Concrete strength prediction method based on heterogeneous ensemble learning algorithm

Abstract

The invention discloses a concrete strength prediction method based on a heterogeneous ensemble learning algorithm, which comprises the following steps: constructing an initial data set according to concrete material parameters, curing age and strength, preprocessing data through feature selection and standardization, and training a base learner in parallel; using a Bayesian optimization algorithm to optimize hyper-parameters, and training an intensity prediction model; inputting concrete raw material parameters and curing conditions into the model to obtain a strength predicted value; and carrying out model performance evaluation and interpretability analysis. According to the concrete strength prediction method based on the heterogeneous ensemble learning algorithm, deep cooperation of feature engineering and model construction is achieved, and efficiency and model interpretability are considered while high prediction precision is guaranteed.

Description

Technical Field

[0001] This invention belongs to the field of civil engineering materials technology, specifically relating to a method for predicting concrete strength based on heterogeneous ensemble learning algorithms. Background Technology

[0002] Concrete strength is a crucial indicator for structural design and safety assessment. Accurate and rapid prediction of concrete strength is essential for concrete mix design and structural testing. Currently, concrete strength prediction typically relies on strength tests under standard curing conditions and empirical linear regression. However, these methods are time-consuming, costly, and susceptible to uncertainties such as testing methods and pouring quality. Furthermore, linear regression cannot establish a non-linear mapping relationship between material parameters and concrete strength. Therefore, proposing a strength prediction method based on multi-characteristic concrete material parameters and testing conditions is crucial.

[0003] Concrete strength is mainly affected by factors such as material type and dosage, curing age, and testing methods. Among these, the water-cement ratio and curing age are the most important factors affecting strength. The national standard "Standard for Test Methods of Long-Term Performance and Durability of Concrete" (GB50082-2009) specifies the strength testing methods and curing ages. However, the relationship between concrete strength and material dosage and curing age is not a simple linear one. Therefore, it is necessary to comprehensively consider multiple influencing factors and establish a method for predicting concrete strength based on existing mix proportions and different curing ages.

[0004] In his 2016 paper, "Research on Real-Time Strength Prediction of Different Concretes Based on Artificial Neural Networks," Fang Tao established an artificial neural network prediction model for concrete strength based on influencing factors such as material parameters, curing age, and testing conditions. This method uses machine learning algorithms to improve the accuracy of concrete strength prediction; however, the model has a large number of input feature parameters, which fails to meet the requirements for high prediction efficiency.

[0005] In his 2019 paper, "A generalized method to predict the compressive strength of high performance concrete by improved random forest algorithm," Han considered the effects of water-cement ratio and age, and selected key feature parameters to establish a random forest-based model for predicting the strength of ultra-high performance concrete. This method uses feature combinations as model input but fails to consider the correlation between material parameters and strength, leading to an overestimation of the predicted strength for low-strength concrete (<30 MPa).

[0006] The Chinese invention patent, "A Concrete Strength Prediction Method Based on Neural Network," filed on October 25, 2016, by Wuhan University of Technology, with publication number CN106568647A, predicts concrete strength based on a backpropagation (BP) neural network. This method establishes a nonlinear relationship between the mechanical properties and strength of concrete, but it does not consider the influence of factors such as water-cement ratio and material parameters, and requires rebound and core drilling tests on concrete specimens.

[0007] The Chinese invention patent, "Alkali-Activated Tailings Cementitious Activity and Concrete Strength Prediction Method Based on Machine Learning," filed on October 15, 2024, by Nanjing Institute of Technology, with publication number CN119538705A, proposes a machine learning-based concrete strength prediction model. This method simplifies the input feature dimensions of the model but requires measured heat flow data of the cementitious materials.

[0008] The Chinese invention patent, "A Concrete Strength Prediction System and Method Based on SVR," filed on December 24, 2024, by China Coal Third Construction (Group) Co., Ltd., with publication number CN119920375A, achieves rapid prediction of concrete strength based on the support vector regression algorithm. This method considers the concrete curing age but requires data such as rebound test results, compressive strength, carbonation depth, and design strength, and adds steps for regression equation fitting and model hyperparameter adjustment.

[0009] The Chinese invention patent, "A Concrete Strength Prediction Method Based on a Hybrid Model," filed on June 30, 2015, by Huaqiao University, with publication number CN104991051A, optimizes extreme learning machine, artificial neural network, and support vector machine models based on an adaptive weighted decision function to predict concrete strength. This method considers water-cement ratio, material parameters, and curing age, but requires separate training for each model, resulting in low prediction efficiency and poor interpretability.

