Method for predicting stability of basalt fiber asphalt mixture

Abstract

The invention provides a method for predicting the stability of a basalt fiber asphalt mixture, relates to the field of road engineering and intelligent prediction, and aims to optimize the mix proportion design process of the basalt fiber asphalt mixture and accurately predict the stability of the basalt fiber asphalt mixture. By combining feature engineering, regression algorithm comparison, model optimization and interpretability analysis, the mechanical property, the crack resistance and the high temperature resistance of the asphalt mixture are remarkably improved, and meanwhile, the dynamic optimization of the performance of the mixture under different use conditions is realized; according to the method, the problem of accurate prediction of the performance change of the asphalt mixture under different environmental conditions is solved through a more scientific, intelligent and adaptive novel mix proportion design thought, and the comprehensive performance of the asphalt mixture and the sustainability of road engineering are improved; based on the characteristics of the basalt fiber, the mix proportion of the asphalt mixture is optimized, the manual intervention and test times are reduced, and the design efficiency and accuracy are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of road engineering and intelligent prediction, and particularly relates to a basalt fiber asphalt mixture stability prediction method based on machine learning. BACKGROUND

[0002] With the continuous development of social economy and the increasing demand for transportation, the quality and durability of road infrastructure have become particularly important. Especially in important transportation hubs such as highways, urban roads and airport runways, asphalt mixture as the main road paving material, its performance directly affects the service life, maintenance cost and traffic safety of the road. Traditional asphalt mixture often performs poorly in extreme weather conditions, high temperature, high load and other environments, and is prone to rutting, cracking and other problems, bringing great challenges to road management. Therefore, how to optimize the mix design of asphalt mixture and improve its stability and durability has become an urgent need in current road engineering. Especially in the face of climate change, reducing road maintenance costs and extending road service life, it is particularly important to propose a more intelligent and accurate design method.

[0003] Currently, the common asphalt mixture optimization technologies on the market mainly include traditional mix design methods, hot mix asphalt technology, fiber reinforcement technology, etc. Traditional mix design methods usually rely on experience and engineer's judgment to select the appropriate ingredient ratio through test data. These methods mostly rely on manual operation, lack of flexibility and automation. In recent years, more and more research has begun to apply mineral fiber reinforced asphalt mixture to improve its mechanical properties and crack resistance. For example, glass fiber, polyester fiber, etc. have been widely used to improve the stability of asphalt. However, the application research of basalt fiber is still in its early stages, although some studies have shown that basalt fiber can effectively improve the mechanical properties of asphalt, but its mix design method and stability prediction model have not been widely standardized.

[0004] The existing asphalt mixture design technology still has many shortcomings and deficiencies. First, the traditional mix design method relies on a large amount of test data and manual adjustment, resulting in a long design cycle, high cost, and poor adaptability to environmental changes. Second, although research on fiber reinforced asphalt has made some progress, the influence of different types and proportions of fibers on asphalt performance has not been fully studied, especially the application of basalt fiber is still in the exploratory stage. Existing technology has not yet combined an intelligent prediction system to dynamically adjust the mix ratio of asphalt mixture, lacking a precise prediction method for the performance changes of asphalt mixture under different environmental conditions. In addition, existing prediction models often ignore the comprehensive influence of multiple factors, resulting in an inability to accurately control the stability of asphalt mixture in actual engineering applications, ultimately affecting the long-term use and maintenance effect of the road. Therefore, there is an urgent need for a

[0005] In view of the above, the present application proposes a basalt fiber asphalt mixture stability prediction method, which aims to solve the above problems through a more scientific, intelligent and adaptive new mix proportion design idea to improve the comprehensive performance of asphalt mixture and the sustainability of road engineering. SUMMARY

[0006] As a new type of environmentally friendly reinforcing material, the application of basalt fiber in asphalt mixture is still in the exploratory stage, and the proportion design and stability prediction still lack systematic and standardized solutions. In view of this problem, the present application proposes a basalt fiber asphalt mixture stability prediction method, which aims to optimize the mix proportion design process of basalt fiber asphalt mixture and accurately predict its stability. By constructing a large-scale database containing 560 sets of experimental data, combining feature engineering, regression algorithm comparison, model optimization and interpretability analysis, the present application significantly improves the mechanical properties, crack resistance and high temperature resistance of asphalt mixture, and simultaneously realizes the dynamic optimization of mixture performance under different use conditions.

