Research On Evaluation And Prediction Of Pavement Performance Based On Machine Learning

As the social economy develops,the traffic condition becomes more and more critical,which gradually makes the deterioration of pavements more severe.Carrying out the studies about the evaluation and prediction of pavement service performance can accurately understand the grade and trend of pavement service performance,which can also provide a reference for the determination of pavement maintenance strategies.In order to explore the feasibility of applying machine learning techniques to evaluation and prediction of pavement performance,in this study,three algorithms,namely,random forest(RF),support vector machine(SVM),and back propagation neural network(BPNN),were employed to evaluate and predict pavement performance.Based on the Guangdong Provincial Highway Maintenance and Analysis CMAP platform,investigations and analyses of the common national and provincial trunk highways in Guangdong Province were conducted and the data on their performance conditions were collected in this study.Taking influencing factors such as natural location,age,traffic,and pavement structures into consideration,the research entities of pavement structure were determined.The standard method,entropy weight method,and machine learning model based on entropy weight were utilized to evaluate the pavement performance of the research entities.The pavement performance level evaluated by RF,SVM,and BPNN models were highly consistent with the evaluation results evaluated by entropy weight.Among them,SVM model showed the highest accuracy,while BPNN model was the second,and RF model was the lowest.SVM model has evident advantages in evaluating the road performance with a small amount of sample,while RF model behaved in an opposite way.Comparing the evaluation results of the standard method,the entropy weight method,and the machine learning model based on entropy weight,it was found that the machine learning models based on the entropy weight had better superiority.The entropy weight based-model can not only get evaluation results closer to reality but also reflect the correlation between subindices and pavement performance level more objectively.The sensitivity analysis of model evaluation indexes can judge the influence degree of each sub-index on the comprehensive evaluation grade.The machine learning method based on entropy weight was regarded as a new way to evaluate the performance of pavement.Through five-fold cross-validation,RF,SVM,and BPNN prediction models were established to predict the performance indexes consisting of PCI,RQI,SRI and RDI.According to the prediction results,the above models had good reliabilities of forecasting,and fitted the data well,which verified the feasibility of machine learning in prediction of road performance.Among the three models,BPNN model performed best,followed by SVM model,while RF model performed worst.Besides,BPNN model showed a strong generalization ability,and could achieve higher accuracy with fewer samples.Based on machine learning method,the objective relationship between pavement performance and its influencing factors could be excavated,and the future evolution of pavement performance could be predicted.Sequentially,the key factors affecting pavement performance could be found by analyzing the importance of pavement performance influencing factors,which provides a reference for pavement structure design,construction and maintenance.