A method and device for predicting the compaction degree of asphalt mixture

Abstract

The present application relates to a kind of asphalt mixture compactness prediction method and device, the method includes obtaining the original training data of asphalt mixture and the original parameter of asphalt mixture to be detected, based on the repeated K-fold cross-validation, the prediction of multiple iterations of modeling dataset and baseline model is carried out, obtain full amount modeling dataset, according to full amount modeling dataset and univariate linear regression calibration model training, obtain target calibration model;Based on full amount modeling dataset and baseline model, construct multiple baseline sub-model, predict and equal-weight arithmetic average by multiple baseline sub-model to external core verification set, obtain integrated prediction result;Integrated prediction result is input to target calibration model and is predicted, obtain calibrated external verification prediction density value;To realize the prediction of external verification prediction density value, improve stability by calibration mechanism;According to original parameter and external verification prediction density value, determine the asphalt mixture compactness of asphalt mixture to be detected.

Description

Technical Field

[0001] This invention relates to the field of road engineering testing technology, and in particular to a method and apparatus for predicting the compaction degree of asphalt mixtures. Background Technology

[0002] The bulk density of asphalt mixtures is a crucial indicator for evaluating the compaction quality of asphalt pavements, directly impacting pavement structural strength, porosity, rutting resistance, water damage resistance, and service life. With the development of road engineering data acquisition and machine learning technologies, establishing compaction degree prediction models based on mixture material parameters, mix proportion parameters, and related physical properties has become an important technical direction for assisting in asphalt pavement quality inspection. The XGBoost limit gradient boosting tree model, due to its strong nonlinear fitting ability, built-in regularization mechanism, and high computational efficiency, has been used for predicting relevant road engineering indicators.

[0003] However, existing XGBoost-based methods for predicting asphalt mixture compaction degree still have the following shortcomings: The XGBoost model contains multiple hyperparameters, including the number of trees, learning rate, maximum tree depth, sampling ratio, and regularization coefficient, leading to low efficiency in manual parameter tuning. The training results of a single model are easily affected by random seeds, sample partitioning, and the sampling process, resulting in insufficient predictive stability on external core validation sets. There may be systematic deviations between model predictions and measured compaction degrees, and existing methods lack a non-information leakage calibration mechanism based on out-of-sample prediction results.

[0004] Therefore, there is an urgent need to propose a method and device for predicting the compaction degree of asphalt mixtures, in order to solve the technical problem that the existing methods for predicting the compaction degree of asphalt mixtures rely on a single model for training and lack a calibration mechanism without information leakage based on out-of-sample prediction results, thus resulting in insufficient prediction stability. Summary of the Invention

[0005] In view of this, it is necessary to provide a method and apparatus for predicting the compaction degree of asphalt mixtures, in order to solve the technical problem that the existing methods for predicting the compaction degree of asphalt mixtures rely on a single model for training and lack a calibration mechanism without information leakage based on out-of-sample prediction results, thus resulting in insufficient prediction stability.

[0006] To address the aforementioned problems, in a first aspect, the present invention provides a method for predicting the compaction degree of asphalt mixtures, comprising: Obtain the original training data of asphalt mixture and the original parameters of the asphalt mixture to be tested, and divide the original training data into an external core drilling verification set and a modeling dataset; The modeling dataset and baseline model are subjected to multiple iterations of OOF prediction based on repeated K-fold cross-validation to obtain the full modeling dataset. The target calibration model is then trained based on the full modeling dataset and the univariate linear regression calibration model. Based on the full modeling dataset and the baseline model, multiple baseline sub-models are constructed, and the external core verification set is predicted and averaged using the multiple baseline sub-models to obtain the integrated prediction result. The integrated prediction results are input into the target calibration model for prediction to obtain the calibrated external validation prediction density value. The asphalt mixture compaction degree of the asphalt mixture to be tested is determined based on the original parameters and the externally verified predicted density value.

[0007] In one possible implementation, the OOF prediction based on repeated K-fold cross-validation on the modeling dataset and the baseline model is performed multiple times to obtain the full modeling dataset, including: The modeling dataset and baseline model are subjected to multiple iterations of OOF prediction based on repeated K-fold cross-validation to obtain the globally optimal hyperparameter combination. Based on the globally optimal hyperparameter combination, the baseline model is configured and OOF prediction is performed to obtain the full modeling dataset.

[0008] In one possible implementation, the step of performing multiple iterations of OOF prediction on the modeling dataset and baseline model based on repeated K-fold cross-validation to obtain the globally optimal hyperparameter combination includes: Set the random loop count and maximum number of iterations; In the current iteration, based on the 9-dimensional search space, the first hyperparameter configuration is randomly generated within the baseline XGBoost algorithm framework; The baseline model is configured according to the first hyperparameter configuration to obtain the first candidate intermediate model; The modeling dataset and the first candidate intermediate model are subjected to OOF prediction using repeated K-fold cross-validation with different fold partitions for random seeds, and the random root mean square error of each fold validation corresponding to each random seed is obtained. Based on all random root mean square errors and the objective function, the first average root mean square error is obtained; When the current iteration number is greater than the random loop number, the first average root mean square error of the random loop number is obtained; The second hyperparameter configuration is generated by inferring the first average root mean square error of the random cycle number using a tree-based Parsons estimator. The baseline model is configured according to the second hyperparameter configuration to obtain a second candidate intermediate model; Repeated K-fold cross-validation with different fold partitioning of random seeds is used to perform OOF prediction and root mean square error calculation on the modeling dataset and the second candidate intermediate model to obtain the second average root mean square error; When the current iteration number is greater than the maximum iteration number, the minimum value among all first average root mean square errors and all second average root mean square errors is determined as the optimal average root mean square error, and the hyperparameter configuration corresponding to the optimal average root mean square error is determined as the globally optimal hyperparameter combination.

[0009] In one possible implementation, the baseline model is configured and OOF prediction is performed based on the globally optimal hyperparameter combination to obtain a full modeling dataset, including: Based on the globally optimal hyperparameter combination, the baseline model is configured to obtain candidate intermediate models; Standard K-fold cross-validation with a fixed random seed is used to perform OOF prediction on the modeling dataset and the third candidate intermediate model to obtain the optimized prediction results of each fold validation, and all optimized prediction results are determined as the full modeling dataset; the full modeling dataset includes the original input features of each sample in the modeling dataset, the corresponding OOF predicted hair volume density value, and the corresponding measured hair volume density value.

[0010] In one possible implementation, the step of training the target calibration model based on the full modeling dataset and the univariate linear regression calibration model includes: Using the full modeling dataset as the independent variable and the corresponding measured bulk density value as the dependent variable, a univariate linear regression calibration model is fitted. The target calibration model is obtained by training the univariate linear regression calibration model using the full modeling dataset.