[0007] A basalt fiber asphalt mixture stability prediction method, characterized in that it comprises the following steps:

[0008] S1: Collect the material composition and performance characteristic parameters of different basalt fiber asphalt mixtures as input feature parameter data sets, including:

[0009] Tensile strength, content, length and diameter of basalt fiber;

[0010] Penetration, softening point and content of asphalt;

[0011] Pass rate of aggregate at different sieve holes;

[0012] Construct a target feature parameter data set with the corresponding Marshall stability of basalt fiber asphalt mixture;

[0013] S2: Calculate the Spearman correlation coefficient of the input feature parameter data set and the target feature parameter data set, remove the feature parameters with an absolute value of the Spearman correlation coefficient with the target feature parameter data set less than 0.1~0.15, and obtain the screened feature parameter data set;

[0014] S3: Based on the deep learning regression algorithm, a basalt fiber asphalt mixture Marshall stability prediction model is established. The screened feature parameter data set and the target feature parameter data set in step S2 are divided into a training set and a test set, the training set is used to train the Marshall stability prediction model, the model parameters are optimized, and the ten-fold cross-validation method is used to test the model, and the trained basalt fiber asphalt mixture Marshall stability prediction model is obtained.

[0015] S4: Perform an explainability analysis on the trained Marshall stability prediction model using the SHAP algorithm, calculate the SHAP values corresponding to the feature parameters of the screened samples, and represent the influence degree of different features on the prediction results through the mean SHAP values of different feature parameters;

[0016] S5: Select 2-4 feature parameters with the largest mean SHAP values as typical feature parameters, and analyze the influence of the typical feature parameters on the Marshall stability results of basalt fiber asphalt mixture;

[0017] S6: Set the target value of the Marshall stability of basalt fiber asphalt mixture, import the initial material composition and performance characteristic parameters of basalt fiber asphalt mixture into the trained stability prediction model, and calculate the SHAP values, and adjust the values of the typical feature parameters to make the prediction result reach the set stability target value.

[0018] The above technical scheme has the following beneficial effects:

[0019] 1. Optimization design of asphalt mixture mix proportion: The traditional design method of asphalt mixture mix proportion relies on experience and a large amount of test data, and the design period is long and the cost is high; the present application can intelligently optimize the mix proportion of asphalt mixture based on the characteristics of basalt fiber by introducing intelligent prediction technology, reduce manual intervention and test times, thereby greatly improving the design efficiency and accuracy.

[0020] 2. Intelligent prediction of asphalt mixture stability: The traditional stability prediction method lacks comprehensive consideration of multiple variables such as climate conditions, traffic load, temperature changes, etc., and the prediction result is not accurate enough; the present application can consider the influence of different factors on the stability of asphalt mixture through intelligent prediction algorithm, provide more accurate prediction results, and thus improve the service performance and durability of the road.

[0021] 3. Application and optimization of basalt fiber reinforced asphalt mixture: The application of basalt fiber in asphalt mixture is still in the exploratory stage, and there is a lack of standardized mix design method for basalt fiber optimization in existing technology; the present application fills this gap, optimizes the use of basalt fiber in asphalt mixture through intelligent design and stability prediction, and improves the mechanical properties, crack resistance and high temperature resistance of the material.

[0022] 4. Environmental adaptability improvement: Traditional asphalt mixture performs poorly in extreme climates such as high temperature and low temperature, and is prone to problems such as rutting and cracking; through the intelligent prediction technology of the present application, the mix proportion of asphalt mixture can be dynamically adjusted according to different climate conditions, improving its adaptability and long-term stability.

[0023] 5. Reduced road maintenance costs: Traditional asphalt mixtures are prone to aging, cracking, rutting, and other issues over long-term use, leading to frequent repairs and increased road maintenance costs. The optimization of the mix ratio and stability prediction of the present application can effectively improve the durability of asphalt mixtures, reduce the frequency of road maintenance and repair, and thus save long-term use costs.

[0024] 6. Improved construction precision: Traditional asphalt mixture design methods lack precision and standardization, leading to uneven mix ratios in actual construction, which affects road surface quality. The intelligent mix ratio design of the present application improves the precision and consistency of the mix ratio during construction, ensuring the stability and consistency of road surface quality.

