基于机器学习算法的沥青路面密度现场预测模型

Regression and electromagnetic (EM) mixing theory models are frequently implemented to establish the correlation between density and dielectric constant when ground-penetrating radar (GPR) is used to perform in-situ density monitoring of asphalt pavements. The current models exhibit limited precision in parameter calibration and accuracy. To further improve the accuracy of the in situ density prediction models for asphalt pavements, a density prediction model based on a machine learning algorithm was investigated. First, a GPR test platform was constructed to collect the reflected EM wave signals from indoor asphalt mixture specimens and in-situ asphalt pavement using a 2. 2 GHz GPR antenna. The zero bias and environmental noise of the signals were eliminated by employing an average subtraction algorithm and a finite impact response bandpass filter. The dielectric constants were calculated using the reflection amplitude method. Total 124 sets of density data were collected from the available literature, indoor tests, and field inspections. In this study, an extreme gradient boosting (XGBoost) algorithm was used to construct a density prediction model. The dataset was trained using 5-fold cross-validation, and a Bayesian hyperparameter optimization (BHPO) algorithm was used to obtain the optimal combination of model hyperparameters. Finally, the model was tested using field inspection data, and the significance of each input parameter was analyzed. The results demonstrate that after processing the signal to eliminate zero bias and environmental noise, the percentage error of the density prediction decreases significantly, reaching 2. 9% in the indoor test and 3. 0% in the field test. The main difference between the reflected signals of the different asphalt mixtures lies in the location and value of the peak amplitude, and the dielectric constant values of the dense-graded asphalt mixtures are greater than those of the open-graded asphalt mixtures. BHPO effectively improved the accuracy of the XGBoost algorithm, and the error of the XGBoost model was significantly lower than that of the density prediction model based on EM mixing theory on the test set. An analysis of the parameters revealed that the design of the asphalt mixture ratio and accurate detection of the dielectric constant based on GPR are the key factors affecting the density results.