Method for analyzing and predicting water stability of asphalt mixture

Abstract

The invention provides a method for analyzing and predicting the water stability of an asphalt mixture, which comprises the following steps of: acquiring multi-dimensional service environment, material composition and standard water stability test data, and carrying out standardization and stratified sampling division on the multi-dimensional service environment, the material composition and the standard water stability test data; a context sensing multi-layer perceptron network supporting feature parameter sharing and gradient accumulation is constructed, and the importance weight of each dimension of features under different data distributions is automatically evaluated and normalized. An attention mechanism and a graph convolutional network are introduced, association between features is quantified, and a key high-order interaction relationship is deduced; by means of K-fold cross validation and meta-learning optimization, the generalization ability of the model for new distribution data is improved, prediction accuracy and feature interpretation are improved, and intelligent analysis and proportion design of the water stability of the asphalt mixture are promoted.

Description

Technical Field

[0001] This invention relates to the field of water stability modeling and prediction technology for road materials, and particularly to a method for analyzing and predicting the water stability of asphalt mixtures. Background Technology

[0002] Currently, in the field of water stability prediction for road engineering materials, especially asphalt mixtures, machine learning models based on multivariate regression, support vector machines, and traditional neural networks are widely used to model the relationship between service environment parameters, material composition characteristics, and water stability indicators. Most existing methods rely on static feature weighting or global feature selection, and K-fold cross-validation is the primary evaluation method during model training and validation. Typically, the model assigns fixed weights to all features globally, and cross-validation improves robustness to changes in the training-test split, aiming to achieve stable prediction capabilities for unknown samples. In recent years, some studies have introduced attention mechanisms and adaptive feature selection to improve the interpretability and generalization ability of the models; however, in cross-validation practice, the ability to dynamically perceive changes in data distribution and adjust feature importance in real time remains limited. With the increasing diversity of engineering site environments and material combinations, data distribution exhibits a dynamic trend. Mainstream technologies primarily rely on pre-defined feature importance indices or posterior analysis using feature importance metrics such as SHAP and LIME to address this heterogeneity. However, these methods typically assume static feature contributions, making it difficult to respond to significant differences in potential data distributions across different rounds of cross-validation and various sampling methods. Particularly in the modeling of water stability in asphalt mixtures, the synergistic effects of different service environments (such as temperature difference, humidity, and traffic load) and multiple material parameters can easily lead to a shift in feature dominance, resulting in structural biases in the modeling results. In practical applications, existing methods for predicting the water stability of asphalt mixtures, such as fixed feature subset selection, bagging, and ensemble learning, are suitable for scenarios where features are dominant and do not fluctuate with data batches. However, when modeling complex material systems with strong contextual dynamics, these approaches struggle to accurately identify the truly critical features under various working conditions, impacting the interpretability and generalization ability of the prediction model. Furthermore, while traditional cross-validation effectively reduces overfitting, the model lacks the ability to adaptively weight features for different data distributions during the K-fold process, and it is also difficult to accumulate knowledge transfer across different scenarios or quickly adapt to challenges in multiple scenarios. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method for water stability analysis and prediction of asphalt mixtures.

[0004] The technical solution of this invention is implemented as follows: a method for analyzing and predicting the water stability of asphalt mixtures, comprising: S1: Collect multi-dimensional service environment parameters, material composition characteristics and water stability test data of asphalt mixture samples, generate a standardized dataset and divide it into training set and validation set; S2: Construct a context-aware subnetwork (CAS) based on a multilayer perceptron architecture, connect each hidden layer node through a parameter sharing mechanism, and configure the gradient accumulation buffer of the meta-learning framework MAML. The input of the CAS network receives the feature matrix of the training set, and the output is connected to the normalized weight vector generation module. S3: Input the current training set into the CAS network, perform the differentiable feature importance evaluation operation, and output a dynamic feature weight vector. The value range of each dimension of the weight vector is [0,1] and satisfies Σwi=1, which is used to characterize the difference in contribution of different features under the current data distribution. S4: Perform element-wise multiplication on the training set feature matrix based on the dynamic feature weight vector to generate a weighted feature space, and at the same time construct a feature correlation matrix based on the attention mechanism to quantify the strength of the interaction between feature dimensions. S5: The base model is trained on the weighted feature space using the K-fold cross-validation strategy. The inner loop parameter update and outer loop gradient accumulation are performed. The inner loop adjusts the base model parameters based on the current fold cross-validation loss, and the outer loop corrects the initial parameters of the CAS network according to the accumulated loss, thereby optimizing the cross-scene generalization capability. S6: Calculate the mean and variance of the prediction error of each fold validation set, and determine whether they meet the preset convergence threshold. If they do not meet the threshold, return to S2 to perform parameter iteration and update. If they meet the threshold, output the final prediction model optimized by dynamic feature weighting and its feature importance distribution map.

[0005] The method for analyzing and predicting the water stability of asphalt mixtures provided by this invention has the following beneficial effects: (1) By introducing a differentiable context-aware subnetwork (CAS) and a dynamic feature weighting mechanism, this invention significantly improves the model's adaptability under different data distributions, effectively overcoming the model bias problem caused by static setting of feature weights in traditional cross-validation methods. By embedding CAS into the cross-validation process, normalized dynamic feature weights are generated according to the current data distribution before each training fold, and the input feature space is weighted and scaled, thereby realizing the enhanced expression of dominant features and the active suppression of irrelevant features. This ensures that the base model is always learning and validating under the optimal feature representation, thus greatly improving prediction accuracy and stability. This mechanism breaks through the limitations of "data-blind" modeling, enabling the cross-validation process to have context-aware capabilities, and significantly enhancing the robustness and applicability of the model in heterogeneous data scenarios. (2) Combining the ideas of the meta-learning framework MAML, this invention constructs an outer loop optimization mechanism for multi-task adaptation, enabling the context-aware subnetwork to have rapid generalization ability across data folds, solving the problem of traditional methods requiring parameter readjustment and poor adaptability in new environments. Unlike conventional approaches that only independently process each cross-validation fold, this invention treats each fold as a task with similar priors, updating the base model parameters through the inner loop and backpropagating and optimizing the initial parameters of the CAS network through the outer loop, thus realizing meta-knowledge learning of "how to weight". This two-layer optimization structure enables CAS to quickly converge to the initial state suitable for the current overall data distribution at the start of a new round of K folds, without the need to manually adjust the network structure or hyperparameters for different material combinations or service conditions, greatly improving modeling efficiency and automation. At the same time, the CAS network adopts a parameter sharing and cumulative gradient update strategy, performing backpropagation only once after a complete round of K folds, effectively controlling computational overhead, avoiding the surge in time complexity caused by frequent differential operations, and ensuring the deployability and real-time response capability of the method in practical engineering applications; (3) This invention constructs a closed-loop adaptive weighted cross-validation system (CAS-CV), realizing the collaborative evolution of the entire process from feature expression optimization to model evaluation, and providing a system-level solution for high-precision modeling of complex material systems. Compared with the traditional CV approach that only focuses on the average performance of the model and ignores the differences in data context, CAS-CV, by integrating three core technologies—context awareness, dynamic weighting, and meta-learning—not only improves the predictive consistency of the model under diverse working conditions, but also enhances the interpretability of the entire modeling process. The output dynamic weight vector can intuitively reflect the evolution trend of the contribution of each feature under different environmental conditions, assisting researchers in identifying key influencing factors and their interaction patterns. In addition, the framework has good model independence and modular design, and can be seamlessly integrated into various machine learning base models (such as random forests, support vector machines, neural networks, etc.), and has high scalability. It is particularly suitable for application scenarios such as asphalt mixtures for campus football fields that require efficient connection between on-campus and off-campus resources. It can accurately match the material performance requirements under different climate zones, usage intensities, and maintenance strategies, support the intelligent recommendation and dynamic optimization of material formulations and construction schemes, and promote the integrated development of intelligent construction, management, and maintenance of transportation infrastructure. Attached Figure Description