[0025] 7. Popularization and application of intelligent road material design: The present application provides a new approach to intelligent road material design, which not only optimizes the mix ratio of asphalt mixtures but also provides an intelligent design demonstration for other types of road materials, promoting the application of intelligent technology in road construction. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a schematic diagram of the complete process of the method of the present application;

[0027] Figure 2 is a schematic diagram of the correlation matrix between the feature values and target values in the specific embodiment;

[0028] Figure 3 is a schematic diagram of the comparison of model performance index quantification evaluation results in the specific embodiment;

[0029] Figure 4 is a schematic diagram of the performance of three indicators calculated using the test set in the specific embodiment;

[0030] Figure 5 is a schematic diagram of the SHAP value interpretation in the specific embodiment;

[0031] Figure 6 is a schematic diagram of the feature dependence of X9 and X5 in the specific embodiment;

[0032] Figure 7 is a schematic diagram of individual interpretation of certain sample data in the specific embodiment;

[0033] Figure 8 is a schematic diagram of the comparison results of SFO and AO optimization algorithms in the specific embodiment;

[0034] Figure 9 is a schematic diagram of the comparison results of SHAP value analysis of SFO and AO optimization algorithms in the specific embodiment;

[0035] Figure 10 This is a schematic diagram illustrating the impact analysis results of Combination 1 on the SHAP value in a specific implementation method.

[0036] Figure 11 This is a schematic diagram illustrating the impact analysis results of Combination 2 on the SHAP value in a specific implementation method.

[0037] Figure 12 This is a schematic diagram illustrating the impact analysis results of combination 3 on the SHAP value in a specific implementation method.

[0038] Figure 13 This is a schematic diagram illustrating the impact analysis results of combination 4 on the SHAP value in a specific implementation method.

[0039] Figure 14 This is a schematic diagram illustrating the impact analysis results of combination 5 on the SHAP value in a specific implementation method. Detailed Implementation

[0040] The foregoing and other technical contents, features and effects of the present invention will be clearly presented in the following detailed description of the embodiments with reference to the accompanying drawings. All contents mentioned in the following embodiments are based on the accompanying drawings.

[0041] Example 1, such as Figure 1 The diagram shows the overall process of this method. This implementation method can explain the intelligent prediction and optimization method for the stability of basalt fiber asphalt mixture mix design empowered by machine learning models. Specifically, it is implemented according to the following steps:

[0042] S1: Collect material composition and performance characteristics of different basalt fiber asphalt mixtures as input feature parameter datasets, mainly including: tensile strength, content, length, and diameter of basalt fibers; penetration, softening point, and content of asphalt; and the passing rate of aggregates through different sieves (2.36 mm, 4.75 mm, 9.5 mm, etc.). Construct target feature parameter datasets based on the Marshall stability (MS) of the corresponding basalt fiber asphalt mixtures.

[0043] S2: Calculate the Spearman correlation coefficient between the input feature parameter dataset and the target feature parameter dataset, remove feature parameters whose absolute value of the Spearman correlation coefficient with the target feature parameter dataset is lower than 0.1~0.15, and obtain the filtered feature parameter dataset;

[0044] S3: A basalt fiber asphalt mixture stability prediction model is established based on the XGBoost algorithm. The screened feature parameter dataset and the target feature parameter dataset in step two are divided into a training set and a test set. The training set is used to train the stability prediction model. The model parameters are optimized using the marlin and the ten-fold cross-validation method. The trained basalt fiber asphalt mixture stability prediction model is obtained.

[0045] S4: The trained stability prediction model is analyzed for interpretability using the SHAP algorithm. The SHAP values corresponding to the screened feature parameters of all samples are calculated. The influence of different features on the prediction results is represented by the mean SHAP values of different feature parameters.

[0046] S5: Select 2-4 feature parameters with the largest mean SHAP values as typical feature parameters. Analyze the influence of typical feature parameters on the stability of basalt fiber asphalt mixture.

[0047] S6: Set the target value of the stability of basalt fiber asphalt mixture. Import the initial material composition and performance characteristic parameters of the asphalt mixture into the trained stability prediction model and calculate the SHAP value. Adjust the typical feature parameters to make the prediction result reach the set stability target value.

