Improving Prediction Accuracy of Bridge Element Deterioration for Decision Support

Deterministic models can aid decision-making in the maintenance, repair, and rehabilitation of bridges. However, they are often static and exhibit a memory-less property, thereby failing to account for the dynamic or chronic characteristics of bridges. This study established a robust framework for a bridge maintenance model grounded in deterioration curves of individual elements and the corresponding element-level inspection data sourced from the National Bridge Inventory. Specifically, we identified influential variables through a comprehensive literature review and laid the foundation for subsequent analysis. Then, we employed a comparative analysis of machine learning and deep learning algorithms facilitated feature extraction and pattern recognition, with model performance assessed through the accuracies of the Decision Tree (69.50%), XGBoost (63.5%), Random Forest (69.3%), Logistic Regression (71.0%), and ANN (68.34%) models. The resultant methodology exhibited enhanced predictive accuracy for bridge deterioration, substantiated by the scrutiny of model predictions against real-world outcomes. Notably, the Logistic Regression model coupled with Dempster Shafer Theory algorithms demonstrated the highest accuracy, attaining a commendable 71.0% and the capability to achieve the data fusion of micro electromechanical systems, sensor technologies, and historic datasets. Validation tests conducted on the resulting deterioration curves underscored their efficacy in empowering transportation agencies to make informed decisions. By advancing the precision of bridge deterioration prediction, this research seeks to furnish decision-makers with invaluable support in optimizing infrastructure management strategies.