Construction and verification method of compressive strength prediction model of road concrete in high-cold and humid area

Abstract

The application discloses a kind of high-cold damp area road concrete compressive strength prediction model construction and verification method, belongs to concrete performance prediction technical field, obtains the concrete mixing proportion parameter and compressive strength data of mineral admixture total replacement rate not less than 30%, according to mineral admixture activity index, construct activity index feature, and construct water-gel ratio-mineral admixture replacement rate interactive feature;Training is carried out using XGBoost model, and the trend that the predicted compressive strength is abnormally increased with the increase of water-gel ratio is limited by physical constraint penalty term.It is verified that the model obtained by training has prediction performance, and the compressive strength prediction value and prediction uncertainty interval of the sample to be measured are output, when the prediction uncertainty is larger, physical test is carried out, and the model is updated according to the new sample data.The application is favorable to improve the accuracy, physical rationality and application reliability of large-dosage mineral admixture concrete compressive strength prediction.

Description

Technical Field

[0001] This invention belongs to the field of concrete performance prediction technology, and specifically relates to a method for constructing and verifying a prediction model for the compressive strength of road concrete in cold and humid regions. Background Technology

[0002] With the increasing demands for resource conservation and low-carbon development in concrete engineering, mineral admixtures such as fly ash and granulated blast furnace slag are widely used to partially replace cement, thereby reducing cement consumption and improving the resource utilization of industrial solid waste. When the replacement rate of cementitious materials by mineral admixtures reaches more than 30%, it is usually referred to as high-volume mineral admixture concrete.

[0003] Compared to ordinary cement concrete, the compressive strength development of concrete with high mineral admixture content is more complex. On the one hand, the activity of mineral admixtures such as fly ash develops relatively late, and their contribution to the early and later strength of concrete differs from that of cement. On the other hand, under high admixture conditions, there is a complex coupling effect between mineral admixtures and components such as cement and water-reducing agents, resulting in a complex nonlinear relationship between the compressive strength of concrete and various mix proportion parameters. Traditional linear or nonlinear regression methods are difficult to accurately reflect the changes in its mechanical properties.

[0004] Currently, the mix design of concrete with high mineral admixture content still relies heavily on physical mix design tests. This typically requires preparing specimens based on different mix proportions, and then conducting compressive strength tests after curing for 7 days, 28 days, or even longer. The mix proportions are then adjusted based on the test results, resulting in a long testing cycle and low efficiency in mix proportion selection. Furthermore, the strength development of concrete with high mineral admixture content exhibits significant dispersion, and traditional empirical prediction methods such as the Bolomey formula have high prediction errors in this type of concrete. Abrams' law, primarily based on the empirical relationship between water-cement ratio and concrete strength, is also insufficient to fully reflect the impact of mineral admixtures on strength development when directly applied to concrete containing high mineral admixture content.

[0005] With the development of machine learning technology, establishing predictive models using historical concrete mix proportion parameters and corresponding compressive strength data provides a new technical approach to shorten the trial mixing cycle and improve the efficiency of compressive strength prediction. Machine learning models can handle the relatively complex nonlinear relationship between mix proportion parameters and compressive strength. However, existing machine learning prediction methods are mostly designed for ordinary concrete or concrete with low mineral admixture content. For high-volume mineral admixture concrete with a mineral admixture replacement rate of not less than 30%, their applicability still needs to be improved, and there is a lack of specific treatment for the strength development characteristics of high-volume mineral admixture concrete.

[0006] Furthermore, existing machine learning prediction models are typically trained primarily on statistical relationships within sample data, lacking direct constraints on the fundamental physical laws governing concrete materials. This can lead to inconsistencies between predicted results and these fundamental material properties. For example, under similar conditions, the model might exhibit an unusual trend where an increase in the water-cement ratio results in a predicted increase in compressive strength. While such predictions may possess some predictive power for overall statistical evaluation indicators, their patterns lack physical plausibility, thus affecting the reliability of the predictions in actual concrete mix design.

