Structural fatigue life prediction considering model uncertainties through a novel digital twin-driven approach

In this work, a digital twin (DT)-driven approach is proposed to accurately predict structural fatigue life by establishing effective dual-information communication between a DT virtual model and a physical model of the structure of interest. The proposed DT virtual model consists of three modules (namely one crack tracking model, one high-precision approximating model and one dynamic Bayesian network (DBN) inference model) and operates in offline and online stages. The offline stage employs the extended finite element method (XFEM) to establish the crack tracking model, which will generate sufficient labeled datasets to train the high-precision approximating model. In the online stage, the model parameters are updated by the DBN inference model based on the well-trained approximating model, where real-time information exchange from the physical model of the structure is performed. As a result, unexpected uncertainties of the model parameters can be significantly reduced. Numerical examples are carried out to evaluate the performance of the proposed DT-driven approach and the analysis results demonstrate that the fatigue crack growth can be efficiently and accurately predicted.