Pasupunuri, Sampath Kumar
(2024)
Development of prediction models for various distresses in jointed plain concrete pavement using machine learning and finite element modelling.
PhD thesis, University of Nottingham.
Abstract
Concrete pavements deteriorate over time due to various factors, including traffic loading, climate conditions, construction quality, and initial designs. While the rate of deterioration in rigid pavements is generally slower than that of flexible pavements, the failure mechanism is more complex. Maintaining and rehabilitating concrete pavements to a serviceable condition presents a significant challenge for engineers. To devise effective maintenance strategies, a reliable pavement deterioration model is essential. This research aims to understand jointed plain concrete pavement deterioration adopting Machine Learning (ML) and numerical modelling techniques.
Firstly, this study undertook a comprehensive review of existing literature on the analysis of concrete pavement deterioration mechanisms, various types of prediction models, numerical analysis techniques and the application of machine learning algorithms in pavement modelling. Through this review, insights into the failure mechanisms associated with each distress were gained, helping in the identification of relevant feature variables for each specific type of distress.
By utilising the comprehensive datasets from the Long-Term Pavement Performance (LTPP) and National Highway (NH) databases, this study carefully extracts, processes and analyses the data to develop ML for predicting five distinct distress types: transverse cracking, longitudinal cracking, spalling, faulting and roughness. The feature variables include the different causation factors covering pavement characteristics, traffic loading, climatic properties and performance data. The utilisation of various ML algorithms including Random Forest Regression (RFR), Extreme Gradient Boost (XGBoost/XGB), Support Vector Machine (SVM) and Deep Neural Networks (DNN), yields robust predictive models, with notable performance enhancement observed through optimisation techniques. Further analysis of the models has been conducted on each of the developed models by performing the sensitivity analysis and parametric study to gain deeper insights into the model performance.
The results show that XGB with Optuna optimiser is the preferred model for longitudinal and transverse cracking, utilising 13 and 15 features respectively. For faulting, XGB with BOT optimiser is used, requiring 17 features. DNN with Optuna optimiser perform well for spalling and roughness, using 14 and 8 features respectively. Performance indices such as Coefficient of determination (R2), Mean Absolute Error (MAE), Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) vary across distress types and models. Parametric analysis was conducted in this study to provide insights into the model's effectiveness with a reduced number of features from an application standpoint. The minimum number of features required to get the acceptable performance for transverse cracking, longitudinal cracking, faulting, spalling and roughness are 8, 9, 9,10 and 8, respectively.
The prediction models for UK roads based on the National Highway database for four defects which include transverse cracking, longitudinal cracking, spalling and roughness were also developed in the study. The dataset is constructed using only two years of distress data, and it contains a limited set of features in comparison to the extensive LTPP database. It has been reflected in the overall performance of the developed model. The transverse cracking prediction was developed based on DNN model with the mean absolute error of 1.11 numbers of cracks per 100 meters. For longitudinal cracking, the mean absolute error was lowest for the XGB model, indicating 1.32 longitudinal cracks per 100-meter section, while the mean squared error (MSE) and root mean squared error (RMSE) were lowest for the Random Forest Regression (RFR) model, measuring 5.82 and 2.41, respectively. The MSE for spalling prediction was 4.15 spall per 100 m. The performance of the roughness prediction model developed for the NH dataset is better than other defects. Both the tree-based algorithms, RFR and XGB, when optimised with the Optuna, have given good predictions.
Furthermore, a novel approach Physics-Informed Neural Networks (PINNs) was proposed to predict roughness by integrating domain-specific knowledge and physical principles into the model. By combining the strengths of deep learning with interpretability and physical constraints based on the Mechanistic-Empirical Pavement Design Guide (MEPDG) prediction model, PINNs provide more interpretable outputs and enhanced trust with users. The tuned model demonstrates exceptional performance, with a mean absolute error of 0.134 and a coefficient of determination of 0.90 for the testing dataset.
A desktop application using Tkinter Library was also developed and presented in this study. This application helps the users to interact and access the machine learning models developed in Python language.
Additionally, finite element modelling (FEM) was used to understand concrete pavement behaviour at the joints, providing insights into stress distributions and load transfer mechanisms. In the finite element model developed for this study, 8-noded linear brick elements were employed to represent the concrete pavement, underlying layers, and dowel bars. Parametric studies were performed to explore how certain factors like joint spacing, slab thickness, dowel presence, and load placement impact pavement performance. These studies offered important insights into load transfer efficiency, vertical displacements, and stress patterns within the pavement. The simulation explored how dowels affect load transfer efficiency (LTE) in concrete pavement. It was observed that as friction coefficients increased from 0.2 to 1 in non-dowelled joints, LTE rose significantly from 6.5% to 44%, whereas in dowelled joints, there was only a slight increase from 81.53% to 82.87%. Moreover, dowels substantially decreased vertical displacements, resulting in reduced displacement at transverse joints. Dowel damage analysis provided valuable insights into the vulnerability of the concrete pavement joints to deterioration over time. These findings pave the way for more informed decision-making in infrastructure maintenance and design, enabling stakeholders to interventions and allocate resources effectively.
Overall, this study represents a significant advancement in the field of pavement engineering, offering a paradigm shift towards more accurate data-driven prediction models and deeper insights into the performance of jointed plain concrete pavement.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)