Rapid inversion method for mesoscopic fracture parameters of cement-stabilized crushed stone

Abstract

The present application discloses a method for rapid inversion of cement-stabilized gravel mesoscopic fracture parameters. By constructing an inversion process of a discrete element model-XGBoost model-particle swarm optimization model framework, it achieves multi-dimensional improvements such as model scale selection, automatic data processing, machine learning model construction, and parameter result optimization, thereby improving the efficiency of inversion and reducing the time and effort consumed by repeated manual trial calculations. It fully utilizes the multi-objective optimization advantages of the vectorized particle swarm optimization algorithm and combines it with the XGBoost model to achieve the optimal solution to the target vector. Compared with the manual trial and error method, the present method can achieve simultaneous optimization of multi-dimensional targets, thereby approaching the actual test results with the fastest efficiency and greatly improving the accuracy of the mesoscopic fracture parameters of cement-stabilized gravel.

Description

Technical Field

[0001] The present application relates to the field of mesoscopic numerical technology of cement-stabilized gravel, and in particular to a method for rapid inversion of mesoscopic fracture parameters of cement-stabilized gravel. Background Art

[0002] The mesostructure of cement-stabilized crushed stone is considered to be a three-phase heterogeneous material consisting of cement mortar, aggregate and cement mortar-aggregate interface transition zone. When the numerical model is established using the discrete element method, the accuracy of the mesoscopic parameters determines the reliability of the numerical model analysis results.

[0003] At present, the traditional method uses manual trial and error to invert mesoscopic parameters, which requires researchers to have certain experience and repeat the trial calculations many times. This method not only consumes time and energy, but also the accuracy of the obtained mesoscopic parameters is often low. In addition, the manual trial and error method has the disadvantages of low efficiency, low accuracy, and reliance on manual experience, which makes it difficult to meet the needs of actual engineering applications. In summary, the shortcomings of this method are mainly reflected in the following aspects: (1) Low efficiency. The manual trial and error method requires researchers to constantly try to adjust parameters and perform multiple numerical simulations, which consumes a lot of time and energy; (2) Low accuracy. The manual trial and error method is affected by the experience and subjective judgment of researchers, and it is difficult to ensure the accuracy of the inversion parameters; (3) Reliance on manual experience. The manual trial and error method requires researchers to have rich experience and professional knowledge, which is difficult to promote and apply. In addition, the manual trial and error method is difficult to deal with the complex mesoscopic structure and the mutual influence between parameters, resulting in the inversion results being difficult to accurately reflect the actual mechanical properties and failure process of the material. Summary of the invention

[0004] In order to solve the above technical problems, the embodiment of the present application provides a rapid inversion method for the mesoscopic fracture parameters of cement-stabilized gravel based on a discrete element model to more accurately simulate the mechanical properties and failure process of the material, improve the accuracy of pavement design and construction, extend the service life of the pavement, and ensure traffic safety.

[0005] The embodiment provided in this application relates to a method for rapid inversion of cement-stabilized crushed stone mesoscopic fracture parameters, including:

[0006] Step S1: obtaining a load-displacement curve according to a cement-stabilized crushed stone semicircular bending test, and obtaining a characteristic value of the load-displacement curve;

[0007] Step S2: establishing a cement-stabilized crushed stone discrete element model, and constructing an adjustable modeling system based on the underlying code data of the discrete element model; the adjustable parameters of the modeling system at least include specimen size, particle radius, arrangement mode, aggregate gradation and model dimension;

[0008] Step S3: Based on the discrete element cement-stabilized gravel model, multiple mesoscopic fracture parameters are screened out and the value range of each mesoscopic fracture parameter is obtained, and the XGBoost model learning samples are constructed by continuously and randomly obtaining a combination of several groups of mesoscopic fracture parameters;

[0009] Step S4: Perform model training based on the learning samples constructed in step S3 to obtain an XGBoost model of mesoscopic parameters of the cement-stabilized gravel discrete element model;