[0010] The Chinese invention patent, "A Concrete Strength Prediction Method Based on Deep Learning," filed on January 6, 2021, by Hangu Cloud Intelligence (Wuhan) Technology Co., Ltd., with publication number CN112684156A, employs autoencoder deep learning to predict concrete strength. This method uses gradient descent to solve the loss function of the prediction model, resulting in low prediction efficiency.

[0011] In 2023, Xiao Binggang published an invention patent entitled "Prediction Model, Construction Method, Medium, and Electronic Equipment for Concrete Compressive Strength". Based on feature reconstruction and multi-layer basis learner, a prediction model was established for the relationship between concrete age, material dosage, and compressive strength.

[0012] The aforementioned machine learning methods are fragmented in their feature engineering and prediction model construction stages, making it difficult to achieve an optimal balance between accuracy, efficiency, and interpretability. Therefore, there is an urgent need for a prediction method that can deeply integrate feature selection and ensemble learning mechanisms to overcome the shortcomings of existing methods in terms of accuracy, efficiency, and interpretability. Summary of the Invention

[0013] The purpose of this invention is to provide a concrete strength prediction method based on a heterogeneous integrated learning algorithm, which solves the technical problems of traditional mechanical testing methods being time-consuming and labor-intensive, unable to handle nonlinear relationships, and traditional machine learning methods being difficult to balance prediction accuracy and efficiency.

[0014] The technical solution adopted in this invention is a concrete strength prediction method based on heterogeneous ensemble learning algorithms, as detailed below: S1. Based on the experimental mix design and publicly available databases, construct an initial dataset containing concrete raw material parameters, curing conditions, and corresponding measured strengths. D ; S2. Perform data preprocessing on the initial dataset and divide it into training and test sets; S3. Perform feature selection and dimensionality reduction on the preprocessed dataset to obtain a reconstructed dataset composed of key features and concrete strength. S4. Train an ensemble learning model based on an optimization algorithm. The ensemble learning model integrates multiple heterogeneous base learners. S5. After processing the concrete parameters to be predicted through S2 and S3, input them into the model trained in S4 to obtain the predicted strength values. S6. Evaluate and interpretability analysis of the prediction model.

[0015] The invention is further characterized by: The specific method for S1 is as follows: The input consists of features covering raw material parameters and curing conditions; the output is the concrete compressive strength. (1) In the formula, For sample feature vectors, This represents the measured strength of the concrete, expressed in MPa.

[0016] Features encompass raw material parameters and curing conditions, including cement dosage, mineral admixture dosage, fine aggregate dosage, coarse aggregate dosage, water-cement ratio, admixture dosage, curing age, and temperature.

[0017] The specific method for S2 is as follows: The initial dataset is processed using a distance- and density-based anomaly detection method—Local Outlier Factor—and then Z-score normalization is performed. The Z-score normalization method is as follows: (2) In the formula, and These are the initial dataset and the normalized dataset, respectively. and These are the mean and standard deviation of the initial data, respectively.

[0018] The specific method for S3 is as follows: The high-dimensional feature parameters were reduced using nearest-neighbor component analysis to obtain a reconstructed dataset consisting of key features and concrete strength. (3) (4) (5) In the formula, Use as a reference point; These are sample values; Features Weighting coefficients; For regularization parameters; For similarity functions; It is a distance function; Randomly select a machine learning classifier and calculate the sample points. and reference points The nearest neighbor probability is used to obtain the key feature, which is expressed as: (6).

[0019] The specific method for S4 is as follows: Training of heterogeneous ensemble learning models based on Bayesian optimization: Parallel training of extreme gradient boosting, lightweight gradient boosting machines, random forests, or shallow fully connected neural network models; using the predictions of all base learners on the validation set along with key features in S3 as new input features; and simultaneously optimizing the hyperparameters of all base learners using the Bayesian optimization algorithm. The objective function is expressed as follows: (7) In the formula, L The loss function; Represents the total number of test specimens; For the first i The predicted value of each target; T Indicates the number of decision trees; It is a regular term; This is the combination function for the decision tree.

[0020] The surrogate function constructed by the Bayesian optimization algorithm represents the objective function, and its expression is: (8) In the formula, For proxy functions, For parameter space.