[0011] In one possible implementation, the construction of multiple baseline sub-models based on the full modeling dataset and the baseline model, and the prediction and equal-weighted arithmetic average of the external core validation set using the multiple baseline sub-models to obtain an integrated prediction result, includes: Multiple random seeds are randomly generated based on the full modeling dataset, and multiple baseline sub-models are constructed based on the multiple random seeds; Based on each random seed, the corresponding baseline sub-models are trained using the original input features and the corresponding measured hair volume density values ​​in the full modeling dataset to obtain the integrated model set. By using each baseline sub-model in the integrated model group to predict the external core verification set, all prediction results are obtained. The integrated prediction result is obtained by performing an equal-weighted arithmetic average on all the prediction results.

[0012] In one possible implementation, after fitting a univariate linear regression calibration model with the full modeling dataset as the independent variable and the corresponding measured bulk density value as the dependent variable, the method further includes: Repeated K-fold cross-validation with a fixed random seed is used to perform OOF prediction and root mean square error calculation on the full modeling dataset and the univariate linear regression calibration model to obtain the calibrated parameters; the calibrated parameters include the calibration root mean square error. After obtaining the integrated prediction result by predicting and equally weighted arithmetic averaging the external core verification set using the multiple baseline sub-models, the method further includes: The predicted values ​​in the integrated prediction results are summarized with the measured bulk density values ​​corresponding to the external core verification set, and the uncalibrated root mean square error is calculated. After inputting the integrated prediction result into the target calibration model for prediction to obtain the calibrated external validation prediction density value, the method further includes: The calibration root mean square error is compared with the uncalibrated root mean square error to obtain the comparison result; When the comparison results meet the prediction accuracy, output the detailed prediction results of the external core verification set.

[0013] In one possible implementation, the calibrated parameters include the calibration mean relative error; all group samples in the external core drilling verification set prediction result details are sorted from largest to smallest according to the calibration mean relative error, and the original data row number and theoretical maximum density are labeled; determining the asphalt mixture compaction degree of the asphalt mixture to be tested based on the original parameters and the external verification predicted density value includes: Based on the original parameters, the prediction results of the external core verification set are searched to determine the target theoretical maximum density; The compaction degree of the asphalt mixture to be tested is obtained by calculating the target theoretical maximum density and the external verification predicted density value.

[0014] In one possible implementation, after dividing the original training data into an external core validation set and a modeling dataset, the method further includes: The modeling dataset is standardized to obtain standardized parameters and the standardized modeling dataset. The external core verification set is standardized using the standardized parameters to obtain the standardized external core verification set.

[0015] Secondly, the present invention also provides an asphalt mixture compaction degree prediction device, comprising: The data acquisition module is used to acquire the original training data of asphalt mixture and the original parameters of the asphalt mixture to be tested, and divides the original training data into an external core drilling verification set and a modeling dataset. The model training module is used to perform multiple iterations of OOF prediction on the modeling dataset and the baseline model based on repeated K-fold cross-validation to obtain the full modeling dataset, and to train the target calibration model based on the full modeling dataset and the univariate linear regression calibration model. An integrated prediction module is used to construct multiple baseline sub-models based on the full modeling dataset and the baseline model, and to perform prediction and equal-weighted arithmetic averaging on the external core verification set through the multiple baseline sub-models to obtain the integrated prediction result. The density calibration module is used to input the integrated prediction results into the target calibration model for prediction, and obtain the calibrated external verification prediction density value; The density calculation module is used to determine the asphalt mixture compaction degree of the asphalt mixture to be tested based on the original parameters and the externally verified predicted density value.

[0016] The beneficial effects of this invention are as follows: It acquires the original training data of asphalt mixtures and the original parameters of the asphalt mixture to be tested, and divides the original training data into an external core validation set and a modeling dataset; it performs multiple iterations of OOF prediction on the modeling dataset and baseline model based on repeated K-fold cross-validation to obtain a full modeling dataset, and trains the target calibration model based on the full modeling dataset and a univariate linear regression calibration model, thereby achieving automatic parameter tuning without manual tuning; it constructs multiple baseline sub-models based on the full modeling dataset and baseline model, and performs prediction and equal-weighted arithmetic averaging on the external core validation set through multiple baseline sub-models to obtain an integrated prediction result; thus, it performs OOF prediction on the baseline model using the modeling dataset and the external core validation set; it can also increase the stability of the external core validation set prediction by integrating multiple baseline sub-models; it inputs the integrated prediction result into the target calibration model for prediction to obtain the calibrated external validation predicted density value; thus, it can realize the prediction of the external validation predicted density value and improve stability through the calibration mechanism; it determines the asphalt mixture compaction degree of the asphalt mixture to be tested based on the original parameters and the external validation predicted density value. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of an embodiment of the asphalt mixture compaction degree prediction method provided by the present invention; Figure 2 A schematic diagram of an embodiment of the triangular correlation coefficient heatmap provided by the present invention; Figure 3A schematic diagram of an embodiment of the globally optimal hyperparameter combination provided by the present invention; Figure 4 A schematic diagram of coordinates for an embodiment of the 5-fold RMSE comparison chart provided by the present invention; Figure 5 For the present invention Figure 1 A schematic flowchart of an embodiment of step S103; Figure 6 A coordinate schematic diagram of one embodiment of the pavement density diagram at different station numbers provided by the present invention; Figure 7 This is a schematic diagram of an embodiment of the asphalt mixture compaction degree prediction device provided by the present invention. Detailed Implementation

[0018] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0019] like Figure 1 As shown, a specific embodiment of the present invention discloses a method for predicting the compaction degree of asphalt mixtures, comprising: S101. Obtain the original training data of the asphalt mixture and the original parameters of the asphalt mixture to be tested, and divide the original training data into an external core drilling verification set and a modeling dataset.

[0020] In this embodiment of the invention, the original training data of asphalt mixtures can be obtained. The file of the original training data can be obtained through a data interface. The original training data can be data of different asphalt mixtures in historical storage data. The data contains several sets of samples, each set of samples consists of 7 input features and 1 target output, as shown in Table 1: Table 1. Description of Input Features and Target Variables

[0021] Input features can include asphalt density, asphalt dielectric constant, aggregate density, aggregate dielectric constant, asphalt content, asphalt mixture dielectric constant, and theoretical maximum density; the target output can be bulk density. After obtaining the original training data, data cleaning can be performed to obtain cleaned original training data, thereby removing samples with missing input features or target values ​​and retaining complete and valid samples. The cleaned original training data is then divided into an external core drilling validation set and a modeling dataset. For example, 20 samples can be used as the external core drilling validation set to simulate the scenario of core drilling sampling of completed road sections in actual engineering. The remaining samples constitute the modeling dataset for model training, hyperparameter optimization, and extrapolation prediction. During use, the original parameters of the asphalt mixture to be tested can also be obtained. The asphalt mixture to be tested can be the asphalt mixture to be predicted, and the original parameters can include asphalt mixture type, target mix proportion, production mix proportion, and construction batch data.

