Asphalt Pavement Performance Prediction Based On Machine Learning

The data was as one of production factors from the China State Council since April9^th,2020.The country stepped in big data time.Accurate prediction for pavement condition would save money each year in pavement engineering.However,the results of the mechanic-empirical method were not precise.Using machine learning method to predict pavement performance,that could not only build correlation between pavement performance and influencing factors,but also offer maintenance advice for pavement management if pavement performance would be combined with machine learning.At the same time,applying suitable pavement database from abroad to pavement performance prediction could accumulate experiences to build domestic performance database as well as full knowledge to different kinds of machine learning models to save necessary man work and time because China was lack of big data,various types and mostly application pavement performance database.So,this paper studied domestic and abroad researchers contribution in pavement performance to found deficiency in prediction methods.At the same time,it was necessary to recommend American Long Term Pavement Performance database（LTPP）in details because abroad researchers were not familiar with it.Further,extracted data from LTPP database about IRI and rut depth was shown in graphic form and Pearson coefficients analysis to fully master pavement condition and elaborate artificial neural networks,support vector machine and gradient boosting decision tree in structure and theory.The IRI and rut depth datasets were trained by making influencing factors as input variables and IRI as well rut depth as prediction objects.Combining grid search with cross validation to search for different combinations of hyper parameters to analysis and obtain best hyper parameters.After that,CAM and one feature out method were used to obtain sensitive influencing factors and final input variables were determined.Final,the model was sorted by pavement prediction performance using root mean square error（RMSE）,mean absolute error（MAE）and coefficient of determination R² after models were retrained,then,the reason of error of models prediction was analyzed.The test results showed that extracted IRI and rut depth datasets described good pavement condition with less pavement distress.And hyper parameters by grid search with 5-fold cross validation improve model prediction performance.Accurate prediction model was built with machine learning methods and GBDT had the best prediction performance.The R²of model were 0.90 and 1.62,RMSE were 0.19 and 1.62 as well as MAE 0.11,1.12 for IRI and rut depth dataset.GBDT could predict pavement performance accurately and offer advice for pavement management as well as be inserted in transportation information platform.