A concrete performance intelligent prediction method based on multi-source data reinforcement

Abstract

This invention discloses an intelligent prediction method for concrete performance based on multi-source data enhancement, belonging to the interdisciplinary field of concrete design and artificial intelligence. The method first constructs a multi-source heterogeneous database and uses an unsupervised anomaly detection algorithm to clean abnormal samples. For missing data, an adaptive imputation method based on an ensemble meta-learning model is constructed to achieve differentiated and accurate data completion. A two-dimensional feature weighting framework is constructed based on nonlinear contribution and linear redundancy as a penalty factor. Simultaneously, sample-level confidence labels are calculated based on three dimensions: completion error, feature importance, and missing rate, quantifying the reliability of the completed data. Finally, feature weights and sample confidence are jointly embedded into model training to achieve intelligent prediction of concrete performance driven by feature and sample collaboration. This invention significantly improves the prediction accuracy and robustness of the model under multi-source heterogeneous data through progressive data enhancement, providing high-quality data support for mix design optimization with low experimental costs.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of concrete design and artificial intelligence, specifically involving an intelligent prediction method for concrete performance based on multi-source data enhancement. It is particularly suitable for high-precision prediction and mix proportion optimization of properties such as compressive strength, slump, chloride ion diffusion coefficient, shrinkage rate, and porosity of cement-based materials. Background Technology

[0002] Data-driven concrete performance prediction and intelligent mix design are current research hotspots and key technological directions in the field of civil engineering materials. By collecting a large amount of concrete mix proportion and performance data, and using machine learning models such as support vector machines, random forests, and extreme gradient boosting for performance prediction, combined with optimization algorithms for multi-objective mix proportion optimization, this approach can effectively overcome the problems of low efficiency and large bias in traditional empirical formulas, as well as the difficulty in balancing multiple performance and objectives in standard-based mix design methods.

[0003] However, existing technologies generally focus on improving the algorithm of the prediction model itself, while neglecting the fundamental factor affecting the upper limit of model performance: the quality of the underlying data. Current concrete machine learning prediction databases all face serious data quality challenges: First, data sources are diverse, including academic papers, engineering measurements, and laboratory test results. Data from different sources has inconsistent formats, recording methods, and precision, and missing values ​​are common. Second, there are both linear and monotonic relationships between various mix proportion parameters and concrete performance, as well as complex nonlinear interactions. Traditional single imputation methods (such as mean imputation, regression imputation, and expectation-maximization imputation) cannot effectively handle different types of missing data simultaneously. Imputation of highly discrete and strongly nonlinear features, such as water-reducing agent dosage, seriously affects the prediction accuracy of subsequent models. Third, most existing technologies ignore the quality differences between imputed data and real data, treating them equally in model training, leading to low-quality imputed data misleading model training. Furthermore, existing methods do not perform feature weighting based on the differences in the true contribution of mix proportion parameters to concrete performance, making it difficult to establish a high-quality input-high-precision output mapping relationship.

[0004] Therefore, there is an urgent need to develop a method that can address the issues of high-precision completion of missing values ​​in multi-source heterogeneous concrete data, quantitative assessment of the quality of completed data, and collaborative model training driven by high-dimensional features and sample accuracy, starting from the source of data quality, so as to fundamentally break through the data quality bottleneck that restricts the performance of machine learning prediction models for concrete performance. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent prediction method for concrete performance based on multi-source data enhancement, which solves the problems of low accuracy of missing data completion, loss of high-dimensional feature information, and inability to quantify and utilize the quality of completed data in existing multi-source concrete data. This invention systematically enhances the prediction model from the source of data quality, thereby significantly improving the accuracy and robustness of the concrete performance prediction model.

[0006] The present invention adopts the following technical solution:

[0007] A method for intelligent prediction of concrete performance based on multi-source data enhancement includes the following steps:

[0008] Step 1: Collect concrete mix proportion parameters and corresponding performance data from different data sources to construct a multi-source heterogeneous database of concrete performance.

[0009] Step 2: Perform outlier detection and cleaning on the complete data sample in the database to obtain the cleaned database;

[0010] Step 3: For the incomplete data samples in the database, an adaptive interpolation method is used to complete the data and output a complete database without missing data; the incomplete data samples are those that are missing at least one feature data.