[0021] The specific method for S5 is as follows: The concrete mix proportions and curing parameters are successively standardized in S2 and feature selected in S3, then input into the heterogeneous integrated learning model trained in S4, and the output is the predicted value of concrete strength.

[0022] The beneficial effects of this invention are: This invention presents a concrete strength prediction method based on a heterogeneous ensemble learning algorithm. It abandons the traditional machine learning process where feature selection and model training are separated. By performing feature dimensionality reduction through nearest neighbor component analysis and reconstructing the dataset of key features using the prediction results of the base learner, the feature information and model structure achieve optimal matching. This enables deep collaboration between feature engineering and model building, allowing for accurate and rapid prediction of concrete strength. It improves the prediction efficiency of traditional experimental testing methods and solves the problems of low accuracy and poor generalization of traditional machine learning methods. Attached Figure Description

[0023] Figure 1 This is a flowchart of the concrete strength prediction method based on heterogeneous ensemble learning algorithm of the present invention; Figure 2 This is a schematic diagram of the heterogeneous ensemble learning algorithm model structure of the present invention; Figure 3 This is a schematic diagram of the concrete strength prediction results involved in Embodiment 6 of the present invention. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0025] This invention relates to a concrete strength prediction method based on heterogeneous ensemble learning algorithms, such as... Figure 1 As shown, the specific method is as follows: S1. Based on the experimental mix design and publicly available databases, construct an initial dataset containing concrete raw material parameters, curing conditions, and corresponding measured strengths. D The specific method is as follows: The input consists of features covering raw material parameters and curing conditions; the output is the concrete compressive strength. (1) In the formula, For sample feature vectors, This represents the measured strength of the concrete, expressed in MPa.

[0026] Features encompass raw material parameters and curing conditions, including cement dosage, mineral admixture (such as fly ash and mineral powder) dosage, fine aggregate dosage, coarse aggregate dosage, water-cement ratio, admixture dosage, curing age, and temperature.

[0027] S2. Data Preprocessing: An anomaly detection method based on distance and density is used to preprocess the initial dataset. The preprocessed data is then standardized, and subsequently divided into training and test sets. The specific method is as follows: The initial dataset is processed using the Local Outlier Factor (LOF) and then Z-score normalization is performed. The Z-score normalization method is as follows: (2) In the formula, and These are the initial dataset and the normalized dataset, respectively. and These are the mean and standard deviation of the initial data, respectively.

[0028] S3. Feature selection and dimensionality reduction: The nearest neighbor component analysis method is used to reduce the dimensionality of the high-dimensional feature parameters after step S2 and select features to screen out key features, resulting in a reconstructed dataset composed of key features and concrete strength. The importance weight of each feature is calculated by optimizing the weight coefficients to maximize the classification probability of a sample among its nearest neighbors. Key features are then selected based on these weights. The specific method is as follows: The high-dimensional feature parameters were reduced using nearest-neighbor component analysis to obtain a reconstructed dataset consisting of key features and concrete strength. (3) (4) (5) In the formula, Use as a reference point; These are sample values; Features Weighting coefficients; For regularization parameters; For similarity functions; It is a distance function; Randomly select a machine learning classifier and calculate the sample points. and reference points The nearest neighbor probability is used to obtain the key feature, which is expressed as: (6).

[0029] S4. Training a Heterogeneous Ensemble Learning Model: An ensemble learning model is trained based on an optimization algorithm. This model integrates multiple heterogeneous base learners. The specific method is as follows: Training of heterogeneous ensemble learning models based on Bayesian optimization: Parallel training of Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Random Forest (RF), or Shallow Fully Connected Neural Network (FCNN) models (at least two). The predictions of all base learners on the validation set are combined with the key features in S3 as new input features. The hyperparameters of all base learners are simultaneously optimized using the Bayesian optimization algorithm. The objective function is expressed as follows: (7) In the formula, L The loss function; Represents the total number of test specimens; For the first i The predicted value of each target; T Indicates the number of decision trees; It is a regular term; This is the combination function for the decision tree.

[0030] The surrogate function constructed by the Bayesian optimization algorithm represents the objective function, and its expression is: (8) In the formula, For proxy functions; For parameter space.

[0031] S5. Concrete Strength Prediction: The concrete mix proportions and curing parameters to be predicted are processed sequentially through S2 and S3, then input into the model trained in S4 to obtain the predicted strength value. The output is the predicted concrete strength value. The specific method is as follows: The concrete mix proportions and curing parameters are successively standardized in S2 and feature selected in S3, then input into the heterogeneous integrated learning model trained in S4, and the output is the predicted value of concrete strength.

