A digital twin-enabled domain adaptation network for cross-space fault diagnosis of roller bearings

Abstract

The accuracy of roller bearing diagnosis is crucial for ensuring the reliability and safety of mechanical systems. Data-driven methods, such as deep transfer learning (DTL), have been widely applied in fault diagnosis. However, the performance of these models is still limited in industrial scenarios due to the scarcity of labeled data for roller bearings and the complex operational conditions. This study introduces a digital twin-enabled domain adaptation network (DTDA) for the fault diagnosis of roller bearings across digital and physical space. In the digital space, a numerical model of a cylindrical roller bearing with its support housing is created based on the Augmented Lagrange multibody dynamics methodology. Typical fault modes, including raceway defects and cage pillar fracture, are accurately described in this model. A large amount of labeled pedestal vibration signals for corresponding defective roller bearings is then generated. For the physical space, defective roller bearings are machined, and a series of bench tests are carried out to obtain the vibration data of the bearing pedestal. A closed-set domain adaptation network based on DTL is developed to minimize discrepancies in data distribution from these two spaces. A bearing cross-space diagnosis task is constructed using labeled simulation data and unlabeled measurement data. The proposed digital twin-enabled fault diagnosis framework is experimentally validated, and the results demonstrate its superiority over the latest published methods.