Data Analytics in Asset Management: Cost-Effective Prediction of the Pavement Condition Index

Abstract

Understanding the deterioration of roads is an important part of road asset management. In this study, the long-term pavement performance (LTPP) data and machine learning algorithms were used to predict the deterioration in the pavement condition index (PCI) over 2, 3, 5, and 6 years. In selecting the attributes for conducting the analysis, we targeted ones that are freely available. This approach can help smaller municipalities, which could be short on money or required expertise. For larger ones and transportation agencies, this can save the increasingly significant costs for collecting field data and any associated safety or traffic implications. In addition, we used this category of attributes to better examine the role of data analytics in asset management. Without considering a causal model, can trends in data help assess deterioration in the PCI? Several models using combinations of 15 attributes were learned and tested. The algorithms used in this study were two types of decision trees and their boosted models based on gradient boosted trees. The accuracy of the ensemble of boosted classifiers was considerably higher than their base learners, with some reaching over 80% in predicting unseen data. We also found that dividing data into different climatic zones can change the relative importance of attributes and the overall accuracy of the models. Increasing the prediction span reduces accuracy, while reducing the number of prediction classes (levels of deterioration) increases the accuracy. In addition to automating the calculation and prediction of PCI, this study presented informative or important attributes for prediction. Such analyses could help municipalities and departments of transportations with forming a more effective policy for data collection and management.