Asphalt mixture design method based on machine learning and continuous targeted prediction
By using a continuous targeted prediction method based on machine learning, the design model of asphalt mixtures is optimized, which solves the problems of low efficiency of traditional methods and insufficient accuracy of machine learning models. This achieves efficient and accurate prediction of asphalt mixture design and improves the prediction accuracy of void ratio and asphalt saturation.
Patent Information
- Application Number
- CN202510982132.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-16
- Publication Date
- 2025-11-04
AI Technical Summary
Traditional asphalt mixture design methods are inefficient, and machine learning models lack accuracy and interpretability, making it difficult to achieve efficient and accurate prediction of asphalt mixture performance.
A continuous targeted prediction method based on machine learning is adopted. Through multi-level linkage hyperparameter optimization and 5-fold cross-validation, combined with data preprocessing and model feature pruning, the asphalt mixture design model is optimized. The optimal hyperparameters of various machine learning models are used for training and fine-tuning to achieve accurate prediction of various design indicators.
It significantly improved the prediction accuracy of porosity and asphalt saturation, with the coefficient of determination R2 increasing by 2.66% and 1.45% respectively, while reducing experimental costs and improving design efficiency and scientific rigor.
Smart Images
Figure CN120895145A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of road material design and engineering technology, and particularly relates to a bituminous mixture design method based on machine learning and continuous targeted prediction. BACKGROUND
[0002] With the rapid development of the transportation industry, the service performance and service life of road structures are increasingly valued. Among them, bituminous mixture as a key component of road structure, its macro performance directly affects the durability, comfort and maintenance cost of the pavement. The performance of bituminous mixture is influenced by many design parameters such as raw material properties, mixture composition, mix proportion, etc. There is usually a significant nonlinear coupling relationship between these parameters, making it extremely challenging to systematically analyze and accurately model them based on traditional methods. Although many researchers have tried to use formula methods to determine the gradation and asphalt content, these methods still follow a labor-intensive trial-and-error process, which is time-consuming and labor-intensive and difficult to promote. Therefore, it is urgent to develop an efficient and intelligent prediction tool to assist bituminous mixture design and explore the potential relationship between variables to improve design efficiency and scientificity. Machine learning (ML) has become a new tool in the field of material engineering due to its strong nonlinear modeling capability and high-dimensional data processing advantage. Compared with traditional experimental methods, machine learning not only significantly reduces experimental cost and time consumption, but also shows higher adaptability and prediction ability in complex modeling tasks. However, machine learning models still have poor prediction performance in predicting some indicators and lack of interpretability. SUMMARY
[0003] In order to solve the problems existing in the prior art, the present application provides a bituminous mixture design method based on machine learning and continuous targeted prediction, which solves the problems of low efficiency of traditional trial-and-error method, insufficient precision of machine learning model and weak interpretability.
[0004] A bituminous mixture design method based on machine learning and continuous targeted prediction, comprising:
[0005] S1: dividing a bituminous mixture dataset into a training set and a test set, using a multi-level linkage hyperparameter optimization method and k-fold cross-validation to find the best hyperparameters of a plurality of machine learning models for 6 targets on the training set, the 6 targets including gross bulk density Gmb, void ratio VV, mineral aggregate void ratio VMA, asphalt saturation VFA, Marshall stability MS and flow value MF, using the best hyperparameters of the plurality of machine learning models, training the training set, and performing performance evaluation on the test set to obtain a determination coefficient R2, horizontally comparing the determination coefficients R2 of the best hyperparameters of the plurality of machine learning models on the test set, saving the machine learning model with the highest R2 for each target and determining it as the best model for each target;
[0006] S2: according to the determination coefficient R2 of the best model of VV, VFA and VMA on the test set, from high to low, the empirical formula based on VMA, VFA and VMA is sequentially arranged, the input features of the best model corresponding to the target with low determination coefficient are fine-tuned according to the continuous target prediction method, and the prediction model corresponding to the target with low determination coefficient is retrained;
[0007] S3: the model feature importance of the low determination coefficient target prediction model is obtained, the SHAP bee colony diagram is drawn, the features are gradually added to the low determination coefficient target prediction model according to the model feature importance ranking, and the corresponding model is trained, the change of the determination coefficient R2 on the test set is observed, the number of input parameters of the determination coefficient R2 corresponding model is reserved, and the input feature pruning of the low determination coefficient target prediction model is realized.