[0010] Step S5: randomly select a group of mesoscopic parameters within the range of mesoscopic parameters determined in step S3 as initial values, and input them into the XGBoost model after normalization;

[0011] The output of the XGBoost model is used as the predicted load-displacement curve eigenvalues, and these eigenvalues ​​are saved as a list for subsequent calculation of the target vector;

[0012] Step S6: construct a vectorized particle swarm optimization model framework, take the relative error between the eigenvalue of the load displacement curve in step S1 and the eigenvalue output by the XGBoost model in step S5 as the target vector, and use the relative error of each eigenvalue as a target element of the target vector to iteratively generate various micromechanical parameters;

[0013] Step S7: Obtain the output result of the XGBoost model in step S5, and feed it back to the vectorized particle swarm optimization model framework to calculate the target vector. If the accuracy requirement is met, the iteration is terminated, and the cement-stabilized gravel mesoscopic fracture parameters obtained at this time are used as the parameters that best meet the mechanical properties of cement-stabilized gravel in the experiment; if the accuracy requirement is not met, the particle swarm optimization model searches and optimizes a new set of mesoscopic parameters, and performs the next iterative calculation until the conditions for iteration termination are met.

[0014] Furthermore, the characteristic values ​​in step S1 at least include peak load, peak displacement, pre-peak slope and post-peak slope.

[0015] Furthermore, step S3 specifically includes:

[0016] Step S31: Based on the discrete element cement stabilized gravel model, traverse various parameters, take multiple values ​​for each parameter and obtain the corresponding load-displacement curve and its characteristic value;

[0017] Step S32: taking the ratio of the change rate of the characteristic value to the change rate of the parameter as the parameter sensitivity evaluation index, sorting the average values ​​of the corresponding indicators of each characteristic value and selecting a plurality of mesoscopic fracture parameters;

[0018] Step S33: optimizing and screening the value range of the above-selected mesoscopic fracture parameters based on the sensitivity analysis results;

[0019] Step S34: randomly select several groups of micro-parameter combinations and automatically delete duplicate combinations through the set data structure, substitute these combinations into the model to calculate the corresponding load-displacement curve, directly obtain the load and displacement data through the model and draw the load-displacement curve;

[0020] Step S35: extract the eigenvalues ​​of the load-displacement curve in MATLAB, and integrate them with the corresponding mesoscopic parameters to form several sets of data sets containing mesoscopic parameters and eigenvalues; divide the data sets into training sets and validation sets according to a certain ratio to form XGBoost model learning samples.

[0021] Furthermore, step S4 specifically includes:

[0022] S41: Establish input layer and set input layer feature X 1 ~X n1 ; The mesoscopic parameters selected in step S3 are used as input features, and the number of input features is the same as the number of selected mesoscopic parameters;

[0023] S42: Set the XGBoost model parameter range: "Tree depth (max_depth): range 3-10; Number of trees (n_estimators): range 50-500; Learning rate (learning_rate): range 0.01-0.3; Threshold (delta): range 0.1-1; Regularization parameters (reg_alpha, reg_lambda): range 0-1";

[0024] S43: Establish the output layer and set the output layer feature Y 1 ~Y n2 ; The load displacement curve eigenvalue extracted in step S3 is used as the output layer, and the number of output features is equal to the number of eigenvalues ​​extracted;

[0025] S44: constructing an XGBoost model objective function to measure the difference between the model prediction value and the true value, and preventing overfitting by introducing a regularization term; the true value is the characteristic value of the load-displacement curve in step S35;

[0026] S45: Use the gradient boosting algorithm to build multiple decision trees to form an XGBoost model.

[0027] Further, step S44 specifically includes:

[0028] S441: A smooth loss function based on absolute error is used to measure the difference between the model prediction value and the true value. The formula is as follows: ,in represents the sample size, is the true value of the i-th sample, is the predicted value of the ith sample, The threshold controls the sensitivity of the loss function to outliers;

[0029] S442: To prevent overfitting, a regularization term is introduced: , where M represents the number of leaf nodes in the tree, is the value of the jth leaf node, , are regularization coefficients;

[0030] S443: Combining the smooth loss function and the regularization term, the objective function of the XGBoost model is finally formed as follows: .

