Out-of-distribution challenges in intelligent fault diagnosis: a survey of generalization and detection

Deep learning (DL)-based intelligent fault diagnosis (IFD) has achieved strong performance in mechanical systems, but its reliability can deteriorate when deployment data differ from the training distribution because of changes in operating conditions, equipment, sensing environments, or fault semantics. These shifts create two related but distinct requirements: preserving accurate predictions for known fault categories under unseen yet relevant conditions and identifying inputs that should not be assigned to any known category. To support this survey, we conducted a reproducible search of IEEE Xplore, the Web of Science Core Collection, and Scopus for studies published from January 1, 2020, to July 16, 2026. The search returned 280 database records, of which 181 remained after deduplication and 161 primary mechanical OOD-related studies were included and coded according to their deployment roles and shift types. The survey focuses on the operational boundaries between generalization-oriented methods and detection-oriented methods in mechanical fault diagnosis. It does not aim to provide an exhaustive review of closed-set domain adaptation or domain generalization without an unseen mechanical-domain evaluation, general anomaly detection unrelated to unknown-fault rejection, non-mechanical industrial process monitoring, or general-purpose computer-vision OOD methods except where they provide necessary conceptual foundations. We introduce a task-relative operational definition based on effective training support, training label space, and a declared deployment envelope, and use it to distinguish condition-related (C-shift), fault-semantic (F-shift), and combined C+F settings. Based on these distinctions, we organize representative methods by deployment role and mechanism, compare their assumptions, data requirements, outputs, strengths, and limitations, and synthesize literature-reported experimental and evaluation practices. The resulting deployment-oriented perspective emphasizes that reliable diagnosis requires known-class generalization, calibrated OOD detection, and an explicit escalation or human-referral pathway to be evaluated separately and integrated coherently.