Hybrid machine learning applications in pavement engineering: predicting spalling with PSO-GBM

Abstract

The longevity and serviceability of Continuously Reinforced Concrete Pavement (CRCP) are compromised by spalling in longitudinal joints, a significant pavement distress. Existing predictive models frequently fail to effectively estimate spalling despite a wealth of data on pavement performance because of the complexity of contributing elements, including as traffic, climate, and structural conditions. The predictive accuracy of traditional statistical and machine learning models is unsatisfactory due to their limitations in capturing nonlinear interactions and optimizing hyperparameters. By using data taken from the Long-Term Pavement Performance database, this work fills this gap by developing and assessing predicting models for spalling using cutting-edge machine learning techniques. A robust dataset comprising 395 observations across 33 CRCP sections was analyzed, encompassing diverse environmental and operational conditions. The Particle Swarm Optimization (PSO)-Gradient Boosting Machine (GBM) model was implemented to enhance prediction accuracy and was benchmarked against baseline GBM and Linear Regression models. Results demonstrate that the PSO-GBM outperforms conventional models, achieving the lowest Root Mean Square Error (RMSE) and the highest R-squared (R2) across five-fold cross-validation. Feature importance analysis identified Age, KESAL, and Initial IRI as the most influential variables. Additionally, sensitivity analysis of hyperparameters highlighted the impact of optimization on model performance, while normalized sensitivity analysis and Partial Dependence Plots (PDPs) provided valuable insights into key predictors. This study contributes to pavement engineering by introducing an optimized hybrid machine learning approach that enhances spalling prediction accuracy, offering a data-driven decision-making tool for proactive CRCP maintenance and management.