[0031] Further, step S45 specifically includes:

[0032] S451: Calculate the gradient of the objective function to the predicted value to guide the construction direction of the next tree;

[0033] S452: Build a tree structure using a greedy algorithm;

[0034] S453: After the tree structure is constructed, the value of each leaf node is calculated;

[0035] S454: weighted fusion of prediction results of all trees to obtain a final prediction result;

[0036] S455: By continuously iteratively adding new decision trees, the XGBoost model can gradually improve the prediction accuracy and eventually obtain a model with accurate predictions.

[0037] Furthermore, the optimization target vector of step S6 is as follows: , , where: is the characteristic value of the load-displacement curve obtained from the test, is the feature value output by XGBoost, is the relative error value of various eigenvalues, These are the target elements of the target vector of the vectorized particle swarm optimization algorithm.

[0038] Beneficial effects of the present invention:

[0039] 1) The inversion method formed can replace the traditional manual trial and error method, avoiding the disadvantages of the existing manual trial and error method of inverting mesoscopic parameters, which is not only inefficient but also has large errors in the obtained mesoscopic parameters;

[0040] 2) The mesoscopic parameter inversion method based on the discrete element model-XGBoost model-particle swarm optimization model framework constructed by this method can realize model scale selection, automatic data processing, machine learning model construction and parameter result optimization, greatly improving the efficiency of inversion and reducing the time and energy consumed by repeated manual trial calculations;

[0041] 3) Make full use of the multi-objective optimization advantages of the vectorized particle swarm optimization algorithm and combine it with the XGBoost model to achieve the optimal solution to the target vector; compared with the manual trial and error method, this method can achieve multi-dimensional target optimization at the same time, thereby approaching the actual test results with the fastest efficiency and greatly improving the accuracy of the mesoscopic fracture parameters of cement-stabilized gravel. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0043] Figure 1 is a flow chart of the method of the present invention;

[0044] Figure 2 yes Figure 1 The corresponding logic block diagram;

[0045] Figure 3 is the load displacement curve of cement stabilized crushed stone semicircular bending test;

[0046] Figure 4 It is the XGBoost model used in the present invention;

[0047] Figure 5 It is a comparison chart between the test load-displacement curve and the load-displacement curve of the parameter inversion scheme;

[0048] Figure 6 It is a comparison diagram between the test specimen failure diagram and the parameter inversion scheme failure diagram. DETAILED DESCRIPTION

[0049] In order to make the purpose, features, and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the embodiments described below are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0050] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments.

[0051] In the description of the present application, it should be understood that the terms "upper", "lower", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0052] The present application is described below in conjunction with specific embodiments:

[0053] like Figure 1 As shown, the inversion method in this application mainly includes the following steps:

[0054] Step S1: Obtain a load-displacement curve according to a cement-stabilized crushed stone semicircular bending test, and obtain a characteristic value of the load-displacement curve.

[0055] As a preferred solution, the characteristic values ​​of the load-displacement curve are extracted through MATLAB.

[0056] As a more preferred solution, the characteristic values ​​include peak load, peak displacement, pre-peak slope and post-peak slope.

[0057] Step S2: Establish a cement-stabilized crushed stone discrete element model, and construct an adjustable modeling system based on the underlying code data of the discrete element model.

[0058] As a preferred solution, the modeling code is saved as a txt file, and a modeling system that can edit the txt file is generated through Python. The modeling system here can modify the specimen size, particle radius, arrangement method, aggregate gradation and model dimension.

[0059] Step S3: Based on the discrete element cement-stabilized gravel model, multiple mesoscopic fracture parameters are screened out and the value range of each mesoscopic fracture parameter is obtained (a maximum value and a minimum value can be selected to form a value range, or a suitable parameter range is determined through literature research and experience), and XGBoost model learning samples are constructed by continuously and randomly obtaining a combination of several groups of mesoscopic fracture parameters.

