Decoupled probabilistic modeling method for seismic-induced track irregularity on CRTS Ⅲ track-bridge system: Prediction and uncertainty quantification

Abstract

Seismic-induced track irregularity (STI) is the key factor affecting train safety in high-speed railway bridges. Due to its high-dimensional spatial characteristics, direct prediction of STI may reduce modeling efficiency and prediction accuracy. More importantly, few existing models quantify the reliability of their outputs, which limits their applicability in post-earthquake. To address this issue, this study proposes a probabilistic model for STI prediction and uncertainty quantification by integrating principal component analysis (PCA) with temporal convolutional network (TCN). First, STI is decoupled into two physical components representing the response amplitude and spatial morphology, respectively. PCA is subsequently introduced to capture the spatial characteristics of STI and establish an efficient low-rank representation. On this basis, TCN is employed to learn the nonlinear relationship between ground-motion inputs and the reduced STI features. Furthermore, an ensemble framework incorporating multiple sources of perturbation is developed to quantify epistemic and aleatory uncertainties. The proposed framework integrates decoupled dimensionality reduction, temporal feature extraction, and probabilistic uncertainty estimation into a unified surrogate modeling strategy. The results show that the proposed TCN-PCA effectively reduces the dimensionality of STI while preserving its dominant spatial characteristics, enabling rapid prediction. The uncertainty quantification results further indicate that the proposed method yields reasonable prediction intervals, providing a basis for evaluating the operational capacity of high-speed railway bridges.