Modified asphalt design method integrating cross-scale series prediction and multi-objective optimization

Abstract

The invention provides a modified asphalt design method integrating cross-scale series prediction and multi-objective optimization, and relates to the technical field of modified asphalt materials.The method comprises the steps that modified asphalt is tested, and a completely-associated cross-scale data set is obtained; inputting the completely associated cross-scale data set into a cross-scale series prediction model, and training by using a multi-task learning loss function based on task difficulty self-adaption to obtain a trained cross-scale series prediction model; performing statistical calculation on the completely associated cross-scale data set to obtain a microcosmic parameter feasible region; setting a macroscopic performance target value of the modified asphalt, performing multi-objective optimization on the preparation parameters by combining the deviation between the macroscopic performance predicted value and the target value and the feasible region constraint of the microcosmic parameters to obtain a final preparation parameter design result, and performing experimental verification according to the final preparation parameter design result to complete the design of the modified asphalt. The problems that a traditional trial and error method is low in efficiency, poor in precision and difficult in multi-target collaboration are solved.

Description

Technical Field

[0001] This specification relates to the technical field of modified asphalt materials, and in particular to a modified asphalt design method integrating cross-scale serial prediction and multi-objective optimization. Background Art

[0002] Asphalt binders are widely used in roads, highways, and airport pavements due to their excellent road performance. With the increase in traffic loads and volume, as well as the influence of adverse climatic conditions, asphalt pavements are prone to damage due to the limitations of the material properties. Therefore, asphalt binders need to have better performance to meet road use requirements during the service phase. In order to obtain high-performance asphalt materials, the base asphalt needs to be modified accordingly. Among them, adding polymer modifiers to form a multiphase system with the asphalt binder is the most common and effective method. SBS modifiers have become a commonly used modifier for modified base asphalt due to their good processability and excellent mechanical properties. The performance of SBS modified asphalt is affected by factors such as dosage, shear temperature, shear rate, and shear time. Therefore, in order to obtain modified asphalt that meets the requirements, a large number of orthogonal experiments are often required to determine the preparation parameters that meet the requirements. This trial-and-error forward design method has a long R&D cycle, high cost, and wastes a large amount of material resources. In addition, existing forward prediction models for asphalt material properties ignore the shear process, microstructure, and the transmission mechanism of macroscopic properties, resulting in low prediction accuracy and generalization capabilities. However, existing inverse optimization methods for asphalt materials ignore the coupling relationship between multiple performance indicators and are prone to falling into local optimality. Summary of the Invention

[0003] In response to the above-mentioned deficiencies in the prior art, the modified asphalt design method provided by the present invention, which integrates cross-scale serial prediction and multi-objective optimization, solves the problems of low efficiency and difficulty in multi-objective coordination of traditional trial and error methods.

[0004] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization, comprising:

[0005] S1: By testing modified asphalt, a fully correlated cross-scale dataset is obtained;

[0006] S2: Construct a cross-scale cascade prediction model based on a shared feature extraction module, a micro-parameter prediction branch, and a macro-performance prediction branch;

[0007] S3: Inputting the fully correlated cross-scale dataset into the cross-scale cascade prediction model, and training the model using a multi-task learning loss function adaptive to task difficulty to obtain a trained cross-scale cascade prediction model; wherein the trained cross-scale cascade prediction model is used to analyze the fully correlated cross-scale dataset to obtain a macro performance prediction value;

[0008] S4: performing statistical calculations on the fully correlated cross-scale data set to obtain a feasible domain of microscopic parameters;

[0009] S5: Set the macro-performance target value of the modified asphalt, combine the deviation between the macro-performance predicted value and the target value and the feasible domain constraints of the micro-parameters, perform multi-objective optimization on the preparation parameters, and obtain the final preparation parameter design results. Based on the final preparation parameter design results, conduct experimental verification to complete the design of the modified asphalt.

[0010] The present invention provides a modified asphalt design method that integrates cross-scale cascade prediction and multi-objective optimization. Based on a fully correlated cross-scale dataset, this method constructs a cross-scale cascade prediction model and introduces a multi-task adaptive learning loss function based on task difficulty. This method achieves precise correlation prediction of preparation parameters, microscopic parameters, and macroscopic performance, overcoming the low efficiency and high cost of traditional trial-and-error methods. Combined with a multi-objective optimization algorithm, this method rapidly optimizes the optimal preparation parameters within the feasible domain of microscopic parameters, ensuring that the macroscopic performance error of the modified asphalt is less than 5%. This significantly improves the efficiency and precision of modified asphalt design, achieving a modified asphalt design guided by a multi-objective performance balance.

