A method and system for mix design that integrates unbiased estimation and non-dominated ordination.
By combining the Kriging interpolation model and the non-dominated sorting genetic algorithm, the problems of resource waste and long cycle in traditional material mix design are solved, realizing efficient and accurate mix design of large-scale hydraulic seepage prevention structures, and improving the global search efficiency and accuracy of multi-dimensional performance indicators.
Patent Information
- Application Number
- CN202611306334.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional material mix design suffers from resource waste, high costs, and long design cycles in the construction of large-scale hydraulic seepage prevention structures. It is difficult to perform systematic global intelligent optimization in a high-dimensional parameter space, and it cannot meet the precise control requirements of modern complex engineering for seepage prevention and deformation resistance limit performance.
A combination ratio design method integrating unbiased estimation and non-dominated sorting is adopted. A prediction variance feedback operator is constructed through a Kriging interpolation model to dynamically guide the optimization process of the third-generation non-dominated sorting genetic algorithm NSGA-III. Combined with an entropy weight multi-attribute decision model, multi-dimensional relative proximity selection is achieved, thereby improving the efficiency and accuracy of global search.
It achieved dynamic balance and optimal decision-making for various key seepage prevention and mechanical performance indicators, shortened the design cycle, reduced physical testing costs, and improved the scientific nature and engineering implementation value of the design results.
Smart Images

Figure CN123120140A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer-aided design technology, and in particular to a method and system for mix design that integrates unbiased estimation and non-dominated ordering. Background Technology
[0002] In current engineering construction, especially in the construction of large-scale hydraulic seepage prevention structures, material mix design is a complex systems engineering project involving the coordinated management of multiple physical and mechanical properties. With the evolution of engineering service environments towards extremes and multifunctionality, structural materials not only need to have extremely low porosity to ensure seepage prevention performance, but also need to have excellent tensile and flexural strain capabilities to adapt to complex stress deformation.
[0003] However, these key performance parameters often exhibit nonlinear constraints. Finding the optimal parameter combination that simultaneously satisfies multiple conflicting objectives within computer-aided design software has become a core challenge urgently needing to be overcome in the field of digital R&D for engineering materials. Traditional material mix design primarily relies on manual physical experiments and basic data regression analysis to explore the parameter space. R&D designers typically select qualified raw materials and, based on past engineering experience, artificially set several initial parameter combination gradients, such as configuring different gradation indices, filler contents, and asphalt-aggregate ratios. Then, for each set of parameters, physical specimens are prepared and subjected to long-term destructive tests of mechanical properties such as splitting, tensile, and water stability. After obtaining limited test sample data, single-factor variation curves or basic computer trial-and-error algorithms are used to attempt to select a usable mix design that barely meets the engineering technical specifications.
[0004] This traditional design approach, which combines manual physical verification with simple computer-aided analysis, has significant drawbacks. It heavily relies on the fabrication of massive amounts of physical specimens and complex indoor operations, resulting in substantial resource consumption and trial-and-error costs, and making the response cycle for adjusting the scheme extremely long. Due to the lack of high-precision data prediction models and efficient multi-objective collaborative optimization algorithms, traditional single-factor empirical trial-and-error methods struggle to perform systematic global intelligent optimization in a vast high-dimensional parameter space. This often leads to design solutions that are only locally feasible, failing to truly uncover the deep synergistic effects between the various material components. This is particularly problematic for hydraulic asphalt concrete core wall engineering, where destructive mechanical testing specimens are difficult to fabricate, time-consuming, and costly, resulting in highly sparse historical data. Conventional surrogate models, such as neural networks, are prone to overfitting under such small sample sizes and cannot quantify and predict risks, making it difficult to meet the precise control requirements of modern complex engineering projects for seepage prevention and deformation resistance limits. Summary of the Invention
[0005] To address the aforementioned issues, this invention provides a method and system for mix design that integrates unbiased estimation and non-dominated sorting. It employs an uncertainty feedback operator constructed using the prediction variance based on a Kriging interpolation model to dynamically guide the optimization process of the third-generation non-dominated sorting genetic algorithm NSGA-III. Furthermore, it combines an entropy-weighted multi-attribute decision model to perform multi-dimensional relative proximity selection on the generated Pareto front solution set. This improves the global search efficiency and accuracy of multi-parameter mix design under multiple nonlinear performance index constraints, achieving dynamic balance and optimal decision-making for various key seepage prevention and mechanical performance indicators.
[0006] The above objectives can be achieved through the following approach: A mix design method integrating unbiased estimation and non-dominated sorting includes: acquiring mix proportion parameters and performance index data of historical asphalt concrete core wall projects; constructing an initial dataset containing the correlation between mix proportion parameters and performance indexes; performing standard score detection processing on the initial dataset to remove outliers and generate a target dataset; performing spatial correlation modeling based on the target dataset to construct a Kriging interpolation model capable of simultaneously outputting the performance prediction mean and prediction variance; quantifying the spatial sparsity of the region to be evaluated based on the prediction variance to generate an uncertainty feedback operator, which is a weighted aggregation index that integrates the performance prediction mean and the prediction variance; embedding the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III to perform a multi-dimensional optimization operation dynamically guided by the prediction variance to generate a Pareto front solution set; and performing relative proximity calculation on the Pareto front solution set based on an entropy-weighted multi-attribute decision model to output the target mix proportion parameter combination.
[0007] Optionally, obtaining the mix proportion parameters and performance index data of historical asphalt concrete core wall projects includes: Extract the gradation index parameters characterizing the skeleton structure of the ore, the filler content parameters characterizing the cementing performance, and the asphalt-aggregate ratio parameters from historical engineering projects to generate the mix proportion parameters, wherein the maximum particle size of the ore in the mix proportion parameters is limited to the range of 19 mm to 26.5 mm. The performance index data is generated by extracting bending strain parameters that characterize deformation coordination ability, tensile strain parameters that characterize crack resistance, and porosity parameters that characterize seepage prevention characteristics from historical engineering projects. The target range of the porosity parameters that characterize seepage prevention characteristics is limited to less than or equal to 2.0%.
[0008] Optionally, the step of performing standard score detection processing on the initial dataset to remove outlier samples and generate the target dataset includes: Calculate the deviation of each data point in the initial dataset from the population mean; Divide the deviation value by the standard deviation to generate the standard score for each data point; Data points whose absolute scores exceed a set threshold are identified as outliers and removed. The set threshold is determined based on the 3σ criterion of normal distribution, and the target dataset is generated.
[0009] Optionally, the step of performing spatial correlation modeling based on the target dataset to construct a Kriging interpolation model capable of simultaneously outputting the performance prediction mean and prediction variance includes: Extract the Gaussian semi-variogram or spherical semi-variogram of the target dataset in the decision space; The spatial distance weights between unknown points and known sample points are calculated based on the semi-variogram function. Perform linear unbiased estimation calculations to generate the performance prediction mean, which represents the expected performance value, and generate the prediction variance, which represents the prediction confidence, based on the estimation variance formula.
[0010] Optionally, the step of quantifying the spatial sparsity of the region to be evaluated based on the prediction variance and generating an uncertainty feedback operator, wherein the uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the prediction variance, includes: Extract the predicted variance of the matching individuals to be evaluated in the current population space; The predicted performance mean and the predicted variance are aggregated using a weighted summation method. After normalization, the predicted variance is multiplied by a weighting coefficient and then algebraically summed with the normalized predicted performance mean. For benefit-type performance indicators, positive values of the predicted performance mean are substituted; for cost-type performance indicators, negative values of the predicted performance mean are substituted. This generates the uncertainty feedback operator. The predicted variance increases with the distance between the individual to be evaluated and the known sample points.
[0011] Optionally, the step of embedding the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, and performing a multidimensional optimization operation dynamically guided by the predicted variance to generate a Pareto front solution set includes: For each performance objective, the corresponding dimension's uncertainty feedback operator is used to replace the mean performance prediction, which is then used as an indicator for evaluating the fitness of each objective dimension in the third-generation non-dominated sorting genetic algorithm NSGA-III, in order to determine the non-dominated ranking among individuals. Perform population evolution iteration, and dynamically adjust the search step size based on the predicted variance feedback results during each generation of evolution; The Pareto front solution set is output by performing association screening based on the standardized hyperplane and the reference points generated by the third-generation non-dominated sorting genetic algorithm NSGA-III.
