Research On Roughness Prediction Method Of Asphalt Pavement Based On LTPP

The roughness is an important indicator for judging the performance of pavements,and it affects driving quality,driving safety and road service quality.The asphalt roughness is used as the research object in this paper.The observations in the LTPP database are mined and analyzed.Based on environmental climate,road surface disease,traffic load,pavement structure and other influencing factors,the data mining techniques are used to analyze the data on the level of roughness in the five Canadian provinces(Alberta,Manitoba,Quebec,Ontario,Saskatchewan)for the last 20 years.The selection of indicators that include temperature,freezing index,precipitation,transverse crack,longitudinal crack,block crack,edge cracks,repair and potholes is the focus of this analysis.Then all influencing factors are filtered and the asphalt pavement roughness database is reconstructed.Firstly the influencing factors are analyzed item by item,and the internal correlation of each variable is explored through factor analysis and correlation analysis.Then,logistic model and mixed effect model are selected as the roughness prediction model,and the fitting precision of the two models was improved by adjusting the input parameters.Finally,it is concluded that the complex judgment coefficients R~2 of logistic model and mixed effect model are respectively 0.731 and 0.754,and the accuracy of both models is lower than 80%.In order to make up the problem of the low accuracy of the two models,artificial neural network model is also introduced to predict roughness in this paper.The B-P neural network algorithm is adopted.The model uses sigmoid as an activation function.In order to reduce the complexity,three layers of network training model is designed.The roughness factors are used as input and output samples,then the root-mean-square error(RMSE)and the mean absolute error(MAE)under different conditions are calculated by adjusting the number of input variables and the number of hidden layer neurons.Comparing the experimental results,when the number of input variables is 9 and the number of implied neurons is 4.the value of RMSE and MAE are 0.881 and 0.682 respectively.The highest fitting accuracyR~2 is 0.876The results show that the prediction method based on neural network is low complexity,high efficiency,high precision.It can evaluate the pavement performance effectively.