[0006] Figure 1 This is a flowchart of the asphalt mixture water stability analysis and prediction method of the present invention; Figure 2 This is a sub-flowchart of the asphalt mixture water stability analysis and prediction method of the present invention; Figure 3 This is another sub-flowchart of the asphalt mixture water stability analysis and prediction method of the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this application provides a method for predicting the water stability of asphalt mixtures, specifically including: S1: Collect multi-dimensional service environment parameters, material composition characteristics and water stability test data of asphalt mixture samples, generate a standardized dataset and divide it into training set and validation set; S2: Construct a context-aware subnetwork (CAS) based on a multilayer perceptron architecture, connect each hidden layer node through a parameter sharing mechanism, and configure the gradient accumulation buffer of the meta-learning framework MAML. The input of the CAS network receives the feature matrix of the training set, and the output is connected to the normalized weight vector generation module. S3: Input the current training set into the CAS network, perform the differentiable feature importance evaluation operation, and output a dynamic feature weight vector. The value range of each dimension of the weight vector is [0,1] and satisfies Σwi=1, which is used to characterize the difference in contribution of different features under the current data distribution. S4: Perform element-wise multiplication on the training set feature matrix based on the dynamic feature weight vector to generate a weighted feature space, and at the same time construct a feature correlation matrix based on the attention mechanism to quantify the strength of the interaction between feature dimensions. S5: The base model is trained on the weighted feature space using the K-fold cross-validation strategy. The inner loop parameter update and outer loop gradient accumulation are performed. The inner loop adjusts the base model parameters based on the current fold cross-validation loss, and the outer loop corrects the initial parameters of the CAS network according to the accumulated loss, thereby optimizing the cross-scene generalization capability. S6: Calculate the mean and variance of the prediction error of each fold validation set, and determine whether they meet the preset convergence threshold. If they do not meet the threshold, return to S2 to perform parameter iteration and update. If they meet the threshold, output the final prediction model optimized by dynamic feature weighting and its feature importance distribution map.

[0010] Step S1: Collect multi-dimensional service environment parameters, material composition characteristics, and water stability test data of asphalt mixture samples, generate a standardized dataset, and divide it into training and validation sets. Specifically, this includes: S1.1: Based on the engineering application background of asphalt mixture, multi-dimensional environmental parameters of samples under different service environments are collected. The environmental parameters include temperature gradient, humidity change rate, number of traffic load cycles and number of freeze-thaw cycles. The above parameters are collected synchronously using an environmental simulation test chamber and on-site monitoring equipment to obtain input characteristic data that reflects the actual service conditions. The input conditions are the environmental and working condition parameters of the asphalt mixture sample under long-term service conditions, which are derived from the multi-channel output signals of the field monitoring equipment and the laboratory environmental simulation test chamber, including multiple physical quantities such as temperature, humidity, traffic load, and freeze-thaw cycle. A multi-point temperature acquisition method (parameters: high-precision PT100 platinum resistance thermometer, sampling frequency 1Hz) is used to simultaneously measure the temperature at various depths on the upper and lower surfaces of the sample and to obtain temperature gradient data between different layers. Furthermore, using the difference calculation formula:

[0011] in, This is the temperature value. For the depth of the measuring point, For depth Temperature value at that location, For depth The temperature value at that location is used to generate a temperature gradient matrix that reflects the intensity of heat flow transfer up and down. A capacitive humidity sensor array (parameters: range 0%RH~100%RH, accuracy ±1%RH) is deployed on the sample surface and at key internal locations to achieve real-time acquisition of humidity time-series curves. Furthermore, the humidity change rate was calculated using the sliding window statistical method (window length 3600s), as shown in the following formula:

[0012] in, This is the humidity value. For time intervals, For time The humidity value at that location For time The humidity value at that location is used to obtain a humidity change rate sequence; A traffic load simulation platform (parameters: wheel load force 50kN, loading frequency 1Hz) was used in combination with on-site piezoelectric sensors to collect the number of traffic load cycles. A time-load cycle relationship data matrix was generated by a counting accumulation algorithm to quantify the history of mechanical stress. A low-temperature cyclic control system and a freeze-thaw test chamber (parameters: temperature range) are adopted. Freeze-thaw cycles were performed on the samples at a temperature of 20℃ to +20℃ (with a cycle control accuracy of ±0.5℃), and the freezing point and melting point changes during each cycle were collected using a phase change temperature sensor. Furthermore, the number of freeze-thaw cycles was accumulated using a cycle counter to create a freeze-thaw cycle history table. By using a feature synchronous acquisition controller, temperature gradient, humidity change rate, traffic load cycle data and freeze-thaw cycle data are aligned to a unified timestamp, and a multi-dimensional service environment parameter matrix is ​​established to achieve spatiotemporal consistency of input features. Through the above algorithms and processing methods, the original sensor signals from the previous step are transformed into structured temperature gradient matrices, humidity change rate sequences, traffic load cycle relationship matrices, and freeze-thaw cycle history tables, enabling service environment input feature data that can be directly entered into the standardized dataset generation module. For example, when collecting environmental parameters of asphalt mixture samples in a test section of a highway, the PT100 temperature sensor was configured with three layers: surface, 5cm depth, and 10cm depth, with two points per layer, for a total of six temperature measurement points. The recorded data showed that the temperature difference between the surface and the 10cm depth was 8℃. The temperature gradient was calculated using the formula as follows: =80℃ / m; the humidity sensor array was set with a sampling interval of 10s and a sliding window length of 3600s. The observed humidity increased from 70%RH to 85%RH, and the rate of change was calculated as follows: =0.00417%RH / s; the traffic load simulation table recorded 600 cycles within 10 minutes, and the cycle relationship matrix showed a positive correlation between the number of loading cycles and the strain amplitude in the sample; the freeze-thaw test chamber was set to... A cycle of holding at 15℃ for 3 hours and then heating to 15℃ and holding for 3 hours constitutes one cycle, which is repeated 20 times. The freeze-thaw cycle history is fully recorded. After processing by the synchronous acquisition controller, all four types of data are aligned to a unified timestamp and output in the form of a feature matrix, which is then input into the subsequent material composition acquisition step S1.2, achieving seamless integration from data acquisition to model input. S1.2: Based on materials science and road engineering standards, collect material composition characteristic parameters of asphalt mixture samples. The parameters include asphalt type, asphalt content, aggregate gradation, porosity, asphalt film thickness and mineral powder type. Through laboratory mix design and physicochemical analysis of core samples taken from the field, obtain structured data of material composition characteristics. S1.3: Based on the standard method for water stability testing, the asphalt mixture samples were subjected to immersion Marshall test, freeze-thaw splitting test and surface permeability test. Water stability indicators such as strength retention rate, splitting strength ratio and permeability coefficient of the samples under different water loss were collected to form a target variable dataset. S1.4: Based on the data standardization processing method, the collected service environment parameters, material composition characteristics and water stability test data are processed by missing value filling, outlier removal and normalization. The Z-score standardization algorithm is used to perform unified dimensional transformation on multidimensional features to generate a structured standardized dataset that can be directly used for modeling. S1.5: Based on the K-fold cross-validation data partitioning principle, the standardized dataset is divided into a training set and a validation set. The training set is used to construct the learning process of the context-aware sub-network and the base model, and the validation set is used to evaluate the generalization performance of the model on unknown data. The partitioning adopts a stratified sampling strategy to maintain the consistency of class distribution. Based on the collected and standardized multi-dimensional feature matrix of asphalt mixture and the corresponding water stability label data, the K-fold cross-validation partitioning algorithm (parameter: K value is set according to the sample size of the prediction task and the model complexity, generally 5 to 10) is used to realize the function of splitting the training set and validation set of the dataset. Furthermore, by using a stratified sampling method (parameters: the stratification variable is the water stability label category, and the sample ratio within the stratum is set according to the original data distribution), the sample ratio of the validation set in each category is consistent with the overall dataset, and the indexes of the segmented training and validation subsets are obtained; Furthermore, by utilizing the index mapping mechanism (parameters: index uniqueness and mutual exclusion constraints), a subset is extracted from the standardized dataset to generate a training set feature matrix and corresponding labels for model construction, as well as a validation set feature matrix and label matrix for performance evaluation. Furthermore, the stratified sampling results are fine-tuned by a sample equalization algorithm (parameter: the class sample difference threshold is set to no more than 5 samples) to ensure the balance of the training set and validation set in terms of classes, and to generate the equalized data segmentation results. By using a hierarchical K-fold cross-validation data partitioning method, the standardized feature matrix from the previous step is transformed into training-validation data pairs that can be used in multiple rounds, thereby achieving the effect of generalization performance stability evaluation technology during model training. For example, in the task of predicting the water stability of asphalt mixtures, K=5 is set, and the dataset containing 500 standardized sample records is stratified. The stratification variable is the water stability level corresponding to the porosity, and there are 3 classes, with each class accounting for 0.4, 0.35, and 0.25% of the original dataset, respectively. During the stratification sampling stage, the proportion of each class in the training set and the validation set is kept constant, that is, the number of samples of each class in the training set is 160, 140, and 100, respectively, and the number of samples of each class in the validation set is 40, 35, and 25, respectively. An index mapping mechanism is applied to generate the data partitioning index for each layer, ensuring that the training set and validation set samples are mutually exclusive between different layers. For the training set of a certain layer, the difference between class samples is calculated, and it is found that the maximum difference in sample size between classes is 4, which is lower than the threshold of 5, so no additional equalization processing is required. The training set feature matrix and labels of this fold are used as inputs to the CAS network and base model, and the model parameters are updated in subsequent cross-validation iterations. The validation set feature matrix and labels are used as performance evaluation data, and the mean squared error of each fold is calculated using the following formula:

[0013] in, To verify the true label values ​​of the set, To predict the output value for the model, The number of samples in the validation set is specified. The 5-fold dataset generated according to this process supports a significant improvement in generalization performance in subsequent steps of this task.