[0008] Further, the process of preparing the asphalt mixture data set in S1 is: determining the design parameters and design indexes of the asphalt mixture in the asphalt mixture design data set; collecting design parameters and design index data of the asphalt mixture from multiple sources to fill the asphalt mixture design data set; based on data missing value filling, data anomaly value detection, data standardization and data dimensionality reduction, the asphalt mixture design data set is preprocessed to obtain the asphalt mixture data set.
[0009] Further, the design parameters of the asphalt mixture include coarse aggregate index, fine aggregate index, asphalt index, mineral powder index and mixture gradation, and the design indexes of the asphalt mixture include asphalt mixture bulk density, void ratio, mineral gap ratio, asphalt saturation, Marshall stability and flow value.
[0010] Further, the multi-stage linkage hyperparameter optimization method comprises: first using the tree structure Parzen estimation TPE to obtain the approximate range of the best hyperparameters of the machine learning model, and then using the grid search method to fine-tune the parameters.
[0011] Further, the tree structure Parzen estimation TPE models the hyperparameter distribution through kernel density estimation, and searches using Bayesian theorem, and all evaluated hyperparameters are used with a threshold y * All data is divided into two parameter subsets, the first is a better parameter subset corresponding to a smaller target value of the hyperparameter x, the probability distribution l(x), and the second is a worse parameter subset corresponding to a larger target value of the hyperparameter x, the probability distribution g(x), the best machine learning model hyperparameter is the hyperparameter corresponding to the maximum x.
[0012] l(x)=p(x|y<y * )
[0013] g(x) = p(x|y >= y * )
[0014]
[0015] Further, the grid search method defines a new range of super parameter combinations by setting the super parameters to positive and negative 5 units based on the best super parameters obtained by the tree structure Parzen estimation TPE optimization, and the grid search method performs an exhaustive search on the super parameter combinations and evaluates the effect of each super parameter combination through the test set.
[0016] Further, the input features of the best model with low coefficient of determination in S2 include: splicing the target prediction value of the best model corresponding to the target with high coefficient of determination into the input features of the best model corresponding to the target with low coefficient of determination to form the input features of the new low coefficient of determination target.
[0017] Further, the retraining of the prediction model corresponding to the low coefficient of determination target in S2 includes: retraining based on the machine learning model using the input features of the new low coefficient of determination target, the tree structure Parzen estimation and the grid search method, and saving the best model super parameters.
[0018] Further, the model input feature pruning in S3 refers to eliminating the input features with low importance and SHAP values around 0.
[0019] Further, the K-fold cross-validation adopts 5-fold cross-validation, and the 5-fold cross-validation refers to dividing the training data set into 5 data sets of approximately equal size, performing 5 training and validation, in which each time one of the 5 data sets is selected as a validation set and evaluated, and the average value of the results of the 5 validation sets is taken as the final evaluation performance of the model.