[0022] S102. Based on repeated K-fold cross-validation, perform multiple iterations of OOF prediction on the modeling dataset and baseline model to obtain the full modeling dataset. Then, train the target calibration model based on the full modeling dataset and the univariate linear regression calibration model.

[0023] In this embodiment of the invention, standard K-fold cross-validation is performed only once. The data is divided into K parts, trained K times, and each sample serves as a validation set exactly once. The result is one group (K) of error values. Repeated K-fold cross-validation, on the other hand, independently repeats the complete process of "standard K-fold cross-validation" R times, and before each execution, the dataset is randomly shuffled again (using a different random seed). Finally, the model is evaluated by combining R×K results. The linear correlation between all input features and the target variable is quantitatively analyzed using the Pearson correlation coefficient matrix, and a triangular correlation coefficient heatmap is plotted as follows. Figure 2 As shown, the heatmap allows engineers to visually determine the strength of the linear correlation between each input feature and the volumetric density, as well as potential multicollinearity issues among the features. This provides a data foundation for subsequent feature understanding and result interpretation. The specific Pearson correlation coefficient formula is shown in Equation 1. (1) In the formula, Representing variables With variables The Pearson correlation coefficient between them.

[0024] The baseline model is the baseline XGBoost model. The XGBoost hyperparameters are configured as follows: number of trees n_estimators=100, maximum tree depth max_depth=6, learning rate learning_rate=0.1, row sampling ratio subsample=1.0, column sampling ratio colsample_bytree=1.0, minimum leaf node sample weight min_child_weight=1, minimum split loss gamma=0, L1 regularization coefficient reg_alpha=0, L2 regularization coefficient reg_lambda=1. The baseline XGBoost model is shown in Table 2. Table 2. Default XGBoost Model Hyperparameter Configuration

[0025] Standard 5-fold cross-validation with fixed random seeds was used to perform OOF prediction on the modeling dataset. The RMSE of each fold was calculated, and the OOF prediction results of all folds were summarized to calculate the four evaluation indicators of the baseline model: RMSE, MAE, MRE, and R², as shown in Equations 2-5. The calculation process of the four evaluation indicators of the baseline XGBoost model in subsequent processes is consistent with this process, serving as a benchmark for comparing the subsequent optimization effects.

[0026] (2) (3) (4) (5) In the formula, For the first The measured bulk density of each sample To predict the bulk density of the hair in the model, This represents the average measured bulk density of the wool. This represents the number of samples.

[0027] This allows for multiple iterations of OOF prediction on the modeling dataset and baseline model based on repeated K-fold cross-validation. Through these repeated predictions, the full modeling dataset can be output.

[0028] Univariate linear regression calibration is a post-processing technique: using the original predicted values ​​from the model output as the sole input features and the actual observed values ​​as the target, it fits a simple linear function to correct systematic biases in the model's predictions (such as overall overestimation or underestimation), thereby improving prediction accuracy. The target calibration model is obtained by training the univariate linear regression calibration model with the full set of modeling data.

[0029] S103. Based on the full modeling dataset and baseline model, construct multiple baseline sub-models, and use multiple baseline sub-models to predict and perform equal-weighted arithmetic averaging on the external core verification set to obtain integrated prediction results.

[0030] In this embodiment of the invention, multiple random seeds can be determined by the full modeling dataset, thereby constructing multiple baseline sub-models for multiple random seeds. Then, the external core verification set can be predicted and averaged using the multiple baseline sub-models to obtain the integrated prediction result.

[0031] S104. Input the integrated prediction results into the target calibration model for prediction to obtain the calibrated external validation prediction density value.

[0032] In this embodiment of the invention, after obtaining the trained target calibration model, the integrated prediction results can be input into the target calibration model for prediction. The target calibration model can output the calibrated external verification prediction density value. The calibrated external verification prediction density value corresponds one-to-one with each core sample to be tested in the external core verification set. The calibration process does not change the sample object, but only corrects the integrated prediction density value of each core sample to be tested.

[0033] S105. Based on the original parameters and the external verification predicted density value, determine the compaction degree of the asphalt mixture to be tested.

[0034] In this embodiment of the invention, the external core drilling verification set contains 20 sets of samples, and also outputs the external verification predicted density values ​​of 20 sets of samples. This allows for matching of the external core drilling verification set with the original parameters. The external verification predicted density values ​​corresponding to samples with consistent parameters are determined, and then the original parameters and the external verification predicted density values ​​are used to calculate the asphalt mixture compaction degree of the asphalt mixture to be tested.

[0035] The asphalt mixture compaction degree prediction method provided in this invention can be applied to an asphalt mixture compaction degree prediction system. The asphalt mixture compaction degree prediction can be based on a software system running on a terminal device. The terminal device can be a server, tablet computer, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), mobile phone, etc. This application embodiment does not impose any restrictions on the specific type of terminal device.

[0036] Compared with existing technologies, this embodiment provides the following method: Obtaining the original training data and original parameters of the asphalt mixture to be tested, and dividing the original training data into an external core validation set and a modeling dataset; performing multiple iterations of OOF prediction on the modeling dataset and baseline model based on repeated K-fold cross-validation to obtain a full modeling dataset; training the target calibration model based on the full modeling dataset and a univariate linear regression calibration model to achieve automatic parameter tuning without manual tuning; constructing multiple baseline sub-models based on the full modeling dataset and baseline model, and using these multiple baseline sub-models to predict the external core validation set and perform equal-weighted arithmetic averaging to obtain an integrated prediction result; thus, performing OOF prediction on the baseline model using the modeling dataset and external core validation set; furthermore, increasing the stability of the external core validation set prediction through the integrated prediction of multiple baseline sub-models; inputting the integrated prediction result into the target calibration model for prediction to obtain the calibrated external validation predicted density value; thereby enabling the prediction of the external validation predicted density value and improving stability through the calibration mechanism; and determining the asphalt mixture compaction degree of the asphalt mixture to be tested based on the original parameters and the external validation predicted density value.

[0037] In some embodiments of the present invention, after step S101, the method further includes: The modeling dataset is standardized to obtain standardized parameters and the standardized modeling dataset.

