Stochastic Performance and Maintenance Optimization Models for Pavement Infrastructure Management
Highway infrastructure, including roads/pavements, contributes significantly to a country’s economic growth, quality of life improvement, and negative environmental impacts. Hence, highway agencies strive to make efficient and effective use of their limited funding to maintain their pavement infrastructure in good structural and functional conditions. This necessitates predicting pavement performance and scheduling maintenance interventions accurately and reliably by using appropriate performance modeling and maintenance optimization methodologies, while considering the impact of influential variables and the uncertainty inherent in pavement condition data.
Despite the enormous research efforts toward stochastic pavement performance modeling and maintenance optimization, several research gaps still exist. Prior research has not provided a synthesis of Markovian models and their associated methodologies that could assist researchers and highway agencies in selecting the Markov methodology that is appropriate for use with the data available to the agency. In addition, past Markovian pavement performance models did not adequately account for the marginal effects of the preventive maintenance (PM) treatments due to the lack of historical PM data, resulting in potentially unreliable models. The primary components of a Markov model are the transition probability matrix, number of condition states (NCS), and length of duty cycle (LDC). Previous Markovian pavement performance models were developed using NCS and LDC based on data availability, pavement condition indicator and data collection frequency. However, the selection of NCS and LDC should also be based on producing pavement performance models with high levels of prediction accuracy. Prior stochastic pavement maintenance optimization models account for the uncertainty of the budget allocated to pavement preservation at the network level. Nevertheless, variables such as pavement condition deterioration and improvement that are also associated with uncertainty, were not included in stochastic optimization models due to the expected large size of the optimization problem.
The overarching goal of this dissertation is to contribute to filling these research gaps with a view to improving pavement management systems, helping to predict probabilistic pavement performance and schedule pavement preventive maintenance accurately and reliably. This study reviews Markovian pavement performance models using various Markov methodologies and transition probabilities estimation methods, presents a critical analysis of the different aspects of Markovian models as applied in the literature, reveals gaps in knowledge, and offers suggestions for bridging those gaps. This dissertation develops a decision tree which could be used by researchers and highway agencies to select appropriate Markov methodologies to model pavement performance under different conditions of data availability. The lack of consideration of pavement PM impacts into probabilistic pavement performance models due to absence of historical PM data may result in erroneous and often biased pavement condition predictions, leading to non-optimal pavement maintenance decisions. Hence, this research introduces and validates a hybrid approach to incorporate the impact of PM into probabilistic pavement performance models when historical PM data are limited or absent. The types of PM treatments and their times of application are estimated using two approaches: (1) Analysis of the state of practice of pavement maintenance through literature and expert surveys, and (2) Detection of PM times from probabilistic pavement performance curves. Using a newly developed optimization algorithm, the estimated times and types of PM treatments are integrated into pavement condition data. A non-homogeneous Markovian pavement performance model is developed by estimating the transition probabilities of pavement condition using the ordered-probit method. The developed hybrid approach and performance models are validated using cross-validation with out-of-sample data and through surveys of subject matter experts in pavement engineering and management. The results show that the hybrid approach and models developed can predict probabilistic pavement condition incorporating PM effects with an accuracy of 87%.
The key Markov chain methodologies, namely, homogeneous, staged-homogeneous, non-homogeneous, semi- and hidden Markov, have been used to develop stochastic pavement performance models. This dissertation hypothesizes that the NCS and LDC significantly influence the prediction accuracy of Markov models and that the nature of such influence varies across the different Markov methodologies. As such, this study develops and compares the Markovian pavement performance models using empirical data and investigates the sensitivity of Markovian model prediction accuracy to the NCS and LDC. The results indicate that the semi-Markov is generally statistically superior to the homogeneous and staged-homogeneous Markov (except in a few cases of NCS and LDC combinations) and that Markovian model prediction accuracy is significantly sensitive to the NCS and LDC: an increase in NCS improves the prediction accuracy until a certain NCS threshold after which the accuracy decreases, plausibly due to data overfitting. In addition, an increase in LDC improves the prediction accuracy when the NCS is small.
Scheduling pavement maintenance at road network level without considering the uncertainty of pavement condition deterioration and improvement over the long-term (typically, pavement design life) likely results in mistiming maintenance applications and less optimal decisions. Hence, this dissertation develops stochastic pavement maintenance optimization models that account for the uncertainty of pavement condition deterioration and improvement as well as the budget constraint. The objectives of the stochastic optimization models are to minimize the overall deterioration of road network condition while minimizing the total maintenance cost of the road network over a 20-year planning horizon (typical pavement design life). Multi-objective Genetic Algorithm (MOGA) is used because of its robust search capabilities, which lead to global optimal solutions. In order to reduce the number of combinations of solutions of stochastic MOGA models, three approaches are proposed and applied: (1) using PM treatments that are most commonly used by highway agencies, (2) clustering pavement sections based on their ages, and (3) creating a filtering constraint that applies a rest period after treatment applications. The results of the stochastic MOGA models show that the Pareto optimal solutions change significantly when the uncertainty of pavement condition deterioration and improvement is included.