[0020] The beneficial effects of the present application include:
[0021] A machine learning-based continuous targeted prediction method for asphalt mixtures preprocesses the asphalt mixture design dataset through data missing value imputation, outlier detection, data standardization, and dimensionality reduction. It optimizes the dataset using multiple machine learning models and a multi-level linked hyperparameter optimization method, employing 5-fold cross-validation for evaluation. This achieves accurate prediction of asphalt mixture design indicators, overcoming the inefficiencies and high costs of traditional trial-and-error methods. By combining the physical relationships between design indicators, high-precision prediction values are concatenated with low-precision target input features to form new low-precision target input features. The low-precision target prediction model is then retrained to ensure optimal performance for each design indicator prediction model. The R² values for porosity and asphalt saturation are improved by 2.66% and 1.45% respectively compared to the initial model, significantly improving the accuracy of porosity and asphalt saturation predictions and achieving continuous targeted prediction based on physical relationships. Finally, the method combines model feature importance evaluation and SHAP bee colony graph to prune specific target input features. Attached Figure Description
[0022] Figure 1 This is an exemplary flowchart of the machine learning-based continuous targeted prediction method for asphalt mixtures involved in the embodiments of this application;
[0023] Figure 2 This is a feature correlation heatmap related to the embodiments of this application;
[0024] Figure 3 This is a schematic diagram of the gradient boosting method for multi-level linkage hyperparameter optimization involved in the embodiments of this application;
[0025] Figure 4 This is a schematic diagram illustrating the principle of the continuous target prediction method involved in the embodiments of this application;
[0026] Figure 5 This is an exemplary schematic diagram of the prediction scatter plots of the optimal machine learning model and the continuous targeting model on the test set involved in the embodiments of this application;
[0027] Figure 6 This is an exemplary schematic diagram illustrating the error between the optimal machine learning model and the continuous targeting model on the test set according to an embodiment of this application;
[0028] Figure 7 This is a schematic diagram of the model feature importance and SHAP bee colony graph involved in the embodiments of this application;
[0029] Figure 8 This is a schematic diagram of model input feature pruning involved in an embodiment of this application. Detailed Implementation
[0030] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0031] Example 1
[0032] The following is in conjunction with the appendix Figure 1 Specific embodiments of the present invention will be described in detail;
[0033] Machine learning-based continuous targeted prediction methods for asphalt mixtures, such as Figure 1 As shown, it includes:
[0034] S1: By analyzing the structural composition of asphalt mixtures, the design parameters and design indices of asphalt mixtures in the asphalt mixture design dataset are determined.
[0035] The design parameters for the asphalt mixture include coarse aggregate indices, fine aggregate indices, asphalt indices, mineral powder indices, and mixture gradation. The design indices include bulk density, porosity, aggregate void ratio, asphalt saturation, Marshall stability, and flow value. The coarse aggregate indices include apparent relative density, crushing value, abrasion rate, and moisture content. The fine aggregate indices include apparent relative density, angularity, soundness, and sand equivalent. The asphalt indices include penetration, softening point, and ductility. The mineral powder indices include apparent relative density and water absorption. The mixture gradation includes the sieve passing rate of each particle size and the asphalt-aggregate ratio.
[0036] S2: Collect design parameters and design index data of the asphalt mixture from multiple sources and populate the asphalt mixture design dataset.
[0037] The design parameters and design index data of the asphalt mixture from various sources include data from master's and doctoral dissertations, engineering reports, and indoor test data.
[0038] The data collected is only for the design parameters and design indices of the AC13, AC16 and AC20 series mixtures.
[0039] The indoor test data is obtained by testing the test methods in the specifications "Highway Engineering Aggregate Test Regulations" (JTG 3432-2024) and "Highway Engineering Asphalt and Asphalt Mixture Test Regulations" (JTG E20-2011). For example, the coarse aggregate abrasion rate is an index calculated by simulating vehicle load and friction effect, the fine aggregate sand equivalent is the sand equivalent value determined by simulating the rainwater scouring and particle sedimentation process by water washing separation method, the asphalt penetration is the vertical penetration depth of the standard needle in the asphalt sample determined by using the standard penetration instrument (load 100g, temperature 25℃, time 5s) and the like.
[0040] S3: Based on data missing value filling, data outlier detection, data standardization and data dimensionality reduction, the asphalt mixture design data set is preprocessed.
[0041] The missing value filling refers to filling the missing positions of each feature of the asphalt mixture design data set obtained in S2 using the feature average value method to form a complete asphalt mixture design data vector.
[0042] The data outlier detection refers to removing all data with a variance greater than 3 for each feature in the asphalt mixture design data set based on the mathematical statistics principle 3σ rule after the missing value filling.
[0043] The data standardization refers to adjusting the numerical range of the asphalt mixture design data set to [-1, 1] after the data outlier detection, and the formula is as follows:
[0044]
[0045] Wherein, X represents the original value of the feature, μ represents the mean value of the feature, σ represents the variance of the feature, and X standardized represents the value of the feature after standardization.