[0012] Optionally, the preprocessing step for the relative proximity calculation includes: The Pareto front solution set is standardized to generate a dimensionless decision matrix; The objective weighting coefficients of each performance indicator are determined by calculating the information entropy of each performance indicator. Construct a weighted standardized decision matrix and extract the positive and negative reference solutions from the decision space.
[0013] Optionally, the relative proximity calculation further includes: Calculate the first spatial distance from each mix design to the positive reference solution, and the second spatial distance to the negative reference solution; The relative proximity data is calculated based on the ratio of the second spatial distance to the sum of the first spatial distance and the second spatial distance; The mixing ratio scheme with the highest numerical value in the relative similarity data is extracted as the target mixing ratio parameter combination.
[0014] Optionally, the method further includes: The target mix proportion parameters are associated with the construction quality control system, and a structured quality control suggestion text is output, which includes raw material index requirements, mixing temperature range parameters, paving temperature range parameters, and compaction pass parameters.
[0015] Based on the same inventive concept, this invention also provides a combination ratio design system that integrates unbiased estimation and non-dominated ranking, the system comprising: The data acquisition module is used to acquire mix proportion parameters and performance index data of historical asphalt concrete core wall projects, and to construct an initial dataset containing the correlation between mix proportion parameters and performance indexes. The preprocessing module is used to perform standard score detection processing on the initial dataset, remove outlier samples, and generate the target dataset. The uncertainty modeling module is used to perform spatial correlation modeling based on the target dataset and construct a Kriging interpolation model that can simultaneously output the performance prediction mean and prediction variance. The feedback quantization module is used to quantify the spatial sparsity of the region to be evaluated based on the prediction variance and generate an uncertainty feedback operator. The uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the prediction variance. The optimization module is used to implant the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, perform multi-dimensional optimization operations guided by the prediction variance, and generate Pareto front solution set. The intelligent decision-making module is used to perform relative proximity calculation on the Pareto front solution set based on the entropy-weighted multi-attribute decision-making model and output the target mix ratio parameter combination.
[0016] Compared with the prior art, the present invention has the following advantages: 1. By performing standard score detection on the mix proportion parameters and performance index data of historical projects, outlier samples were effectively eliminated, and a target dataset conforming to normal distribution characteristics was generated. Subsequently, a Kriging interpolation model was constructed based on spatial correlation modeling. This not only improved the quality of the underlying data for computer-aided design but also enabled the constructed model to simultaneously output higher-precision performance prediction mean and prediction variance, laying a solid data foundation and model support for subsequent exploration of globally optimal mix proportion schemes in virtual space.
[0017] 2. By utilizing the spatial sparsity of the prediction variance quantization to generate an uncertainty feedback operator, and embedding it into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, a multi-dimensional optimization operation dynamically guided by the prediction variance is realized. This feedback mechanism can balance global search and local mining in a complex high-dimensional optimization space, effectively preventing the optimization algorithm from getting trapped in local optima, improving the diversity and efficiency of the Pareto front solution set, and making the final parameter optimization calculation process more intelligent and efficient.
[0018] 3. An entropy-weighted multi-attribute decision-making model is employed to calculate the relative proximity of the generated Pareto front solution set. Evaluation weights are determined by objectively calculating the information entropy of various performance indicators, thus avoiding subjective biases caused by human factors in traditional engineering decision-making. Based on the spatial distance calculation of positive and negative reference solutions, the target mix proportion parameter combination can be accurately output, achieving a comprehensive balance of multiple mutually constraining key performance indicators such as seepage prevention characteristics, deformation coordination ability, and crack resistance, thereby enhancing the scientific rigor and engineering implementation value of the mix proportion design results.
[0019] 4. The technical solution of this invention is not a simple linear concatenation of known algorithms, but rather achieves deep coupling and collaborative innovation between modules, producing unexpected technical effects. Specifically, the prediction variance output by the Kriging interpolation model is used as the core variable, and an uncertainty feedback operator is generated by logically reconstructing it with the confidence upper limit function. This operator is then deeply embedded into the fitness evaluation and step size adjustment mechanism at the bottom layer of the NSGA-III algorithm, fundamentally changing the defect of blind iteration in traditional genetic algorithms. Finally, the entropy-weighted multi-attribute decision model takes over the non-dominated solutions to achieve closed-loop decision-making. The combined scheme exhibits synergistic advantages over individual algorithms or simple linear combinations: when processing the same high-dimensional sparse core wall engineering data, the Pareto front hypervolume index (HV) of the deeply coupled scheme of this invention is improved by 22.9% compared to the control scheme using the same NSGA-III algorithm with the same Kriging interpolation model but without the variance dynamic feedback mechanism. Group ablation experiments further demonstrate that retaining the feedback operator alone or the step size adjustment mechanism alone can achieve HV improvements of 12.5% and 10.0%, respectively, proving that the two components have independent contributions and synergistic effects. The iteration speed for the algorithm to reach the convergence target accuracy is improved by approximately 43.3%. These quantitative accuracy improvements and convergence speed comparisons fully demonstrate that the deep fusion of specific algorithms in this invention effectively overcomes the limitations of single algorithms in multi-objective optimization of complex engineering projects, achieving unexpected progress.
[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart illustrating a combination design method that integrates unbiased estimation and non-dominated ranking according to an embodiment of the present invention.
[0023] Figure 2 This is a schematic diagram of the fitting results of the unbiased estimation model of a combination design method and system that integrates unbiased estimation and non-dominated ranking, according to an embodiment of the present invention.
[0024] Figure 3This is a schematic diagram of the Pareto front results of a combination design method and system that integrates unbiased estimation and non-dominated ordination according to an embodiment of the present invention.
[0025] Figure 4 This is a schematic diagram of a combination ratio design system that integrates unbiased estimation and non-dominated sorting according to an embodiment of the present invention. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] Reference Figure 1 One embodiment of the present invention proposes a mix design method that integrates unbiased estimation and non-dominated sorting. It uses the prediction variance based on the Kriging interpolation model to construct an uncertainty feedback operator to dynamically guide the optimization process of the third-generation non-dominated sorting genetic algorithm NSGA-III. It also combines an entropy-weighted multi-attribute decision model to perform multi-dimensional relative proximity selection on the generated Pareto front solution set. This can improve the global search efficiency and accuracy of multi-parameter mix design under multiple nonlinear performance index constraints, and achieve dynamic balance and optimal decision-making for various key seepage prevention and mechanical performance indicators.
[0028] The method described in this embodiment specifically includes: Obtain mix proportion parameters and performance index data of historical asphalt concrete core wall projects, and construct an initial dataset containing the correlation between mix proportion parameters and performance indexes; The initial dataset is processed by standard score detection to remove outlier samples and generate the target dataset. Based on the target dataset, spatial correlation modeling is performed to construct a Kriging interpolation model that can simultaneously output the performance prediction mean and prediction variance. Based on the spatial sparsity of the region to be evaluated quantified by the predicted variance, an uncertainty feedback operator is generated. The uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the prediction variance. The uncertainty feedback operator is implanted into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, and a multi-dimensional optimization operation guided by the prediction variance is performed to generate a Pareto front solution set. The relative proximity of the Pareto front solution set is calculated based on the entropy-weighted multi-attribute decision model, and the target mix ratio parameter combination is output.