[0048] Note: the Bayesian optimization in step S3 of the embodiment refers to updating the posterior distribution of the objective function, i.e. Gaussian Process (GP), by continuously adding sample points until the posterior distribution basically fits the true distribution, given the objective function to be optimized (a general function, only the input and output need to be specified, without knowing the internal structure and mathematical properties). Bayesian optimization (BO) is an effective method for global optimization of black box, expensive or noisy objective functions. BO establishes a proxy model of the objective function (commonly used as Gaussian Process or method based on Tree-structured Parzen Estimator), and defines an acquisition function (Acquisition Function, such as Expected Improvement EI, Upper Confidence Bound UCB, Probability Improvement PI, etc.) on the proxy model to balance exploration and exploitation, so as to find the approximate optimal solution with fewer real evaluation times. A method for optimizing asphalt mixture proportion based on Bayesian optimization, characterized in that: a proxy model is used to model the objective function, and a limited number of real evaluations are selected based on the acquisition function to seek the optimal proportion, the proxy model is a noisy Gaussian process, and the acquisition function is expected improvement (EI) or upper confidence bound (UCB); the method further comprises filtering the engineering hard constraints as a feasible region to ensure that the proposed proportion meets the engineering specifications. After introducing Bayesian optimization into the joint optimization framework described in the application, high-performance proportions can be found faster with fewer real evaluation times, reducing test costs, and ensuring engineering feasibility by constraining BO; at the same time, the cooperation of BO and SFO (Swordfish Optimization Algorithm) / AO (Eagle Optimization Algorithm) further improves the search efficiency and robustness of the mixed space of discrete combinations and continuous parameters.

[0049] In Example 2, as an example of Example 1, the embodiment is different from Example 1 in that, in addition to 2.36mm, 4.75mm, 9.5mm, the sieve aperture size of the aggregate in step S1 can also be 0.075mm, 0.15mm, 0.3mm, 0.6mm, 1.18mm, 13.2mm, 16mm and 19mm in turn.

[0050] Specifically, the fine grade group (0.075 mm, 0.15 mm, 0.3 mm, 0.6 mm, 1.18 mm) has the following technical characteristics: 1) improved filler and binder layer behavior: finer particle size helps to precisely control the mineral filler content and distribution, thereby optimizing the asphalt film thickness, improving the cementation performance and improving the anti-stripping performance; 2) fine-tuning of the void structure: the controllability of the fine grade enables the void ratio (Vv) and the effective void structure to be more accurately designed, which is beneficial to balance the density and drainage, and reduce the risk of capillary water damage; 3) delay of crack initiation: through careful matching of fine particles, the continuity of the bond between aggregates and stress dispersion can be improved, and the low-temperature crack resistance and fatigue life can be improved; 4) reduce fine material migration and bleeding: reasonable control of the 0.075-0.6 mm interval can reduce the migration of fine materials and the tendency of oil bleeding during mixing and transportation, and improve the uniformity and workability of the mixture.

[0051] The coarse grade group (13.2 mm, 16 mm, 19 mm) has the following technical characteristics: 1) enhance the skeleton bearing and anti-rutting ability: the introduction of larger particle size can build a stable stone skeleton, improve the macroscopic interlocking and shear strength, and thus enhance the anti-rutting ability and load transfer ability at high temperature; 2) adjust the surface macroscopic roughness and anti-skid performance: larger aggregate improves the macroscopic texture of the pavement, which is beneficial to the improvement of friction performance and anti-skid ability; 3) improve the porosity control space: through the adjustment of coarse aggregate ratio, sufficient space can be left for fine material / asphalt film under the premise of ensuring the stability of the skeleton, so as to meet the requirements of density and water permeability / anti-frost at the same time.

[0052] In Example 3, as an example of Example 1, the number of sample sets of material composition and performance characteristic parameter data in step S1 is 500-1000, which is different from Example 1. The preferred number of sample sets in this example is 560 data.

[0053] In Example 4, as an example of Example 1, the characteristic parameters with a Spearman correlation coefficient absolute value lower than 0.1 are removed in step S2, which is different from Example 1.

[0054] In Example 5, as an example of Example 1, the calculation formula of the Spearman correlation coefficient in step S2 is as follows, which is different from Example 1. (1) In the formula, p is the Spearman correlation coefficient, n is the number of samples, is the rank difference between samples. In this example, the sample values are arranged in ascending order to determine the rank.

[0055] In Example 6, as an example of Example 1, the ratio of training set and test set in step S3 is 7:3, which is different from Example 1.

[0056] Embodiment 7, as an example of Embodiment 1, the difference between this embodiment and Embodiment 1 is that the SHAP algorithm in step S4 is represented as: (2) where: g(·) is the interpretation model, f(·) is the machine learning method, For simplicity of input, and input There is a mapping relationship; is the prediction mean of the machine learning method on the data set; is the contribution value (i.e., SHAP value) of the ith feature, and M is the total number of features.

[0057] Embodiment 8, as an example of Embodiment 1, the difference between this embodiment and Embodiment 1 is that in step S5, the two feature parameters with the largest SHAP value mean (coarse aggregate 4.75mm passing rate and asphalt penetration) are selected as typical feature parameters.