[0011] Furthermore, the S1 includes:

[0012] Using base asphalt and modifier, modified asphalt is prepared under various preparation parameters to obtain a modified asphalt set;

[0013] Performing Fourier transform infrared spectroscopy on the modified asphalt set to obtain microscopic composition parameters;

[0014] Observe the modified asphalt set using a laser confocal microscope to determine the average particle size of each modifier fluorescent particle in the laser confocal image of the modified asphalt, and obtain microscopic morphological parameters by weighted average particle size;

[0015] The modified asphalt set is subjected to three major index tests and a rotational viscosity test to obtain macroscopic performance parameters; wherein the preparation parameters, the microscopic composition parameters, the microscopic morphology parameters and the macroscopic performance parameters belong to a completely correlated cross-scale data set.

[0016] By combining Fourier transform infrared spectroscopy, laser confocal microscopy, and macroscopic performance testing, a comprehensive cross-scale dataset encompassing preparation parameters, microscopic composition, micromorphology, and macroscopic performance was constructed. This data system not only reveals the deep correlation between the modifier dispersion state and preparation parameters and macroscopic performance, but also provides high-quality cross-scale feature support for model training, avoiding the reliability issues of traditional single-scale data modeling.

[0017] Furthermore, the cross-scale tandem prediction model includes:

[0018] A shared feature extraction module is used to learn and extract features of the preparation parameters to obtain preparation parameter features;

[0019] A micro-parameter prediction branch is used to map micro-composition parameters, micro-morphological parameters and the preparation parameter characteristics to obtain micro-parameter prediction values;

[0020] The macro performance prediction branch is used to splice the micro parameter prediction value and the preparation parameter characteristics to obtain a splicing result; establish a mapping relationship between the splicing result and the macro performance parameter to obtain a macro performance prediction value.

[0021] The cross-scale cascade prediction model constructs an association framework of preparation parameters, micro parameters and macro performance: the high-dimensional nonlinear features of preparation parameters such as shear temperature, shear time and shear rate are uniformly encoded through a shared feature extraction module; the micro prediction branch uses a self-attention mechanism to dynamically focus on key preparation parameters and accurately predict the output micro parameters; the macro prediction branch innovatively splices the micro prediction results with the original features, and uses the cross-attention module to establish micro-macro dynamic associations to form cross-scale connections; compared with ordinary neural networks, it significantly improves the prediction accuracy and generalization ability.

[0022] Furthermore, the shared feature extraction module includes a first fully connected layer and a second fully connected layer;

[0023] The first fully connected layer is used to map the preparation parameters to a high-dimensional feature space, capture the nonlinear relationship between the parameters, and obtain a first mapping result;

[0024] The second fully connected layer is used to perform dimensionality reduction processing on the first mapping result to obtain preparation parameter features.

[0025] Furthermore, the micro parameter prediction branch includes a third fully connected layer, a self-attention mechanism module and a fourth fully connected layer, and the macro performance prediction branch includes a fifth fully connected layer, a cross self-attention module and a sixth fully connected layer;

[0026] The third fully connected layer is used to perform dimensionality reduction processing on the preparation parameter features, filter out non-critical preparation parameter features, and obtain micro-parameter related features;

[0027] The self-attention mechanism module is used to dynamically weight the micro-parameter-related features based on the contribution of different preparation parameters, identify key factors, and obtain weighted features;

[0028] The fourth fully connected layer is used to map the weighted features to a micro parameter space to obtain a micro parameter prediction value;

[0029] The fifth fully connected layer is used to perform splicing and dimensionality reduction processing on the micro-parameter prediction values and the preparation parameter features to obtain a splicing result; and perform dimensionality reduction processing on the splicing result to obtain macro-performance related features;

[0030] The cross self-attention module is used to dynamically focus on the macro-performance related features based on the macro-performance parameters, capture the features most relevant to the current macro-performance target, and obtain a focusing result;

[0031] The sixth fully connected layer is used to map the focusing result to the macro performance space to obtain a macro performance prediction value.

[0032] Furthermore, the expression of the multi-task learning loss function based on task difficulty adaptation is:

[0033] L Totat =α·L micro -β·L macro ;

[0034]

[0035] Among them, L Total 、L micro and L macro Represent the total loss, micro parameter loss and macro performance loss respectively, Y micro and Represent the true value and predicted value of micro parameters, Y macro and denote the true value and predicted value of macro performance respectively, α and β denote the loss weights of micro parameter prediction task and macro performance prediction task respectively, MSE denotes mean square error, ω micro and ω macro They represent the loss reduction ratios of the micro parameter prediction task and the macro performance prediction task, L micro (t) and L micro (0) represent the t-th loss and initial loss of the micro-parameter prediction task, L macro (t) and L macro (0) denote the t-th loss and initial loss of the macro performance prediction task, respectively.