[0029] Specifically, the grade index, filler content, and asphalt-aggregate ratio, along with their corresponding mechanical performance indicators, were collected from historical engineering projects. Standard scores were used to detect and remove outliers from the initial dataset, ensuring the target dataset conformed to a normal distribution. Subsequently, a Kriging interpolation model (i.e., the unbiased estimation model described in this invention, where Kriging interpolation achieves spatial prediction by solving for optimal weights under linear unbiased constraints, is a specific implementation of unbiased estimation methods) was established using the target dataset. This model simultaneously outputs the performance prediction mean and the prediction variance characterizing uncertainty. Due to the high cost of core wall material mix proportion tests leading to data sparsity, the Kriging model, with its spatial autocorrelation interpolation characteristics, can provide a prediction mean under small sample conditions and output a prediction variance characterizing the prediction confidence level. This provides a risk quantification foundation for the design of core wall seepage-proof structures and generates an uncertainty feedback operator. This operator is embedded in the third-generation non-dominated sorting genetic algorithm NSGA-III, which dynamically guides multi-dimensional spatial optimization through prediction variance, generating a Pareto front solution set covering multiple objectives such as seepage resistance and deformation capacity. By combining an entropy-weighted multi-attribute decision-making model, objective weights are determined using index information entropy. The optimal mix proportion is then locked in by calculating relative proximity and linked to a structured construction quality control system. By introducing a Kriging unbiased estimation model and an uncertainty feedback mechanism, the limitations of traditional optimization algorithms in getting trapped in local optima in complex high-dimensional spaces are addressed. Variance feedback improves the algorithm's ability to explore unexplored areas, thus ensuring the diversity of Pareto solutions and solution efficiency. Furthermore, the application of the entropy-weighted multi-attribute decision-making model effectively avoids subjective biases caused by manually determining weights, achieving a balance between conflicting objectives such as porosity and bending / tensile strain. This method ensures the high impermeability of the core wall while enhancing its ability to adapt to dam deformation, shortening the mix design cycle, and reducing physical testing costs.
[0030] Optionally, obtaining the mix proportion parameters and performance index data of historical asphalt concrete core wall projects includes: Extract the gradation index parameters characterizing the skeleton structure of the ore, the filler content parameters characterizing the cementing performance, and the asphalt-aggregate ratio parameters from historical engineering projects to generate the mix proportion parameters, wherein the maximum particle size of the ore in the mix proportion parameters is limited to the range of 19 mm to 26.5 mm. The performance index data is generated by extracting bending strain parameters that characterize deformation coordination ability, tensile strain parameters that characterize crack resistance, and porosity parameters that characterize seepage prevention characteristics from historical engineering projects. The target range of the porosity parameters that characterize seepage prevention characteristics is limited to less than or equal to 2.0%.
[0031] Specifically, the selection of mix proportion parameters is based on a deep understanding of the mapping relationship between the microstructure and macroscopic properties of asphalt concrete. The gradation index parameter, by controlling the mass distribution ratio of coarse and fine aggregates, directly constructs the aggregate skeleton structure of asphalt concrete, affecting the internal friction angle and modulus of deformation of the mixture. The filler content parameter, working in conjunction with asphalt, forms a binder substance that fills the tiny pores between the aggregate skeletons, determining the quality of the bonding performance and the density of the mixture. The asphalt-aggregate ratio parameter serves as a variable controlling the workability and cohesive force of the mixture. Regarding performance indicators, the flexural strain parameter is used to quantitatively assess the core wall's ability to adapt to flexural deformation under uneven dam settlement or water pressure. The tensile strain parameter, by simulating the tensile stress response of the core wall under sudden temperature changes or water cooling conditions, characterizes the material's crack resistance and prevents through-cracks in the structure. The porosity parameter is a direct measure of seepage prevention characteristics, determining the distribution of open and closed pores within the material, and is a rigid indicator for judging the effectiveness of the seepage prevention system. When determining the aggregate gradation, a formula is used to fit the theoretical gradation curve: ; in the formula For the aperture to be sieve pass rate; This refers to the filler content; This refers to the maximum particle size of the ore. denoted by a sieve aperture size; r is the gradation index. This formula, through a nonlinear power function, achieves continuous adjustment of the curvature of the gradation curve, bringing the aggregate bulk density closer to an ideal state. The gradation index is a physical parameter that determines the mass ratio of coarse and fine aggregates and characterizes the spatial skeleton distribution of the aggregate. The filler content is the percentage of alkaline mineral powder with a particle size less than 0.075 mm in the total mass of the aggregate in asphalt concrete. The asphalt-aggregate ratio is the mass ratio of asphalt mass to the total mass of the aggregate, a core parameter for controlling the thickness of the asphalt film. Bending strain is the ultimate strain value generated at the tensile edge when an asphalt concrete beam fails under specified test temperature and loading rate. Tensile strain is the deformation per unit length corresponding to the peak strength of the asphalt concrete specimen in an axial tensile test. Porosity is the percentage of the internal void volume of the asphalt concrete to the total volume of the specimen, a key physical quantity for evaluating seepage prevention effectiveness.
[0032] For example, the core wall dam project of the upper reservoir of a pumped storage power station is selected as the verification object. The average annual temperature of this project is 17.6 degrees Celsius, and according to the requirements of the test procedure for hydraulic asphalt concrete, the test temperature is set at 17.6 degrees Celsius. The selected mix proportion parameter combination is: maximum aggregate size... The thickness is 19 mm, the gradation index r is 0.41, and the filler content is... The content is 13.0%, and the oil-stone ratio is 6.8%. The theoretical sieving rate for a critical sieve aperture size of 4.75 mm is calculated using the formula. Substituting the known parameters, we obtain the calculation process: The calculation results show that the theoretical passing rate of the 4.75 mm sieve is 57.95%, which meets the requirements of the design specifications for the continuity of coarse aggregate skeleton in the asphalt concrete panel and core wall of earth-rock dams.
[0033] Optionally, the step of performing standard score detection processing on the initial dataset to remove outlier samples and generate the target dataset includes: Calculate the deviation of each data point in the initial dataset from the population mean; Divide the deviation value by the standard deviation to generate the standard score for each data point; Data points whose absolute scores exceed a set threshold are identified as outliers and removed. The set threshold is determined based on the 3σ criterion of normal distribution, and the target dataset is generated.
[0034] Specifically, based on the standard score method in statistical methods, the deviation of each data point in the initial dataset from the population mean is calculated; that is, the relative position of a data point is determined by calculating the degree of deviation from the mean. This deviation is divided by the standard deviation to generate the standard score for each data point. The standard score can intuitively reflect the relative position of a data point in the overall distribution, thus being used for outlier rejection. When the absolute value of the standard score of a data point is greater than a set threshold, the data point can be considered an outlier. Finally, data points whose absolute standard score is greater than the set threshold are determined as outliers and are removed. The set threshold is determined based on the 3σ criterion of normal distribution. Outliers caused by data entry errors or measurement errors are deleted from the dataset, thereby generating the target dataset, making the retained data more consistent with the overall normal distribution characteristics. The standard score is calculated using the following formula: ; In the formula, Z represents the standard score; x represents the value of a specific data point in the initial dataset. This represents the overall mean of this type of data; This represents the standard deviation of this type of data. The initial dataset is a collection of original test samples containing multiple sets of matching ratio parameters and their corresponding performance indicators. The target dataset is a dataset whose statistical characteristics meet the training requirements of the subsequent unbiased estimation model after removing outliers. The bias value is the algebraic difference between a single data point's value and the population mean, representing the absolute distance of an individual from the center. The standard score is a dimensionless statistic measuring the distance of a data point from the population mean, scaled by the standard deviation. The threshold is set as the critical boundary for judging whether a data point deviates from the normal distribution range; based on the 3σ criterion of normal distribution, the default value is 3. Outliers are extreme data points that deviate from the overall distribution characteristics due to external factors such as data entry errors or measurement errors.
[0035] For example, the bending strain performance parameters of 129 core wall dam projects from the historical engineering database of southern China were selected for verification. Statistical calculations were performed to determine the overall mean of the initial bending strain dataset for this batch. It is 3.95%, with a standard deviation of 1. The deviation is 1.25%. The dataset contains a data point affected by instrument reading error, with a bending strain value x of 8.50%. Calculate the deviation of this data point from the population mean: Divide the deviation value by the standard deviation to generate the standard score corresponding to the data point. Because the algorithm internally uses a normal distribution and sets a fixed threshold of 3 for identifying outliers, the absolute value of the current standard score (3.64) is found to be greater than the threshold of 3. Therefore, data points representing 8.50% of the total score are identified as outliers. Subsequently, the system performs a removal operation, deleting the outlier from the dataset and generating a target dataset free from extreme error interference. This removal process prevents outliers from misleading the prediction model and provides a high-confidence data foundation for multi-objective optimization algorithms.
[0036] Optionally, the step of performing spatial correlation modeling based on the target dataset to construct a Kriging interpolation model capable of simultaneously outputting the performance prediction mean and prediction variance includes: Extract the Gaussian semi-variogram or spherical semi-variogram of the target dataset in the decision space; The spatial distance weights between unknown points and known sample points are calculated based on the semi-variogram function. Perform linear unbiased estimation calculations to generate the performance prediction mean, which represents the expected performance value, and generate the prediction variance, which represents the prediction confidence, based on the estimation variance formula.