[0014] Step S2: Construct a Context-Aware Subnetwork (CAS) based on a multilayer perceptron architecture, connect the hidden layer nodes through a parameter sharing mechanism, and configure the gradient accumulation buffer of the meta-learning framework MAML. The CAS network receives the training set feature matrix at its input and connects to a normalized weight vector generation module at its output. Specifically, this includes: S2.1: Based on the multidimensional feature matrix of the asphalt mixture water stability prediction task, a context-aware subnetwork (CAS) with a multilayer perceptron structure is constructed. The number of nodes in the input layer matches the feature dimension, and the hidden layer uses the ReLU activation function for nonlinear mapping to extract high-order interaction relationships between features. Based on the multi-dimensional feature matrix formed by the asphalt mixture water stability prediction task, the input end is loaded with the standardized training set feature data generated by step S1 to ensure that each feature dimension has a unified dimension and a denoising state. A context-aware subnetwork (CAS) is constructed using a multilayer perceptron (MLP) structure (parameters: number of input layer nodes = feature dimension n, number of hidden layers ≥ 3, node number decreasing strategy from the first hidden layer to the output layer n×2, n, n / 2 respectively), to realize the high-order nonlinear mapping function of input features and provide deep feature representation for subsequent dynamic feature weight calculation; Furthermore, by introducing the Rectified Linear Unit (ReLU) activation function (parameter: f(x)=max(0,x)) into the computation of each hidden layer node, a nonlinear mapping is performed on the initial linear transformation result, thereby enhancing the ability of the feature space to express the coupling relationship between complex environment and material, and ensuring the stability of gradient propagation; Furthermore, the Xavier initialization strategy (parameters: mean weight = 0, variance = 2 / (number of input nodes + number of output nodes)) is used to initialize the connection weights from the input layer to the first hidden layer, thereby reducing the risk of gradient vanishing or exploding and improving the training convergence speed. Furthermore, the activation output of each layer is centered and variance standardized by using the batch normalization method, so that the network maintains the consistency of activation distribution under different batches of training data, thereby improving the performance stability of the CAS structure in the cross-fold verification process. Through the above-mentioned multilayer perceptron construction and nonlinear mapping processing, the original multidimensional feature matrix is ​​transformed into a high-order interactive feature representation, which forms the structural basis for dynamic evaluation of the feature importance of asphalt mixtures under different service environments. For example, in a certain engineering application scenario, the input feature matrix has a dimension of 60, including temperature gradient, humidity change rate, traffic load cycle count, freeze-thaw cycle count, asphalt content, aggregate gradation, porosity, etc. The CAS network has 60 nodes in the input layer, 120 nodes in the first hidden layer, 60 nodes in the second hidden layer, 30 nodes in the third hidden layer, and 60 nodes in the output layer. The activation function for the hidden layers is set to ReLU, which is calculated as follows: The weights are initialized using the Xavier method. For the first-layer weight matrix, the variance is calculated using the following formula: Where n=60 and m=120, the initial variance is obtained as follows: ≈0.0111. Batch normalization is performed on the output of each layer to ensure that the mean of each feature within the batch is approximately 0.0111. variance is Under this configuration, the intermediate feature representation vector output by the CAS network can effectively capture the nonlinear interaction between temperature gradient and humidity change rate, significantly improving the ability of the subsequent dynamic feature weight calculation module to identify key features, and showing a significant improvement in generalization performance in validation across datasets with different material ratios; S2.2: Implement a parameter sharing mechanism among the hidden layers of the CAS network. By sharing the weight matrix and bias vector, the consistency of feature importance evaluation in the cross-fold verification process is achieved, thereby reducing model parameter redundancy and improving computational efficiency. The context-aware sub-network based on the constructed multilayer perceptron structure takes as input the asphalt mixture training set feature matrix after standardization in step S1, with each input layer matching the feature dimension. The hidden layers have been nonlinearly mapped using the ReLU activation function. This sub-step, located in the main step S2, is used to implement a parameter sharing mechanism between different hidden layers, ensuring consistency in feature importance evaluation during cross-fold verification. A weight matrix sharing method is adopted (parameter: the dimension of the weight matrix matches the number of adjacent hidden layer nodes) to achieve a globally unified configuration of connection parameters between hidden layers, thereby reducing cross-fold parameter redundancy and improving solution efficiency; Furthermore, by using the bias vector sharing method (parameter: the bias vector length is consistent with the number of hidden layer nodes), the consistency of offsets across hidden layers is achieved, so that under different data distribution conditions, the feature importance calculation is based on the same bias benchmark, which greatly reduces the weight fluctuation caused by bias differences. Furthermore, a regularization constraint method is adopted (parameter: L2 regularization coefficient λ is set to 0.001). Based on the shared weights and biases, a smoothing constraint is applied to the parameter space to control the range of weight variation under different tasks, thereby stabilizing the feature importance scoring results. Furthermore, a gradient aggregation method is applied (parameters: the aggregation method is weighted average, and the weight factor is set according to the number of training samples in each fold). During the K-fold cross-validation process, the gradient information on different folds is uniformly applied to the shared parameters to ensure the global consistency of the parameter optimization direction and avoid overfitting caused by single-fold data bias. By using the above parameter sharing, regularization, and gradient aggregation processing methods, the multilayer perceptron mapping result of the previous step is transformed into a cross-fold consistent feature representation, which enables the feature importance assessment to remain stable under changes in data distribution, thereby achieving the expected technical effect of reducing computational overhead and improving the model's generalization ability. For example, in a practical asphalt mixture water stability prediction task, the training set feature matrix contains 20-dimensional input features. The CAS network is constructed as a three-layer hidden layer structure with 64, 32, and 16 nodes respectively. The shared weight matrix has a dimension of [missing information - likely a missing element]. The dimension between the second and third hidden layers is The bias vector lengths are respectively and Using L2 regularization coefficients Gradient information is accumulated during each fold validation process and aggregated proportionally to the number of samples. For example, the number of samples in the first fold is... The second fold is The gradient aggregation weight factors are respectively and In this scenario, the feature importance scores, after parameter sharing and aggregation, show significantly reduced fluctuations across folds, and the model's mean squared error on the validation set stabilizes at [value missing]. Within this range, it significantly improved the stability and generalization ability of predictions; S2.3: Based on the meta-learning framework MAML, a gradient accumulation buffer is configured in the CAS network to temporarily store the gradient information of each fold during the K-fold cross-validation process, so that parameter updates can be performed uniformly after a complete round of validation, thereby improving the model's ability to quickly adapt to new tasks. The hidden layer weight matrix and bias vector synchronization information of the multilayer perceptron output from step S2.2 are received as input conditions, and the gradient accumulation buffer structure of the meta-learning framework MAML is configured to realize the cross-fold gradient temporary storage function. A cache mapping table is used to establish the association between the K-fold cross-validation index and the gradient storage area (parameter: the index is the fold number, and the storage area size is dynamically allocated according to the feature dimension and the number of hidden layer nodes), so as to realize the ordered storage and retrieval of gradient data between different folds; Furthermore, through the gradient capture module (algorithm: backpropagation gradient interception mechanism, parameters: learning rate and gradient cutoff threshold), all differentiable gradient information from the feature weighting module to the output layer in each training phase is extracted and categorized and stored according to the feature dimension; Furthermore, the gradient accumulation calculation method is adopted (formula as follows):