[0038] In this embodiment of the invention, after dividing the modeling dataset and the external core verification set, the modeling dataset and the external core verification set can be compared and standardized. Specifically, the modeling dataset can be Z-score standardized to calculate the standardization parameters, which may include the mean and standard deviation. At this time, the original data of the modeling dataset itself will be transformed into standard normal distribution data with a mean of 0 and a variance of 1.

[0039] The external core verification set is standardized by standardizing the parameters to obtain the standardized external core verification set.

[0040] In this embodiment of the invention, the sample data in the external core verification set are converted to Z-Score standardization based on the mean and standard deviation, as shown in Formula 6: (6) In the formula, Indicates the first in the external core verification set The first sample The original values ​​of each input feature. and They represent the first and second elements in the modeling dataset, respectively. The mean and standard deviation of each feature. This represents the standardized external core verification set. The first sample The feature values ​​of each input feature.

[0041] In some embodiments of the present invention, step S102 includes: By performing multiple iterations of OOF prediction on the modeling dataset and baseline model using repeated K-fold cross-validation, the globally optimal combination of hyperparameters is obtained.

[0042] In this embodiment of the invention, OOF prediction of the modeling dataset and baseline model can be performed multiple times through repeated K-fold cross-validation. The globally optimal combination of hyperparameters can be determined based on the results of multiple iterations of prediction.

[0043] The baseline model is configured and OOF prediction is performed based on the globally optimal hyperparameter combination to obtain the full modeling dataset.

[0044] In this embodiment of the invention, after determining the globally optimal hyperparameter combination, the baseline model can be configured based on the globally optimal hyperparameter combination. The configured parameters can be referred to Table 2. The parameter values ​​of each parameter in Table 2 are the baseline parameters, while the parameter values ​​of each parameter here are the parameter values ​​in the globally optimal hyperparameter combination. OOF prediction is performed using the configured baseline model to obtain the full modeling dataset.

[0045] In some embodiments of the present invention, such as Figure 3 As shown, based on repeated K-fold cross-validation, multiple iterations of OOF prediction are performed on the modeling dataset and baseline model to obtain the globally optimal hyperparameter combination, including: S301. Set the random loop count and maximum iteration count.

[0046] In this embodiment of the invention, a Tree Parsons Estimator (TPE) is used as the Bayesian optimization sampler. The `n_startup_trials=50` parameter is set, meaning 50 trials are performed randomly before formal Bayesian sampling. This provides sufficient basic data for the TPE to establish an initial prior probability model, avoiding convergence bias caused by over-reliance on a small number of observations in the early stages. For example, the number of random loops can be set to 50 and the maximum number of iterations to 300, resulting in a total of 300 optimization trials.

[0047] S302. In the current iteration, based on the 9-dimensional search space, the first hyperparameter configuration is randomly generated within the baseline XGBoost algorithm framework.

[0048] In this embodiment of the invention, the iteration count can be updated once during each iteration to determine the current iteration count. During the current iteration, based on the 9-dimensional search space and within the baseline XGBoost algorithm framework, the first hyperparameter configuration is randomly generated. Since the random loop count can be 50 times, when the iteration count is less than or equal to 50, the first hyperparameter configuration corresponding to the iteration process can be generated in each iteration, i.e., 50 sets of hyperparameter configurations. These 50 sets of hyperparameter configurations are randomly generated, so they are 50 different sets of hyperparameter configurations. In this embodiment of the invention, a 9-dimensional continuous or discrete hyperparameter search space is defined, with the following search ranges for each dimension: the number of trees (n_estimators) is an integer between 80 and 800; the learning rate (learning_rate) is logarithmically uniformly sampled between 0.003 and 0.12 to more fully explore the low learning rate range; the maximum tree depth (max_depth) is an integer between 2 and 8 to control model complexity; the minimum leaf node sample weight (min_child_weight) is an integer between 1 and 20; the row sampling ratio (subsample) is uniformly sampled between 0.55 and 1.0; the column sampling ratio (colsample_bytree) is uniformly sampled between 0.55 and 1.0; and the minimum splitting loss (gamma) is 1×10^2. -8 Logarithmically uniform sampling between 1 and 20.0; L1 regularization coefficient (reg_alpha) is 1×10 -8 Logarithmically uniform sampling between 1 and 50.0; L2 regularization coefficient (reg_lambda) is 1×10 -8 Logarithmically uniform sampling is used between 100 and 100.0. Logarithmic sampling is used for gamma, reg_alpha, and reg_lambda to retain more exploration points in the small value region close to zero, so that the transition interval from no to some regularization effect is fully searched, as shown in Table 3.

[0049] Table 3. Search Space for Hyperparameters of Optuna-TPE

[0050] S303. Configure the baseline model according to the first hyperparameter configuration to obtain the first candidate intermediate model.

[0051] In this embodiment of the invention, taking one iteration process as an example, after randomly generating the first hyperparameter configuration in the current iteration process, the baseline model can be configured through the first hyperparameter configuration. The configuration process is the same as the baseline model configuration process above, so as to obtain the first candidate intermediate model of the current iteration process.

[0052] S304. Use repeated K-fold cross-validation with different folds to divide the random seeds to perform OOF prediction on the modeling dataset and the first candidate intermediate model, and obtain the random root mean square error of each fold validation for each random seed.

[0053] In this embodiment of the invention, different fold partitioning random seeds can be randomly generated. For example, three different fold partitioning random seeds can be randomly generated, and repeated K-fold cross-validation can be repeated 5-fold cross-validation. Then, repeated 5-fold cross-validation with n_repeats=3 different fold partitioning random seeds (n_splits=5, n_repeats=3, a total of 15 folds for evaluation) is used to perform OOF prediction on the modeling dataset and the first candidate intermediate model. The random root mean square error of each fold validation corresponding to each random seed can be output. That is, when repeated 5-fold cross-validation is performed with the first random seed, the RMSE of each fold validation can be output, which is 5 RMSE. Then, the average root mean square error of repeated 5-fold cross-validation with three random seeds can be obtained, which is 15 RMSE, i.e., random root mean square error.

[0054] S305. Based on all random root mean square errors and the objective function, obtain the first average root mean square error.

[0055] In this embodiment of the invention, all random root mean square errors are substituted into the objective function for calculation to obtain the first average root mean square error, and the first average root mean square error is used as the return value of the objective function, as shown in Formula 7: (7) In the formula, This indicates the combination of XGBoost hyperparameters. Indicates the number of folds. Indicates the number of repetitions. Indicates the first The repetition, the... RMSE under verification.

[0056] S306. When the number of current iterations is greater than the number of random loops, the first average root mean square error of the number of random loops is obtained.