[0037] Specifically, the Gaussian semivariogram or spherical semivariogram of the target dataset is extracted in the decision space. During this process, the type of semivariogram must be clearly defined to ensure the accuracy and applicability of spatial correlation modeling. For asphalt concrete performance prediction, this invention preferably uses either the Gaussian or spherical variogram. The spherical variogram is suitable for characterizing physical processes where correlation rapidly decays within a specific parameter range and tends to stabilize and saturate after reaching a certain range, such as the change in porosity with the asphalt-aggregate ratio. The Gaussian variogram is suitable for characterizing highly continuous and extremely smooth performance response surfaces near the origin, such as the gradual evolution of bending strain. Compared to linear variograms that assume unsaturated spatial correlation and do not conform to actual engineering physical processes, using Gaussian or spherical variograms can more realistically reflect the nonlinear mapping boundary between asphalt concrete material components and their macroscopic mechanical properties. The spatial variation characteristics of asphalt concrete performance when mix proportion parameters change are analyzed by establishing a semivariogram structure. The spatial distance weights between unknown points and known sample points are calculated based on the variogram, requiring that the sum of all generated influence weights strictly equal to one. A linear unbiased estimation calculation is performed by weighting and summing the actual physical test performance data of known sample points with the calculated weight parameters to generate the performance prediction mean, which represents the expected performance value. Simultaneously, based on the estimation variance formula, and combining the extracted spatial variability with the auxiliary calculated Lagrange multiplier, a prediction variance representing the confidence level of this prediction is generated. The prediction mean formula is applied as follows: ; Introduce an unbiased condition that constrains the influence of all known points on the proportion: ; Using the unbiased condition, derive and solve the ordinary kriging matrix equations for each weight: ; The formula for the prediction variance of the degree of dispersion using the characterization model is as follows: ; The meanings of the symbols in the formula are as follows: The average performance prediction value for unknown mix proportions to be tested; This represents the actual performance value measured at the i-th known data sample point; The representative algorithm assigns spatial distance weights to the i-th known data sample point; The algorithm assigns spatial distance weights to the j-th known data sample point; N represents the total number of known data sample points participating in the calculation model. The calculated value of the spatial variogram between known data sample point i and known data sample point j; The calculated value of the spatial variation function representing the spatial variation function between known data sample point i and unknown mix proportion points; This represents the Lagrange multiplier introduced to solve extreme value constraints; This represents the derived predicted variance at the unknown mix proportion point. For example... Figure 2 The figure shown is a fitting result diagram of the unbiased estimation model in an embodiment of the present invention. This diagram visually demonstrates the prediction accuracy of the Kriging interpolation model for asphalt concrete performance indicators through scatter plot fitting of the training and test sets and the distribution of the error histogram. The test set data results are relatively concentrated, and the average error line of the histogram is close to the zero error line, proving that the unbiased estimation model has high accuracy and reliability in handling the design of asphalt concrete mix proportion parameters, and can provide a robust data foundation for subsequent multi-objective optimization.
[0038] Among them, the covariance function is a statistical function that measures the degree of coordination in the changes of attribute values of two spatial data points. The variogram function is a semivariance mathematical expression that describes the degree of variability and the weakening trend of autocorrelation of a spatial random variable as spatial distance increases. The spatial distance weight is an influence multiplier coefficient systematically assigned based on the strength of autocorrelation of spatial parameters between known sample points and unknown points to be predicted. Unbiased estimation calculation is a scientific logical inference process that ensures that the statistical expectation of the prediction result is equal to the true physical value of the estimated core wall parameter. The performance prediction mean is the expected output performance value of the representative target mix ratio under specific experimental constraints, obtained by performing matrix weighted summation on known historical sample performance data. The prediction variance is an auxiliary verification statistic that measures the degree to which the predicted estimated value deviates from the objective true value and is specifically used to characterize the uncertainty of spatial calculation models at specific coordinate positions. The Lagrange multiplier is an auxiliary intermediate variable innovatively introduced in the multivariate calculus system specifically for solving the problem of function extrema with pre-constraint conditions.
[0039] For example, two known parameter sample points in the univariate oil-stone ratio coordinate system of the target dataset are selected to predict and reproduce the porosity performance. The coordinate position of the first known sample point is set to 1, and its actual measured porosity value is 1.2. The coordinate position of the second known sample point is set to 3, and its actual measured porosity value is 1.8. The coordinate position of the unknown sample point to be predicted by the algorithm is set to 2. The fitted spherical variogram model is extracted. Considering that the spatial distance between the current local point to be predicted and the known sample point is much smaller than the autocorrelation range of the model, i.e., it is in the approximately linear growth interval near the origin of the spherical variogram, it can be locally equivalently set as the product of a constant 0.5 and the absolute spatial distance in this small neighborhood. The absolute spatial distances of each term are calculated as follows: Extract the variogram values for the corresponding matrix positions based on the variogram model: Since its own distance is zero, and Based on the obtained variogram values, a system of Kriging equations is constructed to solve the spatial distance weights and Lagrange multiplier constraints: the first equation is... The second equation is The third equation is the standard form of the unbiased constraint equation. By performing a simultaneous linear algebraic derivation of three equations, a definite numerical solution is obtained. and and Substitute the calculated parameter weights into the mean formula to estimate the expected performance: Next, the weighting results and the variogram value are substituted into the variance formula to calculate the model confidence level: The calculated mean porosity performance prediction for unknown target mix proportions is 1.5, and the variance of the synchronous model prediction is 0.5. The entire formula substitution calculation process relies entirely on the explicit coordinate space geometric distance and the underlying variation structure function to directly perform weight allocation, eliminating subjective and blind adjustments based on engineering experience. The black-box calculation process of computer-aided optimization is fully disclosed, ensuring the effectiveness and reproducibility of the spatial prediction mechanism for various key performance indicators.
[0040] Optionally, the step of quantifying the spatial sparsity of the region to be evaluated based on the prediction variance and generating an uncertainty feedback operator, wherein the uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the prediction variance, includes: Extract the predicted variance of the matching individuals to be evaluated in the current population space; The predicted performance mean and the predicted variance are aggregated using a weighted summation method. After normalization, the predicted variance is multiplied by a weighting coefficient and then algebraically summed with the normalized predicted performance mean. For benefit-type performance indicators, positive values of the predicted performance mean are used; for cost-type performance indicators, negative values are used. This generates the uncertainty feedback operator. The predicted variance increases with the distance between the individual being evaluated and known sample points, thus improving the optimization gain of the third-generation non-dominated sorting genetic algorithm NSGA-III in the unexplored search space.
[0041] Specifically, the predicted variance of the individual combinations to be evaluated in the current population space is extracted. This predicted variance is output by the unbiased estimation underlying mathematical model, and its magnitude is positively correlated with the spatial span of the individual combinations to be evaluated from the known real physical experimental sample points. The greater the distance, the higher the spatial sparsity in that area, and the larger the predicted variance output by the model will be. Based on the upper confidence limit function logic, a weighted aggregation operation is performed on the performance prediction mean and predicted variance output synchronously by the calculation model. For benefit-type performance indicators, the performance prediction mean is directly substituted into the upper confidence limit aggregation formula; for cost-type performance indicators, the performance prediction mean is negatively evaluated and substituted into the formula, i.e., a weighted aggregation operation is performed. This approach ensures that variance rewards also play a role in encouraging exploration of high-uncertainty regions in cost-based indicators, thereby achieving a unified balance between expected predictions and spatial uncertainty across all objective dimensions in non-dominated ranking, and ultimately generating an uncertainty feedback operator. This underlying mechanism utilizes the spatial uncertainty represented by prediction variance as an exploration reward for the optimization algorithm. The prediction variance increases with the distance between the individual being evaluated and known sample points, ultimately enhancing the optimization gain of the third-generation non-dominated ranking genetic algorithm NSGA-III in the unexplored search space, forcibly guiding computational resources to conduct gridded exploration of sparse, blank areas. The confidence upper limit aggregation formula is constructed as follows: ; In the formula, U represents the final value of the generated uncertainty feedback operator; The average performance prediction given for the underlying model; The prediction variance characterizes the degree of spatial sparsity; The weighting coefficient for adjusting the ratio of global exploration space to local mining accuracy is specifically defined as the ratio of the standard critical value of the confidence interval to the historical statistical standard value of the prediction variance, with the dimension being the reciprocal of the units of prediction variance; prediction variance. pass The reciprocal of the statistical standard value of the predicted variance included in the data is used for normalization. Dimensionless numerical values; mean of performance prediction After normalization, the values are converted to dimensionless values and substituted into the formula. Both the normalized performance prediction mean and the variable are dimensionless values, allowing for direct addition. The generated uncertainty feedback operator U is a dimensionless comprehensive score used for relative ranking among individuals. Introducing this reciprocal is equivalent to normalizing the prediction variance, making the addition of values under different feature dimensions more reasonable and mathematically sound. Compared to the conventional confidence upper limit strategy using standard deviation, this invention constructs the feedback operator using variance: variance, as a secondary measure of prediction uncertainty, can produce a nonlinear response to changes in prediction confidence, providing more significant exploration incentives in sparse regions with large prediction variance. This encourages the transfer of computational resources from known local optima to unknown high-potential spaces, increasing the probability of the algorithm escaping local optima.
