Small sample tunnel low-temperature asphalt optimization design method based on machine learning

Abstract

The invention provides a small sample tunnel low-temperature asphalt optimization design method based on machine learning. According to the method, an asphalt sample is prepared through orthogonal test design, performance indexes are measured, and an initial small sample data set is established; the SMOTER technology is used for data enhancement, and the sample scale is effectively expanded; constructing an AR-CatBoost machine learning model, and introducing an adaptive regularization and gradient weighting mechanism to enhance the learning ability of the nonlinear relationship between the asphalt component mixing amount and the performance; carrying out hyper-parameter automatic optimization by adopting an improved reptile search algorithm fusing Cauchy variation and dynamic boundary adjustment; based on the optimization model, large-scale virtual ratio performance prediction is generated in a component mixing amount range, and the optimal ratio is accurately screened by integrating a scoring function. The method effectively solves the problem of insufficient model generalization ability under the small sample condition, realizes high-precision and low-cost automatic design of the tunnel low-temperature asphalt mix proportion, and is suitable for rapid research and development and performance optimization of tunnel asphalt materials in cold regions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of low-temperature asphalt materials, and in particular to a small-sample tunnel low-temperature asphalt optimization design method based on machine learning. BACKGROUND

[0002] As the main form of high-grade highways, the service performance and durability of asphalt pavement depend largely on the high and low temperature performance of asphalt binder. Especially in tunnels, mountainous areas and northern cold regions, the ambient temperature is low or the temperature difference changes dramatically all year round. Ordinary asphalt materials are prone to embrittlement, which leads to a significant decrease in pavement crack resistance and the occurrence of temperature shrinkage cracks, seriously affecting driving safety and pavement service life. Existing modified asphalt design mostly uses trial-and-error and response surface methods. However, the trial-and-error method usually requires a large number of indoor tests, which has obvious disadvantages of low efficiency and high cost. Although the response surface method can reduce the amount of testing, its accuracy is not high, and it is difficult to find the optimal formula.

[0003] In recent years, the rise of machine learning (ML) technology has provided a revolutionary tool for material performance prediction and reverse design. By training the model to learn the complex rules in historical experimental data, the performance of new formulations can be quickly predicted, thereby reducing the experimental blind area. However, its application in the field of asphalt materials faces two major bottlenecks. First, it is difficult and time-consuming to obtain a large amount of high-quality experimental data, which leads to the "small sample" dilemma in model training, making it prone to overfitting and lacking generalization ability. Second, pure "black box" data fitting may produce prediction results that violate the material constitutive relationship and may output physically impossible formulations, which has low engineering credibility.

[0004] Although patent CN119811560A builds multiple machine learning models for asphalt material performance prediction based on small sample data sets, it does not address the sparsity problem of small sample data, which can lead to model overfitting and significantly reduced generalization ability. Although patent CN117252107A constructs a "component → performance" forward model and a "performance → component" reverse model, the sample data is 800-2000, which is beyond the small sample category. Even with this sample size, there are deficiencies in learning critical samples (such as formulations with performance close to the target value), and the reverse design may output physically impossible component schemes. In summary, there is an urgent need for a small-sample tunnel low-temperature asphalt machine learning design framework that breaks through the existing technical bottlenecks. SUMMARY

[0005] To solve the above problems in the prior art, the present application provides a small-sample tunnel low-temperature asphalt optimization design method based on machine learning.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A small sample tunnel low temperature asphalt optimization design method based on machine learning, the method comprising the following steps:

[0008] S1: determining the raw material components of the tunnel low temperature asphalt and their dosage range, generating a test scheme of the tunnel low temperature asphalt mixture ratio by an orthogonal test design method, and preparing tunnel low temperature asphalt samples according to the test scheme;

[0009] S2: measuring the performance indicators of low temperature asphalt under different proportions, and establishing an initial small sample data set;

[0010] S3: using synthetic minority over-sampling technique (SMOTER) for data enhancement, identifying the sparse area in the target performance space, using linear interpolation for targeted data enhancement in the sparse area, generating new samples conforming to the original data distribution rule, avoiding the imbalance of a few samples, and effectively expanding the diversity and scale of the data;

[0011] S4: establishing an AR-CatBoost machine learning model, using the data set after data enhancement in S3 for model training, using the raw material components as input features and the target performance indicators as output, learning the complex nonlinear mapping relationship between the dosage of each component of low temperature asphalt and multiple performances;

[0012] S5: using the improved CD-RSA to automatically optimize the hyperparameters of the AR-CatBoost machine learning model, and finding the optimal combination of hyperparameters;

[0013] S6: using the AR-CatBoost machine learning model with optimized hyperparameters by CD-RSA to retrain, generating large-scale virtual proportion performance prediction within the dosage range of each component determined in S1, quantitatively evaluating the performance of different proportions by a comprehensive scoring formula, and accurately selecting the optimal proportion with excellent performance and meeting the engineering practice within the dosage range of each component.

