Abstract
Automated pavement monitoring across pavement networks generates large volumes of high-frequency data that are typically non-stationary, multi-scale, and incomplete, making accurate, physically consistent, and robust prediction challenging. This paper presents Wave-PINN, a hybrid framework that integrates wavelet-based multiscale signal representation with a physics-informed neural network (PINN). Using the International Roughness Index (IRI) of asphalt pavements as a representative application, longitudinal pavement elevation profiles from more than 11000 pavement sections in the Long-Term Pavement Performance (LTPP) database were collected and analyzed. In the proposed framework, longitudinal pavement elevations and vehicle dynamic responses are projected into the wavelet domain, and the quarter-car model is embedded in the loss function via staged physics-informed training. Results demonstrate that Wave-PINN achieves high predictive accuracy (R2 > 0.999) while substantially improving physical consistency, with physics loss reduced by more than 80% as the training data fraction increases from 20% to 80%. The hybrid strategy maintains test accuracy comparable to purely data-driven models while, relative to purely physics-driven models, reducing physics loss by nearly 50% and lowering the coefficient of variation of physics loss by more than 44%. Compared with time-domain neural networks, Wave-PINN provides superior interpolation and extrapolation performance, with relative L2 errors consistently below 0.016. Under incomplete-data conditions, it improves prediction reliability over conventional IRI computation and remains applicable in extreme scenarios where the conventional method becomes inapplicable. These results support Wave-PINN as a robust and effective framework for continuous pavement-signal modeling and network-level pavement-condition assessment.
•A Wave-PINN framework was developed for continuous pavement signals.•Quarter-car governing equations were embedded via staged PINN training.•Wavelet-domain learning improved generalization beyond time-domain models.•Robust IRI prediction was maintained under missing-data conditions.