[0057] In this embodiment of the invention, after each iteration ends, the number of the current iteration can be determined. If the number of the current iteration is greater than the number of random loops, if not, the process returns to step S302 to proceed to the next iteration. If it is greater, for example, if the current iteration number is 50, it means that the random iteration has been completed and 50 first average root mean square errors can be obtained.

[0058] S307. The second hyperparameter configuration is generated by inferring the first average root mean square error of the random cycle number using a tree-based Parsons estimator.

[0059] In this embodiment of the invention, after 50 random configuration parameter generation processes, the first average root mean square error generated in the first 50 processes can be used to provide basic data for the iteration process after the 51st iteration, thereby performing Bayesian inference and generating the second hyperparameter configuration.

[0060] S308. Configure the baseline model according to the second hyperparameter configuration to obtain the second candidate intermediate model.

[0061] In this embodiment of the invention, after determining the second hyperparameter configuration, the baseline model can be configured through the second hyperparameter configuration to obtain the second candidate intermediate model. The configuration process is the same as the above process, and will not be described again in this embodiment of the invention.

[0062] S309. Using repeated K-fold cross-validation with different folds to divide the random seed, perform OOF prediction and root mean square error calculation on the modeling dataset and the second candidate intermediate model to obtain the second average root mean square error.

[0063] In this embodiment of the invention, this step is consistent with the process of S304 above. It also uses repeated K-fold cross-validation with different folds to divide the random seeds to perform OOF prediction on the modeling dataset and the second candidate intermediate model, and obtains the root mean square error of each fold. Then, the root mean square error of 15 folds is substituted into Formula 7 for calculation to obtain the second average root mean square error. The specific process will not be repeated.

[0064] Furthermore, after 50 iterations of random parameter configuration, the Optuna median pruner is used from the 51st iteration onwards, when the formal Bayesian sampling phase begins. With n_warmup_steps=5, pruning is not triggered during the first 5 fold validation steps of each experiment. Pruning begins from the 5th fold validation step. Once the model performance stabilizes, it is compared with the historical median. Experiments significantly worse than historical performance are terminated early to save computational resources. For example, for subsequent parameter combinations with extremely poor performance, the experiment is terminated around the 6th fold.

[0065] S310. When the current iteration number is greater than the maximum iteration number, the minimum value among all first average root mean square errors and all second average root mean square errors is determined as the optimal average root mean square error, and the hyperparameter configuration corresponding to the optimal average root mean square error is determined as the global optimal hyperparameter combination.

[0066] In this embodiment of the invention, after each iteration, the number of iterations can be determined. If the number of the current iteration is greater than the maximum number of iterations, and if not, the process returns to step S307 to proceed to the next iteration. If it is greater, it means that 300 iterations have been performed. The goal is to minimize the repeated K-fold average RMSE (i.e., the optimal average root mean square error). The hyperparameter configuration and corresponding target value of each round of experiment are automatically recorded. That is, the minimum value among all the first average root mean square errors of the first 50 iterations and all the second average root mean square errors of the last 250 iterations is the optimal average root mean square error. The hyperparameter configuration corresponding to the optimal average root mean square error is determined as the final output global optimal hyperparameter combination.

[0067] In some embodiments of the present invention, the baseline model is configured and OOF prediction is performed based on the globally optimal hyperparameter combination to obtain a full modeling dataset, including: The baseline model is configured based on the globally optimal hyperparameter combination to obtain the third candidate intermediate model.

[0068] In this embodiment of the invention, after obtaining the globally optimal hyperparameter combination, the baseline model is configured using the globally optimal hyperparameter combination to obtain a third candidate intermediate model. The configuration process is the same as described above.

[0069] Standard K-fold cross-validation with fixed random seeds was used to perform OOF prediction on the modeling dataset and the third candidate intermediate model to obtain the optimized prediction results of each fold validation. All optimized prediction results were determined as the full modeling dataset. The full modeling dataset includes the original input features of each sample in the modeling dataset, the corresponding OOF predicted hair volume density value, and the corresponding measured hair volume density value.

[0070] In this embodiment of the invention, standard K-fold cross-validation with a fixed random seed is then used to perform OOF prediction on the modeling dataset and the third candidate intermediate model, obtaining the optimized prediction results for each fold. This step is the same as the steps described above, except the random seed is different. This step uses a fixed random seed for the same fold partitioning to ensure a fair comparison between the default parameters and the Optuna optimized parameter model under the same data partitioning. All fold OOF prediction results are summarized, and the optimized RMSE, MAE, MRE, and R² are calculated. The individual RMSE for each fold is recorded separately to plot the fold-by-fold RMSE comparison curve, as shown below. Figure 4 As shown, Figure 4This is a fold-by-fold comparison of 5-fold RMSE, with the x-axis representing CV (fold number) and the y-axis representing RMSE. Blue represents the baseline RMSE generated by standard 5-fold cross-validation, and orange represents the 5-fold RMSE from the optimized prediction results of repeated K-fold cross-validation with a fixed random seed. Compared to single 5-fold cross-validation, repeated K-fold validation, by using different fold divisions with random seeds for multiple repetitions, effectively reduces evaluation bias caused by the randomness of a single division, making the hyperparameter evaluation results more robust and reliable. It visually demonstrates the improvement of Optuna hyperparameter optimization compared to default parameters. Then, all optimized prediction results are used as the full modeling dataset; the full modeling dataset can include the original input features of each sample in the modeling dataset, the corresponding OOF predicted hair volume density values, and the corresponding measured hair volume density values.

[0071] In some embodiments of the present invention, step S102 further includes: Using the full modeling dataset as the independent variable and the corresponding measured bulk density as the dependent variable, a univariate linear regression calibration model is fitted.

[0072] In this embodiment of the invention, the full modeling dataset is used as the independent variable, and the corresponding measured bulk density value is used as the dependent variable to fit a univariate linear regression calibration model (linear calibrator). The key advantage of this calibration scheme is that the OOF prediction value is an unbiased prediction generated when the model has not seen the corresponding sample, and it has similar statistical characteristics to the prediction on the test set. Therefore, the calibrator fitted with the OOF result can objectively capture the systematic bias of the model without introducing the problem of overfitting the training set.

[0073] In some embodiments of the present invention, after fitting a univariate linear regression calibration model with the full modeling dataset as the independent variable and the corresponding measured bulk density value as the dependent variable, the method further includes: Repeated K-fold cross-validation with a fixed random seed was used to perform OOF prediction and root mean square error calculation on the full modeling dataset and the univariate linear regression calibration model to obtain the calibrated parameters; the calibrated parameters include the calibration root mean square error. In this embodiment of the invention, repeated K-fold cross-validation with a fixed random seed is used to perform OOF prediction on the full modeling dataset and the univariate linear regression calibration model. The RMSE of each fold is calculated, and then all RMSEs across all folds are calculated to obtain the calibrated parameters. These calibrated parameters may include the calibrated RMSE (root mean square error), MAE (mean absolute error), MRE (mean relative error), and R² (calibration coefficient of determination), verifying the calibration effect. The calibrator also saves calibration predictions for subsequent external validation sets, ensuring consistency of calibration parameters from the modeling set to the external validation set and avoiding secondary fitting to the external validation set.

