Abstract
 

Accurate vibration-based instantaneous angular speed (IAS) estimation is crucial for effective fault diagnosis in rotating machinery operating under nonstationary speed conditions. However, traditional time–frequency ridge-tracking methods are highly sensitive to noise and hyperparameter settings. Moreover, while recent studies have proposed robust end-to-end IAS estimation approaches with deep learning models, they often struggle under a domain shift scenario, exhibiting a large discrepancy between the source and target domains. In this article, we propose a domain-informed deep learning framework for robust IAS estimation. Our proposed method features two key components: a selective multi-order frequency (SMOF) module and a multiharmonic loss function. SMOF module leverages domain-specific knowledge with a masking mechanism, while the multiharmonic loss function enforces harmonic consistency in the frequency domain to enhance label reliability. Compared to existing approaches, the proposed method achieves stable reliability across diverse operating conditions, highlighting its robustness and applicability in real-world machinery monitoring scenarios without target domain annotations.