Development of Machine Learning Based Analytical Tools for Pavement Performance Assessment and Crack Detection

Abstract

Pavement Management System (PMS) analytical tools mainly consist of pavement condition investigation and evaluation tools, pavement condition rating and assessment tools, pavement performance prediction tools, treatment prioritizations and implementation tools. The effectiveness of a PMS highly depends on the efficiency and reliability of its pavement condition evaluation tools. Traditionally, pavement condition investigation and evaluation practices are based on manual distress surveys and performance level assessments, which have been blamed for low efficiency low reliability. Those kinds of manually surveys are labor intensive and unsafe due to proximity to live traffic conditions. Meanwhile, the accuracy can be lower due to the subjective nature of the evaluators. Considering these factors, semiautomated and automated pavement condition evaluation tools had been developed for several years. In current years, it is undoubtable that highly advanced computerized technologies have resulted successful applications in diverse engineering fields. Therefore, these techniques can be successfully incorporated into pavement condition evaluation distress detection, the analytical tools can improve the performance of existing PMSs. Hence, this research aims to bridge the gaps between highly advanced Machine Learning Techniques (MLTs) and the existing analytical tools of current PMSs. The research outputs intend to provide pavement condition evaluation tools that meet the requirement of high efficiency, accuracy, and reliability. To achieve the objectives of this research, six pavement damage condition and performance evaluation methodologies are developed. The roughness condition of pavement surface directly influences the riding quality of the users. International Roughness Index (IRI) is used worldwide by research institutions, pavement condition evaluation and management agencies to evaluate the roughness condition of the pavement. IRI is a time-dependent variable which generally tends to increase with the increase of the pavement service life. In this consideration, a multi-granularity fuzzy time series analysis based IRI prediction model is developed. Meanwhile, Particle Swarm Optimization (PSO) method is used for model optimization to obtain satisfactory IRI prediction results. Historical IRI data extracted from the InfoPave website have been used for training and testing the model. Experiment results proved the effectiveness of this method. Automated pavement condition evaluation tools can provide overall performance indices, which can then be used for treatment planning. The calculations of those performance indices are required for surface distress level and roughness condition evaluations. However, pavement surface roughness conditions are hard to obtain from surface image indicators. With this consideration, an image indicators-based pavement roughness and the overall performance prediction tools are developed. The state-of-the-art machine learning technique, XGBoost, is utilized as the main method in model training, validating and testing. In order to find the dominant image indicators that influence the pavement roughness condition and the overall performance conditions, the comprehensive pavement performance evaluation data collected by ARAN 900 are analyzed. Back Propagation Neural Network (BPNN) is used to develop the performance prediction models. On this basis, the mean important values (MIVs) for each input factor are calculated to evaluate the contributions of the input indicators. It has been observed that indicators of the wheel path cracking have the highest MIVs, which emphasizes the importance of cracking-focused maintenance treatments. The same issue is also found that current automated pavement condition evaluation systems only include the analysis of pavement surface distresses, without considering the structural capacity of the actual pavement. Hence, the structural performance analysis-based pavement performance prediction tools are developed using the Support Vector Machines (SVMs). To guarantee the overall performance of the proposed methodologies, heuristic methods including Genetic Algorithm (GA) and Particle Swarm Optimization (PSO) are selected to optimize the model. The experiments results show a promising future of machine learning based pavement structural performance prediction. Automated pavement condition analyzers usually detect pavement surface distress through the collected pavement surface images. Then, distress types, severities, quantities, and other parameters are calculated for the overall performance index calculation. Cracks are one of the most important pavement surface distresses that should be quantified. Traditional approaches are less accurate and efficient in locating, counting and quantifying various types of cracks initialed on the pavement surface. An integrated Crack Deep Net (CrackDN) is developed based on deep learning technologies. Through model training, validation and testing, it has proved that CrackDN can detect pavement surface cracks on complex background with high accuracy. Moreover, the combination of box-level pavement crack locating, and pixel-level crack calculation can achieve comprehensive crack analysis. Thereby, more effective maintenance treatments can be assigned. Hence, a methodology regarding pixel-level crack detection which is called CrackU-net, is proposed. CrackU-net is composed of several convolutional, maxpooling, and up-convolutional layers. The model is developed based on the innovations of deep learning-based segmentation. Pavement crack data are collected by multiple devices, including automated pavement condition survey vehicles, smartphones, and action cameras. The proposed CrackU-net is tested on a separate crack image set which has not been used for training the model. The results demonstrate a promising future of use in the PMSs. Finally, the proposed toolboxes are validated through comparative experiments in terms of accuracy (precision, recall, and F-measure) and error levels. The accuracies of all those models are higher than 0.9 and the errors are lower than 0.05. Meanwhile, the findings of this research suggest that the wheel path cracking should be a priority when conducting maintenance activity planning. Benefiting from the highly advanced machine learning technologies, pavement roughness condition and the overall performance levels have a promising future of being predicted by extraction of the image indicators. Moreover, deep learning methods can be utilized to achieve both box-level and pixel-level pavement crack detection with satisfactory performance. Therefore, it is suggested that those state-of-the-art toolboxes be integrated into current PMSs to upgrade their service levels.