[0036] A multi-task learning loss function based on task difficulty adaptation is proposed. By real-time monitoring of the loss reduction ratio of micro / macro tasks, the weights are dynamically adjusted, which solves the problem that traditional fixed weights cause the model to be biased towards simple tasks, realizes dual-task collaborative optimization, and significantly improves prediction accuracy and convergence speed.

[0037] Furthermore, the expression of the feasible region of microscopic parameters is:

[0038] L k =μ micro,k -2σ micro,k ;

[0039] U k =μ micro,k +2σ micro,k ;

[0040] Among them, L k and U k They represent the lower and upper limits of the feasible region of the kth microscopic parameter, μ micro，k and σ micro，k represent the mean and standard deviation of the kth micro parameter respectively.

[0041] Based on statistical methods, the feasible domain of microscopic parameters is defined, and the natural distribution law of laboratory historical data is transformed into optimization constraints, which prevents multi-objective optimization from falling into the physically infeasible solution area, significantly reduces invalid calculations, and ensures that the optimization results meet the actual process conditions.

[0042] Furthermore, the S5 includes:

[0043] By setting a target value of the macro performance of the modified asphalt, the deviation between the predicted value of the macro performance and the target value is calculated to obtain the target deviation of the macro performance;

[0044] Taking the target deviation of the macro performance as the objective function and the feasible region of the micro parameters as the hard constraint, a multi-objective grey wolf optimization is performed on the preparation parameters based on the fuzzy membership function and the highest decision function to obtain the final preparation parameter design result;

[0045] Based on the final preparation parameter design results, test verification is carried out. When the relative error between the macro performance test value and the target value is no greater than the threshold, the design of the modified asphalt is completed.

[0046] The fuzzy membership function is combined with the grey wolf algorithm to quantify the trade-off between different objectives through membership measurement, and the highest decision function is used to screen the Pareto optimal solution, thus avoiding the problem of missing high-quality solutions.

[0047] Furthermore, the expression of the target deviation of the macro performance is:

[0048]

[0049] Among them, g m is the target deviation of the mth macro performance, are the true value and predicted value of the mth macro-performance, respectively, and M is the number of macro-performance.

[0050] Furthermore, the expression of the fuzzy membership function is:

[0051]

[0052] in, and denote the fuzzy membership function and actual value of the i-th objective of the h-th solution in the Pareto front solution set, respectively. and They represent the minimum and maximum values of target i in the Pareto front solution set, respectively, and f i h represents the decision function of the hth solution in the Pareto front solution set, M represents the number of macroscopic performance, N represents the number of solutions in the Pareto front solution set, μ h represents the decision function of the hth solution in the Pareto front solution set.

[0053] The fuzzy membership function normalizes the target values in the Pareto solution set, eliminating the influence of dimensional differences on decision-making. Combined with the weighted summation method to calculate the decision function, it can objectively identify the most balanced solution that takes into account all macroeconomic performance factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] This specification will be further described in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, like numbers represent like structures, wherein:

[0055] Figure 1 is an exemplary flow chart of a modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to some embodiments of this specification;

[0056] Figure 2 is an exemplary schematic diagram of predicting branch performance of microscopic parameters in a cross-scale cascade prediction model according to some embodiments of this specification;

[0057] Figure 3 is an exemplary schematic diagram of predicting branch performance through macro performance in a cross-scale tandem prediction model according to some embodiments of this specification;

[0058] Figure 4 This is an exemplary schematic diagram of the generalization capability of the cross-scale cascade prediction model shown in some embodiments of this specification. DETAILED DESCRIPTION

[0059] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0060] Example

[0061] Figure 1 This is an exemplary flow chart of a modified asphalt design method that integrates cross-scale tandem prediction and multi-objective optimization according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process can be executed by a processor.

[0062] S1: A fully correlated cross-scale dataset was obtained by testing modified asphalt.

[0063] A fully correlated cross-scale dataset is a parameter dataset that correlates preparation parameters with corresponding microscopic composition parameters, microscopic morphological parameters, and macroscopic performance parameters. For example, a fully correlated cross-scale dataset may include preparation parameters, the microscopic composition parameters, the microscopic morphological parameters, and the macroscopic performance parameters.