[0007] Meanwhile, when applying the completed prediction model to a new concrete mix proportion, the prediction uncertainty of different test samples may vary. Therefore, in addition to verifying the overall prediction performance of the model, it is also necessary to evaluate the reliability of the prediction results of specific test samples and, if necessary, revise the prediction model in combination with the results of solid compressive strength tests to improve the adaptability and reliability of the model in subsequent applications.

[0008] Therefore, how to improve the accuracy and physical rationality of compressive strength prediction based on the material characteristics of concrete with high mineral admixture content, and how to effectively verify the prediction model and its prediction results to improve the reliability of the prediction model in practical applications, remains a technical problem that needs to be solved. Summary of the Invention

[0009] One objective of this invention is to provide a method for constructing a predictive model for the compressive strength of road concrete in cold and humid regions, comprising the following steps: S1. Obtain the mix proportion parameters and compressive strength data corresponding to the mix proportion parameters of the high-volume mineral admixture concrete sample. The mix proportion parameters include at least the water content, cement content, mineral admixture content, and total mineral admixture substitution rate, wherein the total mineral admixture substitution rate is not less than 30%. S2. The dosage of mineral admixtures is modified according to the activity index of the mineral admixtures to construct an admixture activity index feature that characterizes the effective cementing contribution of the mineral admixtures; an interaction feature of water-binder ratio and mineral admixture substitution rate is constructed based on the water-binder ratio and the total substitution rate of mineral admixtures, and the mix proportion parameters, the admixture activity index feature, and the interaction feature of water-binder ratio and mineral admixture substitution rate are used as model input features; S3. Construct a machine learning basic model for predicting the compressive strength of the high-volume mineral admixture concrete, and set a physical consistency constraint to characterize the relationship between the water-cement ratio and the compressive strength during the training process of the machine learning basic model. When the compressive strength predicted by the model increases with the increase of the water-cement ratio, the physical consistency constraint is used to penalize the model training. S4. Using the model input features as model input and the compressive strength corresponding to the model input features as model output, train the machine learning basic model with the physical consistency constraints to obtain a compressive strength prediction model for high-volume mineral admixture concrete.

[0010] Preferably, in step S1, the mineral admixture includes at least one of fly ash and granulated blast furnace slag; the mix proportion parameters also include the amount of coarse aggregate, the amount of fine aggregate, and the amount of water-reducing agent. The total replacement rate R of the mineral admixtures is determined according to the following formula: R=(FA+GGBS) / (C+FA+GGBS)×100%; Wherein, FA represents the amount of fly ash, GGBS represents the amount of granulated blast furnace slag, and C represents the amount of cement, and 30%≤R≤60%; the compressive strength data includes 7-day compressive strength and 28-day compressive strength; the box plot method is used to identify and remove outliers from the obtained high-volume mineral admixture concrete samples, and the sample data after removing outliers are normalized.

[0011] Preferably, in S2, the activity index characteristics of the admixture include the amount of fly ash activity correction and / or the amount of granulated blast furnace slag activity correction. The fly ash activity correction dosage is the product of the fly ash dosage and the fly ash activity index coefficient, wherein the fly ash activity index coefficient is 0.7 to 0.9; The amount of granulated blast furnace slag used for activity correction is the product of the amount of granulated blast furnace slag used and the activity index coefficient of granulated blast furnace slag, wherein the activity index coefficient of granulated blast furnace slag is 0.9 to 1.05. The activity index coefficients of fly ash and granulated blast furnace slag are determined based on the 28-day activity index test values ​​of the corresponding mineral admixtures.

[0012] Preferably, in S2, the water-cement ratio-mineral admixture substitution rate interaction characteristic is determined according to the following formula: X_interaction = (W / B) × R; Wherein, W / B is the water-cement ratio, and R is the total replacement rate of mineral admixtures; the interaction feature of water-cement ratio-mineral admixture replacement rate is used to characterize the coupling relationship between water-cement ratio and concrete compressive strength under different mineral admixture replacement rates. The model input features also include maintenance age, which includes 7 days and 28 days, and is used as an independent classification input feature.