[0014] Further, the tunnel low temperature asphalt raw material components in step S1 include 70A grade base asphalt, SBS modifier, rubber powder, resin, sulfur crosslinking agent and zinc oxide catalyst, and the specific dosage range is 100 parts of base asphalt, 3-6 parts of SBS modifier, 10-20 parts of rubber powder, 2-5 parts of resin, 0.1-0.5 parts of sulfur and 0.1-0.3 parts of zinc oxide catalyst.

[0015] Further, the test scheme in step S1 is designed by an orthogonal test design method to design 30 tests.

[0016] Further, the performance indicators in step S2 include the penetration value measured by the penetration test, the 5 ductility of asphalt measured by the ductility test, the softening point measured by the softening point test, the 5 ductility value, softening point value determined by asphalt softening point test, and -18 creep stiffness S and -18℃ creep rate m values.

[0017] Further, the initial small sample data set in step S2 includes: an input matrix of specific values of tunnel low temperature asphalt external components (modifiers, rubber powder, resins, crosslinking agents, catalysts) and a target matrix of specific values of performance indicators (25 penetration, 5 ductility, softening point, -18 creep stiffness S, -18 creep rate m) characterizing low temperature asphalt performance .

[0018] Further, in step S3, the initial small sample data set is divided into a training set and a test set before data augmentation, and the ratio of the training set to the test set is 85:15.

[0019] Further, in step S3, the data is augmented using the Synthetic Minority Over-Sampling Technique (SMOTER), including the following steps:

[0020] S31: Kernel Density Estimation (KDE) is used to analyze the probability density distribution of the target performance indicators in the training set, and local minimum points in the probability density distribution curve are found to divide the target value range into M regions , identify sparse regions in the target performance space, and identify samples in the sparse regions as key samples.

[0021] S32: In the input feature space of the entire training set (i.e., the multi-dimensional space composed of component amounts), find k nearest neighbors for each key sample identified in S31, preferentially randomly select a neighbor that is also identified as a key sample from the neighbors, and perform synchronous linear interpolation on the feature vector and the target vector on the line connecting the key sample and the selected neighbor to generate new sample data.

[0022] S33: Randomly select a portion of the sample data generated in S32, add a small amount of random noise conforming to a Gaussian distribution to the feature vector, simulate small fluctuations in real tests, and improve the diversity and authenticity of the generated sample data.

[0023] S34: According to a predetermined global amplification ratio, cyclically execute S31, S32 and S31 to generate sufficient new samples for all key samples, combine the original training set with all newly generated samples (including samples with noise injection), and construct an augmented training set.

[0024] Further, the AR-CatBoost machine learning model established in step S4 introduces an adaptive regularization mechanism and a gradient-based sample weighting mechanism based on the standard CatBoost framework. When the model is trained using data, the following steps are included:

[0025] S41: forward calculation is performed on the training set samples to obtain the prediction results, and then the back propagation mechanism of the loss function is applied, and the first-order gradient and the second-order gradient are calculated at the same time. These gradient information provides input for the dynamic adjustment of the subsequent steps.

[0026] S42: based on the gradient-based sample weighting mechanism, the first-order gradient calculated in S41 is used as input to calculate the gradient influence factor , the gradient influence factor is normalized using the Softmax function and the temperature coefficient to generate the dynamic weight of the sample at the current iteration .

[0027] S43: when the current iteration is an integer multiple of the preset evaluation interval , the adaptive regularization mechanism is triggered, the relative change rate of the current loss to the previous evaluation is calculated , based on the relative change rate and the set threshold , it is judged whether the model is in the overfitting state, according to the judgment result, the regularization coefficient and the decay factor are adjusted according to the reinforcement factor and the decay factor ;

[0028] S44: using the dynamic sample weight of the current iteration and the adjusted regularization coefficient as a constraint condition, a new decision tree is constructed, and then the tree is integrated into the model in a weighted manner, completing the model update of this round of iteration.

[0029] S45: continuously monitor the training process, when the model reaches the preset maximum iteration number, or the training set performance no longer improves in continuous multiple iterations, it is determined that the model has converged. After training is terminated, the current AR-CatBoost model is output for hyperparameter optimization in S5 stage.

