A Two-Stage Machine Learning Framework for Pavement Asset Management

Abstract: Transportation Asset Management (TAM) is essential for maintaining the efficiency of road networks. Transportation agencies rely on TAM to keep road networks and bridge infrastructure in good condition under the constraint of limited resources. In the context of highway transportation, arguably pavement is the most critical asset class that state agencies own and maintain. Due to its inherent importance and extent, Pavement Asset Management (PAM) takes up a large portion of state resources every year. Recent advances in machine learning (ML) and artificial intelligence (AI) have the potential to make the planning and decision-making for PAM more efficient and to bring about significant savings in valuable resources for relevant programs at state agencies. This study focuses on integrating AI and machine learning into PAM to enhance predictive pavement performance modeling and assessment. Specifically, the study aims to address two key challenges: (1) categorizing road segments based on their resilience to facilitate targeted maintenance and (2) predicting future road conditions using environmental, pavement, and traffic data. Utilizing a comprehensive dataset from New Mexico, we propose a two-stage machine learning approach that combines unsupervised and supervised learning models to categorize road segments into different resilience groups and predict future pavement conditions. The dataset, collected by the New Mexico Department of Transportation (DOT) between 2016 and 2022, includes pavement condition records, maintenance activities, Average Annual Daily Traffic (AADT), emergency events, and historical weather data, including average temperature, humidity, precipitation, frequency of hurricanes, freezing index, and freeze-thaw index. In the first stage, we use K-means clustering to categorize road segments into four resilience groups (low, medium, high, and special cases). The first stage highlights important areas that require attention by exposing the patterns and resilience levels. In the second stage, the Random Forest (RF) model predicts future resilience based on new conditions. The K-means results show that clusters with higher resilience are predominantly associated with lower maintenance intensity and costs, while clusters with lower resilience require more intensive maintenance and incur higher costs. The RF model predicts future road resilience with an overall accuracy of 63.20% and a weighted F1-score of 0.63. Additionally, a feature importance analysis is conducted to highlight the relative significance of each predictor variable in classifying the resilience of road segments. The model identifies key factors such as the cracking percentage, AADT, and the frequency of hurricanes as significant predictors. By identifying critical areas with low resilience, DOTs can prioritize maintenance and rehabilitation efforts to enhance the overall resilience of the road network and bridge infrastructure. The methodology and results from this study can be part of a Decision Support System (DSS) for asset management at state DOTs, as they provide a data-driven approach for improving road maintenance decision-making processes.

Keywords: Transportation Asset Management, Pavement Asset Management, Machine Learning, K-means, Random Forest