[0011] Step 4: Based on the completed database, construct a two-dimensional feature weight calculation framework that includes feature nonlinear contribution and feature linear redundancy, and generate feature-level scaling weights that retain high-dimensional information. The feature nonlinear contribution is used to represent the nonlinear contribution of features in the completed database, and the feature linear redundancy is used to represent the linear redundancy between features in the completed database.

[0012] Step 5: For each completed sample in the completed database, calculate the sample-level confidence label to characterize the reliability of the completed sample data;

[0013] Step 6: Introduce the feature-level scaling weights and sample-level confidence labels into the machine learning model training process. Use the feature-level scaling weights obtained in Step 4 to perform weighted scaling on the input features, and use the sample-level confidence labels obtained in Step 5 to perform confidence weighting on the training samples, thereby constructing a concrete performance prediction model based on the collaborative driving of feature space and sample quality; and output the concrete performance prediction results according to the prediction model.

[0014] Furthermore, in step 1, each record in the multi-source heterogeneous database of concrete performance includes multiple mix proportion characteristic parameters and at least one target performance index. The data in the database comes from various sources, including academic papers, engineering measurements, and laboratory test results. The characteristic parameters are selected from mix proportion parameters that affect concrete performance, such as cement type, cement design strength, water-cement ratio, cement dosage, dosage of each mineral admixture, coarse aggregate dosage, coarse aggregate gradation, sand dosage, sand fineness modulus, sand ratio, maximum aggregate particle size, water-reducing agent dosage, water-reducing agent water reduction rate, water consumption, and age. The target performance index is selected from concrete performance indicators such as compressive strength, tensile strength, flexural strength, slump, slump spread, chloride ion diffusion coefficient, creep, shrinkage, or porosity.

[0015] Furthermore, in step 2, the outlier detection and cleaning process is obtained using an unsupervised anomaly detection method adapted to the characteristics of engineering data. The unsupervised anomaly detection method is selected from any one of the following: Isolation Forest algorithm, Local Outlier Factor (LOF) algorithm, clustering-based anomaly detection algorithm, or Mahalanobis distance-based multivariate consistency test algorithm. After cleaning or removing outlier samples, the coefficient of variation of the database decreases, the data distribution becomes more concentrated, and the mean and median tend to be close.

[0016] Furthermore, in step 3, the adaptive interpolation method uses the Lost Forest algorithm as the model basis. Based on the linear correlation between the feature to be completed and other features and physical constraints, it distinguishes between strongly linearly correlated features and nonlinearly complex correlated features. It constructs an ensemble meta-learning model based on Lost Forest, Gradient Boosting Regression (GBR), and K-Nearest Neighbors Regression (KNN) for weighted interpolation, and finally obtains a complete database without missing features.

[0017] Furthermore, the ensemble meta-learning model calculates the prediction error of each base learner through cross-validation, performs a normalized weighted average with the reciprocal of the root mean square error (RMSE) as the weight, and dynamically adjusts the weights of each base learner during the iteration process.

[0018] Furthermore, in step 4, the dual-dimensional feature weight calculation framework uses the nonlinear contribution of features as the base weight and the linear redundancy between features as the penalty term to automatically reduce the weight of highly collinear features to obtain a composite weight; and further performs power smoothing and normalization processing on the composite weight to generate feature-level weights that retain high-dimensional information.

[0019] Furthermore, the nonlinear contribution of the feature is obtained based on the SHAP interpretability analysis results of the surrogate model, which can be any one of robust machine learning models such as Support Vector Machine, Random Forest, Extreme Gradient Boosting, or Lightweight Gradient Boosting Machine. The linear penalty term is obtained based on any one of the correlation analysis methods such as Spearman rank correlation coefficient, mutual information, or Pearson correlation coefficient to obtain the linear monotonicity and linear redundancy of the feature.

[0020] The formula for the linear penalty term is:

[0021]

[0022] in, : The linear penalty term for feature i, : Linear correlation coefficient between features i and j.

