Author ORCID Identifier

https://orcid.org/0000-0003-3025-6798

Semester

Spring

Date of Graduation

2023

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Civil and Environmental Engineering

Committee Chair

Fei Dai

Committee Member

Yoojung Yoon

Committee Member

John P. Zaniewski

Committee Member

Roger Chen

Committee Member

Linbing Wang

Abstract

The effective functioning of pavements as a critical component of the transportation system necessitates the implementation of ongoing maintenance programs to safeguard this significant and valuable infrastructure and guarantee its optimal performance. The maintenance, rehabilitation, and reconstruction (MRR) program of the pavement structure is dependent on a multidimensional decision-making process, which considers the existing pavement structural condition and the anticipated future performance. Pavement Performance Prediction Models (PPPMs) have become indispensable tools for the efficient implementation of the MRR program and the minimization of associated costs by providing precise predictions of distress and roughness based on inventory and monitoring data concerning the pavement structure's state, traffic load, and climatic conditions. The integration of PPPMs has become a vital component of Pavement Management Systems (PMSs), facilitating the optimization, prioritization, scheduling, and selection of maintenance strategies. Researchers have developed several PPPMs with differing objectives, and each PPPM has demonstrated distinct strengths and weaknesses regarding its applicability, implementation process, and data requirements for development. Traditional statistical models, such as linear regression, are inadequate in handling complex nonlinear relationships between variables and often generate less precise results.

Machine Learning (ML)-based models have become increasingly popular due to their ability to manage vast amounts of data and identify meaningful relationships between them to generate informative insights for better predictions. To create ML models for pavement performance prediction, it is necessary to gather a significant amount of historical data on pavement and traffic loading conditions. The Long-Term Pavement Performance Program (LTPP) initiated by the Federal Highway Administration (FHWA) offers a comprehensive repository of data on the environment, traffic, inventory, monitoring, maintenance, and rehabilitation works that can be utilized to develop PPPMs. The LTPP also includes Weigh-In-Motion (WIM) data that provides information on traffic, such as truck traffic, total traffic, directional distribution, and the number of different axle types of vehicles. High-quality traffic loading data can play an essential role in improving the performance of PPPMs, as the Mechanistic-Empirical Pavement Design Guide (MEPDG) considers vehicle types and axle load characteristics to be critical inputs for pavement design.

The collection of high-quality traffic loading data has been a challenge in developing Pavement Performance Prediction Models (PPPMs). The Weigh-In-Motion (WIM) system, which comprises WIM scales, has emerged as an innovative solution to address this issue. By leveraging computer vision and machine learning techniques, WIM systems can collect accurate data on vehicle type and axle load characteristics, which are critical factors affecting the performance of flexible pavements. Excessive dynamic loading caused by heavy vehicles can result in the early disintegration of the pavement structure. The Long-Term Pavement Performance Program (LTPP) provides an extensive repository of WIM data that can be utilized to develop accurate PPPMs for predicting pavement future behavior and tolerance. The incorporation of comprehensive WIM data collected from LTPP has the potential to significantly improve the accuracy and effectiveness of PPPMs.

To develop artificial neural network (ANN) based pavement performance prediction models (PPPMs) for seven distinct performance indicators, including IRI, longitudinal crack, transverse crack, fatigue crack, potholes, polished aggregate, and patch failure, a total of 300 pavement sections with WIM data were selected from the United States of America. Data collection spanned 20 years, from 2001 to 2020, and included information on pavement age, material properties, climatic properties, structural properties, and traffic-related characteristics. The primary dataset was then divided into two distinct subsets: one which included WIMgenerated traffic data and another which excluded WIM-generated traffic data. Data cleaning and normalization were meticulously performed using the Z-score normalization method. Each subset was further divided into two separate groups: the first containing 15 years of data for model training and the latter containing 5 years of data for testing purposes. Principal Component Analysis (PCA) was then employed to reduce the number of input variables for the model. Based on a cumulative Proportion of Variation (PoV) of 96%, 12 input variables were selected. Subsequently, a single hidden layer ANN model with 12 neurons was generated for each performance indicator.

The study's results indicate that incorporating Weigh-In-Motion (WIM)-generated traffic loading data can significantly enhance the accuracy and efficacy of pavement performance prediction models (PPPMs). This improvement further supports the suitability of optimized pavement maintenance scheduling with minimal costs, while also ensuring timely repairs to promote acceptable serviceability and structural stability of the pavement. The contributions of this research are twofold: first, it provides an enhanced understanding of the positive impacts that high-quality traffic loading data has on pavement conditions; and second, it explores potential applications of WIM data within the Pavement Management System (PMS).

Share

COinS