Hua, Zehui2026-05-292026-05-292026-05-29http://hdl.handle.net/10393/51720https://doi.org/10.20381/ruor-32004The health of key rotating components in machinery systems, such as rolling element bearings and gears, is critical for meeting design requirements and ensuring safe operation. These components degrade over time, leading to faults that can cause unplanned downtime, economic loss, or catastrophic accidents. Vibration signal-based intelligent fault diagnosis (IFD) enables real-time condition monitoring with reduced reliance on human expertise. However, traditional machine learning methods often assume that vibration data from a source domain and a target domain share similar feature distributions, an assumption that rarely holds under variable working conditions in industrial settings. Transfer learning mitigates distribution discrepancies, yet important challenges remain: (1) effective IFD under changing operating conditions, (2) learning models that generalize across multiple domains simultaneously, and (3) transferring knowledge to a totally unseen target domain. This thesis investigates domain generalization for vibration-based IFD under distribution shifts induced by variable working conditions and develops four methods. First, by leveraging inter- and intra-domain invariances, condition-robust representations are learned and achieve consistent improvements over strong baselines on two benchmark datasets across diverse cross-condition settings. Second, to better handle operating-condition variations, a multi-subdomain alignment strategy that introduces multiple condition-related subdomains within a single source domain and aligns them to reduce condition-dependent discrepancies is proposed, improving diagnostic performance on two bearing datasets. Third, a feature disentanglement mechanism is introduced to decouple domain-invariant from domain-specific features, enhancing discriminability and robustness under unseen conditions. Extensive experiments, including low-data regimes, show superiority over several state-of-the-art approaches. Finally, the framework for simulation-to-experiment transfer is extended, and a transferable diagnostic model that captures time-varying characteristics is developed, enabling reliable fault detection based on the obtained time-frequency representations.enAttribution-NonCommercial-NoDerivatives 4.0 Internationalhttp://creativecommons.org/licenses/by-nc-nd/4.0/Rotating machineryNonstationary signal processingDomain generalizationIntelligent fault diagnosisTime-frequency analysisThesis