Digital twin-driven zero-shot transfer diagnosis for rolling bearing faults under data-scarce conditions

Abstract Rolling bearing fault diagnosis is essential for ensuring the reliability of rotating machinery. Traditional deep learning methods struggle due to the lack of labeled data and domain shift across operating conditions. To address this concern, a digital twin-driven–zero-shot transfer diagnosis (DTD–ZSTD) method is proposed that integrates virtual-reality feature fusion and triple-model transfer, enabling accurate diagnosis without physical samples. First, a lumped parameter model (LPM) and a finite element (FE) model are developed to generate virtual vibration responses under various fault scenarios. To mitigate the distribution discrepancy between virtual and physical data, a Wasserstein generative adversarial network is applied to fuse virtual features with real-world noise and operational characteristics, creating the DTD virtual-reality feature integration framework. Second, a digital twin-assisted triple-model transfer framework is proposed, enabling sequential knowledge transfer from the lumped parameter domain to the FE model domain and then to the physical domain. This hierarchical transfer mechanism enhances model generalization and noise resilience by aligning multilevel feature representations across different domains. Experimental validation using data from physical test rigs demonstrates that the DTD–ZSTD method achieves an average diagnostic accuracy exceeding 80% in the absence of physical samples, outperforming traditional and state-of-the-art models. These findings underscore the effectiveness of integrating digital twin technology with zero-shot transfer learning to tackle small-sample and cross-domain challenges in bearing fault diagnosis, offering a robust paradigm for intelligent maintenance under data-scarce scenarios.