[0060] As a specific embodiment, this step can be specifically detailed as follows:

[0061] Step S31: Based on the discrete element cement stabilized gravel model, various parameters are traversed, each parameter takes multiple values ​​(6 values ​​can be selected in this embodiment), the corresponding load displacement curve is calculated and MATLAB is called to extract the four eigenvalues ​​of the curve.

[0062] Step S32: taking the ratio of the change rate of the characteristic value to the change rate of the parameter as the parameter sensitivity evaluation index, sorting the average values ​​of the four characteristic values ​​corresponding to the indexes and selecting a plurality of mesoscopic fracture parameters;

[0063] Step S33: Optimize and select the value range of the above-selected mesoscopic fracture parameters based on the sensitivity analysis results.

[0064] As a specific embodiment, the top six micro-fracture parameters are selected as follows: aggregate elastic modulus, cement mortar elastic modulus, cement mortar-aggregate interface transition zone elastic modulus, aggregate tensile strength, cement mortar tensile strength and cement mortar-aggregate interface transition zone tensile strength (other parameter types include: aggregate friction coefficient, cement mortar friction coefficient, cement mortar-aggregate interface friction coefficient, cement mortar friction angle, cement mortar-aggregate interface friction angle, but in general, the six parameters selected in this embodiment are the selection results in most cases).

[0065] Step S34: According to the selected multiple mesoscopic fracture parameters and their value ranges, several groups of mesoscopic parameter combinations are randomly selected and repeated combinations are automatically deleted through the set data structure. These combinations are substituted into the model to calculate the corresponding load-displacement curves, and the load and displacement data are directly obtained through the model and the load-displacement curve is drawn.

[0066] Step S35: Extract the characteristic value of the load displacement curve in MATLAB and integrate it with the corresponding mesoscopic parameters to form a Group datasets;

[0067] The data set is divided into a certain proportion (when When , the ratio of training set to validation set is 7:3; when When , the ratio of training set to validation set is 8:2; when The training set and validation set ratio is 9:1) to form the XGBoost model learning samples.

[0068] Step S4: Perform model training based on the learning samples constructed in step S3 to obtain an XGBoost model of mesoscopic parameters of the cement-stabilized gravel discrete element model.

[0069] As a specific embodiment, this step can be specifically detailed as follows:

[0070] S41: Establish input layer and set input layer feature X 1 ~X n ; The mesoscopic parameters selected in step S3 are used as input features, and the number of input features is the same as the number of selected mesoscopic parameters;

[0071] S42: Set the XGBoost model parameter range: tree depth (max_depth): range 3-10; number of trees (n_estimators): range 50-500; learning rate (learning_rate): range 0.01-0.3; threshold (delta): range 0.1-1; regularization parameters (reg_alpha, reg_lambda): range 0-1;

[0072] S43: Establish the output layer and set the output layer feature Y 1 ~Y n ; The load displacement curve eigenvalue extracted in step S3 is used as the output layer, and the number of output features is equal to the number of eigenvalues ​​extracted;

[0073] S44: construct an XGBoost model objective function to measure the difference between the model prediction value and the true value (the true value is the characteristic value of the load-displacement curve finally obtained in step S3), and prevent overfitting through a regularization term;

[0074] As a specific embodiment, this step may adopt a smooth loss function based on absolute error, specifically including:

[0075] S441: A smooth loss function based on absolute error is used to measure the difference between the model prediction value and the true value. The formula is as follows: ,in represents the sample size, is the true value of the i-th sample, is the predicted value of the ith sample, is the threshold (controls the sensitivity of the loss function to outliers);

[0076] S442: To prevent overfitting, a regularization term is introduced: , where M represents the number of leaf nodes in the tree, is the value of the jth leaf node, YesL 1 Regularization coefficient, YesL 2 Regularization coefficient;

[0077] S443: Combine the smooth loss function and the regularization term to finally form the objective function of the XGBoost model .

