Digital twin-driven source-free adaptation diagnosis for rolling element bearing under data privacy

Intelligent fault diagnosis methods have been widely adopted in industrial fields with impressive successes, while they heavily rely on massive labeled data. Digital twin (DT) technology can provide sufficient data for training intelligent diagnosis models, tackling the dilemma of fault sample scarcity in practice. However, model distortion and data privacy impede the development of DT-driven fault diagnosis, making it difficult to transfer diagnosis knowledge from DT to measurements. In this paper, a DT-driven source-free adaptation diagnosis (DTSF) framework is proposed to solve the above problem. First, a high-fidelity DT model of rolling element bearing (REB) is constructed through dynamic modeling and structural parameter updating, which reflects the operating condition of the physical REB. Then, a source-free unsupervised domain adaptation network is developed to perform knowledge transfer, consisting of two separated stages: (1) the source model generation, where labeled DT data are fed to pre-train the model, and (2) the model adaptation, where both DT model and data are denied access, the model is adapted to unlabeled measured data only. This source-free paradigm is in line with the demands of data privacy protection and lightweight diagnosis. Furthermore, extensive experiments on two different REB datasets demonstrate the effectiveness and superiority of the proposed methodology.