[0023] The formula for calculating the composite weight is:

[0024]

[0025] in, For the nonlinear weights of feature i obtained based on SHAP, is the composite weight of feature i.

[0026] To avoid the model becoming overly reliant on specific variables due to excessively high weights for a single feature, and to improve the balanced contribution of each proportion parameter, power smoothing is introduced, as shown in the formula:

[0027]

[0028] in, The weights are the power-smoothed values ​​of feature i. This is the power transformation term for weight smoothing, with a value of [0,1].

[0029] The final feature weights used for training the machine learning model are:

[0030]

[0031] in, for Assigning normalized weights.

[0032] Furthermore, in step 5, the sample-level confidence label comprehensively considers three dimensions: feature completion error, feature importance, and data missing rate. Based on a preset benchmark value, a "completion error adjustment term (E)" is assigned according to the completion accuracy of each feature. adj "Importance Adjustment Item (F)" is assigned to the global contribution level of each feature to the target performance index. adjThe term "Missing Rate Adjustment Item (M)" is assigned based on the interval between the feature missing rate and the sample missing rate. adj The reliability of each completed data point is quantified using a weighted summation method. The confidence level of the original, non-missing data is set to 1 or the maximum value. The confidence level of the completed data is dynamically determined based on the aforementioned dimensions, as shown in the following formula:

[0033]

[0034] in, Final confidence level; Baseline confidence level; , , The weights of each adjustment item. .

[0035] Furthermore, based on the linear normalization of the root mean square error (RMSE) during each feature interpolation process, the "completion error adjustment term" for the confidence label is calculated using the following formula:

[0036]

[0037] in, To complete the error adjustment term; For maximum positive adjustment; This is the maximum negative adjustment; The result is the linearly normalized root mean square error (RMSE) of the interpolation of feature j.

[0038] Further, based on the feature-level scaling weights obtained in step 4, the importance adjustment term is calculated, and all features are divided into at least two contribution levels according to the scaling weights from high to low. Features belonging to the high contribution level are given positive adjustment to enhance the influence weight of their supplementary data in model training, and features belonging to the low contribution level are given negative adjustment to weaken the noise interference they may introduce.

[0039] The missing rate adjustment term is calculated based on the overall missing rate. The formula for the overall missing rate is:

[0040]

[0041] in, The overall missing rate is %; The feature missing rate is %; The sample missing rate is %; and They are respectively and The weight, ;

[0042] All features are divided into at least two missing levels according to the overall missing rate from high to low. Features belonging to the low missing level are positively adjusted, and features belonging to the high missing level are negatively adjusted.

[0043] Further, in step 6, during the machine learning model training process, the feature-level scaling weights obtained in step 4 are introduced to scale the input features, and the sample-level confidence labels obtained in step 5 are introduced to weight the samples with confidence, thereby achieving feature-sample level collaborative data quality enhancement-driven model training. The machine learning model can be any one of the following suitable models for handling high-dimensional feature prediction: Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), K-Nearest Neighbors algorithm, Kernel Ridge Regression, Extreme Random Tree, Gradient Boosting Decision Tree, etc. Any one of the following hyperparameter automatic tuning methods is used to optimize the hyperparameters of the selected training model: RIME, Bayesian Optimization (BO), Particle Swarm Optimization (PSO), Dung Beetle Optimization (DBO).

[0044] The beneficial effects of this invention are as follows:

[0045] 1. This invention integrates data augmentation throughout the entire process of preprocessing, feature engineering, and model training, forming a five-step progressive data augmentation closed loop of "abnormal data cleaning - micro-precise repair - macro-structure preservation - quality calibration for completion - collaborative bias training". It can be adapted to various concrete performance indicators and machine learning models, providing a high-quality data and model foundation for data-driven intelligent design of concrete mix proportions.

[0046] 2. This invention effectively prevents abnormal data from contaminating subsequent imputation and model training by intelligently cleaning high-dimensional abnormal samples.

[0047] 3. The adaptive fusion imputation method proposed in this invention achieves differentiated and accurate repair of multiple types of missing data, solving the problem that traditional single methods cannot adapt to complex missing patterns.