[0078] S45: Construction of XGBoost model: The XGBoost model uses the gradient boosting algorithm to build multiple decision trees. The core idea of ​​the algorithm is that each tree tries to correct the prediction error of the previous tree.

[0079] As a specific example, for each sample:

[0080] S451: Calculate the gradient of the loss function with respect to the predicted value , the gradient represents the rate of change of the loss function at the current predicted value, which can guide the construction direction of the next tree.

[0081] S452: Use the greedy algorithm to build the tree structure. The basic idea of ​​the greedy algorithm is to select the optimal split node and split feature at each step so that the objective function decreases the fastest. Specifically, the algorithm will traverse all features and all possible split points, calculate the rate of decrease of the objective function after splitting, and select the feature and split point with the largest rate of decrease for splitting.

[0082] S453: After the tree structure is built, the value of each leaf node needs to be calculated. The value of a leaf node represents the predicted value of the sample that falls into the node. XGBoost determines the value of a leaf node by minimizing the objective function.

[0083] S454: Weighted fusion of the prediction results of all trees to obtain the final prediction result ,in is the initial prediction value (usually set to the mean of the target variable in the training set), represents the prediction result of the t-th tree number, is the learning rate.

[0084] S455: By continuously iteratively adding new decision trees, the XGBoost model can gradually improve the prediction accuracy and eventually obtain a model with accurate predictions.

[0085] Step S5: randomly select a group of meso-parameters within the range of meso-parameter values ​​determined in step S3 as initial values, and input them into the XGBoost model after normalization;

[0086] The output results of the XGBoost model are used as the predicted eigenvalues ​​of the load-displacement curve, and these eigenvalues ​​are saved as a list for subsequent calculation of the target vector.

[0087] Step S6: Use Python to build a vectorized particle swarm optimization model framework, take the relative error between the eigenvalue of the test load displacement curve and the eigenvalue output by the XGBoost model as the target vector, and use the relative error of each eigenvalue as a target element of the target vector. Iterate to generate various micromechanical parameters, and optimize the target vector as follows: , , where: is the characteristic value of the load-displacement curve obtained from the test, is the feature value output by XGBoost, is the relative error value of various eigenvalues, These are the target elements of the target vector of the vectorized particle swarm optimization algorithm.

[0088] Step S7: Obtain the output result of the XGBoost model in step S5, and feed it back to the vectorized particle swarm optimization model framework to calculate the target vector. If the accuracy requirement is met, the iteration is terminated, and the cement-stabilized gravel mesoscopic fracture parameters obtained at this time are used as the parameters that best meet the mechanical properties of cement-stabilized gravel in the experiment; if the accuracy requirement is not met, the particle swarm optimization model searches and optimizes a new set of mesoscopic parameters, and performs the next iterative calculation until the conditions for iteration termination are met.

[0089] After continuous iterative optimization of the particle swarm algorithm, the final output target vector elements all meet the requirements and can characterize the cement-stabilized gravel mesoscopic fracture parameters of the cement-stabilized gravel semicircular bending test.

[0090] This example demonstrates a fast inversion of mesoscopic fracture parameters of cement-stabilized crushed stone based on a discrete element model.

[0091]

[0092] Table 1 Characteristic values ​​of load-displacement curves for indoor tests

[0093] Step S2: Establish a discrete element model of the cement-stabilized crushed stone semicircular bending specimen, save the modeling code as a txt file, and generate a program that can edit the txt file through Python. Finally, a hexagonal two-dimensional discrete element model with the same specimen size and aggregate gradation as the indoor test and a particle radius of 0.5 mm is generated.

[0094] Step S3: Within the range of six mesoscopic fracture parameters of cement-stabilized crushed stone selected after sensitivity analysis, as shown in Table 2 below, 1500 groups of mesoscopic parameter combinations are randomly selected and repeated combinations are automatically deleted through the set data structure, and these combinations are substituted into the model to calculate the corresponding load-displacement curves.

