A Novel Approach for Identifying Rotational Stiffness of Semirigid Joints by Physics-Informed Neural Networks

Identifying critical unknown parameters in structures is important for evaluation of structural performance and structural health monitoring. However, like most inverse problems, certain difficulties exist in the identification execution, such as low accuracy in the presence of noise and insufficient measurement data. In this paper, we explore theoretical and experimental studies on the identification of rotational stiffness of semirigid joints by constructing physics-informed neural networks (PINNs). PINNs are well-suited for solving inverse problems because of their superior ability to discover unknown parameters using only the governing equations and limited measurement data. Given that the problem of concern involves numerous boundary conditions and high-order governing equations, two improvements are proposed to enhance the efficiency and accuracy of PINNs. First, a series of modulating functions are derived and concatenated in the PINN configuration, thereby converting the outputs of DNNs to automatically satisfy the boundary conditions without need of loss functions for boundary conditions. Then, to alleviate the difficulty caused by high-order derivatives, the auxiliary physics-informed neural network (A-PINN) is configured to downscale the governing equations. Both numerical and experimental verifications of the improved PINN framework were conducted. The numerical simulation studies demonstrated that the proposed approach can precisely identify the rotational stiffness of semirigid joints using only a small amount of static deflection measurement data. Its tolerance of noise in measurement data was also validated. The experimental study showed that the relative errors between the PINN-identified rotational angles at the semirigid joints and the measured values are less than 5% in all test scenarios.