[0015] in, This represents the gradient vector at the i-th fold. Indicates the number of folds in cross-validation. This is the accumulated average gradient vector, used for subsequent global parameter updates; Furthermore, by calling the cumulative buffer through the MAML outer loop optimizer, a global parameter update is performed (parameter: update step size is set to ), in order to adjust the initial parameters of the CAS network and ensure the model's rapid adaptation characteristics when accepting new tasks; By configuring and using the gradient accumulation buffer as described above, the scattered gradient information during each training process is integrated into a unified update signal, thereby achieving dynamic optimization of the cross-scene feature weight generation strategy. For example, in the training of the asphalt mixture water stability prediction model, a K=5-fold cross-validation buffer is configured, with the capacity of a single storage area in the buffer set to 500×128 floating-point units to match the gradient matrix size of 500 samples and 128-dimensional hidden layer nodes. After each fold of training, the gradient vector captured during backpropagation (j is the fold number) is written to the corresponding storage area. The gradient truncation threshold is set to remove numerical noise, and the learning rate is set to ensure gradient smoothness. After completing the 5-fold validation, according to...

[0016] The cumulative average gradient is calculated, and the MAML outer loop optimizer is invoked to update the initial weights of the CAS network with a step size. The results show that the model converges in fewer iterations under the new environmental parameter distribution, the prediction stability is significantly improved, and the dynamic weight allocation of key feature dimensions is more in line with actual service conditions. S2.4: The training set feature matrix is ​​used as input to the input layer of the CAS network. After multiple nonlinear transformations, an intermediate feature representation is generated, which serves as the basis for dynamic feature weight calculation. S2.5: Introduce the Softmax normalization function at the output of the CAS network to normalize the intermediate feature representation and generate a set of dynamic feature weight vectors, where the value range of each dimension is [0,1] and satisfies Σwi=1, so as to represent the relative contribution of each feature under the current data distribution.

[0017] like Figure 2 As shown, step S3 involves inputting the current training set into the CAS network, performing a differentiable feature importance evaluation operation, and outputting a dynamic feature weight vector. The weight vector has a value range of [0,1] for each dimension and satisfies Σwi=1, used to characterize the difference in contribution of different features under the current data distribution. Specifically, this includes: S3.1: The Context-Aware Subnetwork (CAS) built on the multilayer perceptron architecture receives the original feature matrix of the current training set as input, and enters the first hidden layer in the form of feature vectors for linear transformation to obtain the initial mapping result of the feature space; Receive the original feature matrix of the current training set, and use batch data loading and memory mapping method (parameters: batch size B, memory page size P) to achieve efficient data reading and matrix construction; Linear transformation algorithm is used (parameter: weight matrix) Bias vector The input d-dimensional feature vector is mapped to the first hidden layer. The feature space is divided into 3D, and an initial mapping matrix is ​​obtained. The linear transformation is then performed by matrix multiplication and vector addition. Furthermore, the initial mapping result is calculated using matrix operation formulas. :

[0018] in, To input feature matrices in batches, For the mapping weight matrix, It is the bias vector; Furthermore, by using the batch normalization method (parameters: mean vector μ, variance vector σ², scaling factor γ, translation factor β), the statistics of each dimension of the initial mapping matrix are normalized, effectively eliminating the influence of feature distribution offset on downstream modeling, and obtaining the normalized mapping matrix. Furthermore, a weight regularization method (parameter: L2 norm coefficient λ) is applied to the weight matrix. Norm penalty is applied to reduce the risk of overfitting and maintain the stability of feature importance calculation; Through the above algorithm, the input feature matrix of the previous step is transformed into the initial mapping result of the first hidden layer, realizing the linear embedding of the original features and the transformation of the basic space. For example, in the task of predicting the water stability of asphalt mixtures, the input feature matrix X has a dimension of 64×256 (i.e., 64 samples, each with 256 standardized features), the batch size B is set to 16, and the memory page size P is set to 64KB. Weight matrix The dimension is set to 256×128, and the bias vector is... The dimension is 128, and the initial mapping result is calculated by matrix multiplication. When batch normalization is used, the mean vector μ and variance vector σ² are obtained statistically from the batch samples, and the scaling factor γ and translation factor β are updated iteratively through gradient descent. The weight regularization coefficient λ is set to 0.001 to ensure a smooth distribution of feature embedding weights. After performing this step, the numerical range of the initial mapping matrix output by the first hidden layer is concentrated in [-0.8, 0.9], which effectively suppresses scale imbalance in the input features and improves the stability of the subsequent feature importance calculation module. S3.2: The initial mapping result is subjected to a nonlinear transformation using the ReLU activation function to generate the activation output of the first hidden layer, which is used as the input of the next hidden layer to enhance the nonlinear modeling capability of feature representation; The initial mapping result after linear transformation of the current training set is used as the input tensor. A modified element-wise nonlinear activation module is introduced, employing the ReLU (Rectified Linear Unit) activation function (parameter: threshold = 0) to implement positive half-wave rectification for each element in the input tensor. Perform nonlinear mapping , thus obtaining the activated tensor; Furthermore, by employing a sparsity retention strategy (parameter: activation ratio retention threshold set to 0.3), the zero values ​​corresponding to most negative values ​​in the ReLU output are masked to reduce redundant calculations during gradient propagation and maintain the sparse structure of features to match the characteristics of asphalt mixtures with high feature dimensionality and significant correlation differences. Furthermore, by adding batch normalization processing after the nonlinear transformation (parameters: mean and variance are estimated from the current batch of data), the distribution of each dimension of the activated output is standardized to zero mean and unit variance, as shown in the formula:

[0019] in For batch normalized output values, This is the batch average. To measure variance, To prevent division by zero of small constants; this treatment can significantly improve the convergence speed of gradient descent and stabilize the training process; Furthermore, by introducing the Dropout random deactivation mechanism (parameter: deactivation probability set to 0.2), the normalized activation output is randomly set to zero, which suppresses the overfitting dependence between features, improves the generalization ability of subsequent hidden layers to receive input, and introduces perturbation into the feature importance gradient calculation process to enhance robustness. By using the ReLU nonlinear activation and subsequent batch normalization and Dropout processing, the initial mapping result after linear transformation is converted into the first hidden layer activation output with nonlinear expressive power and stable distribution, thereby improving the diversity of feature expression and the efficiency and stability of gradient propagation. For example, in the asphalt mixture water stability prediction experiment, the initial mapping result matrix of the current training set has a dimension of 128×64, and ReLU activation is performed on each element. The resulting activation matrix has a negative element ratio of approximately 65%. Due to the sparsity preservation strategy, only about 35% of the positive activation values ​​are retained for subsequent calculations. During batch normalization, the mean μ is in the range of [-0.05, 0.04], and the variance σ² is between [0.92, 1.08]. After normalization, the data distribution approximates zero mean and unit variance, and the gradient update stability on the validation set shows that the gradient oscillation amplitude is reduced to about half of its original value. Dropout processing randomly zeros approximately 1638 elements in the activation matrix with a deactivation probability of 0.2, effectively reducing the accidental correlation between features. In subsequent feature importance gradient calculation steps, the activation output of this first hidden layer exhibits a more pronounced key feature highlighting effect, significantly improving the model's stability under complex data distributions. The prediction performance shows a significant improvement in mean squared error compared to the baseline without this nonlinear transformation. S3.3: Introduce a differentiable feature importance calculation module in the intermediate hidden layer of the multilayer perceptron. Based on the label information of the current input sample, calculate the gradient information between each feature dimension and the target variable to obtain the local sensitivity of the feature to the output prediction. S3.4: Based on the local sensitivity information and the feature representations output by each hidden layer, the importance scores of each feature dimension are normalized using the Softmax normalization function to generate feature weight vectors, ensuring that the value range of each dimension is [0,1] and satisfies the following conditions: ; S3.5: Output the dynamic feature weight vector to the next processing module for subsequent weighted scaling of the feature matrix to achieve dynamic adjustment of the feature space and enhanced expression of key features.