[0032] S6. Model Interpretability Evaluation: The prediction model is evaluated using test set data, within the range of root mean square error (RMSE). RMSE ) and coefficient of determination ( R 2 Based on this, the SHAP method is used to perform interpretability analysis on the ensemble learning model.

[0033] In concrete strength prediction, traditional mechanical testing methods consume significant time, manpower, and resources. Linear regression methods cannot establish the nonlinear relationship between concrete material parameters and strength, and traditional machine learning methods require manual selection of key features. The purpose of this invention is to provide a concrete strength prediction method based on a heterogeneous ensemble learning algorithm, solving the technical problems of traditional methods such as the disconnect between feature engineering and prediction model construction, and the difficulty in balancing prediction accuracy and efficiency.

[0034] Example 1 This embodiment uses a concrete strength prediction method based on a heterogeneous ensemble learning algorithm. The specific method is as follows: S1. Based on the experimental mix design and publicly available databases, construct an initial dataset containing concrete raw material parameters, curing conditions, and corresponding measured strengths. D ; S2. Perform data preprocessing on the initial dataset and divide it into training and test sets; S3. Perform feature selection and dimensionality reduction on the preprocessed dataset to obtain a reconstructed dataset composed of key features and concrete strength. S4. Train an ensemble learning model based on an optimization algorithm, wherein the ensemble learning model integrates multiple heterogeneous base learners; S5. After processing the concrete parameters to be predicted through S2 and S3, input them into the model trained in S4 to obtain the predicted strength values. S6. Model Interpretability Evaluation: The prediction model is evaluated using test set data, within the range of root mean square error. RMSE and coefficient of determination R 2 Based on this, the SHAP method is used to perform interpretability analysis on the ensemble learning model.

[0035] Example 2 This embodiment is based on a concrete strength prediction method using a heterogeneous ensemble learning algorithm. Building upon Embodiment 1, the specific method in S1 is as follows: The input consists of features covering raw material parameters and curing conditions; the output is the concrete compressive strength. (1) In the formula, For sample feature vectors, This represents the measured strength of the concrete, expressed in MPa.

[0036] Features encompass raw material parameters and curing conditions, including cement dosage, mineral admixture (such as fly ash and mineral powder) dosage, fine aggregate dosage, coarse aggregate dosage, water-cement ratio, admixture dosage, curing age, and temperature.

[0037] Example 3 This embodiment is based on a concrete strength prediction method using a heterogeneous ensemble learning algorithm. Building upon Embodiment 1, the specific method in S2 is as follows: The initial dataset is processed using the Local Outlier Factor (LOF) and then Z-score normalization is performed. The Z-score normalization method is as follows: (2) In the formula, and These are the initial dataset and the normalized dataset, respectively. and These are the mean and standard deviation of the initial data, respectively.

[0038] Example 4 This embodiment is based on a concrete strength prediction method using a heterogeneous ensemble learning algorithm. Building upon Embodiment 1, the specific method in S3 is as follows: The high-dimensional feature parameters were reduced using nearest-neighbor component analysis to obtain a reconstructed dataset consisting of key features and concrete strength. (3) (4) (5) In the formula, Use as a reference point; These are sample values; Features Weighting coefficients; For regularization parameters; For similarity functions; It is a distance function; Randomly select a machine learning classifier and calculate the sample points. and reference points The nearest neighbor probability is used to obtain the key feature, which is expressed as: (6).

[0039] Example 5 This embodiment is based on a concrete strength prediction method using a heterogeneous ensemble learning algorithm. Building upon Embodiment 1, the specific method in S4 is as follows: Training heterogeneous ensemble learning models based on Bayesian optimization: Parallel training of Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Random Forest (RF), or Shallow Fully Connected Neural Network (FCNN) models (at least two), such as... Figure 2As shown, the predictions of all base learners on the validation set are combined with the key features in S3 as new input features. The hyperparameters of all base learners are simultaneously optimized using the Bayesian optimization algorithm. The objective function is expressed as follows: (7) In the formula, L The loss function; Represents the total number of test specimens; For the first i The predicted value of each target; T Indicates the number of decision trees; It is a regular term; This is the combination function for the decision tree.

[0040] The surrogate function constructed by the Bayesian optimization algorithm represents the objective function, and its expression is: (8) In the formula, For proxy functions; For parameter space.