[0042] Spatial sparsity quantification is a technique that objectively assesses the density of known physical experimental data points in a region by measuring the variance of a mathematical model at a specific multidimensional coordinate location. Uncertainty feedback operator is a comprehensive evaluation index that integrates core information from both the expected performance of the mix design and the prediction of potential risks. The individuals to be evaluated are candidate chromosomes generated by a non-dominated sorting genetic algorithm through new crossover mutations in the current computer evolutionary generation population, containing combinations of parameters such as gradation and oil-stone ratio. The confidence upper bound function logic is a mathematical game strategy widely used in machine learning and high-dimensional search, designed to balance the algorithm's local mining of known high-yield areas with its global search of unknown high-risk areas. Weighted aggregation operation is a calculation method that assigns specific multiplier ratios to evaluation indicators of different dimensions and then performs algebraic summation to obtain a comprehensive score. Optimization gain is a numerical compensation reward obtained by the algorithm during population iterative evaluation for exploring the unknown parameter space, aimed at increasing the survival probability of individuals.
[0043] For example, an individual with a specific mix design to be evaluated, generated by the third-generation non-dominated sorting genetic algorithm NSGA-III during the fiftieth generation of population evolution calculation, is selected for system verification. The combination of the asphalt-aggregate ratio and fine aggregate gradation parameters of this individual falls precisely in the exploration margin region not fully covered by the historical database. The flexural strain performance of this individual is extrapolated using the underlying unbiased estimation model, and the mean of the performance predictions output by the model is extracted. The variance was 4.20%, and the prediction variance of the model output was extracted simultaneously. The value is 0.392. First, divide it by the historical maximum value of the bending strain index, 8.50% (i.e., the normalized baseline value), to normalize it to a dimensionless value of 0.494. Then, calculate the weighted compensation portion of the prediction variance, i.e., extract the weighting coefficients. The value is 0.85, i.e., 0.392 × 0.85 = 0.3332. Next, the normalized performance prediction mean is algebraically summed with the weighted variance compensation term, i.e., 0.494 + 0.3332 = 0.8272. The calculated comprehensive value of the uncertainty feedback operator for this individual combination to be evaluated is 0.8272. As a verification comparison, if another individual combination near an area with extremely dense known data is evaluated concurrently, although its performance prediction mean is as high as 5.00%, its prediction variance is only 0.10 due to the extremely detailed historical verification data in the surrounding area. Through the same weighted aggregation calculation—first dividing the performance prediction mean of 5.00% by the normalized baseline value of 8.50%, normalizing to 0.588, then calculating the weighted variance compensation part of 0.85 × 0.10 = 0.085, the algebraic sum of the two yields a feedback operator value of only 0.673. Data comparison shows that, based on this weighted aggregation calculation logic of spatial sparsity quantification, even if the absolute predicted mean of the mechanical performance of individuals in the unexplored search space is relatively low, they still obtain a comprehensive feedback operator score that surpasses that of high-scoring conservative individuals thanks to the high variance reward compensation given by the system.
[0044] Optionally, the step of embedding the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, and performing a multidimensional optimization operation dynamically guided by the predicted variance to generate a Pareto front solution set includes: For each performance objective, the corresponding dimension's uncertainty feedback operator is used to replace the mean performance prediction, which is then used as an indicator for evaluating the fitness of each objective dimension in the third-generation non-dominated sorting genetic algorithm NSGA-III, in order to determine the non-dominated ranking among individuals. Perform population evolution iteration, and dynamically adjust the search step size based on the predicted variance feedback results during each generation of evolution; The Pareto front solution set is output by performing association screening based on the standardized hyperplane and the reference points generated by the third-generation non-dominated sorting genetic algorithm NSGA-III.
[0045] Specifically, embedding the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, and performing multi-dimensional optimization operations dynamically guided by the predicted variance to generate the Pareto front solution set, is the core step in realizing intelligent exploration in high-dimensional space. During population initialization and iterative evolution, for each performance objective such as bending strain, tensile strain, and porosity, the uncertainty feedback operator values calculated by the underlying model in the corresponding dimension are extracted. These values replace the simple performance prediction mean for that dimension in traditional algorithms and serve as the core fitness evaluation index for that objective dimension. This allows for the simultaneous reflection of both the predicted expectations and spatial uncertainties of each dimension in multi-objective non-dominated sorting. Deep implantation of operators is achieved by modifying the definition of dominance relationships and niche selection strategies. During rapid non-dominated ranking, the dominance relationship between individuals is determined using uncertainty feedback operator values that include prediction variance compensation, allowing individuals located in high-dimensional sparse spaces to obtain higher non-dominated ranks. In the critical layer (the last non-dominated layer to be accepted), the standard NSGA-III niche retention strategy is modified by associating the individual's uncertainty feedback operator value with a reference point on the normalized hyperplane. A prediction variance adjustment weight is introduced when calculating the vertical distance from an individual to the reference line, giving individuals in unexplored areas a higher survival priority in niche competition. Based on these evaluation indicators that integrate spatial uncertainty, the current population is ranked by non-dominated rank, prioritizing individuals with excellent overall performance or those in extremely unknown exploration areas for inclusion in the next generation mating pool. A dynamic control mechanism is introduced when performing crossover and mutation operations to generate offspring. The prediction variance of the parent individuals is extracted, and the current search step size for that specific individual is calculated using a dynamic step size adjustment formula. Individuals with larger prediction variances are assigned larger search step sizes, allowing them to make greater strides in the parameter space when subjected to gene mutation perturbations. After a set number of generations of population evolution iterations, the optimal boundary set of individuals with the highest non-dominant level is extracted, and the Pareto front solution set reflecting the optimal trade-off between various engineering objectives is output. A dynamic step size adjustment formula is constructed: ; The formula for calculating parameter mutation is based on a defined search step size: ; In the formula, D is the dynamically adjusted current search step size. The preset base search step size represents the inherent exploration range of the algorithm when there is no variance reward. The dynamic feedback amplification factor is used to control the sensitivity of variance to step size amplification. It is defined as the reciprocal of the statistical standard value of the predicted variance. This reciprocal is introduced as a normalization mechanism to eliminate the dimensional differences between variables. These are the specific parameter values of the new generation of combined individuals generated after mutation operations; The parameter values of the original parent mixing individuals involved in the mutation operation; The random perturbation factor is uniformly distributed between -1 and +1. The non-dominated sorting genetic algorithm is a third-generation evolutionary computational framework based on the reference point mechanism for handling high-dimensional multi-objective optimization problems. The fitness evaluation function is a mathematical criterion for measuring the superiority or inferiority of individuals in the current digital evolutionary environment. The non-dominated ranking is a classification and grading method that divides the population into superior and inferior levels based on whether an individual is comprehensively surpassed by other competing individuals in the multi-dimensional objective space. The search step size is the absolute physical magnitude of the parameter variables shifting relative to the original parent generation when the optimization algorithm performs gene mutation to generate a new generation of individuals. The Pareto front solution set is the set of absolutely optimal non-dominated solutions that cannot be further optimized for other objectives without compromising the performance of at least one objective under rigid constraints where multiple objectives such as seepage prevention and deformation resistance conflict. The default basic search step size is set to 0.1%, which is determined based on the minimum precision limit of conventional adjustments in asphalt core wall mix proportion tests. The dynamic feedback amplification factor is set to 2.0 by default, which is normalized by taking the reciprocal of the standard distribution threshold of the predicted variance (0.5). This value is determined based on empirical fitting of the convergence curve of historical engineering data simulation optimization, effectively avoiding divergence distortion caused by excessive expansion of the variable asynchronous length. Figure 3 The figure shown is a Pareto front result generated in an embodiment of the present invention. This figure illustrates a series of non-dominated solution sets, i.e., optimal mix proportion combinations, captured by a non-dominated sorting genetic algorithm within the three-dimensional target space. These discrete star-shaped points represent the dynamic equilibrium state achieved under mutually constraining performance indicators such as porosity, flexural strain, and tensile strain. This front distribution map visually presents the optimization boundary of multi-objective optimization, thereby supporting the subsequent use of an entropy-weighted multi-attribute decision model to lock in the final target mix proportion parameter combination within the solution set.