[0030] Further, step S5 specifically includes the following sub-steps:

[0031] S51: randomly generate an initial crocodile population according to the hyperparameters to be optimized within the boundary of the hyperparameter settings of the AR-CatBoost model, wherein each individual represents a combination of hyperparameters, and differentiated generation logic is used for hyperparameter type differences (integer / continuous) to avoid invalid values;

[0032] S52: update the initial population for multiple iterations, including the following contents:

[0033] S521: according to the current iteration number , execute four stages of high-altitude walking, abdominal walking, hunting coordination and hunting cooperation, update the position (hyperparameter value) of each individual in the population, and ensure that the hyperparameters do not exceed the upper and lower bounds through clipping;

[0034] S522: according to a preset probability, apply the Cauchy mutation mechanism to the current global optimal individual to generate a mutation candidate solution , calculate the fitness value of the mutation candidate solution, and judge whether to accept the candidate solution by comparing the fitness values. If the fitness of the mutation candidate solution is better than that of the original optimal individual, i.e. , replace the original optimal individual with it; when , the original optimal individual remains unchanged;

[0035] S523: every G generations, calculate the diversity of the current population in each hyperparameter dimension, and dynamically adjust the search boundary of the corresponding dimension according to the diversity size to balance the search efficiency and exploration range;

[0036] S524: for each individual in the population, use the hyperparameter combination represented by the individual to train the AR-CatBoost model on the augmented training set, and calculate the weighted RMSE of multiple performance indicators of the model prediction value and the true value as the fitness value of the individual;

[0037] S53: after the iteration is completed, output the hyperparameter combination represented by the historical optimal individual, and use the optimal hyperparameter combination to retrain the AR-CatBoost model on the augmented data set. The performance of the model is evaluated using the test set to obtain the final proxy model for low-temperature asphalt performance prediction.

[0038] Further, the hyperparameters in step S5 include the number of iterations, the learning rate, the decision tree depth, the regularization coefficient, the weight coefficient, the temperature coefficient, the Huber loss threshold and the evaluation interval.

[0039] Compared with the prior art, the present application provides a small sample tunnel low-temperature asphalt optimization design method based on machine learning, which has the following beneficial effects:

[0040] (1) The AR-CatBoost model proposed in the application improves the learning efficiency and prediction accuracy of the model and dynamically prevents the model from overfitting by introducing an adaptive regularization mechanism and a gradient weighting mechanism, can automatically discover the key component synergy rule, and significantly improves the generalization ability under a small amount of experimental data.

[0041] (2) The improved reptile search algorithm adopted in the application has strong global exploration and local development capabilities for the hyperparameter optimization process through Cauchy variation and dynamic boundary self-adaption, can efficiently and reliably find the hyperparameter configuration close to the global optimum for the complex machine learning model, and ensures the maximization of the performance of the whole system.

[0042] (3) The application seamlessly integrates data generation, data enhancement, intelligent modeling and intelligent optimization, forms a complete and automated new paradigm for material research and development, liberates the material ratio optimization design from tedious trial-and-error experiments, and greatly speeds up the design process of new materials.

[0043] (4) The application can significantly shorten the research and development cycle of new materials, greatly reduce the experimental cost and raw material consumption, effectively improve the service life and safety of the tunnel pavement by precisely designing high-performance low-temperature asphalt materials, and has great economic and social benefits. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is the overall flowchart of the method described in the application.

[0045] Figure 2 is a schematic diagram of the SMOTER data enhancement principle.

[0046] Figure 3 is a schematic diagram of the AR-CatBoost model training process.

[0047] Figure 4 is a schematic diagram of the CD-RSA search algorithm optimization process.

[0048] Figure 5 is a comparison chart of the predicted value and the actual value of the AR-CatBoost model after CD-RSA optimization.

[0049] Figure 6 is a predicted residual chart of the AR-CatBoost model after CD-RSA optimization. DETAILED DESCRIPTION

[0050] To make the purpose, technical scheme and advantages of the application clearer and more apparent, the specific embodiments of the application will be further described in detail below with reference to the drawings. It should be pointed out that the specific embodiments described here are only used to explain the application, and are not used to limit the protection scope of the application.

[0051] The embodiment provides a small sample tunnel low-temperature asphalt optimization design method based on machine learning, and comprises the following steps.

[0052] S1: Tunnel low-temperature asphalt proportioning design and sample preparation

[0053] Based on the multi-raw material component system of the tunnel low-temperature asphalt, the types and content ranges of the components of the tunnel low-temperature asphalt are determined, a test scheme of the tunnel low-temperature asphalt mixing proportion is generated, and a tunnel low-temperature asphalt sample is prepared according to the test scheme.