[0095] The eigenvalues ​​of the load-displacement curves were extracted in MATLAB: peak load, peak displacement, pre-peak slope, and post-peak slope, and were integrated with the corresponding mesoscopic parameters to form 1500 sets of data sets containing mesoscopic parameters and eigenvalues.

[0096] The data set is divided into training set and validation set in a ratio of 7:3, that is, 1050 training sets and 450 validation sets, which constitute the XGBoost model learning samples.

[0097]

[0098] Table 2 Ranges of six mesoscopic fracture parameters

[0099] Step S4: Using the learning samples formed in step S3, model training is performed to obtain an XGBoost model of mesoscopic parameters of a cement-stabilized gravel discrete element model, including the following steps:

[0100] Step S41: Establish the input layer and set the input layer feature X 1 ~X 6 .

[0101] The six mesoscopic fracture parameters selected in step S3 are used as the input layer, and the number of input features is the same as the number of selected mesoscopic parameters.

[0102] Step S42: Set the XGBoost model hyperparameter range:

[0103] The depth of the tree (max_depth): range 3-10;

[0104] Number of trees (n_estimators): range 50-500;

[0105] Learning rate (learning_rate): range 0.01-0.3;

[0106] Threshold (delta): range 0.1-1;

[0107] Regularization parameters (reg_alpha, reg_lambda): range 0-1.

[0108] Bayesian optimization was used to minimize the mean square error of the output results as the objective function, and the expected hyperparameter combination was obtained (the tree depth was 7; the number of trees was 200; the learning rate was 0.01; the threshold was 0.9; the regularization parameter was =0.2 and =0.35).

[0109] Step S43: Establish the output layer and set the output layer feature Y 1 ~Y 4 .

[0110] The four eigenvalues ​​of the load-displacement curve calculated by the discrete element model extracted in step S3 are used as the output layer, and the number of output features is equal to the number of eigenvalues.

[0111] Step S44: Training the XGBoost model: The objective function uses a smooth loss function based on absolute error to measure the difference between the model prediction value and the true value. The formula is as follows: .

[0112] Introduce regularization terms to prevent overfitting: , where M represents the number of leaf nodes in the tree, is the value of the jth leaf node.

[0113] Use gradient boosting to build multiple decision trees:

[0114] First, calculate the gradient of the smoothed loss function with respect to the predicted value .

[0115] Then use the greedy algorithm to build the tree structure and calculate the value of each leaf node.

[0116] Finally, the prediction results of all trees are weighted and fused to obtain the final prediction result: ,in is the initial prediction value (usually set to the mean of the target variable in the training set), Represents the prediction result of the t-th tree number.

[0117] Step S5: Randomly select a set of mesoscopic parameters within the range of mesoscopic parameters determined in step S3 as initial values, and input them into the XGBoost model after normalization. The output of the model is the predicted eigenvalues ​​of the load-displacement curve, and these eigenvalues ​​are saved as a list for subsequent calculation of the target vector.

[0118] Step S6: A vectorized particle swarm optimization algorithm framework is constructed through Python. The relative error between the eigenvalue of the test load displacement curve and the eigenvalue output by the XGBoost model is used as the target vector. The relative errors of the four eigenvalues ​​are respectively used as a target element of the target vector. Each microscopic parameter is generated iteratively, and the optimized target vector is as follows: , , where: is the characteristic value of the load-displacement curve obtained from the test, is the feature value output by the XGBoost model, is the relative error value of various eigenvalues, that is, the target elements of the target vector of the vectorized particle swarm optimization algorithm.

[0119] Step S7: Use Python to call the output result of the XGBoost model in step S5, and feed it back to the vectorized particle swarm optimization algorithm framework constructed in step S6 to calculate the target vector.

