Pavement Performance Prediction Using Machine Learning and Instrumentation in Smart Pavement

Abstract

The optimization of pavement Maintenance and Rehabilitation (M&R) planning and costs has been historically proven as a complex task. In recent years, Artificial Intelligence (AI) and Machine Learning (ML) applications in pavement engineering data analytics have been gaining momentum. These advanced techniques have shown promising results in civil engineering and transportation asset management. Therefore, designing a smart pavement framework that relies on the actual pavement responses and in-service condition can help with utilising the ML approach toward better understanding the performance of pavements. To implement the concept of “Smart Pavement”, constructing an interactive pavement pilot section that provides the necessary data feedback to improve the decision-makings of M&R would be needed. This thesis focuses on some aspects of the design of in-situ pavement monitoring and the applying selected machine learning techniques for pavement performance prediction. In order to design an effective pavement instrumentation plan, a literature review was conducted to identify and evaluate the major in-situ monitoring devices and previous case studies. Innovative technologies of Structural Health Monitoring (SHM) were also discussed as a part of the sensory system. A potential pilot section was identified by the Region of Waterloo, for which the pavement structure and technical details were retrieved. Based on the results from the literature review and the evaluation of the proposed section details, a preliminary instrumentation layout has been proposed. Next, the interaction between the proposed embedded sensors and surrounding pavement structure under traffic loading was further studied to evaluate the effect of pavement instrumentation on actual structural responses. Therefore, a series of finite element analysis (FEA) scenarios were defined, and modelling was conducted using ABAQUS to quantify the artefact impacts of the sensors on the pavement responses. Based on the FEA results, high stress- and strain-concentration areas were located which can be used to optimize the design of sensor layout, leading to capturing representative critical pavement responses. Consequently, sensor spacing criteria were suggested to avoid device interference for the response measurement. Furthermore, it would be informative to know how, and which AI/ML methods have been previously used for pavement performance prediction purposes. A systematic literature review iii was conducted indicating that majority of studies used Artificial Neural Network (ANN) of which the prediction process is unexplainable to predict International Roughness Index (IRI) resulting in high prediction accuracies (R2 >= 0.9). A Decision Tree (DT) model and a Random Forest (RF) model were developed using the most commonly used input data retrieved from the Long-Term Pavement Performance (LTPP) database to predict IRI. Finally, after the pruning process, the DT model and RF model resulted in a cross-validation accuracy (R^2) of 0.846 and 0.859, respectively. The single tree from the DT model is less complex than the trees from the RF model. Further studies on machine learning model development should be conducted to refine prediction accuracy. Finally, recommendation for future data collection standards from pilot sections were provided to help with developing a pavement response database that can overcome the inconsistencies in the existing LTPP database and potentially improve the reliability of the future pavement performance modelling.