[0046] The data dimensionality reduction refers to deleting the features with a variance less than 0.01 by using the low variance screening method after the data standardization. According to the experimental results, the apparent relative density of the coarse aggregate, the apparent relative density of the fine aggregate, the apparent relative density of the mineral powder, the moisture content of the mineral powder, the 31.5mm sieve hole passing rate and the 26mm sieve hole passing rate feature variance are all less than 0.01. Then, the correlation heat map between the remaining input features is drawn, as shown in Figure 2As shown, each mesh passing rate with high redundancy is reduced in dimension by principal component analysis method, ensuring 90% variance contribution, and three principal components after dimension reduction are obtained, and the input features of 27 dimensions are finally reduced to 13 dimensions, wherein, Coarse Aggregate Apparent Relative Density (CAARD), Coarse Aggregate Crushing Value (CACV); Coarse Aggregate Abrasion Rate (CAAR); Coarse Aggregate Water Absorption (CAWA); Fine Aggregate Apparent Relative Density (FAARD); Fine Aggregate Sand Equivalent (FASE); Fine Aggregate Angularity (FAA); Fine Aggregate Soundness (SFA); Mineral Powder Apparent Relative Density (MPARD); Mineral Powder Moisture Content (MPMC); Oil-rock Ratio (OR); Penetration (PEN); Softening Point (SP); Ductility (DUC), PCA_1, PCA_2, PCA_3 are the three principal components after dimension reduction of each mesh passing rate.
[0047] S4: The asphalt mixture design data is divided into a training set and a test set according to 8:2, the best hyperparameters of a plurality of machine learning models of six targets of Gmb, VV, VMA, VFA, MS and MF are found on the training set based on using a multi-level linkage hyperparameter optimization method and 5-fold cross-validation, the performance evaluation (R 2 ) of the best hyperparameters of the plurality of machine learning models on the test set is performed, the performance evaluation of the best hyperparameters of the plurality of machine learning models on the test set is compared laterally, the machine learning model with the highest R2 of each target is saved and determined as the best model of each target;
[0048] S5: According to the determination coefficient R2 of the best model of VV, VFA and VMA on the test set, from high to low, the empirical formula based on VMA, VFA and VV is sequentially ordered, the input characteristics of the best model corresponding to the target with low determination coefficient are fine-tuned according to the continuous target prediction method, and the prediction model corresponding to the target with low determination coefficient is retrained;
[0049]
[0050] S6: Obtain the model feature importance of the low-determination-coefficient target prediction model, draw the SHAP bee swarm diagram, add features to the low-determination-coefficient target prediction model according to the model feature importance ranking, and train the corresponding model, observe the change of the determination coefficient R2 on the test set, retain the input parameter number of the determination coefficient R2 corresponding model, and realize the input feature pruning of the low-determination-coefficient target prediction model.
[0051] In another embodiment, the multi-level linkage hyperparameter optimization method refers to using tree structure Parzen estimation (TPE) to obtain the approximate range of the best hyperparameters of the machine learning model, and then using grid search method for fine tuning.
[0052] The tree structure Parzen estimation (TPE) models the hyperparameter distribution through kernel density estimation, and searches using Bayesian theorem, and the specific objective function value y is the minimum validation error, and all evaluated hyperparameters are used with a threshold y * All data is divided into two parameter subsets, the first is a better parameter subset, corresponding to the hyperparameter x with smaller target value, the probability distribution l(x), and the second is a worse parameter subset, corresponding to the hyperparameter x with larger target value, the probability distribution g(x), and the best machine learning model hyperparameter is the hyperparameter corresponding to the maximum x.
[0053] l(x) = p(x|y < y * )
[0054] g(x) = p(x|y ≥ y * )
[0055]
[0056] The grid search method defines a new range of hyperparameters with positive and negative 5 units based on the best hyperparameters obtained by the tree structure Parzen estimation (TPE) optimization, lists the hyperparameter candidate value combinations, and performs exhaustive search on the hyperparameter candidate value combinations, and evaluates the effect of each group of hyperparameters through the test set.
[0057] The plurality of machine learning models includes random forest (RF), support vector machine (SVR), BP neural network, CatBoost and XGBoost.
[0058] In another embodiment, the K-fold cross-validation employs 5-fold cross-validation, which is to divide the data into 5 subsets of similar size, and perform 5 subset training and subset testing. In each subset training and subset testing, 4 subsets are used as the subset training set, and 1 subset is used as the subset test set. The performance index on the subset test set is calculated, and the average of the performance index of the 5 subset test sets is taken as the evaluation of the model performance.
