Digital-twin driven multi-feature adversarial transfer for rolling bearing fault diagnosis

Abstract As a core component in mechanical transmission systems, rolling bearings play a pivotal role in providing support and enabling rotational motion. Their operational condition directly influences the overall performance of the machinery. Currently, fault diagnosis for rolling bearings predominantly relies on cross-device deep transfer learning. However, such methods face the challenge of high data-acquisition costs in the source domain, which significantly limits their practical application. To address this issue, this paper proposes a digital-twin driven multi-feature adversarial transfer learning method (MFAT) for rolling bearing fault diagnosis. Unlike conventional deep transfer learning techniques, the proposed approach enriches the training dataset by incorporating simulation data generated from a digital twin model, and constructs a novel MFAT diagnostic framework to improve accuracy. Specifically, the method commences with constructing a digital twin of the rolling bearing via a four-degree-of-freedom dynamic model. Twin vibration data under diverse operating conditions are generated through Runge–Kutta iterative computation. Subsequently, a MFAT diagnosis model is developed by integrating time-series and multi-scale residual information. Driven by the digital twin, the recognition model is trained using extensive simulated data together with limited real‐world measurements, enabling accurate fault recognition under small‐sample conditions. Experimental results verify that the proposed method achieves high diagnostic accuracy under the scarcity of real samples, and comparative studies with existing and state-of-the-art models further confirm its effectiveness.