Ontology-Based Graph Transformer Network for Robust Bearing Fault Diagnosis under Unseen Operating Conditions

Abstract

Conventional data-driven diagnostic models for rotating machinery often exhibit limited generalization under varying operating conditions. Signal-based approaches typically rely on condition-dependent training data and statistical feature learning, making them vulnerable to domain shifts and complex fault scenarios. In particular, frequency-domain features are commonly treated as unstructured statistical descriptors, without capturing the physical relationships among characteristic frequencies such as harmonic and sideband structures intrinsically linked to bearing fault mechanisms. Consequently, the learned representations tend to overfit to the training distribution, leading to pronounced performance degradation under unseen operating conditions. These limitations highlight the need for diagnostic frameworks that incorporate physically grounded relational structures to achieve robust generalization.

To address these challenges, this study proposes OFG-GTN (Ontology-based Frequency Graph–GTN), an ontology-based bearing fault diagnosis framework that integrates structured frequency representation and relational graph learning. Vibration signals are transformed into ontology-based frequency graphs by encoding physically defined characteristic frequencies, harmonic relations, and sideband dependencies derived from bearing fault mechanics. A GTN is employed to perform graph-level fault classification by capturing higher-order relational dependencies among frequency components. Experimental results demonstrate that OFG-GTN achieves robust and generalizable bearing fault diagnosis across diverse operating conditions, including those not encountered during training.