A concrete performance prediction method based on an orthogonal strategy and machine learning

Abstract

The application discloses a concrete performance prediction method based on an orthogonal strategy and machine learning. The method adopts the orthogonal strategy to obtain real labeled data of concrete performance under different parameter combinations, forms a training data set, and based on the training data set, comprehensively trains and optimizes multiple machine learning models, and evaluates the model prediction performance through an additional test set to build an optimal machine learning model. The final machine learning model is used to realize the concrete performance prediction under any parameter combination within a condition range. The results show that the performance prediction method can establish a performance prediction model with excellent accuracy through a small amount of real data, and can realize high-throughput prediction of concrete performance under different parameter combinations, and has a broad application prospect.

Description

Technical Field

[0001] This invention relates to a method for predicting concrete performance based on orthogonal strategies and machine learning, belonging to the field of concrete performance evaluation technology. Background Technology

[0002] As a multiphase heterogeneous material, concrete performance is closely related to its internal temperature zone (ITZ) parameters and microstructures such as aggregate volume fraction. However, traditional experimental studies can only explore the effect of aggregate volume fraction on concrete strength. In recent years, with the development of computer technology, researchers have begun to use numerical simulations to explore the effects of aggregate volume fraction and ITZ parameters on concrete performance, and have established the correlation between aggregate volume fraction or ITZ parameters and concrete strength. However, the comprehensive impact of these factors on concrete performance remains unclear.

[0003] Machine learning has become a powerful tool in modern materials science, profoundly impacting the design, discovery, and characterization of new materials, and has been widely applied in the prediction of material properties. However, to ensure the accuracy of the model, a huge dataset is usually required. But based on orthogonal strategies, even a small amount of data can guarantee the accuracy of machine learning models. Therefore, there is an urgent need to propose a concrete performance prediction method based on orthogonal strategies and machine learning, aiming to obtain real data on concrete performance with different ITZ parameters and aggregate volume fractions, laying the foundation for a comprehensive understanding of the influence of these micro-components on concrete performance. Summary of the Invention

[0004] Purpose of the invention: In view of the shortcomings of the prior art, the purpose of this invention is to provide a concrete performance prediction method based on orthogonal strategy and machine learning, so as to solve the problem that existing research is unable to accurately predict the macroscopic properties of concrete based on the microscopic components of each phase.

[0005] Technical solution: The present invention provides a method for predicting concrete performance based on orthogonal strategies and machine learning, comprising the following steps:

[0006] (1) Set the interface transition zone (ITZ) and aggregate parameters and their boundary thresholds, select parameter combinations based on the orthogonal strategy, i.e. parameter set A, and use numerical simulation to calculate the corresponding concrete performance as real annotation data to form a model training set, i.e. dataset A.

[0007] (2) Build 29 machine learning models, train the models based on dataset A, and initially select 18 more suitable machine learning models based on the training results;

[0008] (3) The above 18 machine learning models are used to preliminarily predict the trend of concrete performance under the influence of some ITZ and / or aggregate parameters other than the parameter combination in dataset A, i.e. parameter set B. Numerical simulation is used to obtain the real labeled dataset of concrete performance under the influence of each parameter, forming the model test set, i.e. dataset B, to verify the feasibility of the 18 machine learning models.

[0009] (4) Using dataset B as training data, further enhance the training of the above 18 machine learning models. Randomly select parameter combinations that are not in dataset A and dataset B, i.e. parameter set C, and use numerical simulation to calculate the real labeled data of concrete performance to form a model test set, i.e. dataset C, to verify the accuracy of the 18 machine learning models after enhanced training.

[0010] (5) Select the machine learning model with the best prediction effect based on the test results of step (4), and test its accuracy with dataset C;

[0011] (6) Based on the optimal machine learning model, concrete performance values ​​under different combinations of ITZ and aggregate parameters are obtained and cloud maps are formed to realize performance prediction of any combination of parameters within the boundary threshold range.

[0012] Among them, concrete performance refers to the mechanical properties, ion transport properties and / or thermal conductivity of concrete.