[0020] like Figure 3 As shown, step S4 involves performing element-wise multiplication on the training set feature matrix based on the dynamic feature weight vector to generate a weighted feature space. Simultaneously, a feature correlation matrix based on an attention mechanism is constructed to quantify the strength of interactions between feature dimensions. Specifically, this includes: S4.1: Based on the dynamic feature weight vector, perform element-wise multiplication on the original feature matrix of the current training set. Each column of the feature matrix's eigenvector is multiplied by its corresponding weight value using a scalar multiplication operation to generate a weighted feature matrix. This operation enhances the numerical representation of key features and suppresses noise interference from insignificant features, resulting in a more discriminative feature space. The output weighted feature matrix is ​​then used for subsequent modeling processing. Based on the original feature matrix of the current training set and the generated dynamic feature weight vector, the element-wise multiplication operation method is adopted to perform scalar multiplication on each column of feature vectors according to the corresponding weight value. Furthermore, through a matrix broadcasting mechanism (parameter: broadcasting dimension matching n columns of features), the weight vector is repeatedly expanded in the sample rows, and multiplication operations are performed simultaneously on each column of the original feature matrix at the numerical level to obtain a weighted feature matrix; Furthermore, the element-wise product is calculated using the following formula:

[0021] in, For sample index, Indexed by feature dimensions, The element in the i-th row and j-th column of the weighted feature matrix, The value of the element in the i-th row and j-th column of the original feature matrix. The normalized weights of the current feature weight vector in the j-th dimension; Furthermore, floating-point numerical representations are used for weights and eigenvalues, and double-precision calculations are performed to reduce accumulated calculation errors and ensure the accuracy of the weighted results. Furthermore, the element-wise multiplication process is accelerated through a vectorized operation interface, which significantly reduces the loop iteration overhead while keeping the weight-feature mapping relationship unchanged, and the weighted feature matrix G is passed to the subsequent feature space standardization module. By using an element-wise multiplication algorithm, the dynamic feature weight vector from the previous step is transformed into a quantized weighted feature matrix, thereby achieving the expected technical effects of enhancing the expression of key features and suppressing the numerical values ​​of insignificant features. For example, in the task of predicting the water stability of asphalt mixtures, the input original feature matrix F contains 300 samples, each with 10 feature dimensions (temperature gradient, humidity change rate, number of traffic load cycles, number of freeze-thaw cycles, asphalt type code, asphalt content, aggregate gradation parameters, porosity, asphalt film thickness, and mineral powder type code). The dynamic feature weight vector w output by the CAS network is [0.18, 0.12, 0.15, 0.10, 0.05, 0.09, 0.08, 0.07, 0.11, 0.05]. Matrix broadcasting is used to expand the weight vector into a 300×10 matrix, and element-wise multiplication is performed to obtain the weighted feature matrix G. Taking the 3rd feature dimension of the 15th sample as an example (feature = 25.7), its corresponding weight w = 0.15, the weighted result is... = After processing all samples and feature dimensions in this way, the resulting weighted feature matrix shows a significant increase in the value of high-weight dimensions and a significant decrease in the value of low-weight dimensions. After subsequent Z-score standardization, it is used to construct the feature correlation matrix through the attention mechanism, ultimately improving the base model's ability to identify and stably predict key features. S4.2: Perform normalization processing on the weighted feature matrix, and use the Z-score standardization method to center and unify the variance of each feature dimension to eliminate the influence of differences in feature scales on model training, and obtain the standardized weighted feature space to provide a unified scale input for subsequent attention mechanism modeling; S4.3: Based on the multi-head attention mechanism in the Transformer architecture, an attention weight matrix is ​​constructed between feature dimensions. This matrix obtains the correlation strength between each feature dimension by calculating the query-key similarity of feature vectors in the weighted feature space, forming a feature correlation matrix, which is used to characterize the interaction patterns between different features; Based on the standardized weighted feature space matrix as input, the query-key-value structure in the multi-head attention mechanism is adopted (parameters: number of multi-heads h, query vector dimension). Key vector dimension Value vector dimension This enables distributed computation of the correlation strength between feature dimensions; Furthermore, through the linear mapping method (parameter: weight matrix) , , Each column of feature vectors in the weighted feature space is mapped to a query matrix Q, a key matrix K, and a value matrix V, respectively, for subsequent similarity calculation and weighted summation; Furthermore, a scaled dot product attention algorithm is employed (parameter: scaling factor). ), perform matrix multiplication on Q and K to obtain the original similarity matrix; Furthermore, using the Softmax normalization function (parameter: temperature coefficient τ), each row of the similarity matrix is ​​normalized into a probability distribution to ensure the non-negativity of the relevance weights and to satisfy the normalization constraints. We obtain the attention weight matrix A; Furthermore, through multi-head attention connection operation (parameter: h independent attention heads), the weighted sum of h groups of attention weight matrices and value matrices is calculated in parallel, and then linearly concatenated and projected to generate a feature correlation matrix R, which is used to comprehensively characterize the interaction patterns between various feature dimensions. By using a multi-head attention algorithm based on the Transformer architecture, the weighted feature space of the previous step is transformed into a structured feature correlation matrix, thereby achieving explicit quantization of high-order interaction patterns between features. For example, when selecting asphalt mixture feature data with a weighted feature space matrix size of 100×64, the number of heads h is set to 8 in the multi-head attention mechanism, and the dimensions of the query, key, and value are all set to 64. A linear mapping is performed on each column of the feature matrix, resulting in Q, K, and V with dimensions of 100×64. Scaling dot product attention is then applied, with a scaling factor of [missing value]. The initial similarity matrix was calculated to be 100×100. Softmax normalization was applied to the similarity matrix to obtain matrix A, where each row element has a value range of [0,1] and the sum of the row vector elements equals 1. The weighted matrix of eight attention heads was calculated in parallel, concatenated to form a 100×512 matrix, and then linearly projected back to 100×64 as the feature correlation matrix R. Verification showed that the generated R effectively reveals the strong correlation between temperature gradient features and humidity change rate features, and highlights the nonlinear interaction pattern between mineral powder type and freeze-thaw cycle number, significantly improving the discriminative ability of subsequent feature selection and interaction modeling. S4.4: Perform sparsification on the feature correlation matrix and use the Top-K threshold filtering method to retain the K feature dimensions with the highest correlation strength in each row to reduce the interference of redundant information on model training. Output the sparsified feature correlation map as the basis for feature selection and interactive modeling in subsequent model training. Based on the feature correlation matrix data output by S4.3, a sparsity filtering algorithm (parameter: threshold strategy is Top-K, K value is determined based on feature redundancy analysis) is used to reduce redundant feature connections in the correlation matrix. Furthermore, by using a row-level scanning method (parameter: number of features per row = feature space dimension), the correlation strength is sorted row by row, and an index vector is generated to locate the feature dimension with the highest correlation. Furthermore, the Top-K retention algorithm (parameter: K = number of key interaction features) is adopted to retain the feature connection states of the top K positions in each row with the highest correlation strength, and set the remaining feature connection states to zero to generate a binary mask matrix. Furthermore, the binary mask matrix is ​​multiplied element-wise with the original feature correlation matrix to obtain a sparse correlation matrix, which removes low-correlation feature connections and retains the strength of high-correlation connections. By using a matrix structure compression algorithm (parameter: compression ratio is determined based on the proportion of non-zero elements), the sparse matrix is ​​transformed into a data structure in the form of an adjacency list, thereby enabling the output of the sparsed feature correlation map. By using a sparsification algorithm, the dense correlation matrix result from the previous step is transformed into sparse feature correlation map data, thereby achieving the expected technical effects of reducing feature redundancy and increasing the significance of interaction relationships. For example, in the task of predicting the water stability of asphalt mixtures, the feature space has a dimension of 50, and the feature correlation matrix is ​​generated by S4.3, with elements ranging from 0 to 1. The Top-K parameter K is set to 5. Based on the redundancy analysis results, only the connections between each feature dimension and the 5 features with the highest correlation values ​​are retained. A row-level scanning method is used to sort the correlation values ​​of each row in descending order, resulting in an index vector. For example, if a feature row is sorted as [0.92, 0.87, 0.83, 0.80, 0.77, ...], the corresponding feature positions are [12, 7, 25, 3, 41, ...]. A binary mask matrix is ​​generated, assigning a value of 1 to the top 5 index positions and 0 to the remaining positions. This mask matrix is ​​then multiplied element-wise by the original correlation matrix to obtain a sparsified correlation matrix. During compression, the sparse matrix is ​​converted to an adjacency list format, with the proportion of non-zero elements approximately 10% of the original, significantly improving storage and computational efficiency. The final output feature correlation map is an adjacency list data structure, which can significantly improve the speed of feature interaction modeling and reduce noise interference in the subsequent GCN modeling process, and greatly improve the stability of model performance. S4.5: Based on the sparse feature correlation map, the weighted feature space is modeled with a graph structure. The graph convolutional network (GCN) is used to model the high-order interaction relationship between feature nodes, extract the high-order nonlinear correlation between features, and generate an enhanced feature representation vector to improve the feature expression ability and prediction stability of the base model in the subsequent cross-validation process.