[0046] For example, the 60th generation evolutionary process of the non-dominated sorting genetic algorithm in optimizing the oil-stone ratio parameter is selected for reproduction and verification. Parent individuals A and B are extracted from the mating pool for mutation operation. The oil-stone ratio parameter value of parent individual A is given. The variance is 6.50%. The underlying model shows that the prediction variance for individual A, being in the exploration blind zone, is... The value is 0.80. The statistical standard value of the prediction variance corresponding to this oil-stone ratio parameter dimension is extracted (this standard value is determined by the statistical distribution of the prediction variance of the oil-stone ratio parameter in the historical engineering database; because the variance dimension of the oil-stone ratio parameter is different from that of bending strain, its standard value is different from the aforementioned 5.0 for the bending strain dimension). Here, it is set to 0.50, and its reciprocal is taken as the normalization benchmark, i.e., 1 / 0.50. The dynamic feedback amplification factor ω is derived to be 2.0. The preset basic search step size is then substituted into... It is 0.1%, and the dynamic feedback amplification factor. The value is 2.0. Based on the dynamic step size adjustment formula, the dynamic search step size for individual A is calculated using the following formula: D = 0.1% × (1 + 2.0 × 0.80) = 0.1% × 2.6 = 0.26%, resulting in a dynamic search step size of 0.26% for individual A. The perturbation factor currently output by the built-in random number generator is then extracted. The value is 0.8. The parameters of the new offspring generated by individual A are derived based on the formula for calculating parameter variation. The oil-stone ratio of offspring generated by individual A jumped significantly to 6.708%. A simultaneous comparison with parent individual B, which is located in a region with dense historical data, showed the same oil-stone ratio of 6.50%, but its prediction variance... It is only 0.10. Substitute it into the same formula to calculate its dynamic search step size. At the same perturbation factor Under the direct effect of 0.8, the offspring parameters generated by individual B The derivation process is as follows: The oilstone ratio of offspring generated by individual B was only slightly adjusted to 6.596%.
[0047] Optionally, the entropy-weighted multi-attribute decision model performs a relative proximity calculation on the Pareto front solution set, and outputs a combination of target mix ratio parameters including: The Pareto front solution set is standardized to generate a dimensionless decision matrix; The objective weighting coefficients of each performance indicator are determined by calculating the information entropy of each performance indicator. Construct a weighted standardized decision matrix and extract the positive and negative reference solutions from the decision space.
[0048] Specifically, performance index data for each candidate mix design in the Pareto front solution set are extracted to construct a multi-attribute evaluation matrix. Based on the gain or loss attributes of the indicators, data range standardization is performed on each performance index to eliminate the interference of different physical dimensions on the evaluation results, generating a dimensionless decision matrix, i.e., a standardized judgment matrix. The information entropy of each performance index in this matrix is calculated to quantify the degree of variation of each index's data, and the objectively assigned index weight coefficients are derived accordingly. The standardized judgment matrix is then weighted using these index weights to generate a weighted decision matrix. Based on this weighted decision matrix, the maximum and minimum values of each index are extracted to construct a positive reference solution representing the ideal optimal state and a negative reference solution representing the ideal worst state.
[0049] For benefit-type indicators, apply the range standardization calculation formula: ; For cost-related indicators, apply the range standardization calculation formula: ; The formula for calculating the information entropy of the discrete characteristics of performance index data variation is as follows: ; An objective weighting formula based on information theory to eliminate expert bias: ; in the formula The value of the dimensionless performance index after standardization. For the j-th performance index of the i-th candidate mix design in the Pareto front solution set; This is the minimum value of this index among all candidate solutions; This is the maximum value of this index among all candidate solutions; The information entropy is calculated for the j-th performance index; m is the total number of candidate mix proportions in the Pareto front solution set; Let be the feature weight data of the i-th scheme under the j-th indicator, where , The value of the dimensionless performance index after standardization. Let be the objective evaluation weight of the j-th performance index derived from information entropy; M be the total number of performance indices participating in the comprehensive evaluation. The entropy-weighted multi-attribute decision model is a comprehensive mathematical analysis framework that combines the objective weighting principle of information theory with the logic of ranking based on the similarity of ideal solutions to solve the problem of multi-objective conflict evaluation. Data range standardization is a data preprocessing technique that uses the maximum and minimum value intervals of sample data to linearly scale the original variables to unify the feature dimensions. Information entropy is a physical characteristic quantity derived from thermodynamics and introduced into information theory to quantitatively measure the degree of disorder in the system state and the discrete variation characteristics of data. The positive reference solution is the absolute ideal benchmark coordinate point formed by virtually splicing the current optimal values of each performance index in the weighted multi-dimensional decision space. The negative reference solution is the absolute risk avoidance benchmark coordinate point formed by virtually splicing the current worst values of each performance index.
[0050] For example, the Pareto front solution set containing three non-dominated solutions, output by the non-dominated sorting genetic algorithm, is selected for multi-attribute decision data mining reproduction. This solution set includes porosity cost indicators, bending strain benefit indicators, and tensile strain benefit indicators. The original test matrix data is extracted; Scheme 1 is... , , Option 2 is , , Option 3 is , , Extract the extreme value of the first porosity index. , The second extreme value of the bending strain index , The extreme value of the third tensile strain index , Range standardization is performed. To prevent the subsequent logarithm calculation of information entropy from approaching negative infinity and becoming meaningless, a constant bias of 0.01 is set as the basis for preventing overflow. The first indicator's standard value for calculation scheme one is as follows: Similarly, substituting into the formula, we can obtain the result. , The second indicator's standard value , , and the third indicator's standard value , , The sum of the feature weights of the first indicator is calculated to be 0.01 + 0.51 + 1.01 = 1.53. The feature weights of each data element are then calculated separately. , , Substituting the number of extracted candidate solutions m=3, calculate the information entropy of the first indicator: , Due to the mathematical symmetry of the underlying test data, the second indicator, information entropy, can be rigorously calculated similarly. Information entropy, the third indicator The objective weights of each item are allocated according to the formula: Similarly, the calculation yields... , This shared objective weight directly reflects the high consistency in the numerical dispersion of seepage prevention and deformation characteristics within the current population's local optimization space. Perform matrix weighting to calculate the weighted coordinates of Scheme Two: Similarly, the weighted coordinates of Scheme 1 are derived from the extraction process. , , Scheme 3 weighted coordinates are , , By comparing extreme values, a positive reference solution is constructed using absolute coordinate combinations. Combination with absolute coordinates of negative reference solution .
[0051] Optionally, the method further includes: Calculate the first spatial distance from each mix design to the positive reference solution, and the second spatial distance to the negative reference solution; The relative proximity data is calculated based on the ratio of the second spatial distance to the sum of the first spatial distance and the second spatial distance; The mixing ratio scheme with the highest numerical value in the relative similarity data is extracted as the target mixing ratio parameter combination.
[0052] Specifically, the Euclidean distance calculation rule is introduced to calculate the first spatial distance to the positive reference solution and the second spatial distance to the negative reference solution for each mix proportion scheme. Based on the ratio of the second spatial distance to the sum of the first and second spatial distances, relative proximity data is calculated, and the mix proportion scheme with the highest relative proximity data is extracted as the final target mix proportion parameter combination to guide construction.