[0059] The asphalt mixture design data set is divided into a training set and a test set according to a ratio of 8:2. On the training set, for each target, the tree structure Parzen estimation (TPE) is used to obtain each set of machine learning model hyperparameters, and the 5-fold cross-validation is used to evaluate the performance of the machine learning model (R 2 ). The obtained machine learning model hyperparameters are evaluated, and finally, after 50 searches of the tree structure Parzen estimation (TPE) and 50 evaluations of the 5-fold cross-validation, the possible best machine learning model hyperparameters for the target are determined. Based on the possible best machine learning model hyperparameters, a new range of hyperparameters is defined with a positive and negative 5 units to list the hyperparameter candidate value combinations. The grid search method is used to exhaustively search the hyperparameter candidate value combinations, and the 5-fold cross-validation is used to evaluate the hyperparameters. Finally, the best machine learning model hyperparameters for the target are determined. The training set is trained using the best machine learning model hyperparameters for the target, and the best model is saved. The performance of the best model is evaluated on the test set. By repeating the above operations 6 times, the performance evaluation (R 2 ) of the test set of the target by the random forest (RF), the support vector machine (SVR), the BP neural network, the CatBoost, and the XGBoost is obtained, and the performance evaluation (R 2 ) of the test set of the target by each machine learning model is compared. 2 The highest machine learning model determines the best machine learning model for the target.
[0060] The R 2 is the coefficient of determination, and the formula is as follows:
[0061]
[0062] where y i represents the true value of the i-th sample, represents the predicted value of the i-th sample, represents the average value of the sample, and n represents the number of samples.
[0063] The machine R of each target is obtained according to the S4 step 2 , which is summarized in Table 1 below.
[0064] According to the multi-stage cascading hyperparameter optimization method, the gradient of the model generalization ability can be improved, Figure 3 The tree structure Parzen estimation (TPE) is used to show the changes of the evaluation index R 2 ) of the test set using the grid search method, and the grid search method is used on the basis of the tree structure Parzen estimation (TPE) optimization. The evaluation index R 2 of each prediction target is improved by an average of more than 1%.
[0065] Table 1. Evaluation index of machine learning model on training set and test set
[0066]
[0067] In another embodiment, the input feature of the best model with low coefficient of determination in S5 includes: splicing the target prediction value of the best model corresponding to the target with high coefficient of determination into the input feature of the best model corresponding to the target with low coefficient of determination, to form the input feature of the new low coefficient of determination target.
[0068] The retraining of the prediction model corresponding to the low coefficient of determination target in S5 includes: retraining based on the machine learning model using the input feature of the new low coefficient of determination target, the tree structure Parzen estimation and the grid search method, and saving the best model hyperparameter.
[0069] Specifically, Figure 4 The principle of the continuous targeted prediction method is shown. The continuous targeted prediction method refers to using the optimal machine learning model of each target, and on this basis, fine-tuning the input feature, using the high-accuracy prediction target as the input feature of the low-accuracy prediction target according to the physical relationship between target features, and then training the optimal machine learning model of the low-accuracy prediction target and saving the best model.
[0070] There is an empirical formula between the VMA, VFA and VV.
[0071]
[0072] According to Table 1, the coefficient of determination (R 2 ) of the VMA on the test set is as high as 0.9256, and the coefficient of determination (R 2) of 0.8786, the coefficient of determination (R2) of the VV on the test set was at most 0.8643. The best machine learning model of the VFA was XGBoost, and the best machine learning model of the VV was CatBoost. The predicted values of the VMA on the training set and the test set were spliced with the design parameters of the asphalt mixture to form new input features of the VFA, obtaining a VFA training set and a VFA test set. The VFA training set was retrained by the best machine learning model XGBoost of the VFA, and according to the experimental results, the R 2 of the VFA test set was 0.9056, and a continuous targeting VFA model was saved. The continuous targeting VFA model was used to make predictions on the VFA training set and the VFA test set, and the prediction results were spliced with the VFA training set and the VFA test set to obtain a VV training set and a VV test set. The VV training set was retrained by the best machine learning model CatBoost of the VV, and according to the experimental results, the R 2 of the VV test set was 0.9078, and a continuous targeting VV model was saved.