[0058] Embodiment 8, as an example of Embodiment 1, the difference between this embodiment and Embodiment 1 is that in step S5, the SHAP value distribution diagram, dependence diagram and individual contribution analysis diagram of the typical feature parameters are drawn.

[0059] Embodiment 8, as an example of Embodiment 1, the process of Embodiment 1 is described in detail.

[0060] I. Data collection.

[0061] This study constructed a database containing 560 experimental samples, covering multi-dimensional attributes of basalt fiber, asphalt and aggregate. Data collection was completed through laboratory tests and field tests to ensure the authenticity and representativeness of the data. Each sample recorded the following key parameters:

[0062] 1) Basalt fiber: tensile strength (MPa), content (%), length (mm), diameter (μm);

[0063] 2) Asphalt: penetration (0.1mm), softening point (°C), content (%);

[0064] 3) Aggregate: fine aggregate passing 2.36mm sieve (%), coarse aggregate passing 4.75mm sieve (%), coarse aggregate passing 9.5mm sieve (%).

[0065] II. Feature selection.

[0066] Ten key input features were selected from the raw data based on correlation analysis with the target variable (mixture stability) and engineering experience. Feature selection used statistical methods and machine learning techniques (such as recursive feature elimination, RFE) to ensure that the selected features made a significant contribution to model prediction. The asphalt content, asphalt penetration, asphalt softening point, basalt fiber tensile strength, fiber content, fiber diameter, length, and different sieve sizes (2.36 mm, 4.75 mm, 9.5 mm) were collected as material composition and performance feature parameter datasets for different asphalt mixtures. The corresponding asphalt mixture stability was used to construct the target feature parameter dataset. The maximum, minimum, and average values of the feature parameters are shown in Table 1:

[0067] Table 1 Maximum, minimum, and average values of feature parameters

[0068] Variable Median Standard Deviation Minimum 25% (first quartile) 50% (second quartile 75% (third quartile Maximum Fiber tensile light 2863.7 581.0 1080.0 2500.0 3060.0 3200.0 4495.0 Fiber content 0.5 0.7 0.05 0.3 0.4 0.4 7.0 Fiber length 6.5 2.0 3.0 6.0 6.0 6.0 24.0 Fiber diameter 12.9 3.2 5.0 12.0 13.0 15.0 25.0 Asphalt penetration 68.5 13.4 50.4 56.4 65.2 75.0 98.0 Asphalt softening point 60.6 13.4 43.0 48.0 61.0 71.0 85.8 Asphalt content 5.0 0.9 3.1 4.5 5.0 5.5 10.4 2.36 mm passing rate 30.7 8.8 15.8 24.6 30.0 37.0 65.0 4.75 mm passing rate 42.4 12.6 16.0 35.0 41.0 53.0 95.0 9.5 mm passing rate 68.7 10.2 13.6 63.0 71.0 76.5 100.0 Marshall stability 11.8 2.4 6.0 10.0 11.5 13.5 17.7

[0069] The Spearman correlation coefficient between the input feature parameter dataset and the target feature parameter dataset was calculated, and the feature parameters with an absolute value of the Spearman correlation coefficient with the target feature parameter dataset less than 0.1-0.15 were removed to obtain the screened feature parameter dataset as shown in Table 2; the correlation matrix between the feature values and the target values is shown in Table 2. Figure 2

[0070] Table 2 Feature parameters related to stability after screening

[0071] Characteristic name Fiber length Fiber content Fiber length Fiber diameter Asphalt penetration Asphalt softening point Asphalt content 2.36 mm passing rate 4.75 mm passing rate 9.5 mm passing rate Correlation coefficient 0.11 -0.10 -0.05 -0.03 -0.12 0.26 -0.04 0.33 0.42 0.30

[0072] III. Data set preprocessing.

[0073] The dataset was first randomly divided into 2 parts, 70% for training and 30% for testing. The input features of the dataset were standardized to conform to the quantitative range of each parameter. Standardization can make the data distribution uniform and the feature value range consistent, which helps faster convergence during algorithm training and prevents large numerical magnitude features from having a dominant influence on model parameters.

[0074] IV. Model construction. The present application compares five typical regression algorithms to build a prediction model, including:

[0075] • Linear regression (Linear Regression): as a benchmark model;

[0076] • Decision tree regression (Decision Tree Regression): to capture nonlinear relationships;

[0077] ​• Random Forest Regression: Enhances robustness through ensemble learning; • Support Vector Regression (SVR): Suitable for high-dimensional data; • XGBoost: Efficient algorithm for handling complex feature interactions. Each algorithm uses 80% of the training set for training and the remaining 20% for testing.