[0064] In some embodiments, the processor can implement S1 based on the following steps: using matrix asphalt and modifier to prepare modified asphalt under multiple preparation parameters to obtain a modified asphalt set; performing Fourier transform infrared spectroscopy testing on the modified asphalt set to obtain microscopic composition parameters; performing laser confocal microscopy observation on the modified asphalt set to determine the average particle size of each modifier fluorescent particle in the laser confocal image of the modified asphalt, and obtaining microscopic morphological parameters by weighted average particle size; performing three major index tests and rotational viscosity tests on the modified asphalt set to obtain macroscopic performance parameters.

[0065] Base asphalt is the original asphalt material used for modification. For example, base asphalt can include ES70 material and SK70 material.

[0066] Modifiers are reagents used to modify base asphalt. For example, modifiers can include SBS-791H, SBS-792, SBS-796, SBS-161B, etc.

[0067] Preparation parameters are the parameters used in the preparation of modified asphalt. For example, preparation parameters may include shear temperature (170°C, 180°C, 190°C), shear time (30 min, 60 min, 120 min), shear rate (2000 r / min, 4000 r / min, 6000 r / min), and admixture content (3%, 4.5%, 6%, 7.5%).

[0068] In some embodiments, 0.016% of a sulfur stabilizer may be added during the preparation of the modified asphalt.

[0069] The modified asphalt set is a set of modified asphalt types prepared with different preparation parameters, different base asphalts, and different modifier combinations. For example, the modified asphalt set may include SBS-791H modified asphalt.

[0070] Microscopic composition parameters are parameters that reflect the microscopic components of modified asphalt under the influence of different preparation parameters. For example, microscopic composition parameters can include the area of polystyrene characteristic peak, polybutadiene characteristic peak, aliphatic characteristic peak, etc. and the reference peak as the characteristic peak index; wherein the reference peak can be 650-1400 cm -1 The sum of the peak areas in the range.

[0071] In some embodiments, the processor can perform Fourier transform infrared spectroscopy on the SBS modified asphalt, using a wavelength of 650-1400 cm -1 The sum of the peak areas in the range is taken as the reference peak, and the areas of the polystyrene characteristic peak, polybutadiene characteristic peak, aliphatic characteristic peak and the reference peak in the infrared spectrum characteristic peak are calculated as the characteristic peak index, which is used as the microscopic composition parameter.

[0072] The micromorphological parameters are parameters that reflect the micromorphology of the modified asphalt modifier under the influence of different preparation parameters. For example, the micromorphological parameters may include the average particle size of the fluorescent particles of the modifier.

[0073] In some embodiments, the processor can observe the SBS modified asphalt using a laser confocal microscope to obtain the microscopic morphology of the modified asphalt, use image pro plus software to count the average particle size of each SBS modifier fluorescent particle in the laser confocal image of the SBS modified asphalt, and obtain the microscopic morphological parameters by weighted average particle size.

[0074] Macro-performance parameters reflect the road performance of modified asphalt at a macroscopic scale. For example, macro-performance parameters may include 25°C needle penetration, softening point, 5°C elongation, 135°C rotational viscosity, etc.

[0075] In some embodiments, the processor can obtain 25°C needle penetration, softening point, 5°C elongation, and 135°C rotational viscosity based on the three major index tests and rotational viscosity test of SBS modified asphalt to form macroscopic performance parameters.

[0076] In some embodiments, the processor may perform normalization on the fully correlated cross-scale dataset to obtain a fully correlated cross-scale dataset for training and testing; wherein the fully correlated cross-scale dataset for training may divide the normalized parameter dataset into a test set and a training set based on a ratio of 7:3.

[0077] In some embodiments, the expression of the fully correlated cross-scale dataset for training and testing can be:

[0078]

[0079] Among them, X std represents the data in the standardized parameter data set, X represents the data in the parameter data set, μ represents the mean of the data in the parameter data set, and σ represents the standard deviation of the data in the parameter data set.

[0080] S2: Construct a cross-scale cascade prediction model based on the shared feature extraction module, the micro-parameter prediction branch, and the macro-performance prediction branch.

[0081] The cross-scale tandem prediction model is a deep neural network model used to predict macroscopic performance parameters using fully correlated cross-scale datasets.

[0082] In some embodiments, the input of the cross-scale tandem prediction model may be a fully correlated cross-scale data set, and the final output of the cross-scale tandem prediction model may be a macro performance prediction value.