[0053] Formula for calculating the first spatial distance between a candidate individual and the absolute ideal benchmark coordinate point: ; Formula for calculating the second spatial distance of a candidate individual from the absolute avoidance baseline coordinate point: ; Formula for assessing the relative closeness of the overall compromise and balance among various performance objectives: ; in the formula The Euclidean space distance of the i-th candidate mix ratio from the positive reference solution is the first space distance. These are the numerical elements at the corresponding positions in the weighted decision matrix; The ideal optimal coordinate value of the positive reference solution under this index dimension; The Euclidean space distance, i.e., the second space distance, is the distance of the i-th candidate mix ratio from the negative reference solution. The worst-case coordinate value of the negative reference solution under this index dimension is used; The final output is the relative proximity data, which assesses the overall merits of the candidate mix design. The first spatial distance is the linear measure between the coordinates of a candidate mix design and the absolute ideal benchmark coordinates in the Euclidean space composed of various weighted performance indicators. The second spatial distance is the linear measure between the coordinates of the candidate mix design and the absolute risk mitigation benchmark coordinates in the same multidimensional decision space. The relative proximity score is a dimensionless decision score that comprehensively evaluates the dual spatial characteristics of the evaluation object: it is close to the positive reference solution and far from the negative reference solution.
[0054] For example, the weighted coordinates of Scheme 2 are extracted and substituted into the spatial distance formula to calculate the error range: , Substituting into the formula, we deduce the comprehensive relative similarity evaluation score for Scheme Two: The parallel penetration test during the same period yielded the following relative proximity result for Scheme 1: Option 3 is The comparison shows that 0.751 is the highest value. Based on this, it is determined that the data combination of Scheme 2 has the highest level of spatial topological distribution characteristics that approach the ideal optimum and are far from the ideal worst in the current three-dimensional space. Therefore, it is established as the target combination of parameters for external output. To quantify the technical effect of this invention and to verify the core contribution of the dynamic guidance mechanism for prediction variance, an ablation comparison experiment was conducted between the method of this invention and a control scheme that uses the same NSGA-III algorithm and the same Kriging interpolation model but removes the uncertainty feedback operator and dynamic step size adjustment mechanism. The experiment uniformly sets the initial population size to 100, the maximum number of iterations to 200, and the convergence accuracy threshold to an HV change rate of less than 10%. -4 The HV reference point was selected as the historical worst-case combination of each target performance index. The experimental data came from 129 mix proportion test samples of a pumped storage power station core wall dam project. Under the same initial population size and number of iterations, when exploring the mix proportion parameter space of a pumped storage power station core wall, the hypervolume index (HV) of the Pareto front solution set generated by the method of this invention reached 0.875, which is 22.9% higher than the control scheme HV=0.712. The algorithm of this invention only requires 85 iterations to reach the set convergence accuracy, while the control scheme requires 150 iterations, improving the optimization calculation efficiency by about 43.3%. The above experimental data quantitatively confirms that the multidimensional optimization operation guided by the dynamic nonlinearity of the prediction variance effectively overcomes the problem that traditional methods are prone to getting trapped in local optima.
[0055] To further quantify the independent contributions of each core component, two additional groups of ablation experiments were conducted: The first group removed only the uncertainty feedback operator (retaining the dynamic step-size adjustment mechanism), achieving a Pareto front HV of 0.783, a 10.0% improvement over the control scheme, and reaching convergence accuracy in 115 iterations; the second group removed only the dynamic step-size adjustment mechanism (retaining the uncertainty feedback operator), achieving a Pareto front HV of 0.801, a 12.5% improvement over the control scheme, and reaching convergence accuracy in 105 iterations. The data from these grouped experiments demonstrate that both the uncertainty feedback operator and the dynamic step-size adjustment mechanism can independently improve optimization performance, and their combined use exhibits a synergistic effect.
[0056] Optionally, the relative proximity calculation further includes: The target mix proportion parameters are associated with the construction quality control system, and a structured quality control suggestion text is output, which includes raw material index requirements, mixing temperature range parameters, paving temperature range parameters, and compaction pass parameters.
[0057] Specifically, target mix proportion parameters, such as asphalt-aggregate ratio and filler content, are used as input variables, and a pre-defined construction specification database is invoked. For raw material control, the crushing value and limit for needle-like and flaky particle content of the corresponding specifications are extracted based on the maximum aggregate size in the target mix proportion to generate raw material index requirements. For temperature control, based on the viscosity-temperature curve mapping relationship between the target asphalt-aggregate ratio and asphalt grade, mixing temperature range parameters and paving temperature range parameters that ensure the workability of the mixture and prevent thermal aging are derived. For compaction control, based on the target porosity requirements and target asphalt-aggregate ratio data, an empirical conversion model of compaction work is introduced to dynamically calculate and output the number of compaction passes required to meet the final density requirements. Finally, all the calculated and derived parameters are integrated, and according to the basic rules of comprehensive quality management (man, machine, material, method, environment), a structured quality control suggestion text that can be directly viewed and implemented by on-site technical personnel is generated. To achieve dynamic calculation of the number of compaction passes, an empirical conversion formula of compaction work is applied: ; in the formula The recommended number of compaction passes is generated dynamically. The preset baseline number of compaction passes; The porosity influence weighting coefficient has a dimension of 1 / %. The reference porosity is specified in the standard. This refers to the desired target porosity in the performance index data; The weighting coefficient for the influence of the oil-stone ratio is 1%; The target asphalt-aggregate ratio is the target mix proportion parameter. The average benchmark asphalt-aggregate ratio is based on historical engineering statistics. The target mix proportion parameter combination is the mixture formula data that achieves the optimal balance between seepage prevention and deformation, rigorously selected through a multi-attribute decision model. The construction quality control system is a digital regulatory rule library established during the construction period of the water conservancy project, covering the entire lifecycle from raw material arrival to actual compaction. Raw material index requirements are the basic physical performance qualification thresholds set for asphalt binder and various grades of coarse and fine aggregates. The mixing temperature range parameter is the upper and lower limits of the material heating and mixing temperature range to ensure that asphalt and aggregates are fully coated in the mixing drum. The paving temperature range parameter is the minimum working temperature limit that must be maintained when the mixture is unloaded and laid at the dam construction surface. The number of compaction passes parameter is the number of reciprocating compaction passes performed on the paved layer using a specific tonnage vibratory roller to achieve the designed compaction state. The structured quality control recommendation text is a digital guidance document presented in a standardized format, covering various control elements of on-site construction.
[0058] For example, the target mix proportion parameters of the core wall of a pumped storage power station, selected from the pre-output data, are used to derive and reproduce the construction quality control system. The extracted output parameter data includes the target asphalt-aggregate ratio data. The target porosity data is 6.8%. The percentage is 1.5%. The background database is pre-set to call the baseline number of compaction passes for a 22-ton self-propelled vibratory roller. Eight passes. The baseline porosity is set according to the specifications for hydraulic asphalt concrete. The value is 2.0%. The average benchmark oil-stone ratio is based on historical engineering statistics. The result was 6.5%. Based on the compaction data from previous physical engineering projects, which were sourced from a historical engineering database of a pumped storage power station, 215 valid construction samples were extracted. The goodness of fit of this empirical formula was verified. The porosity influence weighting coefficient is set to 0.89. The weighting factor for the influence of the oil-stone ratio is set to 0.5. The value is 0.2. Extract all the above parameters and substitute them into the empirical conversion formula for compaction work to calculate the number of dynamic compaction passes. Calculate the compaction work amplification term caused by porosity differences, and substitute the values to derive the result: The multiplier factor for this term is calculated to be 1.25. Next, the lubrication compensation term resulting from the high oil-stone ratio is calculated, and the derivation is performed by substituting the values: The multiplier factor for this term was calculated to be 0.94. Finally, a multiplication calculation was performed, multiplying the baseline number of compaction passes by the amplification factor and then by the compensation factor. The derivation process is as follows: The theoretical number of compaction passes was calculated to be 9.4. Since the number of compaction passes must represent a complete physical reciprocating motion, a conservative approach was adopted, rounding up to a final recommended number of compaction passes of 10. Simultaneously, based on the 6.8% asphalt-aggregate ratio, the viscosity-temperature curve of Grade 70 A road petroleum asphalt was used to directly map the mixing temperature range to 160°C to 170°C and the paving temperature range to no less than 140°C. The system then input the data into a preset report template and output a structured quality control recommendation document containing a mixing temperature of 165°C and 10 passes of vibratory compaction.