[0022] Step S5: A base model is trained on the weighted feature space using a K-fold cross-validation strategy. This involves updating parameters in the inner loop and accumulating gradients in the outer loop. The inner loop adjusts the base model parameters based on the current fold cross-validation loss, while the outer loop corrects the initial parameters of the CAS network based on the accumulated loss, thereby optimizing cross-scenario generalization capabilities. Specifically, this includes: S5.1: Construct a K-fold cross-validation task partitioning structure based on the weighted feature space. Use the K-fold partitioning algorithm to perform non-repeating random partitioning of the training set samples to generate K mutually exclusive training-validation subset pairs for subsequent multi-round model evaluation and parameter updates. In the weighted feature space, a standardized training sample matrix after feature weighting and correlation sparsification is obtained as the data input benchmark for K-fold cross-validation. The K-fold partitioning algorithm (parameters: K value, random seed, stratified sampling flag) is used to divide the training sample index set into K mutually exclusive subsets while maintaining the consistency of the target variable category distribution, ensuring the statistical consistency of each subset in feature distribution. Furthermore, by employing a non-repeating random partitioning strategy (parameters: sampling ratio, random number generation method), the training set and validation set of each fold are mutually exclusive and cover all samples, ensuring that each sample serves as a validation sample for model evaluation exactly once throughout the entire K-fold process. Furthermore, based on the stratified sampling mechanism (parameters: feature category label, number of strata), the proportion of samples with different material combinations and different service environment categories in each training-validation subset is kept consistent with the proportion of the overall dataset during the partitioning process, so as to improve the fairness and stability of model evaluation; Furthermore, after completing the sample partitioning, a training-validation subset index mapping table is generated to record the training sample number set and validation sample number set corresponding to each fold, providing accurate data positioning for subsequent inner loop parameter updates and outer loop gradient accumulation; By using the above-mentioned K-fold partitioning and index mapping table generation process, the weighted feature space is mapped to K sets of structured training-validation datasets, thereby achieving the repeatability and statistical balance of data partitioning during cross-validation. For example, a standardized training sample matrix in the weighted feature space is selected, with a total of 1200 samples and a feature dimension of 50. The enhanced feature matrix formed after sparse correlation graph modeling is used as input. K is set to 5, the random seed to 2023, and the stratified sampling flag is set to true. Stratification is based on a dual-layer label of material combination category and service environment category. When executing the K-fold partitioning algorithm, a non-repeating random number generation method is used to ensure that each sample enters the validation set only once. The partitioning result generates a training-validation subset index mapping table. For example, the first fold training set contains 960 samples, and the validation set contains 240 samples, with the proportion of material category and environment category in the validation set consistent with the full dataset. This partitioning result is input into the inner and outer loops of cross-validation, verifying the balance between training and evaluation within different folds and improving the stability of the model's generalization performance. In the specific implementation, the parameter sampling ratio of the K-fold partitioning function is set to 80% for the training set and 20% for the validation set. The Mason tween algorithm is used for random number generation to ensure the randomness and reproducibility of the partitioning. S5.2: In each fold cross-validation, the inner loop parameter update operation is performed on the base model based on the current training subset, the gradient descent algorithm is used to calculate the current fold training loss, and the base model parameters are adjusted by backpropagation to improve the model's fitting ability under the current feature distribution; In the current cross-validation task, the training subset and its weighted feature space output by step S5.1 are used as input conditions to perform the inner loop parameter optimization process of the base model; The gradient descent algorithm (parameters: learning rate η, batch size B, number of iterations E) is used to calculate the loss function between the input feature matrix of the training subset and the model prediction output, so as to quantify the degree of deviation of the model prediction under the current feature distribution. Furthermore, the training loss is calculated using the mean squared error loss function, where the loss function is defined as:

[0023] in, The number of training samples. The true label value of the sample. Predict the output value for the base model; Furthermore, the loss value is backpropagated onto the base model parameters. The chain rule is used to decompose the propagation path of the gradient on the parameters of each layer, establishing the mapping relationship between the gradient and the weight matrix and bias vector. The parameter values ​​are updated in each iteration. for:

[0024] in, For learning rate, For the loss function with respect to parameters The gradient; Furthermore, by performing momentum optimization (parameter: momentum coefficient μ) on the gradient vector of each iteration in the batch update strategy, the variance of parameter updates is reduced and convergence is accelerated. At the same time, L2 regularization is performed before weight update to suppress the risk of model overfitting. By using the gradient descent and optimization strategies described above, the weighted feature space from the previous step is mapped to the parameter update path of the base model, thereby achieving a high-fit model parameter configuration for the current feature distribution. For example, in the task of predicting the water stability of asphalt mixtures, the 1200 samples contained in the current training subset are used as input, and the learning rate is... Batch size Number of iterations Using mean squared error as the loss function, the training loss value is calculated as follows: The gradient norm of the weight matrix is ​​calculated through backpropagation. The momentum coefficient μ is set to And add L2 regularization terms to the gradient update weights. After this process is completed, the average difference between the base model's predicted output and the true test label on the training set of the current fold is significantly reduced, the model's fitting ability under this feature distribution is significantly improved, and a stable model parameter basis is provided for the subsequent validation loss calculation in S5.3; S5.3: Calculate the validation loss for each fold based on the current validation subset, use the validation loss value to adjust the feature weight allocation mechanism of the CAS network, and use the meta-gradient accumulation algorithm to cache the gradient information of each fold into the gradient accumulation buffer of the MAML framework to provide a basis for subsequent global parameter optimization. S5.4: After completing one round of K-fold verification, the outer loop parameter update operation is performed based on the accumulated gradient information. The MAML optimization algorithm is used to globally adjust the initial parameters of the CAS network to enhance the model's ability to adapt quickly to new data distributions. After completing one round of K-fold cross-validation, based on the cross-fold gradient information cached in the gradient accumulation buffer of the MAML framework, an outer loop parameter optimization method (parameter settings: learning rate ηouter = 0.001, batch gradient merging mode is full accumulation) is adopted to achieve global adjustment of the initial parameters of the context-aware subnetwork. Furthermore, using a gradient aggregation algorithm (parameter: average normalization factor allocation based on fold number K), the feature weight gradient matrix obtained from each fold is summed element-wise, and the gradient mean vector is calculated. To obtain the global parameter update direction; Furthermore, the MAML optimization formula is used in the outer loop to adjust the initial parameters of the current CAS network. The correction is performed based on the gradient mean vector, and the specific update formula is as follows: ,in For the outer loop learning rate, The global gradient mean vector, To enable rapid adaptation of updated network parameters in the global context; Furthermore, by using the second-order gradient approximation method, a second-order Taylor expansion is performed on the outer loop gradient of MAML to approximately estimate the influence weight of the meta-gradient between different fold tasks, and the parameter update step size is corrected to balance the stability and rapid adaptability of the model under new data distribution. By using a global gradient update process, the local adaptation results of each fold of the previous round of cross-validation are transformed into initial parameters of the CAS network with cross-scenario generalization ability, thereby improving the dynamic feature weighting ability of the model in diverse asphalt mixture service environments. For example, in a cross-validation task with K=5 folds, the gradient accumulation buffer stores the feature weight gradient matrix of each fold, with a size of (64×10), where 64 is the number of hidden layer neurons and 10 is the feature dimension. The mean gradient matrix is ​​obtained by averaging the gradients over the number of folds using a gradient aggregation algorithm. The value range of each matrix element is concentrated in the range of [-0.05, 0.08]. An outer loop learning rate is used. Execute update formula Subsequently, the absolute value range of the initial weights in the CAS network decreased overall, and the weight distribution became more concentrated. After the next round of cross-validation testing, on the new material combination and service environment datasets, the variance of the water stability index predicted by the model significantly decreased, verifying that the external loop parameter update mechanism can improve the model's performance stability and generalization ability under multiple scenario conditions. S5.5: During the multi-round cross-validation iteration process, the mean and variance of the validation error of each fold are continuously calculated. Based on the convergence judgment algorithm, it is determined whether the current model meets the preset convergence threshold. If it does not meet the threshold, return to S2 to perform parameter iteration update to ensure the stability and generalization performance of the model training process.