[0083] In some embodiments, the structure of the cross-scale cascade prediction model is as follows:

[0084] The cross-scale cascade prediction model consists of a shared feature extraction module, a micro-parameter prediction branch, and a macro-performance prediction branch. The output of the shared feature extraction module serves as the input to the micro-parameter prediction branch. The output of the shared feature extraction module and the output of the micro-parameter prediction branch serve as the input to the macro-performance prediction branch. The output of the macro-performance prediction branch serves as the final output of the cross-scale cascade prediction model.

[0085] The shared feature extraction module is used to learn and extract features of the preparation parameters to obtain preparation parameter features. The input of the shared feature extraction module may include preparation parameters, and the output may include preparation parameter features.

[0086] Preparation parameter characteristics are characteristics that reflect the nonlinear relationship between preparation parameters.

[0087] In some embodiments, the shared feature extraction module may include a first fully connected layer and a second fully connected layer; wherein the first fully connected layer is used to map the preparation parameters to a high-dimensional feature space, capture the nonlinear relationship between the parameters, and obtain a first mapping result; the second fully connected layer is used to perform dimensionality reduction processing on the first mapping result to obtain preparation parameter features. For example, the input dimension of the first fully connected layer is 6 and the output dimension is 256, and it is used to map the preparation parameters to a high-dimensional feature space and capture the nonlinear relationship between the parameters; the input dimension of the second fully connected layer is 256 and the output dimension is 128, and it is used to reduce the dimension of the output result of the first fully connected layer to avoid overfitting of subsequent branches and obtain preparation parameter features.

[0088] The micro-parameter prediction branch is used to map the micro-composition parameters, micro-morphology parameters, and the fabrication parameter features to obtain micro-parameter prediction values. The input of the micro-parameter prediction branch may include micro-composition parameters, micro-morphology parameters, and fabrication parameters, and the output may include micro-parameter prediction values.

[0089] The predicted value of microscopic parameters is a numerical value that reflects the predicted situation of microscopic composition parameters and microscopic morphological parameters.

[0090] In some embodiments, the micro-parameter prediction branch includes a third fully connected layer, a self-attention mechanism module, and a fourth fully connected layer; wherein the third fully connected layer is used to perform dimensionality reduction processing on the preparation parameter features, filter out non-critical preparation parameter features, and obtain micro-parameter-related features; the self-attention mechanism module is used to dynamically weight the micro-parameter-related features based on the contribution of different preparation parameters, identify key factors, and obtain weighted features; the fourth fully connected layer is used to map the weighted features to the micro-parameter space to obtain micro-parameter prediction values. For example, the input dimension of the third fully connected layer is 128 and the output dimension is 64, which is used to reduce the dimension and filter non-critical preparation parameter features and focus on micro-parameter-related features; the number of heads of the self-attention module is 4, which is used to dynamically weight the contribution of different preparation parameters based on the micro-parameter-related features and identify key factors; the input dimension of the fourth fully connected layer is 64 and the output dimension is 4, which is used to map the output result of the self-attention module to the micro-parameter space to obtain micro-parameter prediction values.

[0091] Micro-parameter related features are related features among the preparation parameter features that are strongly correlated with the micro-parameters.

[0092] Weighted features are micro-parameter-related features that reflect the contribution of preparation parameters.

[0093] The macro-performance prediction branch is configured to concatenate the micro-parameter prediction values with the preparation parameter characteristics to obtain a concatenated result, and to establish a mapping relationship between the concatenated result and the macro-performance parameters to obtain a macro-performance prediction value. The input of the macro-performance prediction branch may include the micro-parameter prediction values and the preparation parameter characteristics, and the output may include the macro-performance prediction value.

[0094] The macro performance forecast value is a numerical value that reflects the macro performance forecast situation.

[0095] In some embodiments, the macro-performance prediction branch includes a fifth fully connected layer, a cross-self-attention module, and a sixth fully connected layer; wherein the fifth fully connected layer is used to concatenate and reduce the dimension of the micro-parameter prediction value and the preparation parameter feature to obtain a concatenated result; the concatenated result is subjected to dimensionality reduction to obtain macro-performance-related features; the cross-self-attention module is used to dynamically focus on the macro-performance-related features based on the macro-performance parameters, capture the features most relevant to the current macro-performance target, and obtain a focused result; the sixth fully connected layer is used to map the focused result to the macro-performance space to obtain a macro-performance prediction value. For example, the fifth fully connected layer has an input dimension of 128+4 dimensions and an output dimension of 64 dimensions, and is used to concatenate the micro-parameter prediction value and the preparation parameter feature to obtain a concatenated result; the query of the cross-attention module is the macro-performance parameter, and the key / value is the concatenated result, so that the macro-performance prediction branch dynamically focuses on the micro-features most relevant to the current macro-performance target; the sixth fully connected layer has an input dimension of 32 dimensions and an output dimension of 4 dimensions, and is used to map the output result of the cross-attention module to the macro-performance space to obtain a macro-performance prediction value.