[0059] Reference Figure 4 Based on the same inventive concept, this invention also provides a system for designing mix proportions that integrates unbiased estimation and non-dominated ranking, the system comprising: The data acquisition module is used to acquire mix proportion parameters and performance index data of historical asphalt concrete core wall projects, and to construct an initial dataset containing the correlation between mix proportion parameters and performance indexes. The preprocessing module is used to perform standard score detection processing on the initial dataset, remove outlier samples, and generate the target dataset. The uncertainty modeling module is used to perform spatial correlation modeling based on the target dataset and construct a Kriging interpolation model that can simultaneously output the performance prediction mean and prediction variance. The feedback quantization module is used to quantify the spatial sparsity of the region to be evaluated based on the prediction variance and generate an uncertainty feedback operator. The uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the prediction variance. The optimization module is used to implant the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, perform multi-dimensional optimization operations guided by the prediction variance, and generate Pareto front solution set. The intelligent decision-making module is used to perform relative proximity calculation on the Pareto front solution set based on the entropy-weighted multi-attribute decision-making model and output the target mix ratio parameter combination.
[0060] The mix design system that integrates unbiased estimation and non-dominated ordination also includes a quality control output module, which is used to associate the target mix parameters with the construction quality control system and output a structured quality control suggestion text containing raw material index requirements, mixing temperature range parameters, paving temperature range parameters, and compaction pass parameters.
Claims
1. A combination design method integrating unbiased estimation and non-dominated ranking, characterized in that, include: Obtain mix proportion parameters and performance index data of historical asphalt concrete core wall projects, and construct an initial dataset containing the correlation between mix proportion parameters and performance indexes; The initial dataset is processed by standard score detection to remove outlier samples and generate the target dataset. Based on the target dataset, spatial correlation modeling is performed to construct a Kriging interpolation model that can simultaneously output the performance prediction mean and prediction variance. Based on the spatial sparsity of the region to be evaluated quantified by the predicted variance, an uncertainty feedback operator is generated. The uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the prediction variance. The uncertainty feedback operator is implanted into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, and a multi-dimensional optimization operation guided by the prediction variance is performed to generate a Pareto front solution set. The relative proximity of the Pareto front solution set is calculated based on the entropy-weighted multi-attribute decision model, and the target mix ratio parameter combination is output.
2. The combination design method integrating unbiased estimation and non-dominated ranking as described in claim 1, characterized in that, The acquisition of mix proportion parameters and performance index data for historical asphalt concrete core wall projects includes: Extract the gradation index parameters characterizing the skeleton structure of the ore, the filler content parameters characterizing the cementing performance, and the asphalt-aggregate ratio parameters from historical engineering projects to generate the mix proportion parameters, wherein the maximum particle size of the ore in the mix proportion parameters is limited to the range of 19 mm to 26.5 mm. The performance index data is generated by extracting bending strain parameters that characterize deformation coordination ability, tensile strain parameters that characterize crack resistance, and porosity parameters that characterize seepage prevention characteristics from historical engineering projects. The target range of the porosity parameters that characterize seepage prevention characteristics is limited to less than or equal to 2.0%.
3. The combination design method integrating unbiased estimation and non-dominated ranking as described in claim 1, characterized in that, The process of performing standard score detection on the initial dataset to remove outliers and generate the target dataset includes: Calculate the deviation of each data point in the initial dataset from the population mean; Divide the deviation value by the standard deviation to generate the standard score for each data point; Data points whose absolute scores exceed a set threshold are identified as outliers and removed. The set threshold is determined based on the 3σ criterion of normal distribution, and the target dataset is generated.
4. The combination design method integrating unbiased estimation and non-dominated ranking as described in claim 1, characterized in that, The step of performing spatial correlation modeling based on the target dataset to construct a Kriging interpolation model capable of simultaneously outputting the performance prediction mean and prediction variance includes: Extract the Gaussian semi-variogram or spherical semi-variogram of the target dataset in the decision space; The spatial distance weights between unknown points and known sample points are calculated based on the semi-variogram function. Perform linear unbiased estimation calculations to generate the performance prediction mean, which represents the expected performance value, and generate the prediction variance, which represents the prediction confidence, based on the estimation variance formula.
5. The combination design method integrating unbiased estimation and non-dominated ranking according to claim 4, characterized in that, The step of quantifying the spatial sparsity of the region to be evaluated based on the predicted variance to generate an uncertainty feedback operator, wherein the uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the predicted variance, including: Extract the predicted variance of the matching individuals to be evaluated in the current population space; The predicted performance mean and the predicted variance are aggregated using a weighted summation method. After normalization, the predicted variance is multiplied by a weighting coefficient and then algebraically summed with the normalized predicted performance mean. For benefit-type performance indicators, positive values of the predicted performance mean are substituted; for cost-type performance indicators, negative values of the predicted performance mean are substituted. This generates the uncertainty feedback operator. The predicted variance increases with the distance between the individual to be evaluated and the known sample points.
6. The combination design method integrating unbiased estimation and non-dominated ranking according to claim 5, characterized in that, The process of embedding the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, and performing a multidimensional optimization operation dynamically guided by the predicted variance to generate a Pareto front solution set includes: For each performance objective, the corresponding dimension's uncertainty feedback operator is used to replace the mean performance prediction, which is then used as an indicator for evaluating the fitness of each objective dimension in the third-generation non-dominated sorting genetic algorithm NSGA-III, in order to determine the non-dominated ranking among individuals. Perform population evolution iteration, and dynamically adjust the search step size based on the predicted variance feedback results during each generation of evolution; The Pareto front solution set is output by performing association screening based on the standardized hyperplane and the reference points generated by the third-generation non-dominated sorting genetic algorithm NSGA-III.
7. The combination design method integrating unbiased estimation and non-dominated ranking according to claim 6, characterized in that, The preprocessing steps for the relative proximity calculation include: The Pareto front solution set is standardized to generate a dimensionless decision matrix; The objective weighting coefficients of each performance indicator are determined by calculating the information entropy of each performance indicator. Construct a weighted standardized decision matrix and extract the positive and negative reference solutions from the decision space.
8. The combination design method integrating unbiased estimation and non-dominated ranking according to claim 7, characterized in that, The relative proximity calculation also includes: Calculate the first spatial distance from each mix design to the positive reference solution, and the second spatial distance to the negative reference solution; The relative proximity data is calculated based on the ratio of the second spatial distance to the sum of the first spatial distance and the second spatial distance; The mixing ratio scheme with the highest numerical value in the relative similarity data is extracted as the target mixing ratio parameter combination.
9. The combination design method integrating unbiased estimation and non-dominated ranking according to claim 1, characterized in that, The method further includes: The target mix proportion parameters are associated with the construction quality control system, and a structured quality control suggestion text is output, which includes raw material index requirements, mixing temperature range parameters, paving temperature range parameters, and compaction pass parameters.
10. A system for designing match ratios that integrates unbiased estimation and non-dominated ranking, characterized in that, The system includes: The data acquisition module is used to acquire mix proportion parameters and performance index data of historical asphalt concrete core wall projects, and to construct an initial dataset containing the correlation between mix proportion parameters and performance indexes. The preprocessing module is used to perform standard score detection processing on the initial dataset, remove outlier samples, and generate the target dataset. The uncertainty modeling module is used to perform spatial correlation modeling based on the target dataset and construct a Kriging interpolation model that can simultaneously output the performance prediction mean and prediction variance. The feedback quantization module is used to quantify the spatial sparsity of the region to be evaluated based on the prediction variance and generate an uncertainty feedback operator. The uncertainty feedback operator is a weighted aggregation index that integrates the performance prediction mean and the prediction variance. The optimization module is used to implant the uncertainty feedback operator into the fitness evaluation function of the third-generation non-dominated sorting genetic algorithm NSGA-III, perform multi-dimensional optimization operations guided by the prediction variance, and generate Pareto front solution set. The intelligent decision-making module is used to perform relative proximity calculation on the Pareto front solution set based on the entropy-weighted multi-attribute decision-making model and output the target mix ratio parameter combination.