[0025] Step S6: Calculate the mean and variance of the prediction error for each fold validation set, and determine whether they meet the preset convergence threshold. If not, return to S2 to perform parameter iteration and update. If they meet, output the final prediction model optimized by dynamic feature weighting and its feature importance distribution map. Specifically, this includes: S6.1: Based on the predicted output and true label of each fold validation set, calculate the validation loss value for each fold. The validation loss value is either mean squared error or cross-entropy loss, to quantify the prediction bias of the model under the current feature weighting strategy. S6.2: Perform statistical processing on the K-fold validation loss value, calculate its mean and variance to evaluate the stability of the model under different data distributions, and output the mean error vector and error variance matrix as the basis for convergence judgment; The input is the K-fold validation loss sequence obtained in step S6.1. Each element of the sequence corresponds to the validation loss of a cross-validation fold, which is the mean squared error or cross-entropy loss calculated under the current feature weighting strategy. The statistical mean calculation method (parameters: input is the validation loss sequence, statistical mode is arithmetic mean) is used to quantify the central tendency of the prediction error of each fold model and obtain the error mean scalar. Furthermore, by using the variance calculation method (parameters: input is the validation loss sequence, and the bias measure is the squared deviation), the degree of fluctuation of the prediction error of each fold model is quantified, and the error variance scalar is obtained; Furthermore, by using a vectorization method (parameter: mapping the loss value of each fold to a one-dimensional vector index space), the error mean vector is constructed, where each element is the mean loss of the corresponding fold, which is used for subsequent performance comparison analysis across data distributions. Furthermore, by using a matrix encapsulation method (parameter: combining variance values ​​and fold indices into a two-dimensional matrix structure), the error variance matrix is ​​generated, which associates each fold variance with the corresponding data distribution label, making it easier to locate the source of performance instability under different environmental conditions; By combining the mean and variance, the validation loss sequence of the previous step is transformed into an error mean vector and an error variance matrix, thereby realizing a quantitative representation of the model's stability performance under different data distribution scenarios. For example, the input is a sequence of validation loss values ​​[0.042, 0.038, 0.045, 0.037, 0.040] obtained through 5-fold cross-validation, and the mean error is calculated using the arithmetic mean method: ,in The mean of each loss value is 0.0404. The variance formula is then used. ,in The mean error is given, and the variance is 6.16 × 10⁻⁶. -6 The mean values ​​are assembled into an error mean vector [0.0404, 0.0404, 0.0404, 0.0404, 0.0404], and the variance values ​​and fold numbers are encapsulated into an error variance matrix [[1, 6.16×10]. -6 ], [2, 6.16×10 -6 ], [3, 6.16×10 -6 ], [4, 6.16×10 -6 ], [5, 6.16×10 -6 In this embodiment, the quantitative results of the mean and variance show that the validation losses at each fold are highly similar and have very small fluctuations, reflecting that the model has significant stability under the current feature weighting strategy, which can provide accurate quantitative basis for the next step of convergence determination; S6.3: Compare the mean error of the current round with the preset convergence threshold. If the mean error is less than or equal to the convergence threshold and the error variance is lower than the preset fluctuation range, then the model training is determined to be converged. S6.4: If the model does not meet the convergence condition, a gradient update signal is constructed based on the mean and variance of the error. The initial parameters of the context-aware sub-network (CAS) are updated using the gradient descent algorithm to optimize the feature weight generation strategy. Based on the convergence judgment result of step S6, the input error mean vector and error variance matrix are used, and the error gradient mapping method (parameter: error mean) is employed. Error variance This enables the construction of an initial gradient update signal for parameter optimization; Furthermore, a weighted combination algorithm (parameters: weight coefficients α=0.7 for the mean and β=0.3 for the variance) is used to achieve weighted fusion of gradient update signals and obtain vectorized gradient instruction data. Furthermore, a gradient calculation method for the mean square error objective function is adopted, and a formula is introduced to calculate the update direction and magnitude of each feature weight during the gradient update signal generation process; Furthermore, the initial parameter matrix of the context-aware subnetwork (CAS) is updated element-wise using the gradient descent algorithm (parameters: learning rate η=0.005, number of iterations t=10). Furthermore, a forward propagation verification calculation is performed using the updated parameter matrix to generate a new dynamic feature weight vector, and the L2 norm change value between this vector and the previous weight vector is compared to evaluate the optimization magnitude of the weight generation strategy. By combining gradient descent algorithm with error gradient mapping, the convergence failure information from the previous step is transformed into parameter optimization signals for the CAS subnetwork, thereby achieving adaptive adjustment of the feature weight generation strategy and improving the generalization performance of the prediction model. For example, during the training of an asphalt mixture water stability prediction model, the convergence result showed a mean error of 2.35 and a variance of 0.42, while the preset convergence threshold was 1.8. An initial gradient signal was constructed using the error gradient mapping method, where the mean component was 2.35. 1.8 = 0.55, and the variance component is 0.42. The fused gradient instruction vector is calculated using a weighted combination algorithm. The gradient values ​​of each feature weight were calculated based on the mean squared error formula, and the initial parameter matrix of CAS was iteratively updated for 10 rounds with a learning rate of 0.005. The update formula was applied in each round of iteration. The optimized feature weight vector generated in the new forward propagation showed a change of 0.048 in L2 norm compared to the previous round, indicating that the weight generation strategy significantly improved the feature importance capture ability under the current data distribution, and the stability of the prediction model was enhanced in subsequent cross-validation. S6.5: If the model meets the convergence condition, then based on the finally trained context-aware sub-network and base model, a prediction model with dynamic feature weighted optimization is generated, and a feature importance distribution map is output. The map shows the weight change trend of each feature under different validation layers in the form of a heat map. Based on the convergence judgment result of the current round, the parameter matrix of the context-aware sub-network and the weight configuration of the base model after multi-round K-fold cross-validation training are loaded to realize the construction of the dynamic feature weighted optimization prediction model; A weight freezing strategy (parameters: dynamic feature weight vector output by the CAS network and residual weight matrix of the base model) is adopted to ensure the consistency of feature weighting mode during the training phase and during the inference phase. The feature weighting schemes of each validation fold are distributed and aggregated to generate global feature weight distribution data. Furthermore, by using the feature weight time series analysis method (parameters: training round index of each fold, weight value trajectory), the trend of feature weight changes under different validation folds is quantified, and the trend is transformed into a two-dimensional matrix structure for visualization generation; Furthermore, the heatmap drawing algorithm is called (parameters: global feature weight distribution matrix, color mapping interval setting, normalization range [0,1]) to map the values ​​of each unit of the matrix to the color gradient space, generating a feature importance distribution map, where the color depth corresponds to the size of the feature contribution and intuitively displays the weight fluctuation under different folds; By using the feature weight normalization method, the weight matrix is ​​uniformly converted into a relative contribution ratio to ensure that each column of the graph meets the normalization constraint. Through the above weight normalization and visualization processing, the model structure after the convergence judgment in the previous step is transformed into the final prediction model containing the weight distribution map, realizing the explicit output of the feature contribution mode under different material combinations and service environments and verifying the generalization ability. For example, in the scenario of predicting the water stability of asphalt mixtures, after completing multiple rounds of 5-fold cross-validation, the CAS network outputs a final feature weight vector with a dimension of 12. A weight freezing strategy is used to preserve the weight value sequence after each fold training, resulting in a weight matrix of size 12×5. The feature weight time series analysis method is called to calculate the mean and variance of each column of weight sequence and convert it into a heatmap input matrix. The color mapping interval is set to [0,1]. When the average weight of a feature across all folds is 0.85, the corresponding position in the heatmap is close to dark red, while an average weight of 0.15 corresponds to light yellow. After processing with a normalization formula, it is ensured that the sum of the normalization values ​​for each column is 1. The final generated heatmap shows significant fluctuations in the distribution of feature weights under different validation folds, enabling technicians to intuitively identify the consistency and volatility of high-contribution features such as asphalt film thickness and aggregate gradation under different working conditions, thereby significantly improving the interpretability of the prediction model and its ability to guide material optimization.