[0054] The types of the components of the low-temperature asphalt include matrix asphalt and additional components (SBS modifier, rubber powder, resin, sulfur crosslinking agent and zinc oxide catalyst), and the test scheme adopts an orthogonal test design method to design 30 tests, as shown in Table 1.

[0055] The raw material components selected in the embodiment of the tunnel low-temperature asphalt include No. 70 A-grade matrix asphalt, SBS modifier, rubber powder, resin, sulfur crosslinking agent and zinc oxide catalyst, and the specific content ranges are 100 parts of matrix asphalt, 3-6 parts of SBS modifier, 10-20 parts of rubber powder, 2-5 parts of resin, 0.1-0.5 parts of sulfur crosslinking agent and 0.1-0.3 parts of zinc oxide catalyst.

[0056] S2: Establishment of an initial small sample data set

[0057] The performance indexes of the tunnel low-temperature asphalt under different proportions are determined according to the standard test methods in the “Highway Engineering Asphalt and Asphalt Mixture Test Regulations” (JTG 3410-2025), and the initial small sample data set is established.

[0058] The performance indexes of the tunnel low-temperature asphalt include the 25 penetration value determined by the penetration test, the 5 ductility value determined by the asphalt ductility test, the softening point value determined by the asphalt softening point test and the -18℃ creep stiffness S and creep rate m values determined by the bending beam rheological test.

[0059] In the embodiment, the small sample data set is the data for training the tunnel low-temperature asphalt performance prediction model, as shown in Table 1. Specifically, the initial small sample data set comprises an input matrix of specific values of additional components (SBS modifier, rubber powder, resin, sulfur crosslinking agent and zinc oxide catalyst) of the tunnel low-temperature asphalt and a target matrix of specific values of performance indexes (25 penetration, 5 ductility, softening point, -18 creep stiffness S and -18 creep rate m of the tunnel low-temperature asphalt .

[0060] In this embodiment, each row in the small sample dataset represents a sample, containing five input variables and five target variables. The specific doping levels of each added component in each sample are integrated into a feature vector. The corresponding performance index values ​​are integrated into a target vector. Specifically defined as:

[0061] (1)

[0062] In the formula, The dosage of SBS modifier, The amount of adhesive powder added. This refers to the amount of resin added. The dosage of sulfur crosslinking agent, This refers to the dosage of zinc oxide catalyst;

[0063] (2)

[0064] In the formula, 25 Penetration value, 5 Ductility value, This is the softening point value. -18 Creep stiffness, -18 Creep rate.

[0065] Table 1

[0066]

[0067] S3: Data Augmentation Based on Synthetic Minority Oversampling Technique (SMOTER)

[0068] In this embodiment, to address the overfitting problem caused by small sample data, synthetic minority oversampling technique is used to augment the small sample data. The specific process is as follows: Figure 2 As shown, the initial small sample dataset is divided into a training set and a test set before data augmentation, specifically, the ratio of the training set to the test set is 85:15.

[0069] The specific implementation steps in this step are as follows:

[0070] S31: First, calculate the weighted bandwidth to distinguish the importance of the target variable, as shown in Equation (3); use kernel density (KDE) estimation to analyze the probability density distribution of the target variable, as shown in Equation (4); find the local minimum point in the probability density distribution curve and divide the target value range into M regions. identify the sparse regions in the target performance space and label the samples in the sparse regions as key samples.

[0071] (3)

[0072] wherein, is the weighted bandwidth, is the weight of the i-th target variable, is the standard deviation of the i-th target variable, is the number of samples.

[0073] (4)

[0074] wherein, is the probability density function, is the kernel function, is the number of samples, is the value of the i-th target variable, is the measured value of the i-th target variable of the j-th sample.

[0075] S32: In the input feature space of the entire training set (i.e. the multi-dimensional space composed of the amounts of each additional component), find the k nearest neighbors of each key sample identified in S31, wherein the value of k is determined according to the size of the training set and ranges from 3 to 8; preferentially select a neighbor that is also identified as a key sample from the nearest neighbors, and perform simultaneous linear interpolation on the feature vector and the target vector on the line connecting the key sample and the selected neighbor to generate new sample data, wherein the interpolation formulae are shown in formula (5) and formula (6).

[0076] (5)

[0077] wherein, is the feature vector of the new sample, is the feature vector of the key sample, is the feature vector of the neighbor sample of the key sample, is the interpolation coefficient, and the value range is .