[0074] The target calibration model is obtained by training the univariate linear regression calibration model using the full modeling dataset.

[0075] In this embodiment of the invention, after constructing a univariate linear regression calibration model, the univariate linear regression calibration model can be trained using a full modeling dataset to obtain the trained target calibration model.

[0076] In some embodiments of the present invention, such as Figure 5 As shown, step S103 includes: S501. Generate multiple random seeds based on the full modeling dataset, and construct multiple baseline sub-models based on the multiple random seeds.

[0077] In this embodiment of the invention, based on the obtained global optimal hyperparameters, 10 significantly different random seeds (e.g., 7, 11, 21, 42, 66, 88, 123, 2024, 3407, 9999 respectively) are selected from the constructed full modeling dataset to construct the baseline sub-model for each random seed.

[0078] S502. Based on each random seed, the corresponding baseline sub-models are trained using the original input features and the corresponding measured hair volume density values ​​in the full modeling dataset to obtain the integrated model group.

[0079] In this embodiment of the invention, 10 XGBoost sub-models are independently trained using the original input features and corresponding measured hair volume density values ​​from the full modeling dataset, forming an ensemble model group. The selection of the 10 random seeds takes into account small, medium, and large values, fully covering different initialization states and ensuring the diversity of the ensemble model group.

[0080] S503. Predict the external core verification set by each baseline sub-model in the integrated model group to obtain all prediction results.

[0081] In this embodiment of the invention, each baseline sub-model in the integrated model group makes predictions on the external core validation set, thereby obtaining the prediction results output by each baseline sub-model, and thus obtaining all prediction results. The input of the baseline sub-model is the external core validation set, and the output is the prediction result.

[0082] S504. Perform an equal-weighted arithmetic average on all prediction results to obtain the integrated prediction result.

[0083] In this embodiment of the invention, the prediction results of 10 sub-models are averaged using equal weights to obtain the ensemble prediction result. The ensemble strategy effectively offsets the random errors of individual sub-models by averaging the prediction errors of multiple independent models, reducing variance and improving the stability and generalization ability of the ensemble prediction.

[0084] In some embodiments of the present invention, after obtaining the integrated prediction result by predicting and equally weighted arithmetic averaging the external core verification set through multiple baseline sub-models, the method further includes: The predicted values ​​in the integrated prediction results are combined with the corresponding measured bulk density values ​​in the external core verification set, and the uncalibrated root mean square error is calculated.

[0085] In this embodiment of the invention, the predicted values ​​in the integrated prediction results and the corresponding measured bulk density values ​​in the external core verification set are calculated using formulas 2-5 to obtain the RMSE (i.e., uncalibrated root mean square error), MAE, MRE, and R² of the external core verification set of the uncalibrated integrated model.

[0086] In some embodiments of the present invention, after step S104, the method further includes: The calibrated root mean square error is compared with the uncalibrated root mean square error to obtain the comparison results.

[0087] In this embodiment of the invention, after obtaining the calibration root mean square error and the uncalibrated root mean square error through the above process, the calibration root mean square error and the uncalibrated root mean square error before and after calibration can be compared and analyzed. The uncalibrated root mean square error is 0.28309, and the calibration root mean square error is 0.27230. The comparison result shows that the RMSE of the external core verification set is reduced from 0.28309 to 0.27230, a reduction of 0.01079.

[0088] When the comparison results meet the prediction accuracy, output the detailed prediction results of the external core verification set.

[0089] In this embodiment of the invention, the comparison result shows a reduction of 0.01079, with an error reduction of approximately 3.81%. This indicates that the calibration mechanism can further improve the model's prediction accuracy. Therefore, the comparison result can be judged. For example, if the comparison result shows a reduction in error, or the reduction in error is greater than a certain predicted number, such as 2%, or the reduction is greater than a certain value, such as 0.01, it can be determined that the comparison result meets the prediction accuracy. The specific settings can be made according to the actual situation. This embodiment of the invention does not make such settings here. The integrated prediction result is the predicted value of the external core verification set. That is, if there are 20 sets of sample data in the external core verification set, the integrated prediction result also includes the predicted values ​​of the 20 sets of sample data. Step S104 is to input the predicted values ​​of the 20 sets of sample data into the target calibration model for prediction. The target calibration model outputs the calibrated external verification prediction density value. The calibrated parameters can include the calibrated RMSE (root mean square error), MAE (mean absolute error), MRE (mean relative error), and R² (calibration determination coefficient). Therefore, after all the predicted values ​​of all sets of samples in the external core verification set are calibrated in the integrated prediction result, the external verification prediction density value of each set of samples can be obtained. When the comparison result meets the prediction accuracy, the detailed prediction result of the external core verification set can be output.

[0090] In some embodiments of the present invention, the calibrated parameters include the calibration mean relative error; all group samples in the external core verification set prediction results details are sorted from largest to smallest according to the calibration mean relative error, and the original data row number and theoretical maximum density are marked; In this embodiment of the invention, all group samples in the external core drilling verification set prediction result details can be sorted from largest to smallest according to the calibration mean relative error, and labeled with the original data row number, measured density, external verification predicted density value, theoretical maximum density, and corresponding calibration mean absolute error and calibration mean relative error. This facilitates engineers to quickly locate high-error samples for targeted verification. Density prediction maps can be obtained by inputting on-site material, station number, and relevant testing information, as shown in the figure. Figure 7 As shown, Figure 6This diagram illustrates the pavement density at different station numbers for asphalt pavement. The x-axis represents the station number, and the y-axis represents the pavement density (i.e., the externally validated predicted density value). The evaluation metrics summary covers the number of modeling samples, the number of external validation samples, and five dimensions of evaluation metrics (RMSE, MAE, MRE, R²) for the default parameter baseline model, the Optuna optimized model, the calibrated model, the externally integrated uncalibrated model, and the externally integrated calibrated model, as well as the Optuna optimal repeated K-fold average RMSE, providing complete data support for performance comparisons at different optimization stages.

[0091] In some embodiments of the present invention, step S105 includes: Based on the original parameters, the detailed prediction results of the external core verification set are searched to determine the target theoretical maximum density.

