MACHINE LEARNING-BASED PREDICTIVE MODELING FOR ASSESSING BRIDGE LOAD CAPACITY USING REAL-TIME SENSOR DATA

Abstract

This study develops and evaluates a real-time, data-driven framework for predicting bridge load capacity by integrating Internet-of-Things (IoT) sensor streams with advanced machine learning. A multi-site dataset of ~200k five-minute records from 48 bridges across temperate, tropical, and continental climates was compiled, cleaned (k-NN imputation; IQR outlier filtering), normalized, and time-synchronized. Models—Random Forest (RF), Gradient Boosting (GBM), Support Vector Regression, and Deep Neural Networks—were trained using blocked time-series cross-validation (80/20 split; five folds) and benchmarked with MAE, RMSE, and R2R^2R2. Ensemble approaches consistently outperformed single learners. RF achieved the best balance of accuracy and stability (MAE ≈ 12.4 kN; RMSE ≈ 18.1 kN; R2R^2R2 ≈ 0.958; fold SD < 2.1%), with an RF–GBM blend reaching R2R^2R2 ≈ 0.962 and 11% lower residual skewness. Incorporating environmental covariates (temperature, humidity) and dynamic features (vibration, deflection rate) improved accuracy by ~10% over structural-only baselines. Sensitivity and correlation analyses identified strain and deflection as dominant predictors (strain rrr ≈ 0.88; deflection rrr ≈ 0.80), with temperature exerting material-dependent moderating effects, particularly in steel bridges. Real-time deployment tests demonstrated operational feasibility with sub-2 s inference latency (RF 1.82 s average), >99% system uptime, and superior accuracy under dynamic loading (RF MAPE ≈ 3.6%). Sliding-window retraining (7-day refresh) mitigated temporal drift and reduced error by ~6–7% relative to static models. Early-warning simulations showed high detection reliability for load-exceedance events (RF true-positive rate 97.2% with low false alarms). Findings establish that harmonized sensing plus ensemble learning yields accurate, robust, and responsive estimates of bridge load capacity, advancing structural health monitoring from periodic inspection toward continuous, anticipatory asset management and providing a reproducible blueprint for physics-aware, data-driven infrastructure decision support.