Multi-source machine learning driven asphalt fatigue life prediction model construction method

Abstract

The invention relates to the technical field of road engineering material performance prediction, and discloses a multi-source machine learning driven asphalt fatigue life prediction model construction method, which comprises the following steps: obtaining a characteristic parameter set; obtaining a converted characteristic parameter set; obtaining a processed training set and a test set; the output of the feature data coding layer is extracted to serve as enhanced feature representation and feature importance scores; obtaining a preliminary prediction result of the base model; and inputting the preliminary prediction result and the feature importance score into an adaptive weighted ridge regression meta-model for training and fusion to obtain a modified asphalt fatigue life prediction model. According to the method, the model prediction precision and the model stability are improved, the limitation that a traditional regression model is low in universality and depends on artificial experience is effectively overcome, efficient and accurate prediction of the fatigue life of the modified asphalt is achieved, and a reliable data driving solution is provided for engineering material performance evaluation.

Description

Technical Field

[0001] This invention relates to the field of road engineering material performance prediction technology, and more specifically, to a method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model. Background Technology

[0002] Asphalt pavements are subjected to the combined effects of traffic loads and environmental factors over a long period of time, leading to fatigue damage and eventually cracking. Fatigue life is a key indicator for evaluating the service performance of asphalt pavements. To improve the fatigue performance of asphalt, various admixtures, such as polymers, rubber powder, fibers, or whiskers, are often added.

[0003] Current research on fatigue life prediction of modified asphalt mainly relies on laboratory tests. Empirical or regression models established based on experimental data to predict fatigue life and influencing factors often lack sufficient accuracy and generalization ability. Furthermore, the models exhibit poor stability when the material system or external conditions change, requiring recalibration.

[0004] Machine learning provides powerful support for handling complex nonlinear relationships. However, a single machine learning model struggles to balance accuracy and stability. Therefore, there is an urgent need for a deep learning-machine learning model coupling framework that can integrate the advantages of multiple models and possess high accuracy and high stability.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model. This modified asphalt fatigue life prediction model possesses high prediction accuracy and high stability, effectively overcoming the limitations of conventional data regression fitting models. Furthermore, it solves the problems of existing technologies in establishing asphalt fatigue life prediction models, such as being time-consuming and labor-intensive, having weak prediction accuracy and universality.

[0007] (II) Technical Solution To achieve the high prediction accuracy and high stability of the modified asphalt fatigue life prediction model, and to effectively overcome the limitations of conventional data regression fitting models, the specific technical solution adopted in this invention is as follows: A method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model includes: Obtain a dataset of fatigue life of modified asphalt, perform one-hot encoding on the feature parameters in the dataset, calculate the correlation between each feature parameter to construct a feature correlation matrix, and obtain a set of feature parameters. Using fatigue life as the target variable, and processing the target variable to complete the transformation of the target variable, we obtain the transformed set of feature parameters. The transformed feature parameter set is divided into a training set and a test set according to a preset ratio, and the training set and test set are standardized using a standardization method to obtain the processed training set and test set. The feature data in the processed training set is embedded and positionally encoded, and then input into a deep learning model for training. The output of the feature data encoding layer is extracted as an enhanced feature representation and feature importance score. Based on the enhanced feature representation, the gradient boosting regression model and the extreme gradient boosting regression model are trained using the five-fold cross-validation method to obtain the preliminary prediction results of the base model. The preliminary prediction results and feature importance scores are input into an adaptive weighted ridge regression meta-model for training and fusion to obtain a modified asphalt fatigue life prediction model.

[0008] Furthermore, the transformed feature parameter set is divided into a training set and a test set according to a preset ratio, and the training set and test set are standardized using a standardization method. The processed training set and test set include: The transformed feature parameter set is divided into a training set and a test set according to a preset ratio; Calculate the mean and standard deviation of the training set, and then use the mean and standard deviation to standardize the training set according to the standard deviation standardization formula to obtain the standardized training set data; Based on the mean and standard deviation of the training set, the test set is standardized according to the standard deviation standardization formula to obtain the standardized training set data.

[0009] Furthermore, the feature data in the processed training set is embedded and positionally encoded, and then input into a deep learning model for training. The output of the feature data encoding layer is extracted as an enhanced feature representation and feature importance score, including: The feature data in the processed training set is embedded to obtain the embedding vector, and the embedding layer dimension of the deep learning model is set. The position encoding matrix is ​​generated based on the formulas for sine and cosine functions. The position encoding matrix is ​​added to the embedding vector and then input into the deep learning model. A multi-head self-attention mechanism with several attention heads is set up. The objective function is to minimize the mean squared error between the predicted and true values. The deep learning model is trained using adaptive moment estimation, and the output of the feature data encoding layer is extracted as an enhanced feature representation and feature importance score.

