Two-Scale Physics-Informed Neural Networks for Structural Dynamics Parameter Inversion: Numerical and Experimental Validation on T-Shaped Tower Health Monitoring

We present a two-scale physics-informed neural network (TSPINN) algorithm to address structural parameter inversion problems involving small parameters. The algorithm’s core mechanism directly embeds small parameters into the neural network architecture. By constructing a two-scale neural network architecture, this approach enables the simultaneous analysis of structural dynamic responses and local parameter perturbation effects, which effectively addresses challenges posed by high-frequency oscillations and parameter sensitivity. Numerical experiments demonstrate that TSPINNs significantly improve prediction accuracy and convergence speed compared to conventional physics-informed neural networks (PINNs) and maintain robustness in high-stiffness scenarios. The T-shaped tower shaking table test results confirm that the model’s identification errors for stiffness reduction coefficients and mass parameters remain below 10% under lower noisy conditions, demonstrating high precision and strong generalization capability for multi-damage scenarios and random load excitations.