Digital twin-assisted graph contrastive domain adaptation for small-sample bearing fault diagnosis

Rolling bearing fault diagnosis is crucial for ensuring the safe and reliable operation of mechanical equipment. However, the scarcity of labeled fault data often prevents data-driven diagnostic models from being sufficiently trained, thereby limiting their effectiveness and generalizability in practical applications. To address this issue, this study introduces a digital twin-assisted graph contrastive domain adaptation (GCDA) network for bearing fault diagnosis under small-sample conditions. Specifically, a lumped-parameter dynamic model of the bearing is established in the digital space, covering representative fault modes associated with raceways and rolling elements. Nevertheless, the significant distribution discrepancy between the dynamic responses and measured signals leads to performance degradation if applied directly. To mitigate this issue, a cross-domain distribution coupling strategy is proposed at the data level. By leveraging modal decomposition and energy matching, the simulated dynamic responses are fused with the background components of measured signals, thereby generating twin samples that simultaneously retain physical mechanisms and realistic distribution characteristics. At the model level, a GCDA network is further developed to achieve domain-invariant feature extraction and discriminative representation learning. Finally, validation on both a self-collected laboratory dataset and the publicly available XJTU-SY dataset demonstrates that the proposed method outperforms several state-of-the-art approaches across multiple evaluation metrics.