[0092] In this embodiment of the invention, the external core drilling validation set prediction result details can include all data from 20 sets of samples. This allows for the search for corresponding samples in the external core drilling validation set prediction result details after obtaining the original parameters of the asphalt mixture to be tested. For example, the search can be conducted to find the asphalt mixture type, target mix proportion, production mix proportion, and construction batch of the asphalt mixture to be tested, thereby obtaining the target theoretical maximum density corresponding to the current testing point. The theoretical maximum density can be derived from road design documents, mix proportion design reports, production mix proportion verification data, or construction quality control documents. For different mixture types, different structural layers, different production batches, or different construction sections, the corresponding theoretical maximum density should be used separately. It is not advisable to use the same theoretical maximum density for uniform calculation across all road sections. The units of the predicted bulk density and the theoretical maximum density should be consistent, preferably using g / cm³.

[0093] The compaction degree of the asphalt mixture to be tested is obtained by calculating the target theoretical maximum density and the external verification predicted density value.

[0094] In this embodiment of the invention, the bulk density of the asphalt mixture reflects the mass per unit volume of the mixture after compaction, while the theoretical maximum density reflects the maximum compaction state under the condition that the internal air voids of the mixture are completely eliminated. The ratio of the externally verified predicted density value to the target theoretical maximum density can be calculated to obtain the compaction degree of the asphalt mixture at each test point. As shown in Formula 8: (8) In the formula, Compaction degree of asphalt mixture, in percentage (%) The bulk density of asphalt mixture can be obtained by using the bulk density predicted and calibrated by the model (i.e., the externally verified predicted density value). : The target theoretical maximum density of the asphalt mixture to be tested.

[0095] After obtaining the compaction degree of the asphalt mixture, the road construction quality can be evaluated based on the magnitude of the compaction degree. According to the construction quality control specifications, the compaction degree of ordinary asphalt mixture pavement can be evaluated using the following benchmarks: a theoretical maximum density compaction degree of not less than 93%, a laboratory standard density of not less than 97%, and a test section density of not less than 99%. After calculating the predicted compaction degree for each testing point, it is compared with the construction quality control limits. For asphalt mixtures, when the predicted compaction degree is not less than 93%, the compaction degree of the testing point is deemed to meet the requirements; when it is lower than the corresponding limit, the testing point is deemed to have a risk of insufficient compaction and is included in the key verification points. The system further summarizes the compaction degree results according to road station number, lane, and construction section, quickly identifying continuous low-compaction areas, providing a basis for on-site core drilling verification, rolling process adjustment, and construction quality evaluation.

[0096] Based on the prediction of the bulk density of asphalt mixtures, this invention further combines the predicted bulk density with the theoretical maximum density determined by the design or mix proportion to obtain the predicted compaction degree, which directly reflects the compaction state of the road. This transforms the model output from density prediction to construction quality evaluation. This embodiment of the invention can perform single-point and section evaluations of compaction quality for different station numbers, lanes, and construction sections, quickly identifying insufficient compaction points, critical points, and continuous low-compaction risk sections. This provides data support for construction units to determine key core drilling locations, optimize rolling processes, adjust construction parameters, and conduct quality verification.

[0097] This invention establishes a complete intelligent prediction technology system for the bulk density of asphalt mixtures by organically integrating Bayesian hyperparameter automatic optimization, multi-random seed ensemble modeling, and OOF unbiased linear calibration. This invention employs the Optuna-TPE Bayesian optimization algorithm to replace traditional grid search and random search, achieving efficient global search within a 9-dimensional hyperparameter space through 300 rounds of finite trials. The TPE algorithm constructs a surrogate probability model between parameter values ​​and the objective function, prioritizing the sampling of parameter regions more likely to improve performance in each iteration, resulting in significantly higher search efficiency than random search without prior information. Combined with repeated K-fold cross-validation, the reliability and stability of the obtained optimal hyperparameters are significantly improved compared to a single K-fold evaluation, providing a high-quality parameter foundation for subsequent ensemble models. The ensemble strategy effectively improves external generalization stability, and the OOF calibration mechanism effectively eliminates systematic prediction bias. This invention forms a complete technical closed loop from raw data reading, anomaly handling, standardization, correlation analysis, baseline modeling, Bayesian hyperparameter optimization, ensemble training, and calibration prediction, with each step automated to minimize human intervention.

[0098] To better implement the asphalt mixture compaction degree prediction method in the embodiments of the present invention, correspondingly, the embodiments of the present invention also provide an asphalt mixture compaction degree prediction device, such as... Figure 7 As shown, the asphalt mixture compaction degree prediction device 700 includes: The data acquisition module 701 is used to acquire the original training data of asphalt mixture and the original parameters of the asphalt mixture to be tested, and divides the original training data into an external core drilling verification set and a modeling dataset. The model training module 702 is used to perform multiple iterations of OOF prediction on the modeling dataset and the baseline model based on repeated K-fold cross-validation to obtain the full modeling dataset, and to train the target calibration model based on the full modeling dataset and the univariate linear regression calibration model. The integrated prediction module 703 is used to construct multiple baseline sub-models based on the full modeling dataset and baseline models, and to perform prediction and equal-weighted arithmetic averaging on the external core verification set through multiple baseline sub-models to obtain the integrated prediction results. The density calibration module 704 is used to input the integrated prediction results into the target calibration model for prediction, and obtain the calibrated external validation prediction density value. The density calculation module 705 is used to determine the compaction degree of the asphalt mixture to be tested based on the original parameters and the externally verified predicted density value.

[0099] The asphalt mixture compaction degree prediction device 700 provided in the above embodiments can realize the technical solutions described in the above embodiments of the asphalt mixture compaction degree prediction method. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the asphalt mixture compaction degree prediction method, which will not be repeated here.

[0100] The above provides a detailed description of the asphalt mixture compaction degree prediction method and apparatus provided by the present invention. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting the compaction degree of asphalt mixtures, characterized in that, include: Obtain the original training data of asphalt mixture and the original parameters of the asphalt mixture to be tested, and divide the original training data into an external core drilling verification set and a modeling dataset; The modeling dataset and baseline model are subjected to multiple iterations of OOF prediction based on repeated K-fold cross-validation to obtain the full modeling dataset. The target calibration model is then trained based on the full modeling dataset and the univariate linear regression calibration model. Based on the full modeling dataset and the baseline model, multiple baseline sub-models are constructed, and the external core verification set is predicted and averaged using the multiple baseline sub-models to obtain the integrated prediction result. The integrated prediction results are input into the target calibration model for prediction to obtain the calibrated external validation prediction density value. The asphalt mixture compaction degree of the asphalt mixture to be tested is determined based on the original parameters and the externally verified predicted density value.