[0010] Furthermore, based on the enhanced feature representation, the gradient boosting regression model and the extreme gradient boosting regression model were trained using the five-fold cross-validation method, respectively, and the preliminary prediction results of the base model were obtained, including: Select the gradient boosting regression model and the extreme gradient boosting regression model as the base models, and initialize the base models according to the preset parameters; The training set is divided into five folds using the five-fold cross-validation method. The base model is trained using four folds in sequence, and predictions are made for the remaining fold. After repeating the loop several times, the predicted values ​​of the training set on each base model are obtained to complete the training of the base model. The trained base model is used to predict the enhanced feature matrix of the test set, generating preliminary prediction results.

[0011] Furthermore, the preliminary prediction results and feature importance scores are input into an adaptive weighted ridge regression meta-model for training and fusion, resulting in a modified asphalt fatigue life prediction model, including: The predicted values ​​of the base model and the feature importance scores are input into the adaptive weighted ridge regression meta-model, and the adaptive weighted ridge regression meta-model is trained with the target variable as the label to obtain the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model. Using a search algorithm, the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model are optimized to determine the optimal weight coefficients. Based on the optimal weighting coefficients, a fatigue life prediction model for modified asphalt is determined, and the performance of the modified asphalt fatigue life prediction model is evaluated on an independent validation set to obtain the final modified asphalt fatigue life prediction model.

[0012] Furthermore, using a search algorithm, the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model are optimized to determine the optimal weight coefficients, including: Initialize the parameters of the search algorithm, set the maximum number of iterations, define the optimization space of the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model, and configure a set of preset candidate values ​​for each weight coefficient; Establish the traversal order of the weight coefficient combinations and use a dynamic adjustment algorithm to select the current weight coefficient combination to be evaluated from the parameter space; The predicted values ​​of the base model are weighted and summed using the current combination of weight coefficients, and then multiplied by the feature importance score to obtain the predicted values ​​of the validation set samples. Calculate the mean square error between the predicted values ​​and the actual values ​​of the validation set samples, and use this mean square error as the performance evaluation index of the current weight coefficient combination. Repeat the process of selecting, calculating, and evaluating the weight coefficient combination. If the maximum number of iterations is reached, the weight coefficient combination with the smallest mean square error is output as the optimal weight coefficient.

[0013] Furthermore, the traversal order of the weight coefficient combinations is established, and the current weight coefficient combination to be evaluated is selected from the parameter space using a dynamic adjustment algorithm, including: Initialize the parameters of the dynamically adjusted algorithm and set the maximum number of iterations; Initialize the weight coefficient search space, generate multiple weight coefficient combinations as candidate solutions, calculate the initial performance evaluation value of each weight coefficient combination and store it in the memory bank; The first random number is generated randomly. If the first random number is less than the preset memory selection probability, a weight coefficient combination is randomly selected from the memory. Otherwise, a new weight coefficient combination is generated randomly. A second random number is generated. If the second random number is less than the preset coefficient adjustment probability, the selected weight coefficient combination is mutated and adjusted to obtain a new weight coefficient combination; otherwise, the original weight coefficient combination is retained. Calculate the performance evaluation value of the new weight coefficient combination. If the evaluation value is greater than the worst evaluation value in the memory, replace it; otherwise, keep the memory unchanged. If the maximum number of iterations is reached, the optimal combination of weight coefficients is output as the current combination of weight coefficients to be evaluated; otherwise, the iteration continues.

[0014] Furthermore, a first random number is generated. If the first random number is less than the preset memory selection probability, a weight coefficient combination is randomly selected from the memory; otherwise, a new weight coefficient combination is randomly generated, including: Initialize the random number generator and generate the first random number by uniformly sampling within a preset interval; The first random number is compared with the preset memory selection probability to determine whether to select a weight coefficient combination from the memory. If the first random number is less than the preset memory selection probability, then an existing combination of weight coefficients is uniformly and randomly drawn from the memory as the current candidate solution. If the first random number is greater than or equal to the preset memory selection probability, then independent random sampling is performed within the domain of each weight coefficient combination to generate a new weight coefficient combination as a candidate solution.

[0015] Furthermore, a second random number is generated. If the second random number is less than the preset coefficient adjustment probability, the selected weight coefficient combination is mutated and adjusted to obtain a new weight coefficient combination; otherwise, the original weight coefficient combination is retained unchanged. A second random number is generated, and the second random number is compared with the preset coefficient adjustment probability to determine whether the mutation condition is met. If the second random number is less than the preset coefficient adjustment probability, then the genetic algorithm will perform a reverse mutation operation on the currently selected weight coefficient combination to generate a mutated intermediate solution. Boundary checks and parameter repairs are performed on the intermediate solutions to ensure that all weight coefficient combinations fall within the preset feasible range. After the verification is completed, the final new weight coefficient combination is obtained. If the second random number is greater than or equal to the preset coefficient adjustment probability, the original weight coefficient combination is directly retained without change.

