Research On The Prediction Method Of Pavement Roughness Based On Multi-source Data And Model Fusion

As China’s transportation development enters a new stage,the mileage of road maintenance and its proportion in the total mileage of national roads increases year by year,and large-scale roads need to carry out maintenance work.Pavement roughness is an important index to judge the performance of road surface,and the quality of pavement roughness will not only affect the safety and comfort of driving,but also affect the service quality and service life of the road.The construction of an accurate and effective pavement roughness prediction model can provide technical support for road maintenance and repair decisions.In the thesis,based on the acquisition and integration of multi-source data of pavement roughness inspection data and its influencing factors of XH highway from 2018-2021,the important feature factors of pavement roughness are screened out,various machine learning methods are selected for pavement roughness prediction,and finally an inverse variance weight fusion model based on H-PO optimization is constructed for pavement roughness prediction.Firstly,by analyzing the influencing factors of pavement roughness,the thesis obtains the pavement roughness inspection data of XH highway as well as five influencing factors of road foundation information,road structure material,climate environment,traffic load and pavement disease,and extracts 17 influencing factors from the data of five influencing factors,and integrates the above multi-source data for spatial and temporal scale alignment to get the pavement roughness multi-source data set.Based on the pavement roughness multi-source dataset,a combination of feature importance analysis and feature correlation analysis is used to screen the essential feature factors of pavement roughness,the average weight fusion method is first used to fuse the results of the two feature importance analysis methods of mutual information and ranking importance for preliminary screening,and then the Spielman correlation coefficient is used to conduct feature correlation analysis on the initially screened influencing factors,and finally 10 influencing factors are screened as important feature factors of pavement roughness.Secondly,10 important pavement roughness factors and historical pavement roughness inspection data are selected as model inputs,and 2021 pavement roughness data is used as the prediction target to construct the pavement roughness prediction dataset.Five machine learning methods including integrated learning and non-integrated learning are used to construct pavement roughness prediction models,and the parameters of the 5 models constructed are tuned using Bayesian optimization algorithms,and 4 model evaluation indicators are used to assess the model prediction performance.The results show that the optimized Tab Net,Random Forest,XGBoost and Cat Boost models have good prediction accuracy,with R~2reaching 0.8627,0.8306,0.8532 and 0.8737 respectively,indicating that these four models have better pavement leveling prediction effects,and they are considered to be selected as preparatory models for pavement roughness prediction.Finally,in order to further improve the prediction accuracy of the model for pavement roughness,the Hunter-Prey Optimizer algorithm is used to achieve higher accuracy tuning of the model parameters of the 4 pavement roughness prediction preparations,and after optimizing the 4 monomeric models,the inverse variance weight calculation allocation method is used to construct a weight fusion prediction model for pavement roughness with the 4 monomeric models as the base model.Through the comparative analysis of the results of the monomeric and fusion models,the inverse variance weight fusion model has higher pavement roughness prediction accuracy than the ranked weight fusion model and Cat Boost,which is the best performer among the monomeric models,with R~2 improved by 0.0075 and 0.0291,and MSE,RMSE and MAPE reduced by 0.0163,0.0074,0.1815 and 0.0373,0.0271 and 0.4033,respectively,and Ablation experiments are designed to demonstrate that each base model contributed positively to the prediction performance improvement of the constructed H-PO optimized inverse variance weight fusion model.The H-PO optimized inverse variance weight fusion model proposed in this thesis can achieve accurate and efficient prediction of pavement roughness of XH highway,which can provide a theoretical basis for highway maintenance management departments to determine the timing of pavement maintenance and provide a theoretical basis for the formulation of pavement repair and maintenance strategies.