[0078] (6)

[0079] wherein, is the target vector of the new sample, is the target vector of the key sample, is the target vector of the neighbor sample of the key sample.

[0080] ​​​​​S33: Randomly select a part of the samples generated in S32, usually 20%-40%, and add a small amount of random noise conforming to Gaussian distribution to the feature vectors of the selected samples, as shown in formula (7), to simulate the slight fluctuations in real tests and improve the diversity and authenticity of the generated samples.

[0081] (7)

[0082] In the formula, is the new feature vector generated after injecting Gaussian noise, is a multivariate Gaussian distribution, is a variance, with a value range of [0.01, 0.05], is a 5-dimensional unit matrix.

[0083] S34: According to the preset global expansion ratio, S31, S32 and S31 are executed in a loop to generate sufficient new samples for all key samples, and the global expansion ratio is determined according to the number of initial training set samples, which can make the enhanced training set reach 2-5 times of the original training set. The original training set and all newly generated samples (including samples injected with noise) are combined to construct an enhanced training set.

[0084] S4: Establishing an AR-catboost machine learning model for predicting the performance of tunnel low temperature asphalt

[0085] An AR-CatBoost machine learning model is established, and the training set enhanced in S3 is used for model training, with asphalt additives as input features and target performance indicators as output targets, to learn the complex nonlinear mapping relationship between the dosage of each component of low temperature asphalt and multiple performances.

[0086] In this embodiment, the AR-CatBoost machine learning model introduces an adaptive regularization mechanism and a gradient-based sample weighting mechanism based on the standard CatBoost framework, and the specific process is as shown in Figure 3 The training and prediction of the tunnel low temperature asphalt dataset are realized by the following steps:

[0087] S41: Use the input matrix of the training set samples to perform forward calculation to obtain the prediction result of the current iteration , based on the real target matrix and Calculate the loss function using formulas (8)-(11), and calculate the first-order gradient and the second-order gradient according to the loss function, as shown in formulas (12) and (13).

[0088] (8)

[0089] wherein, is the loss function, is the loss of the th performance indicator, and are the dosage interval constraint and the local smooth constraint, respectively, , and are the weight coefficients.

[0090] (9)

[0091] wherein, is the Huber loss of the th performance indicator, is the Huber loss threshold of the th performance indicator, and the value range is [1, 3].

[0092] (10)

[0093] wherein, is the lower limit of the th performance indicator, is the upper limit of the th performance indicator.

[0094] (11)

[0095] wherein, is is the neighbor sample of the sample , is the Euclidean distance square of the input features of the sample and .

[0096] (12)

[0097] (13)

[0098] S42: based on the gradient, the first-order gradient calculated in S41 is used as the input to calculate the gradient influence factor of the sample, as shown in formula (14); the Softmax function and the temperature coefficient are used to normalize the gradient influence factor to generate the dynamic weight of the sample at the current iteration number , as shown in formula (15), and the value range of the temperature coefficient is [0.5, 2].

[0099] ​ (14)

[0100] (15)

[0101] S43: At the current iteration number is an integer multiple of the preset evaluation interval, trigger the adaptive regularization mechanism, wherein the value range of is 10~50; according to the current loss, calculate its relative change rate to the previous evaluation , as shown in formula (16); based on the relative change rate and the set threshold , judge whether the model is in the overfitting state, according to the judgment result, adjust the regularization coefficient and , as shown in formula (17), the value range of the threshold is [0.05, 0.25], the value range of the reinforcement factor is 1.1~1.5, and the value range of the attenuation factor is 0.5~0.9.

[0102] (16)

[0103] In the formula, is the loss value on the validation set of the iteration , and is the loss value on the validation set of the iteration .

[0104] (17)

[0105] S44: Use the dynamic sample weight of the current iteration and the adjusted regularization coefficient as the constraint condition to construct a new decision tree, and then integrate the tree into the model in a weighted manner to complete the model update of this round of iteration.

[0106] S45: Continue to monitor the training process, when the model reaches the preset maximum iteration number, or the performance of the training set no longer improves in continuous multiple iterations, determine that the model has converged. After the training is terminated, output the current AR-CatBoost model for hyperparameter optimization in S5 stage.

[0107] S5: Hyperparameter optimization based on improved Crawler-based Differential Search Algorithm (CD-RSA)

[0108] ​​​An improved reptile search algorithm (CD-RSA) is used to automatically optimize the hyperparameters of the AR-CatBoost machine learning model, finding the optimal combination of hyperparameters, including the number of iterations, learning rate, decision tree depth, regularization coefficient, weight coefficient, temperature coefficient, Huber loss threshold, and evaluation interval.