[0073] Figure 5 (a) is the prediction scatter plot of the best machine learning model CatBoost of the VV and the continuous targeting VV model on the test set, Figure 5 (c) is the prediction scatter plot of the best machine learning model CatBoost of the VFA and the continuous targeting VFA model on the test set. The continuous targeting VV model reduces the error interval of the best machine learning model CatBoost of the VV from [-2.5, 3.2] to [-1.9, 2.8], and the continuous targeting VV model reduces the root mean square error of the best machine learning model CatBoost of the VV by 6.14%. It can be seen that the prediction scatter of the continuous targeting VFA model on the test set is closer to the R 2 = 1 baseline, and the prediction result is more accurate.
[0074] Figure 6 (a) is the error graph of the best machine learning model XGBoost of the VV and the continuous targeting VV model on the test set, Figure 6(b) is the error chart of the VFA's best machine learning model XGBoost and the continuous targeting VFA model on the test set. The continuous targeting VFA model reduces the error interval of the VFA's best machine learning model XGBoost from [-9.2, 8.1] to [-8.7, 7.9], and the continuous targeting VFA model reduces the root mean square error of the VFA's best machine learning model XGBoost by 9.87%. It can be seen that the error peak and the average value of the continuous targeting VFA model on the test set are less than the VFA's best machine learning model XGBoost.
[0075] In another embodiment, the model input feature pruning in S6 refers to eliminating input features with low importance and SHAP values around 0.
[0076] Specifically, Figure 7 (left) shows the model feature importance, Figure 7 (right) shows the SHAP bee swarm chart. Keeping the new low-determination coefficient target prediction model hyperparameters unchanged, the features are added to the model in order according to the model feature importance ranking and the model is trained, and the evaluation index R 2 The trend chart is shown in Figure 8 When the added features are 9, the evaluation index R 2 is the highest, and the remaining input features have a negative effect on the model accuracy. The remaining input features are eliminated to achieve input feature pruning.
[0077] In some embodiments of the present specification, an asphalt mixture design method based on machine learning and continuous targeting prediction is proposed. First, mixture components, gradation, volume and mechanical property data are collected, and after missing value filling, outlier elimination, principal component analysis dimension reduction and standardization preprocessing, a high-quality data set is constructed. Second, a variety of machine learning models are used to predict the asphalt mixture gross bulk density (Gmb), void ratio (VV), mineral aggregate void ratio (VMA), asphalt saturation (VFA), Marshall stability (MS) and flow value (MF), and the multi-level linkage hyperparameter optimization method is used to optimize the model hyperparameters. Further, based on the physical correlation between volume parameters, a continuous targeting prediction method (LXBX) is proposed, which gradually corrects the low-precision target (VFA, VV) by taking the high-precision prediction target (VMA) as the input feature, realizes the collaborative optimization of the model, and finally realizes the specific target input feature pruning by combining the model feature importance and the SHAP bee swarm chart. Experiments show that the multi-level linkage hyperparameter optimization method makes the model accuracy produce a gradient improvement, and the R 2 value of the key indicators is more than 0.86, among which the R 20.9275 and 0.9256, respectively, LXBX makes the R 2 The additional promotion of 2.66% and 1.45% reduces the model calculation cost while improving the model precision by pruning the input feature of VV. The application significantly reduces the experimental cost and improves the design efficiency and precision.
[0078] The above-described embodiments only express the specific implementation of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the protection scope of the present application. It should be noted that for ordinary skilled persons in the art, without departing from the technical concept of the present application, a number of modifications and improvements can be made, which are within the protection scope of the present application.
