasets with different defect diameters, PGAST leads to 95.96% cross-domain recognition accuracy showing the capability to generalize to unseen compound couplings. This suggests that the use of structured physics-informed alignment is an improvement over the purely statistical baselines.