[0016] Furthermore, the formulas for the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model are as follows: ; In the formula, This represents the weight vector to be learned; These represent the weighting coefficients that indicate the prediction results of the gradient boosting regression model. These represent the weighting coefficients that indicate the prediction results of the extreme gradient boosting regression model; This represents the vector of predicted values ​​from the gradient boosting regression model. This represents the vector of predicted values ​​from an extreme gradient boosting regression model. This indicates that the gradient boosting regression model applies to the first... Predicted values ​​for each training sample; This indicates that the extreme gradient boosting regression model is effective for the first... Predicted values ​​for each training sample; Represents the feature importance score vector; Indicates the first The overall score of feature importance corresponding to each sample; Indicates the first The true fatigue life of each sample; This indicates the number of training samples.

[0017] (III) Beneficial Effects Compared with existing technologies, this invention provides a method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model, which has the following beneficial effects: (1) This invention constructs a Transformer deep learning model for feature extraction and dynamic weighting, combines an integrated learning framework of gradient boosting regression and XGBoost models, and employs an adaptive weighted ridge regression meta-model for multi-level prediction fusion. This improves the model's prediction accuracy and stability, effectively overcomes the limitations of traditional regression models such as weak universality and reliance on human experience, and achieves efficient and accurate prediction of the fatigue life of modified asphalt, providing a reliable data-driven solution for the performance evaluation of engineering materials.

[0018] (2) This invention introduces a dynamic adjustment algorithm to realize intelligent exploration and adaptive optimization of the weight coefficient space. Combined with the systematic traversal mechanism of the search algorithm, it improves the efficiency of weight coefficient configuration of the gradient boosting regression model and the extreme gradient boosting regression model, ensuring that the gradient boosting regression model and the extreme gradient boosting regression model obtain the optimal weight coefficient combination, thereby accurately integrating the prediction results of the base model and the feature importance score, and finally constructing a modified asphalt fatigue life prediction model with higher prediction accuracy and stronger generalization ability, effectively overcoming the problems of low coefficient optimization efficiency and limited prediction performance of traditional methods. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to an embodiment of the present invention. Detailed Implementation

[0021] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0022] According to an embodiment of the present invention, a method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model is provided.

[0023] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to an embodiment of the present invention includes: S1. Obtain the dataset of fatigue life of modified asphalt, perform one-hot encoding on the feature parameters in the dataset, calculate the correlation between each feature parameter to construct the feature correlation matrix, and obtain the feature parameter set. Specifically, an experimental dataset containing the fatigue life data of modified asphalt was collected. The dataset's feature parameter set includes: 1) Asphalt type: such as base asphalt, SBS asphalt, and rubber-modified asphalt, etc., are processed using unique thermal coding to convert categorical variables into numerical variables. Coding is set according to whether different asphalt types are used, for example, base asphalt [1, 0, 0], SBS asphalt [0, 1, 0], and rubber-modified asphalt [0, 0, 1].

[0024] 2) Modifier type and dosage: such as calcium sulfate whiskers, chopped fibers, etc. Different types are individually thermally coded, and the dosage is entered as a continuous numerical value. Different modifier types can be coded as calcium sulfate whiskers [1, 0], chopped fibers [0, 1]. The modifier dosage is entered as a specific value according to the experimental conditions, such as 0, 5, 10, ...

[0025] 3) Aging conditions: such as thermo-oxidative aging, ultraviolet aging, and salt corrosion aging, etc., are set as ordered categorical variables or continuous variables according to the aging cycle or aging duration. For thermo-oxidative aging tests, fill in 0, 1, 2, ... according to the standard test cycle; for ultraviolet aging tests, fill in the number of aging cycles 0, 5, 10, 15, ... according to the set cycle; for salt corrosion aging, fill in the specific aging duration according to the aging duration, such as 0, 24, 48, ...

[0026] 4) Other relevant engineering parameters: such as stress level and test temperature, should be directly entered according to the actual test data.

[0027] S2. Using fatigue life as the target variable, and processing the target variable to complete the transformation of the target variable, the transformed feature parameter set is obtained. Specifically, the output feature is fatigue life. Since fatigue life typically spans multiple orders of magnitude, to improve the training effect of deep learning-machine learning models, the fatigue life value is taken as base 10 or the natural logarithm.

[0028] S3. Divide the transformed feature parameter set into a training set and a test set according to a preset ratio, and use a standardization method to standardize the training set and the test set to obtain the processed training set and test set. Specifically, the StandardScaler function is used to standardize the feature parameters, dividing the training and test sets into 80% and 20% portions, respectively. The training set is used for model training, and the test set is used for model validation. The StandardScaler function calculates key parameters for each feature, specifically the mean (μ) and standard deviation (σ) for each feature in the training set. The results are saved to a scaler object for later retrieval. The parameters (μ and σ) learned from the training set are used to transform the training and test sets. The command code (X_train - μ) / σ is used to transform each feature column in the training set. When transforming the test set, the parameters μ and σ learned from the training set are used, and the transformation is performed using (X_test - μ) / σ. This yields the standardized data.