[0120] If all elements of the target vector are less than 5%, the iteration is terminated, and the cement-stabilized gravel mesoscopic fracture parameters obtained at this time are the parameters that best meet the mechanical properties of cement-stabilized gravel in the test; if not, the particle swarm algorithm will optimize a new set of mesoscopic parameters, and automatically call the XGBoost model through Python to output the eigenvalues ​​for the next iterative calculation.

[0121] After continuous iterative optimization by the particle swarm algorithm, the final output target vector elements are all less than 5%, which can characterize the mesoscopic fracture parameters of cement-stabilized gravel in the semicircular bending test of cement-stabilized gravel.

[0122] like Figure 5 and Table 3, which plots the load-displacement curves obtained by calculating the microscopic parameters by inversion of the present invention and the comparison diagrams of the test results.

[0123] It can be seen intuitively from the figure that the trends of the two curves are the same. At the same time, it can be seen from the table that the relative error of peak load is 0.93%, the relative error of peak displacement is 0.34%, the relative error of pre-peak slope is 3.69%, and the relative error of post-peak slope is 3.53%, all of which are less than 5%. This shows that the cement-stabilized gravel meso-fracture parameter inversion system based on discrete element model proposed in the present invention has extremely high computational efficiency and accuracy, which greatly reduces the workload of scholars in related fields to conduct manual trial algorithm inversion parameters.

[0124]

[0125] Table 3 Comparison of experimental characteristic values ​​and characteristic values ​​of inversion results of the present invention

[0126] exist Figure 6 It can also be seen that when the inversion parameters are applied to the discrete element model, the calculated specimen failure crack diagram is similar to the actual specimen failure diagram, indicating that the present invention is helpful to further study the generation law and development trend of cement-stabilized crushed stone failure cracks at the mesoscale.

[0127] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0128] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations (such as quantity, shape, position, etc.) can be made to the technical scheme of the present invention, and these equivalent transformations are all protected by the present invention.

Claims

1. A rapid inversion method for mesoscopic fracture parameters of cement-stabilized crushed stone, characterized in that: include: Step S1: obtaining a load-displacement curve according to a cement-stabilized crushed stone semicircular bending test, and acquiring a characteristic value of the load-displacement curve; Step S2: establishing a cement-stabilized crushed stone discrete element model, and constructing an adjustable modeling system based on the underlying code data of the discrete element model; the adjustable parameters of the modeling system at least include specimen size, particle radius, arrangement mode, aggregate gradation and model dimension; Step S3: Based on the discrete element cement-stabilized gravel model, multiple mesoscopic fracture parameters are screened out and the value range of each of the mesoscopic fracture parameters is obtained, and XGBoost model learning samples are constructed by continuously and randomly obtaining a combination of several groups of mesoscopic fracture parameters; Step S4: Perform model training based on the learning samples constructed in step S3 to obtain an XGBoost model of mesoscopic parameters of the cement-stabilized gravel discrete element model; Step S5: randomly select a group of meso-parameters within the range of meso-parameter values ​​determined in step S3 as initial values, and input them into the XGBoost model after normalization; The output of the XGBoost model is used as the predicted load-displacement curve eigenvalues, and these eigenvalues ​​are saved as a list for subsequent calculation of the target vector; Step S6: construct a vectorized particle swarm optimization model framework, take the relative error of the eigenvalue of the load-displacement curve in step S1 and the eigenvalue output by the XGBoost model in step S5 as the target vector, and use the relative error of each eigenvalue as a target element of the target vector, and iteratively generate various micromechanical parameters; Step S7: Obtain the output result of the XGBoost model in step S5, and feed it back to the vectorized particle swarm optimization model framework to calculate the target vector. If the accuracy requirement is met, the iteration is terminated, and the cement stabilized gravel mesoscopic fracture parameters obtained at this time are used as the parameters that best meet the mechanical properties of cement stabilized gravel in the experiment; if the accuracy requirement is not met, the particle swarm optimization model searches and optimizes a new set of mesoscopic parameters, and performs the next iterative calculation until the conditions for iteration termination are met; Wherein, the step S3 specifically includes: Step S31: Based on the discrete element cement stabilized macadam model, traverse various parameters, take multiple values ​​for each parameter and obtain the corresponding load displacement curve and its characteristic value; Step S32: taking the ratio of the change rate of the characteristic value to the change rate of the parameter as the parameter sensitivity evaluation index, sorting the average values ​​of the corresponding indicators of each characteristic value and selecting a plurality of mesoscopic fracture parameters; Step S33: optimizing and screening the value range of the above-selected mesoscopic fracture parameters based on the sensitivity analysis results; Step S34: randomly select several groups of micro-parameter combinations and automatically delete duplicate combinations through the set data structure, substitute these combinations into the model to calculate the corresponding load-displacement curve, directly obtain the load and displacement data through the model and draw the load-displacement curve; Step S35: extract the eigenvalues ​​of the load-displacement curve in MATLAB, and integrate them with the corresponding mesoscopic parameters to form several sets of data sets containing mesoscopic parameters and eigenvalues; divide the data sets into training sets and validation sets according to a certain ratio to form XGBoost model learning samples.