[0078] Five, Model Evaluation. Model performance is quantitatively evaluated by the following indicators:

[0079] 1) Mean Squared Error (MSE): Measures the average squared difference between predicted and actual values;

[0080] 2) Root Mean Squared Error (RMSE): Represents error size in target variable units;

[0081] 3) Coefficient of Determination (R²): Reflects the model's ability to explain data variation. For example, Figure 3 Evaluation results show that the XGBoost model performs best, with MSE <0.04, RMSE <0.2, and R² >0.98, demonstrating excellent prediction accuracy and generalization ability.

[0082] Six, Prediction and Optimization. Based on the optimal model (XGBoost), the stability of new input mixture proportion parameters can be predicted. To further improve performance, an optimization algorithm is used to adjust model parameters and feature weights.

[0083] Note: XGBoost is a second-order gradient optimization-based boosting tree model that achieves high-performance prediction and good generalization ability through fine target function design and efficient splitting algorithm. The objective function of XGBoost is as follows:

[0084] (3)

[0085] where, represents the prediction result of the ith sample after the tth iteration; yi represents the true value of the ith sample; represents the loss function of the model; Ω (fi ) is the regularization term of the tth iteration, representing the complexity of the base learner added in the tth iteration. Continue to expand the regularization term , we have:

[0086] (4)

[0087] where, The regularization term for the previous t-1 rounds, which is a constant value at the tth iteration, is denoted as constant.

[0088] When performing model prediction expression, the prediction value of the tth round model can be written as the sum of the prediction value of the (t-1) th round and the prediction of the base learner constructed in this round:

[0089] (5)

[0090] Substitute the total objective function calculation, formula (4), (5) into (3), and the following formula (6) can be obtained:

[0091] (6)

[0092] In order to facilitate optimization, Taylor expansion of the loss function is performed, and XGBoost uses Taylor expansion to perform second-order approximate expansion of the objective function at the prediction value of the (t-1) th round :

[0093] (7)

[0094] Where, is the first-order derivative (gradient); is the second-order derivative; is the output of the tth base learner. Ignore the constant term irrelevant, and the objective function becomes:

[0095] (8)

[0096] The conversion of the objective function in the form of leaf nodes is performed, and the base learner constructed in the tth round is a tree that divides the input samples into T leaf nodes, and the output value of each leaf node j is . Use to represent the leaf node number to which the sample belongs ( ), then . Substitute it into formula (8) to obtain:

[0097] (9)

[0098] Where represents the sample set falling into the jth leaf node. The regularization term is expressed as follows:

[0099] (10)

[0100] Substitute it into formula (9) and combine to obtain the final optimization objective:

[0101] (11)

[0102] where , . The optimal leaf weight setting, for derivation and let the derivative be zero, the optimal output weight of each leaf node is obtained:

[0103] (12)

[0104] Calculate the optimal target value of this round, substitute into formula (11) to obtain the optimal target value of the tth round:

[0105] (13)

[0106] Derive the node splitting gain formula. In the process of growing the tree structure, the basis for judging whether to split a certain node is the gain of the target function brought by splitting. Let a certain node be split into a left child node L and a right child node R, then the gain is calculated as follows:

[0107] (14)

[0108] Example 9, as an example of example 8, specific experimental verification is carried out on the process of example 8. I. Data description.

[0109] The data set used in the present application contains 560 experimental samples, covering various characteristics of basalt fiber asphalt mixture. The data collection process follows strict experimental specifications to ensure the repeatability and consistency of the parameters. The feature list is as follows:

[0110] Basalt fiber properties:

[0111] · Tensile strength (MPa): reflects the mechanical strength of the fiber, ranging from 500 to 3000 MPa;

[0112] · Content (%): the proportion of fiber in the mass of the mixture, ranging from 0.1% to 1.0%;

[0113] · Length (mm): physical length of the fiber, ranging from 6 to 24 mm;

[0114] · Diameter (μm): fiber cross-sectional dimension, ranging from 10 to 20 μm.

[0115] Asphalt properties:

[0116] · Penetration (0.1mm): measures the hardness of asphalt, ranging from 40 to 100;

[0117] · Softening point (°C): indicates the temperature resistance of asphalt, ranging from 40 to 60°C;

[0118] • Content (%): Proportion of bitumen in the mixture mass, ranging from 4% to 6%.