[0029] S4. Embed and positionally encode the feature data in the processed training set, and input it into the deep learning model for training. Extract the output of the feature data encoding layer as the enhanced feature representation and feature importance score. S5. Based on the enhanced feature representation, the gradient boosting regression model and the extreme gradient boosting regression model are trained using the five-fold cross-validation method to obtain the preliminary prediction results of the base model. S6. Input the preliminary prediction results and feature importance scores into the adaptive weighted ridge regression meta-model for training and fusion to obtain the modified asphalt fatigue life prediction model.

[0030] Specifically, the coupled model (modified asphalt fatigue life prediction model) consists of three layers: 1) Transformer Model (Deep Learning Model) Training: The input embedding layer of the Transformer model is set to 64 dimensions, employing a multi-head self-attention mechanism with 8 attention heads. After embedding and positional encoding of the standardized training set feature data, it is input into the Transformer model for training. During training, the objective function is to minimize the mean squared error between the predicted and true values, using the Adam optimizer (adaptive moment estimator) with a learning rate of 0.001, and training for 200 iterations. After training, the output of the Transformer model's encoding layer is extracted as the enhanced feature representation. Before training the model, the standardized training set feature data undergoes embedding and positional encoding. Embedding converts the original feature data into a vector representation suitable for model processing. Positional encoding allows the model to perceive feature order information, generating a positional encoding matrix based on sine and cosine function formulas, which is then added to the embedding vector.

[0031] Assume the maximum length of the input sequence is Embedding dimension is =64, for position (0≤) <1) and dimension (0≤) <1), Position Encoding The calculation is as follows: when When it is even: ; when When it is an odd number: ; Generate size is × After the position encoding matrix, take the first... OK( The actual length of the input sequence is added to the embedding vector.

[0032] 2) Base Model Training: The gradient boosting regression model and XGBoost regression function are invoked, and the model is initialized according to preset parameters. Five-fold cross-validation is used to train the base model. The gradient boosting regression model has high training stability and anti-overfitting ability, while the XGBoost regression model (extreme gradient boosting regression model) has excellent prediction accuracy and computational efficiency. The training set is divided into five folds. The model (base model) is trained sequentially using four folds, and predictions are made on the remaining fold. This process is repeated five times to obtain the complete prediction results for the training set. The complete base model is then used to predict the enhanced feature matrix of the test set, generating preliminary prediction values. The purpose of cross-validation is to fully utilize the experimental data and generate predictive features that are less prone to overfitting for the third-level meta-model.

[0033] 3) Meta-model layer: The predicted values ​​generated by the two base models (gradient boosting regression model and extreme gradient boosting regression model) in the second layer through cross-validation are used as a new feature matrix and re-inputted into the third meta-model layer. The predicted values ​​of the base models and the feature importance scores output by the Transformer are input into the adaptive weighted ridge regression meta-model. The meta-model (adaptive weighted ridge regression meta-model) is trained with the target variable (logarithmized fatigue life, i.e., the target variable fatigue life value) of the original training set as the label, and the weighted feature coefficients of the gradient boosting regression model are learned.

[0034] Let the prediction vector of the gradient boosting regression model be... The predicted vector of the XGBoost model is, The feature importance score vector is The training process of the meta-model (adaptive weighted ridge regression meta-model) can be represented as minimizing the objective function: ; In the formula, This represents the weight vector to be learned; These represent the weighting coefficients that indicate the prediction results of the gradient boosting regression model. These represent the weighting coefficients that indicate the prediction results of the extreme gradient boosting regression model; This represents the vector of predicted values ​​from the gradient boosting regression model. This represents the vector of predicted values ​​from an extreme gradient boosting regression model. This indicates that the gradient boosting regression model applies to the first... Predicted values ​​for each training sample; This indicates that the extreme gradient boosting regression model is effective for the first... Predicted values ​​for each training sample; Represents the feature importance score vector; Indicates the first The overall score of feature importance corresponding to each sample; Indicates the first The actual fatigue life of each sample; This indicates the number of training samples.

[0035] The final prediction formula for the coupled model (modified asphalt fatigue life prediction model) is: Final predicted value = ( ×Gradient boosting regression predicted values+ ×XGBoost predicted value) × Transformer feature importance weight. Where, and The weight coefficients of the two models (gradient boosting regression model and extreme gradient boosting regression model) reflect the contribution of the base model's prediction results to the final prediction.

[0036] Model evaluation was performed on an independent validation set, and the predictive accuracy and stability of the model were verified by calculating the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²).