[0041] Example 6 This embodiment is based on a concrete strength prediction method using a heterogeneous ensemble learning algorithm. Building upon Embodiment 1, the specific method in S5 is as follows: The concrete mix proportions and curing parameters are sequentially processed through S1~S3 and input into the heterogeneous ensemble learning model trained in S4. The output is the predicted concrete strength value. Figure 3 As shown.

[0042] As can be seen from the above examples, this invention can predict concrete strength based on material usage and age. It effectively solves the problems of disconnect between feature engineering and prediction model construction in traditional methods, and the difficulty in balancing prediction accuracy and efficiency, thus achieving accurate and efficient prediction of concrete strength.

Claims

1. A method for predicting concrete strength based on heterogeneous ensemble learning algorithms, characterized in that, Specifically as follows: S1. Based on the experimental mix design and publicly available databases, construct an initial dataset containing concrete raw material parameters, curing conditions, and corresponding measured strengths. D ; S2. Perform data preprocessing on the initial dataset and divide it into training and test sets; S3. Perform feature selection and dimensionality reduction on the preprocessed dataset to obtain a reconstructed dataset composed of key features and concrete strength. S4. Train an ensemble learning model based on an optimization algorithm, wherein the ensemble learning model integrates multiple heterogeneous base learners; S5. After processing the concrete parameters to be predicted through S2 and S3, input them into the model trained in S4 to obtain the predicted strength values. S6. Evaluate and interpretability analysis of the ensemble learning model.

2. The concrete strength prediction method based on heterogeneous ensemble learning algorithm according to claim 1, characterized in that, The specific method of S1 is as follows: The input consists of features covering raw material parameters and curing conditions; the output is the concrete compressive strength. (1) In the formula, For sample feature vectors, This represents the measured strength of the concrete, expressed in MPa.

3. The concrete strength prediction method based on heterogeneous ensemble learning algorithm according to claim 2, characterized in that, The features include raw material parameters and curing conditions, such as cement dosage, mineral admixture dosage, fine aggregate dosage, coarse aggregate dosage, water-cement ratio, admixture dosage, curing age, and temperature.

4. The concrete strength prediction method based on heterogeneous ensemble learning algorithm according to claim 1, characterized in that, The specific method of S2 is as follows: The initial dataset is processed using a distance- and density-based anomaly detection method—the local outlier factor—and then Z-score normalization is performed. The Z-score normalization method is as follows: (2) In the formula, and These are the initial dataset and the normalized dataset, respectively. and These are the mean and standard deviation of the initial data, respectively.

5. The concrete strength prediction method based on heterogeneous ensemble learning algorithm according to claim 1, characterized in that, The specific method of S3 is as follows: The high-dimensional feature parameters were reduced using nearest-neighbor component analysis to obtain a reconstructed dataset consisting of key features and concrete strength. (3) (4) (5) In the formula, Use as a reference point; These are sample values; Features Weighting coefficients; For regularization parameters; It is a similarity function; It is a distance function; Randomly select a machine learning classifier and calculate the sample points. and reference points The nearest neighbor probability is used to obtain the key feature, which is expressed as: (6)。 6. The concrete strength prediction method based on heterogeneous ensemble learning algorithm according to claim 1, characterized in that, The specific method of S4 is as follows: Training of heterogeneous ensemble learning models based on Bayesian optimization: Parallel training of extreme gradient boosting, lightweight gradient boosting machines, random forests, or shallow fully connected neural network models; using the predictions of all base learners on the validation set along with key features in S3 as new input features; and simultaneously optimizing the hyperparameters of all base learners using the Bayesian optimization algorithm. The objective function is expressed as follows: (7) In the formula, L The loss function; Represents the total number of test specimens; For the first i The predicted value of each target; T Indicates the number of decision trees; It is a regularization term; This is the combination function for the decision tree.

7. The concrete strength prediction method based on heterogeneous ensemble learning algorithm according to claim 6, characterized in that, The surrogate function constructed by the Bayesian optimization algorithm represents the objective function, and its expression is: (8) In the formula, For proxy functions, For parameter space.

8. The concrete strength prediction method based on heterogeneous ensemble learning algorithm according to claim 1, characterized in that, The specific method of S5 is as follows: The concrete mix proportions and curing parameters are successively standardized in S2 and feature selected in S3, then input into the heterogeneous integrated learning model trained in S4, and the output is the predicted value of concrete strength.

Images