A physics-informed neural network for predicting structural fatigue damage of orthotropic bridge deck through updating model uncertainties

Abstract

Orthotropic bridge deck (OBD) is commonly used in long-span bridges and is the critical structure prone to high-cycle fatigue (HCF) damage. To solve the difficulties in fatigue assessment on full-scale engineering structures, this paper proposed a modern physical-informed neural network (PINN) for conducting the structural fatigue modeling on the OBD structures through updating model uncertainties. Firstly, the three-stage fatigue crack growth (FCG) model for structural components was presented, especially with the multiple uncertainties introduced. Then, the PINN that consists of input layers, output layers, and interpretable hidden layers never involved in the previous model was built and named as StructFatigueNet. Subsequently, following the workflow of the StructFatigueNet, the structural fatigue models on three ODB structures were carried out for the prediction of the local crack and structural displacement change during the life, which showed not only the exceeding 90% accuracy, but also the uses of the structural fatigue behavior prediction, missing crack data acquisition and real-time health monitoring. Compared with a normal LSTM, the StructFatigueNet improved accuracy by 33% owing to the three-stage FCG physics information. Furthermore, it was also extended for other shapes of critical fatigue detail in the long-span bridge, illustrating the relative accuracy.