[0048] 4. The dual-dimensional feature weight calculation framework proposed in this invention adopts a weight synthesis strategy of "nonlinear basis + linear penalty", which solves the collinearity problem while fully preserving the high-dimensional nonlinear information.

[0049] 5. The sample-level confidence label proposed in this invention enables refined quantification and differentiated utilization of the reliability of the completed data, avoiding interference from low-quality completed data on the model. Attached Figure Description

[0050] Figure 1 This is a flowchart of the intelligent prediction method for concrete performance based on multi-source data enhancement according to the present invention.

[0051] Figure 2 This is a comparison chart of errors for different data interpolation methods. (a) represents the mean absolute error (MAE); (b) represents the root mean square error (RMSE).

[0052] Figure 3 This is a feature importance graph of SHAP based on the RF proxy model.

[0053] Figure 4 A heatmap showing the Spearman rank correlation results among the features.

[0054] Figure 5 The figures show a comparison of the prediction accuracy of the concrete compressive strength model. (a) shows the effectiveness verification results based on RIME automatic hyperparameter tuning; (b) shows a comparison between the model's predicted values ​​and the actual values. Detailed Implementation

[0055] like Figure 1 As shown in the figure, this invention provides a method for intelligent prediction of concrete performance based on multi-source data enhancement, comprising the following steps:

[0056] Step 1: Obtain concrete mix proportion parameters and corresponding performance data from different data sources to construct a multi-source heterogeneous database of concrete performance.

[0057] Each record in the database contains multiple mix proportion characteristic parameters and at least one target performance index. The data in the database comes from various sources, including academic papers, engineering measurements, and laboratory test results. The characteristic parameters include: cement type, cement design strength, water-cement ratio, cement dosage, dosage of each mineral admixture, coarse aggregate dosage, coarse aggregate gradation, sand dosage, sand fineness modulus, sand ratio, maximum aggregate particle size, water-reducing agent dosage, water-reducing agent water reduction rate, water consumption, and age—all mix proportion parameters affecting concrete performance. The target performance index includes: concrete compressive strength, tensile strength, flexural strength, slump, slump spread, chloride ion diffusion coefficient, creep, shrinkage, and porosity—all concrete performance indicators.

[0058] Step 2: Perform outlier detection and removal on the complete data sample in the database to obtain the cleaned database.

[0059] The outlier detection and removal employs an unsupervised anomaly detection method adapted to the characteristics of engineering data, including any one of the following: the Isolation Forest algorithm, the Local Outlier Factor (LOF) algorithm, a clustering-based anomaly detection algorithm, or a multivariate consistency test algorithm based on Mahalanobis distance.

[0060] Step 3: For the multiple types of missing data in the database, an adaptive imputation method based on model fusion is used to complete the data and output a complete database without missing data.

[0061] The adaptive imputation method uses the Miss-Forest algorithm as its model base. Based on the linear correlation between the feature to be imputed and other features, and physical constraints, it distinguishes between strongly linearly correlated features and nonlinearly complex correlated features. An ensemble meta-learning model based on Miss-Forest, Gradient Boosting Regression (GBR), and K-Nearest Neighbors (KNN) is constructed for weighted imputation. This ultimately yields a complete database without missing features. During this process, the ensemble meta-learning model calculates the prediction errors of each base learner through cross-validation, performs a normalized weighted average using the reciprocal of the root mean square error (RMSE) as the weight, and dynamically adjusts the weights of each base learner during iteration.

[0062] Step 4: Based on the completed database, construct a two-dimensional feature weight calculation framework that includes feature nonlinear contribution and feature linear redundancy, and generate feature-level scaling weights that retain high-dimensional information.

[0063] The dual-dimensional feature weight calculation framework uses the nonlinear contribution of features as the base weight and the linear redundancy between features as the penalty term. It automatically reduces the weight of highly collinear features to obtain a composite weight. Furthermore, it performs power smoothing and normalization on the composite weight to generate feature-level weights that retain high-dimensional information.

[0064] The nonlinear contribution of the feature is obtained based on the SHAP interpretability analysis results of the surrogate model, which can be any one of robust machine learning models such as Support Vector Machine, Random Forest, Extreme Gradient Boosting, or Lightweight Gradient Boosting Machine. The linear penalty term is obtained based on any one of the correlation analysis methods such as Spearman rank correlation coefficient, mutual information, or Pearson correlation coefficient to obtain the linear monotonicity and linear redundancy of the feature.

