Failures in bearings and gears are among the most critical issues in mechanical systems. Deep learning (DL) methods have achieved remarkable success in fault diagnosis owing to their ability to extract informative representations from data. However, most existing models assume consistent domains and predefined fault categories, which limits their applicability under variable operating conditions and unseen fault types. To address these challenges, we propose Multi-Source Open-set Domain Adaptation (MSODA), a multi-source domain adversarial graph convolutional framework for intelligent fault diagnosis in open-set and cross-domain scenarios. Our study reveals that DL-based diagnostic models often experience performance degradation when encountering previously unobserved faults or diverse working environments. MSODA overcomes this limitation by employing multi-source domain adaptation, which transfers diagnostic knowledge across domains through shared and transferable representations, thereby enhancing generalization and the recognition of unknown faults. Furthermore, traditional graph construction methods that rely solely on time–frequency or distance metrics lack adaptability. To overcome this, we design an adaptive graph generation mechanism that dynamically learns structural relationships from data in a domain-aware manner. Combined with advanced graph convolutional networks, this approach enables robust and discriminative feature extraction, leading to effective cross-domain open-set fault detection. The proposed framework can be seamlessly integrated into real-world prognostics and health management systems for rotating machinery—such as bearings, gears, and rolling mills—ensuring adaptive and generalizable fault diagnosis across complex operational environments. Extensive experiments on three benchmark datasets demonstrate that MSODA consistently outperforms state-of-the-art methods in diagnostic accuracy, generalization ability, and robustness to domain shifts.
Zhao et al. (Sat,) studied this question.