Machine learning of Truck Traffic Classification groups from Weigh-in-Motion data

Abstract The pavement Mechanistic-Empirical (ME) design requires high-dimensional traffic feature inputs by categories, including Vehicle Class Distributions (VCD), Monthly Distribution Factors (MDF), Hourly Distribution Factors (HDF), and Normalized Axles Load Spectra (NALS). In simplifying the Pavement ME design practice, Truck Traffic Classification (TTC) groups are commonly used for characterizing traffic inputs. Thus, properly defining TTC groups is critical for state-specific pavement ME design practice. In this study, the truck traffic data from existing Weight-in-Motion (WIM) stations were mined to develop specific TTC groups to assist with pavement ME design practice in Georgia. An effective data analytics procedure was developed by leveraging unsupervised machine learning techniques to reduce the high-dimensional traffic features by stratified Principal Component Analysis (PCA), followed by K-means clustering to establish appropriate TTC groups. For a case study, the performance of two typical designs was evaluated using the AASHTOWare pavement mechanistic-empirical (ME) design software with respect to two scenarios of traffic inputs: (1) the derived cluster-based groups, and (2) the national default TTC groups. The results indicated that direct application of the national default TTC groups resulted in over-design of pavement structure in Georgia. Therefore, it is highly recommended that customized TTC groups should be developed using state-specific WIM data.