Machine learning-based damage identification in stay cables using anchorage vibration responses under moving loads

While existing machine-learning structural health monitoring (SHM) methods predominantly focus on macro-level, severe damage scenarios (>10%), this study presents a baseline feasibility study targeting the theoretical resolution limit for early-stage micro-damage (0–10%) in stay cables. Unlike ambient stochastic excitation, moving vehicle-bridge interaction (VBI) provides a deterministic, broadband, and high-energy impulse that makes micro-damage signatures mathematically extractable. This study aims to quantify such minor cable degradation for a cable-stayed bridge (Shaozhou Bridge) using a Bayesian Optimized Least Squares Support Vector Machine (BO-LSSVM) framework. Firstly, a finite-element model was established to obtain anchorage acceleration responses under varied two-axle truck parameters and multiple minor-damage scenarios (0–10%). To address the computational bottleneck associated with high-dimensional time-history data, Principal Component Analysis (PCA) was employed to compress the raw acceleration signals, which serves as a mathematical prerequisite to preserve damage-sensitive variance while significantly improving computational efficiency. Results demonstrate that the PCA-compressed BO-LSSVM framework can accurately localize and quantify early-stage cable damage. By avoiding local minima and requiring minimal empirical tuning, the BO-LSSVM yields superior predictive accuracy for highly nonlinear micro-damage features compared to traditional algorithms. Overall, this study provides a robust, computationally efficient baseline numerical framework for the SHM of complex infrastructure under operational traffic conditions.