[0026] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting water stability of asphalt mixtures, characterized by, The method comprises the following steps: S1: Collecting multi-dimensional service environment parameters, material composition characteristics and water stability test data of asphalt mixture samples, generating a standardized data set, and dividing the training set and the validation set; S2: Based on the multi-layer perception architecture, a context-aware subnetwork is constructed, each hidden layer node is connected through a parameter sharing mechanism, and a gradient accumulation buffer of the meta-learning framework MAML is configured, the input end of the context-aware subnetwork receives the feature matrix of the training set, and the output end is connected with a normalized weight vector generation module; S3: Input the current training set into the context-aware subnetwork, perform a differentiable feature importance evaluation operation, and output a dynamic feature weight vector; S4: Based on the dynamic feature weight vector, perform element-by-element multiplication operation on the training set feature matrix to generate a weighted feature space, and construct a feature correlation matrix based on the attention mechanism; S5: Training the base model on the weighted feature space using K-fold cross-validation strategy, performing inner loop parameter update and outer loop gradient accumulation, wherein the inner loop adjusts the base model parameters based on the current fold validation loss, and the outer loop corrects the initial parameters of the context-aware subnetwork according to the accumulated loss; S6: Calculate the mean and variance of the prediction error of each fold validation set, and determine whether it meets the preset convergence threshold, if not, return to step S2 to perform parameter iteration update, if yes, output the final prediction model optimized by dynamic feature weighting and its feature importance distribution map.

2. The method of asphalt mixture water stability analysis prediction of claim 1, wherein, The step S1 specifically comprises: Based on the engineering application background of asphalt mixture, multi-dimensional service environment parameters of the sample under different service environments are collected, and the multi-dimensional service environment parameters are synchronously collected by using an environmental simulation test box and a field monitoring equipment to obtain input feature data reflecting the actual service conditions; Based on the material science and road engineering standards, the material composition characteristic parameters of the asphalt mixture sample are collected, and the structured data of the material composition characteristics are obtained through physical and chemical analysis of the laboratory mix design and the field core sample; Based on the water stability test standard method, the immersion Marshall test, freeze-thaw splitting test and surface permeability test are performed on the asphalt mixture sample to collect water stability test data of the sample under different water damage, and form a target variable data set; Based on the data standardization processing method, the service environment parameters, the material composition characteristic parameters and the water stability test data are subjected to missing value filling, outlier removal and normalization processing, the multi-dimensional features are subjected to unified dimension conversion by using Z-score standardization algorithm, and a standardized data set is generated; Based on the data division principle of K-fold cross-validation, the standardized data set is divided into a training set and a validation set.

3. The method of asphalt mixture water stability analysis prediction according to claim 2, characterized in that, The multi-dimensional service environment parameters include temperature gradient, humidity change rate, traffic load cycle number and freeze-thaw cycle number, and the corresponding data are obtained by using multi-point temperature sensor, capacitive humidity sensor array, traffic load simulation table and freeze-thaw test tank respectively, and the parameter data of various types are aligned by using a synchronous acquisition controller.

4. The method of asphalt mixture water stability analysis and prediction of claim 1, wherein, The step S2 specifically comprises: Based on the multi-dimensional feature matrix of the asphalt mixture water stability prediction task, a context perception subnetwork with a multilayer perceptron structure is constructed; A parameter sharing mechanism is performed between each hidden layer of the context perception subnetwork, and the feature importance in the cross-validation process is consistently evaluated by sharing the weight matrix and bias vector; Based on the meta-learning framework MAML, a gradient accumulation buffer is configured in the context perception subnetwork to temporarily store the gradient information of each fold in the K-fold cross-validation process; The feature matrix of the training set is input into the input layer of the context perception subnetwork as input, and after a plurality of nonlinear transformations, an intermediate feature representation is generated. A Softmax normalization function is introduced at the output end of the context perception subnetwork to normalize the intermediate feature representation and generate a set of dynamic feature weight vectors.

5. The method of asphalt mixture water stability analysis prediction of claim 4, wherein, The context perception subnetwork is a multilayer perceptron structure, the number of input layer nodes is consistent with the feature dimension, the number of hidden layers is not less than 3, and the initial linear transformation result is nonlinearly mapped by using a ReLU activation function in the hidden layer to extract high-order interaction relationships between features.

6. The method of asphalt mixture water stability analysis and prediction of claim 1, wherein, The step S3 specifically includes: The context perception subnetwork constructed based on the multilayer perceptron architecture receives the original feature matrix of the current fold training set as input in the form of a feature vector, performs linear transformation in the first hidden layer to obtain the initial mapping result of the feature space; The initial mapping result is nonlinearly transformed by using a ReLU activation function to generate the activation output of the first hidden layer as the input of the next hidden layer; A differentiable feature importance calculation module is introduced in the middle hidden layer of the multilayer perceptron to calculate the gradient information between each feature dimension and the target variable based on the label information of the current input sample, and obtain the local sensitivity of the feature to the output prediction; According to the local sensitivity information and the feature representation output by each hidden layer, the importance score of each feature dimension is normalized by using a Softmax normalization function to generate a dynamic feature weight vector.

7. The method of asphalt mixture water stability analysis and prediction of claim 1, wherein, The step S4 specifically includes: Based on the dynamic feature weight vector, an element-wise multiplication operation is performed on the original feature matrix of the current fold training set to generate a weighted feature matrix; The weighted feature matrix is normalized by using a standardization method to centralize and uniformize each feature dimension, and a standardized weighted feature space is obtained; Based on the multi-head attention mechanism in the Transformer architecture, an attention weight matrix between feature dimensions is constructed, which obtains the correlation strength between each feature dimension by calculating the Query-Key similarity of the feature vectors in the weighted feature space to form a feature correlation matrix; The feature correlation matrix is sparsified to output a sparse feature correlation graph; Based on the sparse feature correlation graph, the standardized weighted feature space is modeled as a graph structure, the high-order interaction relationships between feature nodes are modeled by using a graph convolution network, the high-order nonlinear correlation between features is extracted, and an enhanced feature representation vector is generated.

8. The method of asphalt mixture water stability analysis prediction of claim 7, wherein, The step S4 further comprises element-wise weighting of the training set feature matrix based on the dynamic feature weight vector, wherein each column feature vector of the feature matrix is scalar multiplied with the corresponding weight value to obtain a weighted feature matrix, and the weighted feature matrix is further standardized using the Z-score method, then the feature correlation is calculated through the multi-head attention mechanism, and the Top-K sparsification threshold screening method is used to retain the K feature dimensions with the highest correlation intensity in each row, thereby generating a sparsified feature correlation graph.

9. The method of asphalt mixture water stability analysis and prediction of claim 1, wherein, The step S5 specifically comprises: Based on the weighted feature space, a K-fold cross-validation task division structure is constructed, and the K-fold division algorithm is used to perform non-repeated random division on the training set samples to generate K mutually exclusive training-validation subset pairs; In each fold cross-validation, the base model is executed based on the current training subset to perform inner loop parameter update operation, the gradient descent algorithm is used to calculate the current fold training loss, and the base model parameters are adjusted through back propagation; The validation loss of each fold is calculated based on the current validation subset, the feature weight distribution mechanism of the context-aware subnetwork is feedback adjusted using the validation loss value, and the meta-gradient accumulation algorithm is used to cache the gradient information of each fold to the gradient accumulation buffer of the MAML framework; After completing a round of K-fold validation, the outer loop parameter update operation is executed based on the accumulated gradient information, and the MAML optimization algorithm is used to globally adjust the initial parameters of the context-aware subnetwork; During the multi-round cross-validation iteration process, the mean and variance of the validation error of each fold are continuously calculated, and the convergence judgment algorithm is used to judge whether the current model meets the preset convergence threshold, if not, return to step S2 to perform parameter iteration update.

10. The method of asphalt mixture water stability analysis and prediction of claim 1, wherein, In the step S6, if the multi-round iteration converges, the dynamic feature weighting optimized asphalt mixture water stability prediction model is finally output, and the feature importance distribution under different cross-validation folds is output in the form of a heat map for visual display of the change trend of each feature weight.

Images