[0037] It needs to be explained that, taking the prediction of fatigue life of calcium sulfate whisker-modified asphalt under aging as an example: 1) 324 sets of linear amplitude scanning test results were collected and preprocessed for use in generating training and test sets in subsequent steps. Feature parameters were processed as follows: Asphalt type was individually thermally coded, with participating asphalt coded as 1 and non-participating asphalt coded as 0. Base asphalt was coded as [1, 0, 0], SBS asphalt as [0, 1, 0], and rubber asphalt as [0, 0, 1]. The modifier was calcium sulfate whiskers, with dosage set as a continuous value, filled with 0, 0.05, 0.10, and 0.15 according to the test results. Aging conditions included three types: UV aging, thermo-oxidative aging, and salt corrosion aging. Aging node data was set as ordered categories according to the cycle. For UV aging, the corresponding column data was entered as 0, 5, 10, and 15 based on the cycle number; for thermo-oxidative aging, 0, 1, and 5 were entered based on different aging cycles; and for salt corrosion aging, 0, 5, 10, and 15 were entered based on the aging cycle. Pearson correlation coefficients were calculated using the pandas library to construct the feature correlation matrix.

[0038] 2) Use fatigue life value as the output feature and take its natural logarithm to eliminate the impact of large numerical range on model training.

[0039] 3) Use the StandardScaler function to perform data standardization. The training set is set up with 80% of the data, and the test set with 20%. The training set is used for model training, and the test set is used to test the constructed prediction model. The data is standardized using the Z-score standardization method. First, the original dataset is divided into a training set (X_train) and a test set (X_test). Then, the mean (μ) and standard deviation (σ) are calculated only from the training set data. Next, μ and σ are used to standardize the training set data according to (X_train - μ) / σ. The test set data is then standardized using the command (X_test - μ) / σ.

[0040] 4) Set the input embedding layer dimension of the Transformer model to 64-dimensional, meaning each feature maps to a 64-dimensional vector space. Employ a multi-head self-attention mechanism with 8 attention heads to capture various correlations. Input the processed data into the Transformer model and perform inverse training. The training objective is to minimize the mean squared error between the predicted and true values. Use the Adam optimizer with a learning rate of 0.001 and 200 training iterations. After training, extract the Transformer model's output as the enhanced feature representation. Call the gradient boosting regression function and XGBoost regression function from the function library to construct the first base model. Initialize the base model, using the same settings for both the gradient boosting regression and XGBoost models: 100 decision trees, a learning rate of 0.1, a maximum tree depth of 4, and 42 random seeds. Use 5-fold cross-validation to train the base model to prevent overfitting and improve generalization ability. Divide the training set into 5 folds, train the model sequentially with 4 folds, and predict the remaining fold. Repeat this process 5 times to obtain the complete prediction result for the training set. The full base model is used to predict the enhanced feature matrix of the test set, generating preliminary predictions. Feature importance scores are calculated using the attention weights of the Transformer model. These attention weights are then integrated to obtain the importance score for each feature. The predicted values ​​from the base model and the feature importance scores are input into an adaptive weighted ridge regression model, and the meta-model is trained using the target variable from the original training set as the label. The meta-model continuously adjusts the weight coefficients to make the final prediction results approximate the actual values.

[0041] 5) Based on the above, complete the model run and obtain the weight coefficients of the gradient boosting model and the XGBoost model: 0.6 and 0.4, respectively. The final predicted value = (0.6 × gradient boosting regression predicted value + 0.4 × XGBoost predicted value) × Transformer feature importance weight.

[0042] 6) Calculate the coefficient of determination R², mean square error (MSE), and mean absolute error (MAE). The model's evaluation metrics are: coefficient of determination R² = 0.98, mean square error (MSE) = 0.05, and mean absolute error (MAE) = 0.09, indicating that the constructed calcium sulfate whisker-modified asphalt fatigue life prediction model has high prediction accuracy and stability.

[0043] In this optional embodiment, the transformed feature parameter set is divided into a training set and a test set according to a preset ratio, and the training set and test set are standardized using a standardization method to obtain the processed training set and test set, which include: The transformed feature parameter set is divided into a training set and a test set according to a preset ratio; Calculate the mean and standard deviation of the training set, and then use the mean and standard deviation to standardize the training set according to the standard deviation standardization formula to obtain the standardized training set data; Based on the mean and standard deviation of the training set, the test set is standardized according to the standard deviation standardization formula to obtain the standardized training set data.

[0044] In this optional embodiment, the feature data in the processed training set is embedded and positionally encoded, and then input into a deep learning model for training. The output of the feature data encoding layer is extracted as an enhanced feature representation and feature importance score, including: The feature data in the processed training set is embedded to obtain the embedding vector, and the embedding layer dimension of the deep learning model is set. The position encoding matrix is ​​generated based on the formulas for sine and cosine functions. The position encoding matrix is ​​added to the embedding vector and then input into the deep learning model. A multi-head self-attention mechanism with several attention heads is set up. The objective function is to minimize the mean squared error between the predicted and true values. The deep learning model is trained using adaptive moment estimation, and the output of the feature data encoding layer is extracted as an enhanced feature representation and feature importance score.