[0013] In step (1), the ITZ parameters are one or more of the following: ITZ thickness, ITZ strength, ITZ thermal conductivity, and ITZ moisture / ion diffusion coefficient.

[0014] In step (1), the aggregate parameters are one or more of aggregate gradation, shape, and volume fraction.

[0015] In step (1), setting the ITZ parameter and aggregate parameter and their boundary threshold parameter refers to setting the interval and boundary of the ITZ and aggregate parameter.

[0016] In steps (1), (3), and (4), the numerical simulation is mainly implemented using COMSOL and / or ABAQUS numerical simulation software.

[0017] In steps (2), (3), (4), and (5), the accuracy of the machine learning model is measured by the mean squared error (MSE) and / or the coefficient of determination (R²). 2 To measure it.

[0018] In step (2), all 29 machine learning models are built based on five commonly used machine learning algorithms: linear regression (LR), support vector machine (SVM), artificial neural network (ANN), Gaussian process regression (GPR), and decision tree (DT).

[0019] In step (5), the best-performing machine learning model is the one with the smallest MSE and R. 2 The model closest to 1.

[0020] In step (6), the concrete performance cloud map under different combinations of ITZ and aggregate parameters is formed based on MATLAB and / or Origin plotting software.

[0021] Invention principle: This invention is based on commonly used machine learning models at present. It uses an orthogonal dataset with parameter boundary thresholds to optimize step by step. Finally, based on the machine learning model with the best performance, it realizes accurate prediction of concrete performance under different parameter combinations.

[0022] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0023] This invention employs orthogonal datasets and considers the impact of boundary thresholds on concrete performance, overcoming the drawback of existing machine learning models requiring massive datasets to ensure prediction accuracy. It also addresses the difficulty in accurately predicting the macroscopic properties of concrete based on its micro-components. This performance prediction method can establish a highly accurate performance prediction model using a small amount of real data and achieve high-throughput prediction of concrete performance under different parameter conditions, demonstrating broad application prospects. Attached Figure Description

[0024] Figure 1 This is a cloud map showing the predicted compressive strength of concrete. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0026] Example 1

[0027] (1) Set the ITZ thickness, ITZ strength, and aggregate volume fraction and their boundary thresholds. Since the compressive strength of concrete is mainly related to the ITZ thickness, ITZ strength, and aggregate volume fraction, a concrete compressive strength prediction model is proposed based on these three parameters. According to the literature, the ITZ thickness is usually 0-500μm with 50μm intervals; the ITZ strength is usually 50-100% of the mortar strength with 5% intervals; and the aggregate volume fraction is usually 0-60% with 5% intervals. Based on the orthogonal strategy, parameter combinations are selected, i.e., parameter set A (see Table 1). The ABAQUS software is used to calculate the corresponding concrete compressive strength based on the concrete damage plasticity model as the real labeled data to form the model training set, i.e., dataset A. The specific parameter settings in the simulation process are shown in Table 2.

[0028] Table 1 Parameter Set A (94 points in total)

[0029]

[0030] Table 2 Parameters of various microstructure components in concrete

[0031]

[0032] Note: In the table, ρ represents density; E represents elastic modulus; μ represents Poisson's ratio; f represents density. c f is the compressive strength; t This represents the tensile strength.

[0033] (2) Build 29 machine learning models (as shown in Table 3), train the models based on dataset A, and preliminarily select 18 more suitable machine learning models based on the training results (as shown in Table 4).

[0034] Table 3. Structure of 29 machine learning models

[0035]

[0036]

[0037] Table 4. Structure of 18 machine learning models

[0038]

[0039]

[0040] (3) The above 18 machine learning models were used to preliminarily predict the trend of concrete compressive strength under the influence of ITZ thickness, ITZ strength and aggregate volume fraction in parameter set B shown in Table 5. The numerical simulation was used to obtain the real labeled dataset of concrete compressive strength under the influence of these parameters, forming the model test set, i.e. dataset B, to verify the feasibility of the 18 machine learning models.

