Ensemble Learning Models For Non-contact Evaluation And Decay Prediction Of Asphalt Pavement Skid Resistance

The decrease of skid resistance of asphalt pavement is the main reason of frequent traffic accidents.How to evaluate and predict skid resistance of asphalt pavement scientifically and effectively is the key of road traffic safety performance testing.The traditional road skid resistance testing technology is mainly carried out by contact friction measuring equipment,which is affected by testers,test method and environment,and has weak repeatability and time-consuming.Previous studies have shown that high-precision pavement texture information can be obtained by non-contact measuring equipment based on optical principles such as laser and optical stereo camera,and the relationship between pavement texture and skid resistance is analyzed by combining the roughness characteristics of pavement texture,and the evaluation and prediction model of skid resistance is established.It is of great significance for accurate evaluation and prediction of skid resistance of pavement,rational decision of pavement maintenance time and reduction of traffic accident rate.In this paper,effective macro/micro structure information of pavement is extracted by high-precision 3D laser measurement equipment,and the influence characteristics of asphalt pavement technical conditions are deeply excavated and analyzed.The evaluation model of pavement skid resistance based on XGBoost and the decay prediction model of pavement skid resistance based on GRU are established.It provides reference for the research of intelligent detection technology of skid resistance.Based on these above,the thesis carries out the following work:（1）Based on Butterworth filter,macro and micro texture separation was carried out for the original 3d pavement texture containing noise.MAD method was used to remove local outliers of pavement texture,and discrete wavelet threshold method was used to remove high frequency noise synchronously,and the 3d asphalt pavement texture model was reconstructed accurately by slope correction.On this basis,the effective contact texture projection area of SMA-13 road surface is determined in the interval of[20%,90%]based on Hertz theory（2）The statistical characteristics,fractal characteristics and signal characteristics of asphalt pavement texture are comprehensively analyzed.Aiming at the problem of over-counting and missing box counting in traditional DBC fractal estimation method,an improved omni-directional differential box dimension counting method（IMDBC）for grid displacement was proposed to optimize and improve the existing grid displacement mechanism,which significantly reduced the fitting error in evaluation process.Then,correlation analysis and cluster analysis were used to select effective characteristics closely related to skid resistance and independent evaluation at effective macro and micro scales.The two points Slope Variance(SV_2p)and the Height average（R_a）which overlapped with the Mean Profile Depth（MPD）at the micro level were excluded.（3）Based on the optimized effective contact characteristics of asphalt pavement texture,an intelligent evaluation model for skid resistance of asphalt pavement was established based on machine learning algorithm.The results show that when the effective area is 50%,the evaluation result of XGBoost model is significantly better than other models（R²=0.85）,and the evaluation error is the lowest（MAE=2.386,RMSE=3.095）.（4）The input characteristics of XGBoost-based evaluation model are scientifically and transparently explained by SHapley Additive ex Planations（SHAP）.The analysis results show that the road temperature after sprinkling is the key factor affecting the evaluation results of the model,and the influencing factor is as high as 3.422.Among the numerous texture features,R_ku（Micro）is the most influential feature for evaluating micro texture,reaching 1.023,indicating that R_ku（Micro）fully reflects the change of micro effective contact texture under different wear conditions,and is very suitable for evaluating BPN.The influence coefficients of FD（Upper）,λq（Micro）,Δq（micro）,Cq（5.07mm）and R_ku（Macro）are all below 0.1,indicating that the invalid texture information included by these features is more than valid and replaced by other features.（5）In order to establish an accurate and effective decay prediction model for asphalt pavement skid resistance,the paper collected 48 groups of pavement texture wear data of 11wear stages,and analyzed the prediction effect under long asynchronous conditions based on Gate Recurrent Unit（GRU）time series model.It is found that with the increase of set step size,the model relies on more historical information,and the early stage of rapid attenuation of friction coefficient is gradually unable to learn.Most of the data sets learned by subsequent step size are usually in a stable stage,and the predicted results are also biased to the adjacent values,which promotes the high precision and low error of prediction results.However,setting too long step size will not only cause a sharp decrease in sample size,but also reduce the prediction space of the model.The results show that when the step size is 3,the prediction accuracy of the model is higher and the generalization ability is stronger.The R² values of the prediction results of training set,test set and verification set are 93%,88%and 87%,respectively.