[0109] In this embodiment, the CD-RSA optimization algorithm introduces Cauchy mutation and dynamic boundary adjustment mechanisms on the basis of the standard RSA optimization algorithm, enhancing the algorithm's global exploration capability and local fine-grained search capability in the hyperparameter space, avoiding getting trapped in local optima and improving the convergence accuracy of optimization. The specific process is shown in Figure X. The hyperparameter optimization of the AR-CatBoost machine learning model is achieved through the following steps:

[0110] S51: Randomly generate an initial crocodile population based on the hyperparameters to be optimized, where each individual represents a set of hyperparameter combinations, all of which fall within a preset upper bound. and the lower realm Meanwhile, differential generation logic is adopted for the differences in hyperparameter types (integer / continuous) to avoid invalid values, as shown in equation (18).

[0111] (18)

[0112] In the formula, A uniformly random number in the interval [0,1]. This is for rounding down.

[0113] S52: Perform multiple rounds of iterative updates on the initial population, as detailed below:

[0114] S521: Based on the current iteration number Within the range, the population performs four stages: high-altitude walking, abdominal walking, hunting coordination, and hunting cooperation, updating the position (hyperparameter value) of each individual in the population. At the same time, the hyperparameter is ensured not to exceed the upper and lower bounds through pruning, as shown in Equations (19) and (20).

[0115] (19)

[0116] In the formula, For the clipping function, For stage functions, For the first One hyperparameter in The global optimal value at the nth iteration. It is a contraction factor.

[0117] (20)

[0118] In the formula, is a hunting operator, is a cooperation probability coefficient, is an evolution perception function, is a shock coefficient.

[0119] S522: with a preset probability A Cauchy mutation mechanism is applied to the current global optimal individual to generate a mutation candidate solution , as shown in equation (21); the fitness value of the mutation candidate solution is calculated using equation (22), and whether to accept the candidate solution is judged by comparing the fitness; if the fitness of the mutation candidate solution is better than that of the original optimal individual, i.e. , the original optimal individual is replaced by the mutation candidate solution; when , the original optimal individual remains unchanged.

[0120] (21)

[0121] In the formula, is a differential mutation step size, is a standard Cauchy random distribution.

[0122] (22)

[0123] S523: every interval G generations, the diversity of the current population in each hyperparameter dimension is calculated, as shown in equation (23), and the G is 5-15 generations; according to the size of the diversity, the search boundary of the corresponding dimension is dynamically adjusted to balance the search efficiency and the exploration range, as shown in equations (24) and (25).

[0124] (23)

[0125] (24)

[0126] (25)

[0127] In the formula, , is a diversity coefficient, and the value range is [0.5, 3].

[0128] S524: for each individual in the population, use the hyperparameter combination represented by it to train the AR-CatBoost model on the augmented training set, and calculate the multi-performance index weighted RMSE of the model prediction value and the true value as the fitness value of the individual (the smaller the value, the better the hyperparameter combination).

[0129] S53: After the iteration, output the hyperparameter combination represented by the historical optimal individual, retrain the AR-CatBoost model on the augmented dataset using the optimal hyperparameter combination, and evaluate the performance of the model on the test set to obtain the final surrogate model for predicting the performance of the tunnel low-temperature asphalt.

[0130] S6: Low-temperature asphalt proportioning optimization design

[0131] Using the AR-CatBoost model trained in S4 and S5, generate large-scale virtual proportions within the component dosage range determined in S1 to predict the performance of low-temperature asphalt under different characteristic parameters, and quantitatively evaluate the performance of different proportions through the comprehensive scoring formula. In the component dosage range, accurately select the optimal proportion with excellent performance and meet the engineering practice, and provide a reliable technical solution for efficient and accurate design of tunnel low-temperature asphalt.

[0132] Score (26)

[0133] In the formula, Score is the comprehensive score of low-temperature asphalt performance, the higher the score, the better the comprehensive performance, is the target value of creep stiffness at -18℃, is the predicted value of creep stiffness at 18℃, is the target value of creep rate at -18℃, is the predicted value of creep rate at -18℃. is the target value of ductility at 5℃, is the predicted value of ductility at 5℃, is the target value of softening point, is the predicted value of softening point, is the target value of penetration at 25℃, is the predicted value of penetration at 25℃, is the weight coefficient.