[0065] The formula for the linear penalty term is:

[0066]

[0067] in, : The linear penalty term for feature i, : Linear correlation coefficient between features i and j.

[0068] The formula for calculating the composite weight is:

[0069]

[0070] in, For the nonlinear weights of feature i obtained based on SHAP, is the composite weight of feature i.

[0071] To avoid the model becoming overly reliant on specific variables due to excessively high weights for a single feature, and to improve the balanced contribution of each proportion parameter, power smoothing is introduced, as shown in the formula:

[0072]

[0073] in, The weights are the power-smoothed values ​​of feature i. This is the power transformation term for weight smoothing, with a value of [0,1].

[0074] The final feature weights used for training the machine learning model are:

[0075]

[0076] in, for Assigning normalized weights.

[0077] Step 5: Calculate the sample-level confidence label for each completed sample in the completed database.

[0078] The sample-level confidence label comprehensively considers three dimensions: feature completion error, feature importance, and data missing rate. Based on a preset benchmark value, it assigns a "completion error adjustment term (E)" according to the completion accuracy of each feature. adj "); Based on the global contribution level of each feature to the target performance index, an "importance adjustment item (F)" is assigned. adj "); Based on the intervals to which the feature missing rate and sample missing rate belong, assign a "missing rate adjustment term (M)"; adj The reliability of each completed data point is quantified using a weighted summation method. The confidence level of the original, non-missing data is set to 1 or the maximum value. The confidence level of the completed data is dynamically determined based on the aforementioned dimensions, as shown in the following formula:

[0079]

[0080] in, Final confidence level; Baseline confidence level; , , The weights of each adjustment item. .

[0081] In this process, based on the linear normalization of the root mean square error (RMSE) during each feature interpolation, the "completion error adjustment term" for the confidence label is calculated, using the following formula:

[0082]

[0083] in, To complete the error adjustment term; For the maximum positive adjustment, the assigned value range is +0.5 to +0.15; For the maximum negative adjustment, the value range is -0.5 to -1.5; The result is the linearly normalized root mean square error (RMSE) of the interpolation of feature j.

[0084] The "feature importance adjustment term" is based on the feature-level scaling weight values ​​obtained in step 4. It divides all features into four levels: core features (top 20%), important features (20%–50%), minor features (50%–80%), and marginal features (bottom 80%). High-contribution features are given positive adjustments to enhance their impact on model training, while low-contribution features are given negative adjustments to reduce potential noise interference. The assigned value ranges for the four levels are: +0.03 ~ +0.06, +0.01 ~ +0.03, -0.01 ~ -0.03, and -0.03 ~ -0.06, respectively.

[0085] The calculation method for the "missing rate adjustment item" is as follows: First, based on the missing proportion of each feature in all samples of the database and the missing proportion of features in each sample, the overall missing rate is calculated using the following formula:

[0086]

[0087] in, The overall missing rate is %; The feature missing rate is %; The sample missing rate is %; and They are respectively and The weight, .

[0088] Furthermore, the "Missing Rate Adjustment Item" is divided into four levels: MR < 10%; 10 ≤ MR < 20%; 20 ≤ MR < 30%; and MR ≥ 30%. A lower missing rate indicates richer available information for completion, thus receiving a positive adjustment; a higher missing rate indicates scarcer information for completion, thus receiving a negative adjustment. The assigned value ranges for the four levels are: +0.03 ~ +0.06, +0.01 ~ +0.03, -0.01 ~ -0.03, and -0.03 ~ -0.06, respectively.

[0089] Step 6: Introduce the feature-level scaling weights and sample-level confidence labels into the machine learning model training process. Weight the input features and weight the training samples with confidence, thereby constructing a concrete performance prediction model driven by the synergy of feature space and sample quality. The prediction model then outputs concrete performance prediction results or is used for mix design optimization.