[0045] In this optional embodiment, based on the enhanced feature representation, a gradient boosting regression model and an extreme gradient boosting regression model are trained using five-fold cross-validation to obtain the preliminary prediction results of the base model, including: Select the gradient boosting regression model and the extreme gradient boosting regression model as the base models, and initialize the base models according to the preset parameters; The training set is divided into five folds using the five-fold cross-validation method. The base model is trained using four folds in sequence, and predictions are made for the remaining fold. After repeating the loop several times (5 times), the predicted values ​​of the training set on each base model are obtained to complete the training of the base model. The trained base model is used to predict the enhanced feature matrix of the test set, generating preliminary prediction results.

[0046] In this optional embodiment, the preliminary prediction results and feature importance scores are input into an adaptive weighted ridge regression meta-model for training and fusion to obtain a modified asphalt fatigue life prediction model, including: The predicted values ​​of the base model and the feature importance scores are input into the adaptive weighted ridge regression meta-model, and the adaptive weighted ridge regression meta-model is trained with the target variable as the label to obtain the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model. Using a search algorithm, the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model are optimized to determine the optimal weight coefficients. Based on the optimal weighting coefficients, a fatigue life prediction model for modified asphalt is determined, and the performance of the modified asphalt fatigue life prediction model is evaluated on an independent validation set to obtain the final modified asphalt fatigue life prediction model.

[0047] In this optional embodiment, a search algorithm is used to optimize the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model, and the optimal weight coefficients are determined by: Initialize the parameters of the search algorithm, set the maximum number of iterations, define the optimization space of the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model, and configure a set of preset candidate values ​​for each weight coefficient; Establish the traversal order of the weight coefficient combinations and use a dynamic adjustment algorithm to select the current weight coefficient combination to be evaluated from the parameter space; The predicted values ​​of the base model are weighted and summed using the current combination of weight coefficients, and then multiplied by the feature importance score to obtain the predicted values ​​of the validation set samples. Calculate the mean square error between the predicted values ​​and the actual values ​​of the validation set samples, and use this mean square error as the performance evaluation index of the current weight coefficient combination. Repeat the process of selecting, calculating, and evaluating the weight coefficient combination. If the maximum number of iterations is reached, the weight coefficient combination with the smallest mean square error is output as the optimal weight coefficient.

[0048] Specifically, the search algorithm is a label propagation algorithm, a community detection and node classification algorithm based on graph semi-supervised learning. It iteratively propagates label information through the similarity between nodes until convergence. In this invention, this algorithm is applied to optimize the weight coefficient combination of gradient boosting regression and extreme gradient boosting regression models: the weight coefficient combination is regarded as a node in the graph, and its performance evaluation value is used as the initial label. A relationship network is constructed by defining the similarity between weight coefficient combinations, and the label propagation mechanism is used to explore in the coefficient space. The algorithm iteratively updates the performance labels of each weight coefficient combination, ultimately guiding the search process to quickly converge to the optimal weight coefficient combination configuration region, thereby replacing traditional grid or random search and achieving more efficient adaptive weight coefficient optimization.

[0049] In this optional embodiment, establishing the traversal order of the weight coefficient combinations and selecting the current weight coefficient combination to be evaluated from the parameter space using a dynamic adjustment algorithm includes: Initialize the parameters of the dynamically adjusted algorithm and set the maximum number of iterations; Initialize the weight coefficient search space, generate multiple weight coefficient combinations as candidate solutions, calculate the initial performance evaluation value of each weight coefficient combination and store it in the memory bank; The first random number is generated randomly. If the first random number is less than the preset memory selection probability, a weight coefficient combination is randomly selected from the memory. Otherwise, a new weight coefficient combination is generated randomly. A second random number is generated. If the second random number is less than the preset coefficient adjustment probability, the selected weight coefficient combination is mutated and adjusted to obtain a new weight coefficient combination; otherwise, the original weight coefficient combination is retained. Calculate the performance evaluation value of the new weight coefficient combination. If the evaluation value is greater than the worst evaluation value in the memory, replace it; otherwise, keep the memory unchanged. If the maximum number of iterations is reached, the optimal combination of weight coefficients is output as the current combination of weight coefficients to be evaluated; otherwise, the iteration continues.

[0050] Specifically, the dynamic adjustment algorithm is a harmony search algorithm, a metaheuristic optimization algorithm that simulates the improvisational process of musicians in musical performance. In this invention, the algorithm stores historically optimal weight coefficient combinations in a memory bank (analogous to a musician's memory), selects new weight coefficient combinations from the memory bank with a preset probability or generates them randomly (analogous to improvisation), and generates new candidate weight coefficient combinations through probabilistic pitch adjustment (i.e., mutation operation). After each iteration, the memory bank is updated according to fitness, maintaining a balance between exploration and development in the search process, ultimately guiding the weight coefficient optimization to quickly converge to the high-performance configuration region, achieving optimization of weight coefficients in gradient boosting regression models and extreme gradient boosting regression models.

