Damage Importance Analysis for Pavement Condition Index Using Machine-Learning Sensitivity Analysis

Abstract

The Pavement Condition Index (PCI) is a prevalent metric for assessing the condition of rigid pavements. The PCI calculation involves evaluating 19 types of damage. This study aims to analyze how different types of damage impact the PCI calculation and the impact of the performance of prediction models of PCI by reducing the number of evaluated damages. The Municipality of León, Gto., Mexico, provided a dataset of 5271 records. We evaluated five different decision-tree models to predict the PCI value. The Extra Trees model, which exhibited the best performance, was used to assess the feature importance of each type of damage, revealing their relative impacts on PCI predictions. To explore the potential for reducing the complexity of the PCI evaluation, we applied Sequential Forward Search and Brute Force Search techniques to analyze the performance of models with various feature combinations. Our findings indicate no significant statistical difference in terms of Mean Absolute Error (MAE) and the coefficient of determination (R2) between models trained with 13 features compared to those trained with all 17 features. For instance, a model using only eight damages achieved an MAE of 4.35 and an R2 of 0.89, comparable to the 3.56 MAE and 0.92 R2 obtained with a model using all 17 features. These results suggest that omitting some damages from the PCI calculation has a minimal impact on prediction accuracy but can substantially reduce the evaluation’s time and cost. In addition, knowing the most significant damages opens up the possibility of automating the evaluation of PCI using artificial intelligence.

Keywords

• pavement condition index
• pavement management system
• machine learning
• decision trees
• feature importance
• feature sensitivity analysis