[0119] Aggregate properties:

[0120] • Fine aggregate passing 2.36 mm sieve (%): Proportion of fine particles, ranging from 20% to 40%;

[0121] • Coarse aggregate passing 4.75 mm sieve (%): Proportion of medium particles, ranging from 30% to 50%;

[0122] • Coarse aggregate passing 9.5 mm sieve (%): Proportion of large particles, ranging from 40% to 60%.

[0123] The selection of these features is based on research results in the field of road engineering and preliminary experimental analysis of the invention, ensuring their strong correlation with the stability of the mixture.

[0124] II. Algorithm selection.

[0125] To determine the optimal prediction model, the performance of five regression algorithms was systematically evaluated:

[0126] • Linear regression: Assumes a linear relationship between features and stability, serving as a performance benchmark.

[0127] • Decision tree regression: Captures non-linear patterns through recursive splitting, but is prone to overfitting.

[0128] • Random forest regression: Integrates multiple decision trees, reducing the risk of overfitting and improving stability.

[0129] • Support vector regression (SVR): Utilizes kernel functions to handle high-dimensional data and is robust to outliers.

[0130] • XGBoost: Based on the gradient boosting framework, excels at handling complex interactions between features.

[0131] Through cross-validation and performance indicator comparison, XGBoost was chosen as the final model due to its low error (MSE < 0.04, RMSE < 0.2) and high explanatory power (R² > 0.98). Its advantages lie in efficient parallel computing capabilities and good handling of missing values and non-linear relationships.

[0132] III. Model training and evaluation.

[0133] The training process of the XGBoost model includes the following steps:

[0134] • Data split: The dataset is divided into training set (448 groups) and test set (112 groups) with a ratio of 80:20.

[0135] • Performance evaluation: After training, the test set is used to calculate MSE, RMSE and R². The results show that the prediction error of the model on the test set is extremely low, and can effectively generalize to unseen data.

[0136] As shown in Figure 4 , the specific performance indicators are as follows:

[0137] MSE: 0.0012, indicating that the average deviation between predicted and actual values is extremely small;

[0138] RMSE: 0.034, consistent with the stability unit, the error is controlled within an acceptable range;

[0139] R²: 0.982, indicating that the model explains 98.2% of the stability variation.

[0140] Four, explainability and visual analysis.

[0141] To enhance the engineering applicability of the model, the SHAP (SHapley Additive exPlanations) method is used to analyze the explainability of the XGBoost model:

[0142] Global explanation: Calculate the SHAP value mean of each feature to reveal its overall contribution to stability. The results show that the coarse aggregate 4.75mm passing rate and asphalt penetration are the two most influential features.

[0143] Individual explanation: Generate SHAP value decomposition for a single sample to show the specific influence of each feature.

[0144] Feature dependency: Draw partial dependence plot (PDP) to find that coarse aggregate 4.75mm passing rate and asphalt penetration have significant interaction effects, and have nonlinear influence on stability.

[0145] In addition, SHAP value distribution chart and prediction ability chart are generated to intuitively show feature importance and model output distribution, providing data-driven insights for engineering design.

[0146] The SHAP algorithm is used to analyze the explainability of the trained stability prediction model, calculate the SHAP value corresponding to the feature parameters after screening of all samples, and represent the influence degree of different features on the prediction result through the SHAP value mean of different feature parameters as Figure 5The SHAP value of the selected 2-4 feature parameters is the largest, and the typical feature parameters are selected as the typical feature parameters. Figure 6 The feature dependence graph is obtained by analyzing the influence of the typical feature parameters on the basalt fiber asphalt mixture stability result.

[0147] In order to further show the influence of data on the basalt fiber asphalt mixture stability result, the individual contribution dependence graph is drawn, as shown in Figure 7

[0148] Five, optimization technology.

[0149] In order to further improve the model performance, the application introduces two kinds of optimization algorithms:

[0150] SFO (Swordfish Optimization Algorithm): by simulating the interaction model algorithm of swordfish-sardine, the feature selection and model parameters are iteratively adjusted, and the prediction error is reduced by about 5%.

[0151] AO (Hawk Optimization Algorithm): dynamically adjust the learning rate and regularization parameters, so that the model can better adapt to the data distribution characteristics.

[0152] The optimized XGBoost model has significantly improved in MSE and RMSE, which are 0.035 and 0.187 respectively, further enhancing the reliability of the prediction.