Claims
1. A machine learning and continuous targeting prediction based asphalt mixture design method, characterized by, The method comprises the following steps: S1: divide the asphalt mixture dataset into a training set and a test set, use a multi-level linkage hyperparameter optimization method and k-fold cross-validation to find the best hyperparameters of multiple machine learning models for six targets of the asphalt mixture on the training set, the six targets including the gross bulk density Gmb, the void ratio VV, the mineral aggregate void ratio VMA, the asphalt saturation VFA, the Marshall stability MS and the flow value MF, use the best hyperparameters of the multiple machine learning models to train the training set, perform performance evaluation on the test set to obtain the determination coefficient R2, compare the determination coefficients R2 of the best hyperparameters of the multiple machine learning models on the test set, save the machine learning model with the highest R2 for each target and determine it as the best model for each target; S2: according to the determination coefficients R2 of the best models of VV, VFA and VMA on the test set, sequentially sort them from high to low, based on the empirical formula between VMA, VFA and VV, fine-tune the input features of the best model corresponding to the target with a low determination coefficient according to the continuous target prediction method, and retrain the prediction model corresponding to the target with a low determination coefficient; S3: obtain the model feature importance of the prediction model with a low determination coefficient, draw a SHAP bee swarm chart, add features to the prediction model with a low determination coefficient according to the model feature importance ranking and train the corresponding model, observe the change of the determination coefficient R2 on the test set, retain the input parameter number of the model corresponding to the determination coefficient R2, and realize the pruning of the input features of the prediction model with a low determination coefficient.
2. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 1, characterized in that, The production process of the asphalt mixture dataset in S1 is as follows: determining the design parameters and design indexes of the asphalt mixture in the asphalt mixture design dataset; collecting design parameters and design index data of the asphalt mixture from multiple sources to fill the asphalt mixture design dataset; and pre-processing the asphalt mixture design dataset based on data missing value filling, data anomaly value detection, data standardization and data dimension reduction to obtain the asphalt mixture dataset.
3. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 2, characterized in that, The design parameters of the asphalt mixture include coarse aggregate indexes, fine aggregate indexes, asphalt indexes, mineral powder indexes and mixture gradation, and the design indexes of the asphalt mixture include the gross bulk density, the void ratio, the mineral aggregate void ratio, the asphalt saturation, the Marshall stability and the flow value.
4. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 1, characterized in that, The multi-level linkage hyperparameter optimization method comprises: first using a tree structure Parzen estimation TPE to obtain the approximate range of the best hyperparameters of the machine learning model, and then using a grid search method to finely tune the parameters.
5. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 4, characterized in that, The tree structure Parzen estimation TPE models the hyperparameter distribution through kernel density estimation and uses Bayesian theorem for search, using a threshold y for all evaluated hyperparameters * All data is divided into two parameter subsets, the first is a better parameter subset corresponding to the hyperparameter x of the smaller target value, the probability distribution l(x), and the second is a worse parameter subset corresponding to the hyperparameter x of the larger target value, the probability distribution g(x), and the best machine learning model hyperparameter is the hyperparameter corresponding to the maximum x. l(x) = p(x | y < y * ) g(x) = p(x|y ≥ y * ) 6. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 4, characterized in that, The grid search method defines a new range of hyperparameters based on the best hyperparameters obtained by the tree structure Parzen estimation TPE, lists the candidate value combinations of the hyperparameters, exhaustively searches the candidate value combinations of the hyperparameters, and evaluates the effect of each group of hyperparameters through the test set.
7. The method of claim 1, wherein, The input feature of the best model with low coefficient of determination in S2 includes: splicing the target prediction value of the best model corresponding to the target with high coefficient of determination into the input feature of the best model corresponding to the target with low coefficient of determination, to form the input feature of the new target with low coefficient of determination.
8. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 7, characterized in that, The retraining of the prediction model corresponding to the target with low coefficient of determination in S2 includes: retraining based on the machine learning model by using the input feature of the new target with low coefficient of determination, the tree structure Parzen estimation and the grid search method, and saving the best model hyperparameter.
9. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 1, characterized in that, The model input feature pruning in S3 refers to eliminating the input features with low importance and SHAP values around 0.
10. The asphalt mixture design method based on machine learning and continuous targeting prediction according to claim 1, characterized in that, The K-fold cross-validation in S1 adopts 5-fold cross-validation, which refers to dividing the training data set into 5 data sets with approximately equal size, performing 5 times of training and verification, in which each time a data set in the 5 data sets is selected as a verification set and evaluated, and the average value of the results of the 5 verification sets is taken as the final evaluation performance of the model.