[0090] In the machine learning model training process, the feature weights obtained in step 4 are used to scale the input features, and the sample-level confidence labels obtained in step 5 are used to weight the samples with confidence, achieving feature-sample level collaborative data quality enhancement-driven model training. The machine learning model can be any one of the following suitable models for handling high-dimensional feature prediction: Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), K-Nearest Neighbors, Kernel Ridge Regression, Extreme Random Tree, Gradient Boosting Decision Tree, etc. Hyperparameter optimization is performed on the selected training model using any one of the following automatic hyperparameter tuning methods: RIME, Bayesian Optimization (BO), Particle Swarm Optimization (PSO), Dung Beetle Optimization (DBO), etc.

[0091] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

[0092] This embodiment uses concrete compressive strength as the target performance index; the characteristic parameters include: cement type, cement design strength, specimen strength standardization coefficient, water-cement ratio, cement content, fly ash content, mineral powder content, silica fume content, other mineral admixture content, fiber content, sand ratio, sand content, coarse aggregate content, maximum aggregate particle size, water consumption, water-reducing agent content, water-reducing agent / binder ratio, and concrete age; the radial basis function kernel function support vector machine (SVR) is used as the prediction model algorithm.

[0093] Concrete mix proportions and compressive strength data were collected from peer-reviewed academic literature, engineering trial mix records, and laboratory test results to construct an initial database. A statistical analysis of the completeness of the 1116 data records in the initial database revealed 279 missing records. These included: 24 missing records for cement type (2.15%); 50 missing records for cement design strength (4.48%); 21 missing records for specimen strength standardization coefficients (1.88%); 110 missing records for maximum aggregate size (9.86%); and 112 missing records for water-reducing agent dosage (10.04%).

[0094] The Isolation Forest algorithm was used as the model for outlier removal in the database. By substituting 1116 sets of original missing data into the model, outlier removal was performed on the complete samples with a removal rate of 5%. Ultimately, the coefficient of variation of the database decreased from the initial 37.1% to 33.57%, a reduction of 9.51% (Table 1). The results clearly show that after outlier removal, the data distribution is more concentrated, there are fewer outliers, and the mean and median are closer, indicating that the outlier removal effect is good.

[0095] Table 1. Database outlier removal status

[0096]

[0097] Note: Coefficient of variation = Standard deviation / Mean × 100%

[0098] For the various types of missing data in the database, an adaptive imputation method based on a fusion model of Lost Forest, Gradient Boosting Regression (GBR), and K-Nearest Neighbor Regression (KNN) is used to complete the data. Figure 2 It can be seen that the completion error of this method is much lower than that of other traditional completion methods. This fusion interpolation method can cover linear, nonlinear, local, and global patterns, and can offset errors made by each model on different samples, achieving complementary synergy and robust generalization. Using random forest as a surrogate model, combined with the interpretability analysis results of SHAP (… Figure 3 The absolute mean of the SHAP value for each feature is calculated and normalized to represent the nonlinear contribution of each feature to the target performance. For the input features and output performance used to train the prediction model, the Spearman rank correlation coefficient is calculated between them. Figure 4 The linear monotonic relationship between each feature and the compressive strength, as well as the linear redundancy among features, are obtained. Finally, based on the nonlinear SHAP contribution as the base weight, a penalty term is calculated using the Spearman rank correlation coefficient, and the synthesized weights are further subjected to power smoothing and normalization to generate feature-level weights that retain high-dimensional information.

[0099] When calculating sample-level confidence labels, a baseline confidence level of 0.9 is set to balance the credibility of the completed data and avoid overweighting. The completion error adjustment term (E...) adj Importance adjustment item (F) adj ), missing rate adjustment item (M) adj The weights for ) were set to 0.5, 0.3, and 0.2 respectively to highlight the core impact of data completion accuracy. The confidence level was ultimately constrained within the range of [0.8, 1.0], with the confidence level of all original, non-missing data fixed at 1.0. Among these, the completion error adjustment term... The value is assigned to +0.1. The value is assigned as -0.08; the four levels of the feature importance adjustment term are assigned values ​​in order of importance: +0.03, +0.01, -0.02, -0.04; the four levels of the missing rate adjustment term are assigned values ​​in order of missing rate from low to high: +0.03, +0.01, -0.02, -0.04. By assigning a greater weight to the feature missing rate when calculating the missing rate adjustment term, the overall missing rate can more accurately reflect the overall reliability level of the completed data. and weight and The confidence scores are set to 0.6 and 0.4 respectively. The completed data for different features are calculated based on the above three dimensions to obtain differentiated confidence scores.

