Explainable ensemble learning framework for estimating corrosion rate in suspension bridge main cables

Ensuring the safe operation of suspension bridges is paramount to prevent unwanted events that can cause failures. Therefore, it is crucial to continuously monitor their operational status to uphold safety and reliability levels. However, natural deterioration caused by the surrounding environment, primarily due to corrosion, inevitably impacts these structures over time, particularly the main cables made of steel. In this study, a robust framework is proposed to predict the annual corrosion rate in main cables of suspension bridges, while investigating the impact of the surrounding environmental factors on this process. To do so, the implementation of four regression models and four machine learning techniques are used in the first phase for modeling the annual corrosion rate based on a comprehensive database containing various environmental factors. The modeling performance is evaluated through a range of statistical and graphical metrics. After that, Shapley Additive Explanations (SHAP) is utilized to explain the model and to extract the impact of each variable on the final modeling results. Overall, the findings demonstrate the effectiveness of the proposed framework for addressing this issue. The Extreme Gradient Boosting (XGB) emerged as the top-performing model, achieving an overall R 2 of 0.982. Moreover, the SHAP findings highlight the impact of CL − on the annual corrosion rate as the factor with the highest influence during the modeling process. The high performance of the proposed model suggests its potential utility in further research concerning the reliability of suspension bridge main cables. • A machine learning framework is developed to accurately estimate the annual corrosion rate of suspension bridge main cables. • Comparative results between machine learning and regression-based predictive models are presented. • The Shapley Additive Explanations technique explains the optimal ensemble learning approach based on environmental effect. • Extreme Gradient Boosting outperformed all proposed techniques, with an overall R 2 of 0.982.