Data-Driven Analysis On Pavement Maintenance Management Optimization

Traffic first.The highway mileage in China is close to 5.2 million kilometers.The increasing mileage and age are accelerating the demand for highway maintenance.Improving efficiency is the critical task of the highway industry,which ensures the high-quality supply of infrastructure and the premium service in transportation.Limited by the cost of data collection and computing resources,the current management methods have barriers in improving efficiency: First,rough data.The decision making is mostly based on statistical performance indicators,which fails to reflect the difference of pavement refinement state and support the right remedy.Second,crude management.The decision making is highly dependent on manual experience,and the efficiency improvement lacks effective theoretical support.In recent years,the rapid development of distributed sensing,data storage,and network communication has promoted the emergence of massive data.The problem of rough data is gradually being solved.Organizing massive data to achieve efficient strategies is an important topic in the field of highway maintenance management.This paper proposes a data-driven maintenance management optimization system based on the theory of highway engineering and data science.Technical methods like ensemble learning and deep reinforcement learning are applied to form executable solutions in the system.This paper sorts out the current management methods and points out that the model-driven and "condition-triggered" management are respectively limited by the efficiency and optimization problem.The data-driven maintenance management framework is established,where four management links are defined: information collection,state deduction,strategy evaluation and intelligent optimization.Three datadriven tasks are conducted to promote efficiency improvement: pavement state deduction in dense data environment,refined data-sensitive maintenance strategy evaluation,and multi-dimensional spatiotemporal strategy optimization.For the generalization in "state deduction",we construct a pavement performance prediction model in dense data environment by ensemble learning.Influencing factors like road structure,environment and traffic loads are extracted as input parameters to reduce the workload of parameter calibration in different scenes.Based on the random forest algorithm,road section data in LTPP dataset are adopted to train the roughness prediction model.Ten common and accessible factors are taken as input.The results show that the proposed algorithm demonstrates better accuracy than the deterministic model.The absolute error rate is lower by 3.01% and the RMSE is lower by 0.026.For the multiple and long-term demand in "strategy evaluation",we propose a refined data-sensitive method for maintenance strategy evaluation.The relationship between internal/external data and the cost-benefit of maintenance is analyzed to reflect the effects on multiple subjects.In the cost calculation,direct economic cost,indirect economic cost and ecological cost are included,which overcomes the limitation of predetermined section division and incomplete coverage in general methods.In the effectiveness calculation,the cost saving based on the refined condition data,and the performance improving based on the value theory,are carried out,which realizes the characterization and calculation in the benefits of multiple subjects.For the solution quality and efficiency in "intelligent optimization",we first study the basic scene of the multi-dimensional spatiotemporal strategy: single section.We propose to apply the deep Q-learning algorithm framework to optimize the maintenance strategy.The life-cycle optimization is described as the Markov decision process to control the state of pavement.The random forest algorithm for performance prediction proposed in this paper is taken as a part of the reinforcement learning environment.The variation of the cost effectiveness value is set as the reward obtained by the selected maintenance measure.The objective is to find the sequence of maintenance measures achieving the maximum of cost effectiveness.In the numerical cases,scenarios of different budget constraints and decision preferences are set to discuss the effect of maintenance strategy.The results show that compared with the "condition-triggered" method,the proposed method will improve the cost-effectiveness by 18%.To broaden the scene of "intelligent optimization ",we expand the single section to multiple continuous subsections and conduct the research on the work zone optimization.We propose to divide road sections,select maintenance measures for each subsection and set up working areas for maintenance.The subsections add the spatial dimension to the state and the action space.The deep Q-learning model will encounter‘dimension explosion’.Therefore,we apply the Deep Centralized Multi-agent Actor Critic framework to construct the work zone optimization model.A centralized actor network is used to output the probability distribution of different actions.The dimension of the action is linear to the number of subsections by setting the conditional independence.The "scale effect" of cost is introduced to the reward calculation.In the numerical cases,we compare and analyze the performance of work zone optimization and single section optimization.The results show that,compared with the single section optimization,work zone optimization strategy obtains higher cost-effectiveness by 1.44%and maintain more stable pavement condition.The work zone optimization can find a better balance between the conversion rate,the scale effect and the overall effectiveness.This research is recommended in long-distance sections with spatial performance difference and uneven disease distribution,which further improves the refinement of road maintenance management.This study promotes the interdisciplinary intersection of road engineering and computer engineering,contributes to efficiency improvement of maintenance projects and provides theoretical basis and technical support for maintenance management and decision-making in the data-driven environment.