2. The method for predicting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The method involves performing multiple iterations of OOF prediction on the modeling dataset and the baseline model based on repeated K-fold cross-validation to obtain the full modeling dataset, which includes: The modeling dataset and baseline model are subjected to multiple iterations of OOF prediction based on repeated K-fold cross-validation to obtain the globally optimal hyperparameter combination. Based on the globally optimal hyperparameter combination, the baseline model is configured and OOF prediction is performed to obtain the full modeling dataset.

3. The method for predicting the compaction degree of asphalt mixtures according to claim 2, characterized in that, The method of performing multiple iterations of OOF prediction on the modeling dataset and baseline model based on repeated K-fold cross-validation to obtain the globally optimal hyperparameter combination includes: Set the random loop count and maximum number of iterations; In the current iteration, based on the 9-dimensional search space, the first hyperparameter configuration is randomly generated within the baseline XGBoost algorithm framework; The baseline model is configured according to the first hyperparameter configuration to obtain the first candidate intermediate model; The modeling dataset and the first candidate intermediate model are subjected to OOF prediction using repeated K-fold cross-validation with different fold partitions for random seeds, and the random root mean square error of each fold validation corresponding to each random seed is obtained. Based on all random root mean square errors and the objective function, the first average root mean square error is obtained; When the current iteration number is greater than the random loop number, the first average root mean square error of the random loop number is obtained; The second hyperparameter configuration is generated by inferring the first average root mean square error of the random cycle number using a tree-based Parsons estimator. The baseline model is configured according to the second hyperparameter configuration to obtain a second candidate intermediate model; Repeated K-fold cross-validation with different fold partitioning of random seeds is used to perform OOF prediction and root mean square error calculation on the modeling dataset and the second candidate intermediate model to obtain the second average root mean square error; When the current iteration number is greater than the maximum iteration number, the minimum value among all first average root mean square errors and all second average root mean square errors is determined as the optimal average root mean square error, and the hyperparameter configuration corresponding to the optimal average root mean square error is determined as the globally optimal hyperparameter combination.

4. The method for predicting the compaction degree of asphalt mixtures according to claim 2, characterized in that, The baseline model is configured and OOF prediction is performed based on the globally optimal hyperparameter combination to obtain the full modeling dataset, including: The baseline model is configured based on the globally optimal hyperparameter combination to obtain a third candidate intermediate model; Standard K-fold cross-validation with a fixed random seed is used to perform OOF prediction on the modeling dataset and the third candidate intermediate model to obtain the optimized prediction results of each fold validation, and all optimized prediction results are determined as the full modeling dataset; the full modeling dataset includes the original input features of each sample in the modeling dataset, the corresponding OOF predicted hair volume density value, and the corresponding measured hair volume density value.

5. The method for predicting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The step of training the target calibration model based on the full modeling dataset and the univariate linear regression calibration model includes: Using the full modeling dataset as the independent variable and the corresponding measured bulk density value as the dependent variable, a univariate linear regression calibration model is fitted. The target calibration model is obtained by training the univariate linear regression calibration model using the full modeling dataset.

6. The method for predicting the compaction degree of asphalt mixtures according to claim 1, characterized in that, The process involves constructing multiple baseline sub-models based on the full modeling dataset and the baseline model, and then using these multiple baseline sub-models to predict and perform an equal-weighted arithmetic average on the external core validation set to obtain an integrated prediction result, including: Multiple random seeds are randomly generated based on the full modeling dataset, and multiple baseline sub-models are constructed based on the multiple random seeds; Based on each random seed, the corresponding baseline sub-models are trained using the original input features and the corresponding measured hair volume density values ​​in the full modeling dataset to obtain the integrated model set. By using each baseline sub-model in the integrated model group to predict the external core verification set, all prediction results are obtained. The integrated prediction result is obtained by performing an equal-weighted arithmetic average on all the prediction results.

7. The method for predicting the compaction degree of asphalt mixtures according to claim 5, characterized in that, The process of fitting a univariate linear regression calibration model using the full modeling dataset as the independent variable and the corresponding measured bulk density value as the dependent variable further includes: Repeated K-fold cross-validation with a fixed random seed is used to perform OOF prediction and root mean square error calculation on the full modeling dataset and the univariate linear regression calibration model to obtain the calibrated parameters; the calibrated parameters include the calibration root mean square error. After obtaining the integrated prediction result by predicting and equally weighted arithmetic averaging the external core verification set using the multiple baseline sub-models, the method further includes: The predicted values ​​in the integrated prediction results are summarized with the measured bulk density values ​​corresponding to the external core verification set, and the uncalibrated root mean square error is calculated. After inputting the integrated prediction result into the target calibration model for prediction to obtain the calibrated external validation prediction density value, the method further includes: The calibration root mean square error is compared with the uncalibrated root mean square error to obtain the comparison result; When the comparison results meet the prediction accuracy, output the detailed prediction results of the external core verification set.

8. The method for predicting the compaction degree of asphalt mixtures according to claim 7, characterized in that, The calibrated parameters include the calibration mean relative error; all group samples in the external core drilling verification set prediction result details are sorted from largest to smallest according to the calibration mean relative error, and the original data row number and theoretical maximum density are marked; the determination of the asphalt mixture compaction degree of the asphalt mixture to be tested based on the original parameters and the external verification predicted density value includes: Based on the original parameters, the prediction results of the external core verification set are searched to determine the target theoretical maximum density; The compaction degree of the asphalt mixture to be tested is obtained by calculating the target theoretical maximum density and the external verification predicted density value.

9. The method for predicting the compaction degree of asphalt mixtures according to claim 1, characterized in that, After dividing the original training data into an external core verification set and a modeling dataset, the method further includes: The modeling dataset is standardized to obtain standardized parameters and the standardized modeling dataset. The external core verification set is standardized using the standardized parameters to obtain the standardized external core verification set.

10. A device for predicting the compaction degree of asphalt mixtures, characterized in that, include: The data acquisition module is used to acquire the original training data of asphalt mixture and the original parameters of the asphalt mixture to be tested, and divides the original training data into an external core drilling verification set and a modeling dataset. The model training module is used to perform multiple iterations of OOF prediction on the modeling dataset and the baseline model based on repeated K-fold cross-validation to obtain the full modeling dataset, and to train the target calibration model based on the full modeling dataset and the univariate linear regression calibration model. An integrated prediction module is used to construct multiple baseline sub-models based on the full modeling dataset and the baseline model, and to perform prediction and equal-weighted arithmetic averaging on the external core verification set through the multiple baseline sub-models to obtain the integrated prediction result. The density calibration module is used to input the integrated prediction results into the target calibration model for prediction, and obtain the calibrated external verification prediction density value; The density calculation module is used to determine the asphalt mixture compaction degree of the asphalt mixture to be tested based on the original parameters and the externally verified predicted density value.

Images