[0134] Test verification

[0135] To evaluate the prediction accuracy of the model, the model predicted values and actual values are compared in this embodiment. The comprehensive verification is carried out by combining numerical indicators with visual charts, and the results show that the determination coefficients R² of all performance indicators of the model are higher than 0.91, and the root mean square error RMSE is at an extremely low acceptable level, as shown in Table 2; Figure 5 The scatter plot of the predicted value and the actual value of shows that all sample points are closely distributed on both sides of the y=x reference line, which intuitively shows that the predicted value and the true value have high consistency; Figure 6 The residual plot of shows that the residuals are uniformly and randomly distributed near the zero line, indicating that the model maintains stable prediction accuracy in different performance intervals; it is proved that the model constructed in this embodiment has excellent prediction accuracy and generalization ability.

[0136] Based on the above-verified high-precision model, a set of optimal matching schemes is output according to the comprehensive score formula, specifically 4.3% SBS modifier, 14.8% rubber powder, 3.1% resin, 0.32% sulfur crosslinking agent and 0.17% zinc oxide catalyst. The optimal matching is verified by test, and the deviation analysis of the predicted value and the measured value is shown in Table 3. The relative error of all performance indicators is less than 5%, within the allowable error range of engineering.

[0137] Table 2

[0138]

[0139] Table 3

[0140]

Claims

1. A small-sample low-temperature asphalt tunnel optimization design method based on machine learning, characterized in that, Includes the following steps: S1: Determine the raw material components and dosage range of the low-temperature asphalt for tunnels, generate an experimental scheme for the mix proportion of low-temperature asphalt for tunnels through orthogonal experimental design, and prepare low-temperature asphalt samples for tunnels according to the experimental scheme; S2: Determine the performance indicators of low-temperature asphalt for tunnels with different mix proportions and establish an initial small sample dataset; S3: Data augmentation is performed using the synthetic minority oversampling technique SMOTER. By identifying sparse regions in the target performance space, linear interpolation is used to perform targeted data augmentation on the sparse regions, generating new samples that conform to the distribution pattern of the original data, avoiding the imbalance of minority samples, and thus effectively expanding the diversity and scale of the data. S4: Establish an AR-CatBoost machine learning model, use the data augmented dataset in S3 for model training, take raw material components as input features and target performance indicators as output, and learn the complex nonlinear mapping relationship between the admixture content of each component of low temperature asphalt in tunnels and various properties. S5: The improved reptile search algorithm CD-RSA is used to automatically optimize the hyperparameters of the AR-CatBoost machine learning model and find the optimal combination of hyperparameters. S6: The AR-CatBoost machine learning model with optimized hyperparameters using CD-RSA is retrained to generate large-scale virtual proportion performance predictions within the dosage range of each component determined in S1. The performance of different proportions is quantitatively evaluated through a comprehensive scoring formula, and the optimal proportion that achieves excellent performance and conforms to engineering practice is accurately selected within the dosage range of each component.

2. The method according to claim 1, characterized in that, The tunnel low-temperature asphalt raw material components in step S1 include No. 70 Grade A base asphalt, SBS modifier, rubber powder, resin, sulfur crosslinking agent, and zinc oxide catalyst. The specific dosage ranges by mass parts are as follows: 100 parts base asphalt, 3-6 parts SBS modifier, 10-20 parts rubber powder, 2-5 parts resin, 0.1-0.5 parts sulfur, and 0.1-0.3 parts zinc oxide catalyst.

3. The method according to claim 1, characterized in that, In step S1, the experimental design adopts the orthogonal experimental design method to design 30 sets of experiments.

4. The method according to claim 1, characterized in that, The performance indicators in step S2 include 25% penetration test. Penetration value, ductility value at 5℃ determined by asphalt ductility test, softening point value determined by asphalt softening point test, and -18℃ value determined by bending beam rheology test. Creep stiffness S and creep rate m at -18℃.

5. The method according to claim 1, characterized in that, The initial small sample dataset in step S2 includes an input matrix containing the specific values ​​of the tunnel low-temperature asphalt additives, namely modifiers, rubber powder, resin, crosslinking agents, and catalysts. And the performance index of low-temperature asphalt in tunnels, namely 25 Penetration, 5 Ductility, softening point, -18 Creep stiffness S, -18 The target matrix for the specific value of the creep rate m .

6. The method according to claim 1, characterized in that, In step S3, before data augmentation, the initial small sample dataset is divided into a training set and a test set, with the ratio of the training set to the test set being 85:

15.