[0096] The focused results are the preparation parameter characteristics and micro parameter characteristics that are most relevant to the current macro performance goals.

[0097] S3: Input the fully correlated cross-scale data set into the cross-scale cascade prediction model, and train it using a multi-task learning loss function based on task difficulty adaptation to obtain a trained cross-scale cascade prediction model; wherein the trained cross-scale cascade prediction model is used to obtain a macro performance prediction value.

[0098] In some embodiments, the cross-scale cascade prediction model can be trained using training samples. For example, the training samples can be input into the initial cross-scale cascade prediction model. A multi-task learning loss function based on task difficulty adaptation is constructed using the true sample values and the results of the initial cross-scale cascade prediction model. The parameters of the initial cross-scale cascade prediction model are iteratively updated based on the multi-task learning loss function based on task difficulty adaptation. When preset conditions are met, model training is completed, resulting in a trained cross-scale cascade prediction model. The preset conditions may include convergence of the multi-task learning loss function based on task difficulty adaptation, or the number of iterations reaching a threshold.

[0099] In some embodiments, the training samples may include a fully correlated cross-scale dataset.

[0100] In some embodiments, the expression of the multi-task learning loss function based on task difficulty adaptation can be:

[0101] L Total =α·Lmicro -β·L macro ;

[0102]

[0103] Among them, L Total 、L micro and L macro Represent the total loss, micro parameter loss and macro performance loss respectively, Y micro and Represent the true value and predicted value of micro parameters, Y macro and denote the true value and predicted value of macro performance respectively, α and β denote the loss weights of micro parameter prediction task and macro performance prediction task respectively, MSE denotes mean square error, ω micro and ω macro They represent the loss reduction ratios of the micro parameter prediction task and the macro performance prediction task, L micro (t) and L micro (0) represent the t-th loss and initial loss of the micro-parameter prediction task, L macro (t) and L macro (0) denote the t-th loss and initial loss of the macro performance prediction task, respectively.

[0104] In some embodiments, as Figure 2 、 Figure 3 and Figure 4 As shown, the processor can be based on the regression coefficient R 2 The performance of the cross-scale tandem prediction model on the training set and the test set is evaluated by using the mean relative error (MAPE). The generalization ability GA is evaluated based on the relative deviation of the model performance on the training set and the data set. When the R 2 The test is passed when it is not less than 0.90, MAPE is not greater than 5%, and GA is not greater than 10%.

[0105] In some embodiments, the expression for the regression coefficient is:

[0106]

[0107] Among them, y i represents the true value of the i-th sample, represents the predicted value of the i-th sample, represents the sample mean, and n represents the number of samples.

[0108] The mean relative error is an indicator used to evaluate the error of the prediction results of the trained forward prediction model.

[0109] In some embodiments, the expression of the average relative error may be:

[0110]

[0111] In some embodiments, the expression of the generalization ability GA can be:

[0112]

[0113] Among them, P 测试集 is the model performance on the test set, P 训练集 is the model performance on the training set.

[0114] S4: Perform statistical calculations on the fully correlated cross-scale data set to obtain a feasible domain of microscopic parameters.

[0115] The feasible domain of micro parameters is the data that reflects the size of the feasible range of micro parameters.

[0116] In some embodiments, the expression of the feasible domain of microscopic parameters can be:

[0117] L k =μ micro,k —2σ micro,k ;

[0118] U k =μ micro,k +2σ micro,k ;

[0119] Among them, L k and U k They represent the lower and upper limits of the feasible region of the kth microscopic parameter, μ micro，k and σ micro，k represent the mean and standard deviation of the kth micro parameter respectively.

[0120] S5: Set the macro-performance target value of the modified asphalt, combine the deviation between the macro-performance predicted value and the target value and the feasible domain constraints of the micro-parameters, perform multi-objective optimization on the preparation parameters, and obtain the final preparation parameter design results. Based on the final preparation parameter design results, conduct experimental verification to complete the design of the modified asphalt.