[0153] It is not difficult to find that the stability of the prediction model is greatly improved after optimization by the above two algorithms, and the R2 of the training set and the test set is significantly improved compared with the default XGBoost model before, and the results are as follows Figure 8 The SHAP value graph after optimization shows that different representation parameters also change significantly, so that the representation parameters also change in order; the results are as shown in Figure 9

[0154] Based on the above optimized model and influence law of the swordfish algorithm and the whale algorithm, the gradation design is carried out, first the Marshall stability parameter design target is determined, the 4.75mm passing rate and the asphalt penetration and content are selected as combination 1, and the prediction model is introduced to predict the stability index, if it meets the design requirements, the design is ended. If it does not meet the requirements, the SHAP value is drawn according to the prediction result as shown in Figure 10 According to the influence law of the important feature parameters on the stability index in step six, the combination 1 does not meet the requirements, the positive gain provided by the asphalt content is too strong, so the asphalt content needs to be increased, and then combination 2, combination 3 and combination 4 as shown in Figure 11 , Figure 12 , Figure 13 are selected, among which combination 3 and 4 produce a large negative gain of asphalt content, so that they meet the requirements, and combination 4 as shown in Figure 14 ​​The combination 5 adjusts the 2.36 mm pass rate to produce certain negative gain, and reduces the void ratio of the mixture, but still does not meet the requirements.

[0155] The above is only for the purpose of illustrating the present application, and it should be understood that the present application is not limited to the above examples, and various modifications in accordance with the spirit of the present application are within the scope of the present application.

Claims

1. A method of predicting stability of basalt fiber asphalt mixture, characterized by, The method comprises the following steps: S1: collecting material composition and performance characteristic parameters of different basalt fiber asphalt mixtures as input characteristic parameter data set, including: tensile strength, content, length and diameter of basalt fiber; penetration, softening point and content of asphalt; pass rate of aggregate at different sieve holes; target characteristic parameter data set is constructed by using Marshall stability of corresponding basalt fiber asphalt mixture; S2: calculating Spearman correlation coefficient of input characteristic parameter data set and target characteristic parameter data set, removing characteristic parameters with absolute value of Spearman correlation coefficient lower than 0.1-0.15, and obtaining screened characteristic parameter data set; S3: basing on deep learning regression algorithm, a basalt fiber asphalt mixture Marshall stability prediction model is trained and established, the screened characteristic parameter data set and the target characteristic parameter data set in step S2 are divided into training set and test set, the training set is used to train the Marshall stability prediction model, the model parameters are optimized, and the model is tested by using ten-fold cross validation method, and a trained basalt fiber asphalt mixture Marshall stability prediction model is obtained; S4: using SHAP algorithm to perform explainability analysis on the trained Marshall stability prediction model, calculating SHAP values corresponding to the screened characteristic parameters of all samples, and using mean values of SHAP values of different characteristic parameters to represent the influence degree of different characteristics on the prediction result; S5: selecting 2-4 characteristic parameters with the largest mean SHAP values as typical characteristic parameters, and analyzing the influence of the typical characteristic parameters on the basalt fiber asphalt mixture Marshall stability result; S6: setting a target value of basalt fiber asphalt mixture Marshall stability, inputting initial material composition and performance characteristic parameters of the basalt fiber asphalt mixture into the trained stability prediction model, and calculating SHAP values, and adjusting the values of the typical characteristic parameters to make the prediction result reach the set stability target value.

2. The method of claim 1, wherein the basalt fiber asphalt mixture stability is predicted. The process of constructing the target characteristic parameter data set in step S1 comprises: Ten key input characteristics are selected from the original data, and statistical methods and recursive feature elimination techniques are used to ensure that the selected characteristics have significant contribution to the model prediction.

3. The method of claim 1, wherein the basalt fiber asphalt mixture stability is predicted. The deep learning regression algorithm in step S3 comprises: linear regression, decision tree regression, random forest regression, support vector regression and extreme gradient boosting (XGBoost).

4. The method for predicting the stability of basalt fiber asphalt mixture according to claim 3, characterized in that, The grid search in the XGBoost can use Bayesian optimization algorithm.

5. The method of claim 1, wherein the basalt fiber asphalt mixture stability is predicted. The optimization of model parameters in step S3 comprises: Swordfish optimization method (SFO) is used as a global discrete searcher for feature subset selection, discrete proportion candidate generation and coarse-grained hyperparameter space search to generate a Pareto solution set; eagle optimization method (AO) is used as a local continuous fine tuner for fine tuning of continuous hyperparameters and continuous proportion variables, and each Pareto solution set is locally refined; the two algorithms are alternately run to reduce the probability of falling into local optimum, and a more stable and engineered mix proportion solution set is obtained.

Images