[0051] In this optional embodiment, a first random number is generated. If the first random number is less than the preset memory selection probability, a weight coefficient combination is randomly selected from the memory. Otherwise, a new weight coefficient combination is randomly generated, including: Initialize the random number generator and generate the first random number by uniformly sampling within a preset interval; The first random number is compared with the preset memory selection probability to determine whether to select a weight coefficient combination from the memory. If the first random number is less than the preset memory selection probability, then an existing combination of weight coefficients is uniformly and randomly drawn from the memory as the current candidate solution. If the first random number is greater than or equal to the preset memory selection probability, then independent random sampling is performed within the domain of each weight coefficient combination to generate a new weight coefficient combination as a candidate solution.

[0052] In this optional embodiment, a second random number is randomly generated. If the second random number is less than the preset coefficient adjustment probability, the selected weight coefficient combination is mutated and adjusted to obtain a new weight coefficient combination; otherwise, the original weight coefficient combination is retained unchanged. A second random number is generated, and the second random number is compared with the preset coefficient adjustment probability to determine whether the mutation condition is met. If the second random number is less than the preset coefficient adjustment probability, then the genetic algorithm will perform a reverse mutation operation on the currently selected weight coefficient combination to generate a mutated intermediate solution. Boundary checks and parameter repairs are performed on the intermediate solutions to ensure that all weight coefficient combinations fall within the preset feasible range. After the verification is completed, the final new weight coefficient combination is obtained. If the second random number is greater than or equal to the preset coefficient adjustment probability, the original weight coefficient combination is directly retained without change.

[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model, characterized in that, include: Obtain a dataset of fatigue life of modified asphalt, perform one-hot encoding on the feature parameters in the dataset, calculate the correlation between each feature parameter to construct a feature correlation matrix, and obtain a set of feature parameters. Using fatigue life as the target variable, and processing the target variable to complete the transformation of the target variable, we obtain the transformed set of feature parameters. The transformed feature parameter set is divided into a training set and a test set according to a preset ratio, and the training set and test set are standardized using a standardization method to obtain the processed training set and test set. The feature data in the processed training set is embedded and positionally encoded, and then input into a deep learning model for training. The output of the feature data encoding layer is extracted as an enhanced feature representation and feature importance score. Based on the enhanced feature representation, the gradient boosting regression model and the extreme gradient boosting regression model are trained using the five-fold cross-validation method to obtain the preliminary prediction results of the base model. The preliminary prediction results and feature importance scores are input into an adaptive weighted ridge regression meta-model for training and fusion to obtain a modified asphalt fatigue life prediction model.

2. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 1, characterized in that, The transformed feature parameter set is divided into a training set and a test set according to a preset ratio, and the training set and test set are standardized using a standardization method to obtain the processed training set and test set, which include: The transformed feature parameter set is divided into a training set and a test set according to a preset ratio; Calculate the mean and standard deviation of the training set, and then use the mean and standard deviation to standardize the training set according to the standard deviation standardization formula to obtain the standardized training set data; Based on the mean and standard deviation of the training set, the test set is standardized according to the standard deviation standardization formula to obtain the standardized training set data.

3. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 1, characterized in that, The process of embedding and positionally encoding the feature data in the processed training set, and then inputting it into a deep learning model for training, extracting the output of the feature data encoding layer as an enhanced feature representation and feature importance score includes: The feature data in the processed training set is embedded to obtain the embedding vector, and the embedding layer dimension of the deep learning model is set. The position encoding matrix is ​​generated based on the formulas for sine and cosine functions. The position encoding matrix is ​​added to the embedding vector and then input into the deep learning model. A multi-head self-attention mechanism with several attention heads is set up. The objective function is to minimize the mean squared error between the predicted and true values. The deep learning model is trained using adaptive moment estimation, and the output of the feature data encoding layer is extracted as an enhanced feature representation and feature importance score.

4. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 1, characterized in that, The enhanced feature representation is used to train a gradient boosting regression model and an extreme gradient boosting regression model using five-fold cross-validation, respectively. The preliminary prediction results of the base model include: Select the gradient boosting regression model and the extreme gradient boosting regression model as the base models, and initialize the base models according to the preset parameters; The training set is divided into five folds using the five-fold cross-validation method. The base model is trained using four folds in sequence, and predictions are made for the remaining fold. After repeating the loop several times, the predicted values ​​of the training set on each base model are obtained to complete the training of the base model. The trained base model is used to predict the enhanced feature matrix of the test set, generating preliminary prediction results.

5. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 1, characterized in that, The process of inputting the preliminary prediction results and feature importance scores into an adaptive weighted ridge regression meta-model for training and fusion to obtain a modified asphalt fatigue life prediction model includes: The predicted values ​​of the base model and the feature importance scores are input into the adaptive weighted ridge regression meta-model, and the adaptive weighted ridge regression meta-model is trained with the target variable as the label to obtain the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model. Using a search algorithm, the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model are optimized to determine the optimal weight coefficients. Based on the optimal weighting coefficients, a fatigue life prediction model for modified asphalt is determined, and the performance of the modified asphalt fatigue life prediction model is evaluated on an independent validation set to obtain the final modified asphalt fatigue life prediction model.

6. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 5, characterized in that, The process of using a search algorithm to optimize the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model, and determining the optimal weight coefficients, includes: Initialize the parameters of the search algorithm, set the maximum number of iterations, define the optimization space of the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model, and configure a set of preset candidate values ​​for each weight coefficient; Establish the traversal order of the weight coefficient combinations and use a dynamic adjustment algorithm to select the current weight coefficient combination to be evaluated from the parameter space; The predicted values ​​of the base model are weighted and summed using the current combination of weight coefficients, and then multiplied by the feature importance score to obtain the predicted values ​​of the validation set samples. Calculate the mean square error between the predicted values ​​and the actual values ​​of the validation set samples, and use this mean square error as the performance evaluation index of the current weight coefficient combination. Repeat the process of selecting, calculating, and evaluating the weight coefficient combination. If the maximum number of iterations is reached, the weight coefficient combination with the smallest mean square error is output as the optimal weight coefficient.

7. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 6, characterized in that, The step of establishing the traversal order of the weight coefficient combinations and selecting the current weight coefficient combination to be evaluated from the parameter space using a dynamic adjustment algorithm includes: Initialize the parameters of the dynamically adjusted algorithm and set the maximum number of iterations; Initialize the weight coefficient search space, generate multiple weight coefficient combinations as candidate solutions, calculate the initial performance evaluation value of each weight coefficient combination and store it in the memory bank; The first random number is generated randomly. If the first random number is less than the preset memory selection probability, a weight coefficient combination is randomly selected from the memory. Otherwise, a new weight coefficient combination is generated randomly. A second random number is generated. If the second random number is less than the preset coefficient adjustment probability, the selected weight coefficient combination is mutated and adjusted to obtain a new weight coefficient combination; otherwise, the original weight coefficient combination is retained. Calculate the performance evaluation value of the new weight coefficient combination. If the evaluation value is greater than the worst evaluation value in the memory, replace it; otherwise, keep the memory unchanged. If the maximum number of iterations is reached, the optimal combination of weight coefficients is output as the current combination of weight coefficients to be evaluated; otherwise, the iteration continues.

8. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 7, characterized in that, The process of randomly generating the first random number, and if the first random number is less than the preset memory selection probability, then randomly selecting a weight coefficient combination from the memory; otherwise, randomly generating a new weight coefficient combination includes: Initialize the random number generator and generate the first random number by uniformly sampling within a preset interval; The first random number is compared with the preset memory selection probability to determine whether to select a weight coefficient combination from the memory. If the first random number is less than the preset memory selection probability, then an existing combination of weight coefficients is uniformly and randomly drawn from the memory as the current candidate solution. If the first random number is greater than or equal to the preset memory selection probability, then independent random sampling is performed within the domain of each weight coefficient combination to generate a new weight coefficient combination as a candidate solution.

9. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 7, characterized in that, The process of randomly generating a second random number, and if the second random number is less than the preset coefficient adjustment probability, then the selected weight coefficient combination is mutated and adjusted to obtain a new weight coefficient combination; otherwise, the original weight coefficient combination is retained unchanged. This includes: A second random number is generated, and the second random number is compared with the preset coefficient adjustment probability to determine whether the mutation condition is met. If the second random number is less than the preset coefficient adjustment probability, then the genetic algorithm will perform a reverse mutation operation on the currently selected weight coefficient combination to generate a mutated intermediate solution. Boundary checks and parameter repairs are performed on the intermediate solutions to ensure that all weight coefficient combinations fall within the preset feasible range. After the verification is completed, the final new weight coefficient combination is obtained. If the second random number is greater than or equal to the preset coefficient adjustment probability, the original weight coefficient combination is directly retained without change.

10. The method for constructing a multi-source machine learning-driven asphalt fatigue life prediction model according to claim 5, characterized in that, The formula for obtaining the weight coefficients of the gradient boosting regression model and the extreme gradient boosting regression model is as follows: ； In the formula, This represents the weight vector to be learned; These represent the weighting coefficients that indicate the prediction results of the gradient boosting regression model. These represent the weighting coefficients that indicate the prediction results of the extreme gradient boosting regression model; This represents the vector of predicted values ​​from the gradient boosting regression model. This represents the vector of predicted values ​​from an extreme gradient boosting regression model. This indicates that the gradient boosting regression model applies to the first... Predicted values ​​for each training sample; This indicates that the extreme gradient boosting regression model is effective for the first... Predicted values ​​for each training sample; Represents the feature importance score vector; Indicates the first The overall score of feature importance corresponding to each sample; Indicates the first The actual fatigue life of each sample; This indicates the number of training samples.

Images