AI-ENABLED STRUCTURAL HEALTH MONITORING AND SAFETY OPTIMIZATION MODELS FOR HIGH-SPEED RAIL INFRASTRUCTURE IN SEISMIC REGIONS

Abstract

This study had evaluated AI-enabled structural health monitoring (SHM) inference and safety optimization linkages for high-speed rail infrastructure in seismic regions using a corridor-scale quantitative dataset and decision-output records. The analyzed dataset had included 48 assets across 12 corridor segments observed for 365 days, producing 312,480 train-pass windows and 96 seismic event windows. Data integrity had remained high overall, with mean system uptime of 93.6%, while missingness had varied by modality, with accelerometers at 4.2% and displacement/tilt channels at 11.7%. Outcome distributions had been highly imbalanced, with damage-present windows representing 1.8% of all windows and severe outcomes representing 0.2%, confirming rare-event conditions for model evaluation. Correlation analysis had shown strong redundancy among amplitude features, where peak and RMS response had correlated at 0.84, and confounding patterns where train speed had correlated with peak acceleration at 0.63 and temperature had correlated with dominant-frequency proxies at –0.46. Reliability assessment had shown acceptable within-regime stability for key indicators, including antiregime consistency of 0.82 for peak acceleration and 0.76 for dominant-frequency proxies. Collinearity diagnostics had identified inflated overlap in the full feature set (peak acceleration VIF 6.8), which had been reduced after screening (retained amplitude indicator VIF 2.6). Regression results had indicated statistically meaningful contributions from probabilistic risk and selected engineered features to damage detection and condition scoring, with the detection model achieving PR-AUC 0.56 compared with 0.41 for a baseline feature-only specification and the condition-score model achieving adjusted R² 0.48. Operational decision alignment had been supported by correlations between risk score and speed restriction tier (0.71) and inspection priority rank (0.67), indicating coherent translation from inferred risk to safety actions under corridor constraints.