A Clustering Approach to Linking Pavement Mix Design and Safety Outcomes for Pavement Management

Abstract
The goal of this ongoing research is to develop tools and techniques to design, deliver, and manage pavements for safety using continuous friction and macrotexture data. The technique outlined in this paper identifies and groups road segments by factors corresponding to the aggregate ingredient proportions of various pavement mix designs. The results of the analysis are used in a friction prediction model to better understand the impact of various elements of a mix design on in-service friction performance. In this paper, we will discuss data acquisition, preprocessing, clustering methods, results, and current and potential use cases.

Over the last three years, WDM USA collected approximately 40,000 lane miles of continuous pavement friction and macrotexture data on State-maintained roadways in the US state of Kentucky. The dataset includes annual measurements taken in both directions on all Interstate and Parkway roads and in one direction (cardinal or non-cardinal) on all State Primary and State Secondary roads. Road geometry features, continuous pavement friction, and texture values were obtained by WDM's SCRIM road survey vehicle. Road network features (e.g., AADT, speed limit) and aggregate mix design, pavement construction, and age/treatment information was provided by the Kentucky Transportation Cabinet (KYTC).

Preprocessing aggregate mix design and construction record data required significant manual review and consultation with subject matter experts. The team explored several aggregation methods to minimize information loss. Mix design information was assigned to each 0.1-mile (0.16-km) road segment in the WDM survey data.

The exploratory phase of the analysis used a K-Prototype (mixed variable type) algorithm. Aggregate mix design data was represented by a primary and secondary ingredient name and associated mix proportion value. Initial results of this analysis indicated that road geometry features dominated clustering behaviors, with aggregate mix ingredients having minimal impact.

In the subsequent phase of the analysis, aggregate mix design data were restructured to contain only numerical variables. The ingredients in each mix design were aggregated based on particle size and polish resistance, represented as a proportion of the overall design mix. The proportion of reclaimed asphalt pavement (RAP) in each mix was also represented. Various available clustering methods were evaluated, with the final analysis using the K-means clustering algorithm. Initial results showed clear cluster separation driven by particle size and polish resistance values.

The results of the clustering analysis were incorporated into a friction prediction model trained using the Extreme Gradient Boosting (XGBoost) library to predict a friction value represented as the Mean SCRIM Coefficient (MSC). Two methods of incorporating the results of the clustering analysis were compared to a model that did not use any of the results.

KYTC is using this analysis as an input to maintenance programming and pavement friction service level setting. Future implications also include enhanced pavement friction deterioration modelling, surface treatment selection, and safety outcomes through regular/preventative maintenance programming.