[0100] The obtained feature-level scaling weights were linearly mapped to the [0.5, 1.5] interval to adapt to the SVR feature space scaling; the obtained sample-level confidence labels were used as sample confidence weights. The SVR algorithm was used for model training. By introducing sample confidence weights into the loss function, the prediction error contribution of low-confidence imputed samples was appropriately reduced, while the prediction error contribution of high-confidence original samples received greater attention. Automatic hyperparameter tuning employed the Rime Optimization (RIME) algorithm. Results showed that compared to randomly generated hyperparameters (…),… Figure 5 a) The hyperparameters obtained through RIME automatic tuning reduced the average mean square error of the model in predicting concrete compressive strength by 86.97%, demonstrating the high efficiency of RIME automatic hyperparameter tuning. The final concrete compressive strength prediction model R... 2 It can reach 94.75% (Table 2, Figure 5 (b) The training effect is significantly improved compared to the original database, which proves that the method of the present invention has excellent prediction accuracy, generalization ability and robustness under the condition of multi-source heterogeneous concrete data with high missing rate.

[0101] Table 2 Model Prediction Accuracy

[0102]

Claims

1. A method for intelligent prediction of concrete performance based on multi-source data enhancement, characterized in that, Includes the following steps: Step 1: Collect concrete mix proportion parameters and corresponding performance data from different data sources to construct a multi-source heterogeneous database of concrete performance. Step 2: Perform outlier detection and cleaning on the complete data sample in the database to obtain the cleaned database; Step 3: For the incomplete data samples in the database, use an adaptive imputation method to fill in the missing values ​​and output a complete database without missing values. Step 4: Based on the completed database, construct a two-dimensional feature weight calculation framework that includes feature nonlinear contribution and feature linear redundancy, and generate feature-level scaling weights. The feature nonlinear contribution is used to represent the nonlinear contribution of features in the completed database, and the feature linear redundancy is used to represent the linear redundancy between features in the completed database. Step 5: For each completed sample in the completed database, calculate the sample-level confidence label to characterize the reliability of the completed sample data; Step 6: Introduce the feature-level scaling weights and sample-level confidence labels into the machine learning model training process. Use the feature-level scaling weights obtained in Step 4 to perform weighted scaling on the input features, and use the sample-level confidence labels obtained in Step 5 to perform confidence weighting on the training samples, thereby obtaining a concrete performance prediction model based on the collaborative driving of feature space and sample quality; and output the concrete performance prediction results according to the prediction model.

2. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 1, characterized in that: In step 1, each record in the multi-source heterogeneous database of concrete performance contains multiple mix proportion characteristic parameters and at least one target performance index; the characteristic parameters are selected from any one or more of the following parameters: cement type, cement design strength, water-cement ratio, cement dosage, dosage of each mineral admixture, coarse aggregate dosage, coarse aggregate gradation, sand dosage, sand fineness modulus, sand ratio, maximum aggregate particle size, water-reducing agent dosage, water-reducing agent water reduction rate, water consumption, and age; the target performance index is selected from any one or more of the following indices: concrete compressive strength, tensile strength, flexural strength, slump, slump spread, chloride ion diffusion coefficient, creep, shrinkage, or porosity.

3. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 1, characterized in that: In step 2, the outlier detection and cleaning process adopts an unsupervised outlier detection method adapted to the characteristics of engineering data. The method used is selected from any one of the following: Isolation Forest algorithm, Local Outlier Factor algorithm, Cluster-based Outlier Detection algorithm, or Multivariate Consistency Test Algorithm based on Mahalanobis Distance.

4. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 1, characterized in that: In step 3, the adaptive interpolation method uses the Lost Forest algorithm as the model basis. Based on the linear correlation between the feature to be completed and other features and physical constraints, it distinguishes between strongly linearly correlated features and nonlinearly complex correlated features. It constructs an ensemble meta-learning model based on Lost Forest, gradient boosting regression and K-nearest neighbor regression to perform weighted interpolation, and finally obtains a complete database without missing features.

5. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 4, characterized in that: The ensemble meta-learning model calculates the prediction error of each base learner through cross-validation, performs normalized weighted averaging with the reciprocal of the root mean square error as the weight, and dynamically adjusts the weights of each base learner during the iteration process.

6. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 1, characterized in that: In step 4, the dual-dimensional feature weight calculation framework uses the nonlinear contribution of features as the base weight and the linear redundancy between features as the penalty term to automatically reduce the weight of highly collinear features to obtain the composite weight; and further performs power smoothing and normalization on the composite weight to generate feature-level weights that retain high-dimensional information.

7. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 6, characterized in that: The nonlinear contribution of the feature is obtained based on the SHAP interpretability analysis results of the surrogate model, which can be any one of support vector machine, random forest, extreme gradient boosting, or lightweight gradient boosting machine; the linear penalty term is calculated using any one of the correlation analysis methods of Spearman rank correlation coefficient, mutual information, or Pearson correlation coefficient, and is used to quantify the linear monotonic relationship and linear redundancy between features. The formula for the linear penalty term is: in, : The linear penalty term for feature i, : Linear correlation coefficient between features i and j. The formula for calculating the composite weight is: in, For the nonlinear weights of feature i obtained based on SHAP, is the composite weight of feature i. To avoid the model becoming overly reliant on specific variables due to excessively high weights for a single feature, and to improve the balanced contribution of each proportion parameter, power smoothing is introduced, as shown in the formula: in, The weights are the power-smoothed values ​​of feature i. This is the power transformation term for weight smoothing, with values ​​in the range [0,1]. The final feature weights used for training the machine learning model are: in, for Assigning normalized weights.

8. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 1, characterized in that: In step 5, the sample-level confidence label comprehensively considers three dimensions: feature completion error, feature importance, and data missing rate. Based on a preset benchmark value, an adjustment term E for the completion error is assigned according to the completion accuracy of each feature. adj Based on the global contribution level of each feature, an importance adjustment term F is assigned to the target performance index. adj The missing rate adjustment term M is assigned based on the interval between the feature missing rate and the sample missing rate. adj The reliability of each completed data point is quantified using a weighted summation method. The confidence level of the original, non-missing data is set to 1 or the maximum value, while the confidence level of the completed data is dynamically determined based on the aforementioned dimensions, as shown in the following formula: in, Final confidence level; Baseline confidence level; , , The weights of each adjustment item. .

9. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 8, characterized in that: Based on the linear normalization of the root mean square error during each feature interpolation process, the completion error adjustment term for the confidence label is calculated using the following formula: in, To complete the error adjustment term; For maximum positive adjustment; This is the maximum negative adjustment; The result is the linearly normalized root mean square error (RMSE) of feature j during interpolation; Based on the feature-level scaling weights obtained in step 4, the importance adjustment term is calculated. All features are divided into at least two contribution levels according to the scaling weights from high to low. Features belonging to the high contribution level are given positive adjustment to enhance the influence weight of their supplementary data in model training, and features belonging to the low contribution level are given negative adjustment to weaken the noise interference they may introduce. The missing rate adjustment term is calculated based on the overall missing rate. The formula for the overall missing rate is: in, The overall missing rate is %; The feature missing rate is %; The sample missing rate is %; and They are respectively and The weight, ; All features are divided into at least two missing levels according to the overall missing rate from high to low. Features belonging to the low missing level are positively adjusted, and features belonging to the high missing level are negatively adjusted.

10. The intelligent prediction method for concrete performance based on multi-source data enhancement according to claim 1, characterized in that: In step 6, during the training of the machine learning model, the machine learning model is selected from any one of the following: support vector machine, random forest, extreme gradient boosting, lightweight gradient boosting machine, K-nearest neighbor algorithm, kernel ridge regression, extreme random tree, and gradient boosting decision tree; the selected training model is hyperparameter optimized using the rime optimization algorithm, Bayesian optimization, dung beetle optimization algorithm, or particle swarm optimization algorithm.

Images