7. The method according to claim 1, characterized in that, Step S3 employs the synthetic minority oversampling technique SMOTER for data augmentation, including the following steps: S31: Analyze the probability density distribution of the target performance index in the training set using kernel density KDE estimation, find the local minimum points in the probability density distribution curve, and divide the target value range into M regions. Identify sparse regions in the target performance space and mark the samples therein as key samples; S32: In the input feature space of the entire training set, i.e. the multidimensional space composed of the doping of each component, find the k nearest neighbor samples for each key sample identified in S31. Prioritize randomly selecting a neighbor that is also identified as a key sample from the nearest neighbors. On the line connecting the key sample and its selected neighbor, perform synchronous linear interpolation on the feature vector and the target vector to generate new sample data. S33: Randomly select a portion of the samples generated in S32, add a small amount of random noise that conforms to a Gaussian distribution to their feature vectors, simulate the small fluctuations in real experiments, and improve the diversity and authenticity of the generated sample data. S34: Based on the preset global amplification ratio, repeat S31, S32 and S31 to generate a sufficient number of new samples for all key samples. Then merge the original training set with all the newly generated samples, including those injected with noise, to construct the enhanced training set.

8. The method according to claim 1, characterized in that, The AR-CatBoost machine learning model established in step S4 introduces an adaptive regularization mechanism and a gradient-based sample weighting mechanism on top of the standard CatBoost framework. When training this model using data, the following steps are included: S41: Perform forward computation on the training set samples to obtain the prediction results, and then apply the backpropagation mechanism of the loss function to calculate the first-order gradient. and second gradient These gradient information provide input for dynamic adjustments in subsequent steps; S42: Gradient-based sample weighting mechanism, using the first-order gradient calculated in S41. Its gradient influence factor is calculated as input. Using the Softmax function and temperature coefficient The gradient influence factor is normalized to generate the dynamic weights of the samples at the current iteration number. ; S43: At the current iteration number Preset evaluation interval When the loss is an integer multiple of the previous value, an adaptive regularization mechanism is triggered, which calculates the relative rate of change of the current loss with respect to the previous evaluation. Based on the relative rate of change and a set threshold To determine whether the model is overfitting, based on the determination result, adjust the enhancement factor accordingly. and attenuation factor Adjusting the regularization coefficient and ; S44: Using the dynamic sample weights of the current iteration and the adjusted regularization coefficients as constraints, construct a new decision tree, and then integrate the tree into the model in a weighted manner to complete the model update for this iteration. S45: Continuously monitor the training process. When the model reaches the preset maximum number of iterations, or the performance of the training set no longer improves in multiple consecutive iterations, it is determined that the model has converged. After training is terminated, the current AR-CatBoost model is output for hyperparameter optimization in the S5 stage.

9. The method according to claim 1, characterized in that, Step S5 specifically includes the following sub-steps: S51: Within the boundaries of the hyperparameter settings of the AR-CatBoost model, an initial crocodile population is randomly generated according to the hyperparameters to be optimized, where each individual represents a set of hyperparameter combinations. At the same time, differential generation logic is used for differences in hyperparameter types, i.e., integer / continuous, to avoid invalid values. S52: Perform multiple rounds of iterative updates on the initial population, specifically including the following: S521: Based on the current iteration number Within its range, it performs four stages: high-altitude walking, belly walking, hunting coordination, and hunting cooperation, updating the position and hyperparameter values ​​of each individual in the population, while ensuring that the hyperparameters do not exceed the upper and lower bounds through pruning. S522: Apply the Cauchy mutation mechanism to the current globally optimal individual according to a preset probability to generate candidate solutions. Calculate the fitness value of the mutated candidate solutions, and determine whether to accept the candidate solution by comparing fitness values. If the fitness of the mutated candidate solution is better than that of the original optimal individual, then... Replace the original optimal individual with it; when At that time, the original optimal individual remains unchanged; S523: Every G generations, calculate the diversity of the current population in each hyperparameter dimension, and dynamically adjust the search boundary of the corresponding dimension according to the diversity level to balance search efficiency and exploration range. S524: For each individual in the population, train an AR-CatBoost model on the augmented training set using the hyperparameter combination it represents, and calculate the weighted RMSE of the model's predicted values ​​and the true values ​​as the fitness value of that individual. S53: After the iteration is completed, output the hyperparameter combination represented by the best individual in history. Using this optimal hyperparameter combination, retrain the AR-CatBoost model on the augmented dataset. Use the test set to evaluate the model's performance and obtain the final surrogate model for predicting the performance of low-temperature asphalt in tunnels.

10. The method according to claim 1, characterized in that, The hyperparameters in step S5 include the number of iterations, learning rate, decision tree depth, regularization coefficient, weight coefficient, temperature coefficient, Huber loss threshold, and evaluation interval.

Images