Research On Key Technologies For Intelligent Maintenance And Management Of Asphalt Pavements

Asphalt pavement maintenance management is the key to ensure the long and durable service of the road.With the increasing pressure of road network maintenance,intelligent maintenance management technology can improve maintenance efficiency and overcome the limitations of the past maintenance experience-oriented model,which has a broad development prospect.However,the overall intelligent level of the current maintenance management system is still not high.There are a large number of empirical models in the key process of the maintenance system,and the analysis and application of maintenance management technology to data is insufficient,which result in the inability to directly guide the maintenance construction in a reliable,accurate and efficient manner.In view of the above problems,this study starts from the whole process of asphalt pavement maintenance management technology system,and carries out data-driven intelligent upgrading of technology for each key maintenance process.The main research contents and results are as follows:(1)Optimization of computer vision-based pavement distress detection algorithm:i.For the current problems of continuous crack segmentation breakage and background noise misidentification that often occur in semantic segmentation of pavement cracks,a pixel-level semantic segmentation network,called Crack W-Net,using a skip-level round-trip sampling block structure is proposed.The experimental results show that Crack W-Net improves the performance of pavement crack segmentation and is equally applicable to other types of pavement deterioration-type disease semantic segmentation tasks.ii.In order to further improve the performance of computer vision-based pavement disease detection,and simultaneously realize disease identification,classification and geometric feature calculation,an end-to-end convolutional neural network,called YOLO-Crack,with dual-task coupling of target detection and semantic segmentation is designed in this study.Experimental results demonstrate that YOLO-Crack outperforms single-task networks and has the potential ability to correct calibration errors in training data.(2)Optimization of asphalt pavement structure modulus detection algorithm: i.For the disadvantages of poor generalization ability and low convergence accuracy of the traditional artificial neural network based falling weight deflectometer structural modulus backcalculation method in terms of dynamic modulus,the hybrid neural network,called Res RNN-W&D,and its migration learning method is proposed.The experimental results show that Res RNN-W&D achieves fast and highly accurate inverse convergence with strong robustness.ii.To address the problem of low frequency and poor timeliness of structural health detection,in this study,real-time monitoring means are added to the original detection system as data complement,Smart Rock sensors are applied to the road maintenance and management stage,and a real-time backcalculation model of pavement dynamic modulus based on improved genetic algorithm is proposed to achieve timely localization of rapidly deteriorating structural health sections.(3)Intelligent maintenance management data base optimization: A data cleaning framework based on spatio-temporal correlation analysis of asphalt pavement inspection indexes and deep artificial neural network is established to realize the assessment of pavement collection data abnormality metric and intelligent cleaning,which significantly improves the reliability of road inspection data.(4)Asphalt pavement performance prediction model study: i.To address the shortcomings of single theory-driven pavement performance prediction models in terms of prediction accuracy,efficiency and reliability,in this study,the idea of combining mechanics-driven and data-driven index prediction is proposed and MEANN is established which based on mechanics empirical method and artificial neural network to predict rutting depth.The experimental results demonstrate that MEANN improves the prediction accuracy and stability compared with the single driver model.ii.In order to utilize the correlation information between multiple indicators data to improve the accuracy and robustness of prediction,in this study,Transformer FRTS is proposed as a joint prediction model for multiple performance indicators of asphalt pavements,which achieves simultaneous high-precision prediction of multiple indicators by mining the time-series correlation of historical data of single indicators and the information correlation between multiple indicators.The experimental results show that the prediction performance of Transformer FRTS exceeds that of neural network-based prediction models.(5)Project-level asphalt pavement maintenance plan decision study: i.From the idea of machine learning,a data-driven optimization upgrade of the decision tree model commonly used for empirical decision making of maintenance schemes is carried out,and an intelligent maintenance decision model,called IWRF,based on improved weighted random forest algorithm for asphalt pavements is proposed.The experimental results prove that the decision accuracy of IWRF is over 90% and the decision tree generation speed is 4 times faster than the traditional method.ii.A reinforcement learning training mechanism for asphalt pavement maintenance decisions based on proximal policy optimization reinforcement learning algorithms from a deep learning perspective.The experimental results show that the reinforcement learning-based decision algorithm has significantly improved the decision accuracy compared with the improved ANN decision,and the policy optimization class reinforcement learning algorithm can visualize the full life-cycle decision steps,which helps conservation engineers to understand the behavioral process of intelligent body decision making.(6)Network-level asphalt pavement maintenance plan decision study: Aiming at the current situation that the pavement maintenance management system is difficult to fully apply intelligent technology for data-driven maintenance practice in maintenance projects,the study analyzes the key causes of this problem and proposes a set of data-driven methods that combine the assessment of maintenance history data,the calculation of comprehensive benefits of maintenance programs and the application of multi-model fusion to upgrade the intelligence level of the management system and realize the engineering application of intelligent maintenance management technology for asphalt pavements.