Data-driven regression with VIF-based feature selection for cable tension estimation

Abstract

Accurate and interpretable cable tension estimation is a critical component of structural health monitoring (SHM) systems for cable-stayed bridges. While deep learning and regularized regression techniques have achieved high accuracy in previous studies, their reliance on large-scale sensor networks and opaque model structures limits their practical deployment and interpretability. This study asks whether cable strand-averaged forces (MLC1–MLC8) can be estimated from indirect, readily available sensing (resonant frequency, stress/strain aggregates, temperature, humidity, wind) using an interpretable multivariate linear regression screened by variance-inflation factors (VIF). On a real bridge dataset (9004 time steps; 442 channels), iterative VIF elimination (cut-off 20) yields a compact predictor set of ≈150 variables and preserves high out-of-sample accuracy (R²≈0.97), while remaining robust to sensor dropouts and missing data. The result is an operationally transparent virtual-sensing surrogate that reduces dependence on any single instrument class and lowers computational/maintenance burden relative to latent-factor or deep models. Comparative analysis with state-of-the-art approaches—including Elastic Net, Partial Least Squares Regression (PLS), and convolutional neural networks (CNNs)—demonstrates that the proposed MLR + VIF model not only achieves competitive accuracy but also offers superior interpretability, lower computational cost, and greater robustness to sensor faults and missing data. This study highlights the potential of VIF-guided linear modeling as a scalable and explainable alternative to complex black-box models, offering a practical solution for real-time SHM applications where transparency, resource efficiency, and reliability are essential.