[0121] In some embodiments, the processor can implement S5 based on the following steps: by setting a macro-performance target value of the modified asphalt, calculating the deviation between the macro-performance prediction value and the target value, and obtaining the target deviation of the macro-performance; using the target deviation of the macro-performance as the objective function, and the feasible domain of the micro-parameters as the hard constraint, performing multi-objective grey wolf optimization on the preparation parameters based on the fuzzy membership function and the highest decision function, and obtaining the final preparation parameter design result; based on the final preparation parameter design result, conducting experimental verification, and when the relative error between the macro-performance test value and the target value is no greater than a threshold value, completing the design of the modified asphalt.

[0122] The target deviation of macro-performance is data reflecting the deviation between the predicted value of macro-performance and the target value of macro-performance.

[0123] In some embodiments, the expression for the target deviation of the macro performance may be:

[0124]

[0125] Among them, g m is the target deviation of the mth macro performance, are the true value and predicted value of the mth macro-performance, respectively, and M is the number of macro-performance.

[0126] In some embodiments, the expression of the fuzzy membership function may be:

[0127]

[0128]

[0129] in, and denote the fuzzy membership function and actual value of the i-th objective of the h-th solution in the Pareto front solution set, respectively. and They represent the minimum and maximum values of target i in the Pareto front solution set, respectively, and f i h represents the decision function of the hth solution in the Pareto front solution set, M represents the number of macroscopic performance, N represents the number of solutions in the Pareto front solution set, μ h represents the decision function of the hth solution in the Pareto front solution set.

[0130] In some embodiments, the processor may use the feasible region of microscopic parameters as a hard constraint, forcing the microscopic parameters to be within the feasible region during initialization and position update, and projecting out-of-limit values to the nearest boundary.

[0131] In some embodiments, the processor can prepare modified asphalt based on the optimal preparation parameters found and test the actual macro performance test values to determine whether the relative errors between the test values and the target values are less than 5%. If so, the design of the modified asphalt is completed.

[0132] The optimal preparation parameter design results are used to design the preparation parameter results of the modified asphalt that meets the target macroscopic properties. For example, the target properties and preparation scheme can be shown in Table 1.

[0133] Table 1 Target performance and preparation plan

[0134]

[0135]

[0136] In some embodiments of this specification, a modified asphalt design method that integrates cross-scale cascade prediction and multi-objective optimization is proposed. Based on a fully correlated cross-scale dataset, this method constructs a cross-scale cascade prediction model and introduces a multi-task adaptive multi-task learning loss function based on task difficulty. This method achieves precise correlation prediction of preparation parameters, microscopic parameters, and macroscopic performance, overcoming the low efficiency and high cost of traditional trial-and-error methods. Combined with a multi-objective optimization algorithm, this method rapidly optimizes the optimal preparation parameters within the feasible domain of microscopic parameters, ensuring that the macroscopic performance error of the modified asphalt is less than 5%. This significantly improves the efficiency and accuracy of modified asphalt design, achieving a modified asphalt design guided by a multi-objective performance balance.

Claims

1. A modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization, characterized by: include: S1: By testing modified asphalt, a fully correlated cross-scale dataset is obtained; S2: Construct a cross-scale cascade prediction model based on a shared feature extraction module, a micro-parameter prediction branch, and a macro-performance prediction branch; S3: Inputting the fully correlated cross-scale dataset into the cross-scale cascade prediction model, and training the model using a multi-task learning loss function adaptive to task difficulty to obtain a trained cross-scale cascade prediction model; wherein the trained cross-scale cascade prediction model is used to analyze the fully correlated cross-scale dataset to obtain a macro performance prediction value; S4: performing statistical calculations on the fully correlated cross-scale data set to obtain a feasible domain of microscopic parameters; S5: Set the macro-performance target value of the modified asphalt, combine the deviation between the macro-performance predicted value and the target value and the feasible domain constraints of the micro-parameters, perform multi-objective optimization on the preparation parameters, and obtain the final preparation parameter design results. Based on the final preparation parameter design results, conduct experimental verification to complete the design of the modified asphalt.

2. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 1 is characterized in that: Said S1 comprises: Using base asphalt and modifier, modified asphalt is prepared under various preparation parameters to obtain a modified asphalt set; Performing Fourier transform infrared spectroscopy on the modified asphalt set to obtain microscopic composition parameters; Observe the modified asphalt set using a laser confocal microscope to determine the average particle size of each modifier fluorescent particle in the laser confocal image of the modified asphalt, and obtain microscopic morphological parameters by weighted average particle size; The modified asphalt set is subjected to three major index tests and a rotational viscosity test to obtain macroscopic performance parameters; wherein the preparation parameters, the microscopic composition parameters, the microscopic morphology parameters and the macroscopic performance parameters belong to a completely correlated cross-scale data set.

3. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 1 is characterized in that: The cross-scale tandem prediction model includes: A shared feature extraction module is used to learn and extract features of the preparation parameters to obtain preparation parameter features; A micro-parameter prediction branch is used to map micro-composition parameters, micro-morphological parameters and the preparation parameter characteristics to obtain micro-parameter prediction values; The macro performance prediction branch is used to splice the micro parameter prediction value and the preparation parameter characteristics to obtain a splicing result; establish a mapping relationship between the splicing result and the macro performance parameter to obtain a macro performance prediction value.

4. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 3 is characterized in that: The shared feature extraction module includes a first fully connected layer and a second fully connected layer; The first fully connected layer is used to map the preparation parameters to a high-dimensional feature space, capture the nonlinear relationship between the parameters, and obtain a first mapping result; The second fully connected layer is used to perform dimensionality reduction processing on the first mapping result to obtain preparation parameter features.

5. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 3 is characterized in that: The micro parameter prediction branch includes a third fully connected layer, a self-attention mechanism module and a fourth fully connected layer, and the macro performance prediction branch includes a fifth fully connected layer, a cross self-attention module and a sixth fully connected layer; The third fully connected layer is used to perform dimensionality reduction processing on the preparation parameter features, filter out non-critical preparation parameter features, and obtain micro-parameter related features; The self-attention mechanism module is used to dynamically weight the micro-parameter-related features based on the contribution of different preparation parameters, identify key factors, and obtain weighted features; The fourth fully connected layer is used to map the weighted features to a micro parameter space to obtain a micro parameter prediction value; The fifth fully connected layer is used to perform splicing and dimensionality reduction processing on the micro-parameter prediction values and the preparation parameter features to obtain a splicing result; and perform dimensionality reduction processing on the splicing result to obtain macro-performance related features; The cross self-attention module is used to dynamically focus on the macro-performance related features based on the macro-performance parameters, capture the features most relevant to the current macro-performance target, and obtain a focusing result; The sixth fully connected layer is used to map the focusing result to the macro performance space to obtain a macro performance prediction value.

6. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 1 is characterized in that: The expression of the multi-task learning loss function based on task difficulty adaptation is: L Total =α·L micro -β·L macro ; Among them, L Total 、L micro and L macro Represent the total loss, micro parameter loss and macro performance loss respectively, Y micro and Represent the true value and predicted value of micro parameters respectively, Y macro and denote the true value and predicted value of macro performance respectively, α and β denote the loss weights of micro parameter prediction task and macro performance prediction task respectively, MSE denotes mean square error, ω micro and ω macro They represent the loss reduction ratios of the micro parameter prediction task and the macro performance prediction task, L micro (t) and L micro (0) represent the t-th loss and initial loss of the micro-parameter prediction task, L macro (t) and L macro (0) denote the t-th loss and initial loss of the macro performance prediction task, respectively.

7. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 1 is characterized in that: The expression of the feasible region of microscopic parameters is: L k =μ micro，k -2s micro，k ; U k =μ micro，k +2s micro，k ; Among them, L k and U k They represent the lower and upper limits of the feasible region of the kth microscopic parameter, μ micro，k and σ micro，k represent the mean and standard deviation of the kth micro parameter respectively.

8. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 1 is characterized in that: The S5 includes: By setting a target value of the macro performance of the modified asphalt, the deviation between the predicted value of the macro performance and the target value is calculated to obtain the target deviation of the macro performance; Taking the target deviation of the macro performance as the objective function and the feasible region of the micro parameters as the hard constraint, a multi-objective grey wolf optimization is performed on the preparation parameters based on the fuzzy membership function and the highest decision function to obtain the final preparation parameter design result; Based on the final preparation parameter design results, test verification is carried out. When the relative error between the macro performance test value and the target value is no greater than the threshold, the design of the modified asphalt is completed.

9. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 8, characterized in that: The expression for the target deviation of the macro performance is: Among them, g m is the target deviation of the mth macro performance, are the true value and predicted value of the mth macro-performance, respectively, and M is the number of macro-performance.

10. The modified asphalt design method integrating cross-scale tandem prediction and multi-objective optimization according to claim 8, characterized in that: The expression of the fuzzy membership function is: in, and denote the fuzzy membership function and actual value of the i-th objective of the h-th solution in the Pareto front solution set, respectively. and They represent the minimum and maximum values of target i in the Pareto front solution set, represents the decision function of the hth solution in the Pareto front solution set, M represents the number of macroscopic performance, N represents the number of solutions in the Pareto front solution set, μ h represents the decision function of the hth solution in the Pareto front solution set.

Images