[0041] Table 5 Parameter Set B (22 points in total)

[0042]

[0043] (4) Using dataset B as training data, further enhance the training of the above 18 machine learning models. In addition, randomly select parameter combinations that are not in datasets A and B (see Table 6, i.e. parameter set C), and use numerical simulation to calculate the real labeled data of concrete compressive strength to form a model test set, i.e. dataset C, to verify the accuracy of the machine learning model after enhanced training.

[0044] Table 6 Parameter set C (28 points in total)

[0045]

[0046] (5) Select the best machine learning model based on the test results (see Table 7) and test its accuracy using dataset C;

[0047] Table 7. Structure of the optimal machine learning model

[0048]

[0049] (6) Based on the best-performing machine learning model (ANNs1), the compressive strength of concrete with different parameters to be predicted is predicted, and finally a result is formed as follows: Figure 1 The values ​​of concrete compressive strength are shown for different ITZ thicknesses (0-500 μm, in 5 μm intervals), ITZ strengths (50-100%, in 0.5% intervals), and aggregate volume fractions (0-60%, in 0.5% intervals).

Claims

1. A method for predicting concrete performance based on orthogonal strategies and machine learning, characterized in that, Includes the following steps: (1) Set ITZ and aggregate parameters and their boundary thresholds, select parameter combinations based on orthogonal strategy, i.e. parameter set A, and use numerical simulation to calculate the corresponding concrete performance as real annotation data to form a model training set, i.e. dataset A; Setting ITZ and aggregate parameters and their boundary thresholds refers to setting the interval and boundary of ITZ parameters and aggregate parameters. ITZ parameters are one or more of ITZ thickness, ITZ strength, ITZ thermal conductivity, and ITZ moisture / ion diffusion coefficient. (2) Build 29 machine learning models, train the models based on dataset A, and preliminarily select 18 more suitable machine learning models based on the training results; (3) The above 18 machine learning models are used to preliminarily predict the change trend of concrete performance under the influence of some ITZ and / or aggregate parameters other than the parameter combination in dataset A, i.e. parameter set B. The numerical simulation is used to obtain the real labeled dataset of concrete performance under the influence of the parameters, forming the model test set, i.e. dataset B, to verify the feasibility of the 18 machine learning models. (4) Using dataset B as training data, further enhance the training of the above 18 machine learning models, randomly select parameter combinations that are not in dataset A and dataset B, i.e. parameter set C, and use numerical simulation to calculate the real labeled data of concrete performance to form a model test set, i.e. dataset C, to verify the accuracy of the 18 machine learning models after enhanced training. (5) Select the machine learning model with the best prediction effect based on the test results of step (4), and test its accuracy with dataset C; (6) Based on the optimal machine learning model, concrete performance values ​​under different combinations of ITZ and aggregate parameters are obtained and cloud maps are formed to realize performance prediction of any combination of parameters within the boundary threshold range.

2. The prediction method according to claim 1, characterized in that, Concrete properties refer to the mechanical properties, ion transport properties, and / or thermal conductivity of concrete.

3. The prediction method according to claim 1, characterized in that, In step (1), the aggregate parameters are one or more of aggregate gradation, shape, and volume fraction.

4. The prediction method according to claim 1, characterized in that, In steps (1), (3) and (4), the numerical simulation is performed using COMSOL and / or ABAQUS numerical simulation software.

5. The prediction method according to claim 1, characterized in that, In steps (2), (3), (4), and (5), the accuracy of the machine learning model is measured by the mean squared error (MSE) and / or the coefficient of determination (R²). 2 )measure.

6. The prediction method according to claim 1, characterized in that, In step (2), all 29 machine learning models are constructed based on five machine learning algorithms: linear regression (LR), support vector machine (SVM), artificial neural network (ANN), Gaussian process regression (GPR), and decision tree (DT).

7. The prediction method according to claim 1, characterized in that, In step (5), the best-performing machine learning model is the one with the smallest MSE and R. 2 The model closest to 1.

8. The prediction method according to claim 1, characterized in that, In step (6), the concrete performance cloud maps under different combinations of ITZ and aggregate parameters are generated based on MATLAB and / or Origin plotting software.

Images