Enhancing analytical toolboxes of pavement management systems via integration of computational intelligence

摘要

Infrastructure asset management refers to a decision support system that helps authorities in making decisions among alternatives for further development of assets or improvement of existing infrastructure. Pavement management system (PMS) is a subset of transportation infrastructure asset management that covers all the activities involved in providing and maintaining pavements at an adequate level of service. Increased pavement deterioration and increased performance demands along with limited budget and human resources have all made pavement management a critical and challenging task.;Probably the most critical analysis tools in every PMS are condition evaluation, performance prediction, needs assessment, prioritization and optimization toolboxes. However, it is not easy to develop these toolboxes in a reliable fashion. Often, agencies have to deal with incomplete, subjective, ambiguous, uncertain, or erroneous information that makes the developed analysis tools, and, in turn, the decision-making process very unreliable or even irrational. Therefore, there is a need for more capable computational tools that can handle problems such as numerical intensity or ambiguous subjectivity within the collected data.;Computational intelligence methods, including machine learning (ML) techniques, and evolutionary algorithms (EA) have been demonstrated to be promising in addressing these difficulties involved in the development of PMS analysis tools. The general goal of this dissertation is to enhance the analytical toolboxes for pavement condition evaluation and performance prediction via integration of computational intelligence techniques.;Combining the computational efficiency of artificial neural networks (ANN) with the reliability and effectiveness of EA paradigms resulted in superior backcalculation of flexible pavement structural properties, compared to conventional available software. The developed backcalculation methodology results in lower deflection-matching errors and is independent of seed values. In addition, a backcalculation methodology considering pavement behavior under different load levels was developed using a multi-objective evolutionary algorithm (MOEA). The multi-objective approach allows implementation of more available deflection data in order to provide better insight into the complex backcalculation problem. Using MOEA in backcalculation of deflection data from FHWA's long-term pavement performance (LTPP) study resulted in more consistent moduli along each pavement section, compared to conventional single objective routines.;A framework is provided for the development and implementation of an ANN performance prediction model in a network-level PMS. The developed pavement roughness prediction model has superior prediction capability compared to multivariate nonlinear regression models, demonstrated by better generalization of performance trends. An example illustrates the implementation of the roughness model along with life-cycle cost analysis (LCCA) in making future pavement rehabilitation recommendations.;Comparing machine learning techniques, the most suitable for pavement performance prediction is determined through rigorous testing. Preprocessing of data using normalization and principal component analysis simplifies the developed prediction model and avoids correlation of input variables. Feed-forward and cascade-forward ANN, support vector machines (SVM), and radial basis function (RBF) networks are contrasted regarding their prediction capability. The developed models do not require previous knowledge about the equation form and can accommodate noisy data. Variants of these learning paradigms were compared using quantitative and qualitative evaluations, including evaluation of model accuracy, generalization capability, parsimony, sensitivity to changes in input factors, and predicted deterioration progression rates.;Overall, this study provides a framework for integration of computational intelligence in backcalculation and performance modeling towards more effective, efficient, and reliable PMS toolboxes compared to previously available solutions. Superior performance of ML techniques in function approximation compared to multivariate nonlinear regression, and EAs in complex optimization compared to exhaustive search routines, have contributed in this regard.