2. The rapid inversion method for mesoscopic fracture parameters of cement-stabilized crushed stone according to claim 1 is characterized in that: The characteristic values ​​in step S1 at least include peak load, peak displacement, pre-peak slope and post-peak slope.

3. The rapid inversion method for mesoscopic fracture parameters of cement-stabilized crushed stone according to claim 1 is characterized in that: The step S4 specifically includes: S41: Establish the input layer and set the input layer features X1~X n1 ; The mesoscopic parameters selected in step S3 are used as input features, and the number of input features is the same as the number of selected mesoscopic parameters; S42: Set the XGBoost model parameter range; S43: Establish the output layer and set the output layer features Y1~Y n2 ; The load displacement curve eigenvalue extracted in step S3 is used as the output layer, and the number of output features is equal to the number of eigenvalues ​​extracted; S44: constructing an XGBoost model objective function to measure the difference between the model prediction value and the true value, and preventing overfitting by introducing a regularization term; the true value is the characteristic value of the load-displacement curve in step S35; S45: Use the gradient boosting algorithm to build multiple decision trees to form an XGBoost model.

4. The rapid inversion method for mesoscopic fracture parameters of cement-stabilized crushed stone according to claim 3 is characterized in that: The step S44 specifically includes: S441: A smooth loss function based on absolute error is used to measure the difference between the model prediction value and the true value. The formula is as follows: Where N represents the number of samples, y i is the true value of the i-th sample, is the predicted value of the i-th sample, δ is the threshold, and δ controls the sensitivity of the loss function to outliers; S442: To prevent overfitting, a regularization term is introduced: Where M represents the number of leaf nodes in the tree, W j is the value of the jth leaf node, γ and λ are regularization coefficients; S443: Combine the smooth loss function and the regularization term to finally form the objective function of the XGBoost model 5. The rapid inversion method for mesoscopic fracture parameters of cement-stabilized crushed stone according to claim 3 is characterized in that: The step S45 specifically includes: S451: Calculate the gradient of the objective function to the predicted value to guide the construction direction of the next tree; S452: Build a tree structure using a greedy algorithm; S453: After the tree structure is constructed, the value of each leaf node is calculated; S454: weighted fusion of prediction results of all trees to obtain a final prediction result; S455: By continuously iteratively adding new decision trees, the XGBoost model can gradually improve the prediction accuracy and eventually obtain a model with accurate predictions.

6. The rapid inversion method for mesoscopic fracture parameters of cement-stabilized crushed stone according to claim 3 is characterized in that: The optimization target vector of step S6 is as follows: Where: x i is the characteristic value of the load-displacement curve obtained from the test, y i is the eigenvalue output by XGBoost, f(i) is the relative error value of various eigenvalues, These are the target elements of the target vector of the vectorized particle swarm optimization algorithm.

Images