[0013] Preferably, in S3, the machine learning base model is the XGBoost model; In step S4, the physical consistency constraint is set using an outer-layer iteration method, and the trend of the predicted compressive strength f relative to the water-cement ratio (W / B) is determined by numerical difference: When ∂f / ∂(W / B)>0, it is determined that the model's predicted trend does not conform to the preset law of decreasing compressive strength as the water-cement ratio increases, and the physical constraint penalty value for the corresponding training sample is determined. Lphysics=[∂f / ∂(W / B)] 2 ; When ∂f / ∂(W / B)≤0, Lphysics=0; The sample weights of the corresponding training samples are adjusted according to the physical constraint penalty values ​​of each training sample, and the adjusted sample weights are used as the sample weights of the XGBoost model for the next round of training. Through multiple rounds of outer layer iterations, the trends in the model predictions that do not conform to the stated change rules are constrained.

[0014] Preferably, in S3, a particle swarm optimization algorithm is used to optimize the hyperparameters of the XGBoost model within a preset search space. The hyperparameters include the number of cycles, Gamma value, maximum tree depth, and learning rate. The search range for the number of iterations is 20 to 120, the search range for the Gamma value is 0.05 to 0.9, the search range for the maximum tree depth is 3 to 11, and the search range for the learning rate is 0.01 to 0.3. The particle swarm optimization algorithm has a population size of 20 and an iteration count of 20.

[0015] Another objective of this invention is to provide a verification method for a predictive model of the compressive strength of road concrete in cold and humid regions. This verification method is used to verify the predictive model of the compressive strength of high-volume mineral admixture concrete constructed using the above-described method, and includes the following steps: S5. Obtain the compressive strength prediction model of high-volume mineral admixture concrete to be verified and the test samples that have not participated in the training of the prediction model. Input the test samples into the prediction model to obtain the predicted compressive strength value of the test samples. Determine the model prediction performance evaluation index based on the predicted compressive strength value and the measured compressive strength of the corresponding test samples. Determine whether the prediction model meets the preset prediction performance requirements based on the model prediction performance evaluation index. S6. When the prediction model meets the preset prediction performance requirements, obtain the mix proportion parameters of the concrete with high mineral admixture content to be tested, process the mix proportion parameters according to the data processing method and model input feature construction method used during the training of the prediction model, and obtain the model input features of the sample to be tested; input the model input features of the sample to be tested into the prediction model to obtain the predicted compressive strength value and the corresponding prediction uncertainty interval, and compare the width of the prediction uncertainty interval with the preset interval width threshold to evaluate the prediction reliability of the predicted compressive strength value; S7. When the width of the predicted uncertainty interval exceeds the preset interval width threshold, a standard test block is made according to the mix proportion of the concrete with high mineral admixture to be tested, and the compressive strength of the standard test block is tested to obtain the actual compressive strength value of the solid, and the prediction error rate is determined according to the predicted compressive strength value and the actual compressive strength value of the solid. S8. Compare the prediction error rate obtained in step S7 with the preset error rate threshold. When the prediction error rate exceeds the preset error rate threshold, add the mix proportion parameters of the concrete with high mineral admixture content to be tested and the corresponding measured values ​​of solid compressive strength as new sample data to the sample database used to train the prediction model, and update the prediction model based on the supplemented sample database.

[0016] Preferably, the prediction uncertainty interval is determined based on the prediction error distribution of the prediction model on the training samples; the prediction error of each training sample is calculated and sorted according to the magnitude of the prediction error, and the prediction uncertainty interval corresponding to the high-volume mineral admixture concrete to be tested is determined based on the quantile of the prediction error distribution. The width of the predicted uncertainty interval is used as the prediction reliability evaluation parameter. When the width of the predicted uncertainty interval exceeds the preset interval width threshold, a physical test verification of the concrete with high mineral admixture content to be tested is triggered.

[0017] Preferably, the preset interval width threshold is 8 MPa; The standard test block was tested according to the concrete compressive strength test method to obtain the measured value of the solid compressive strength; The prediction error is represented by a prediction error rate, which is determined according to the following formula: Prediction error rate = |Predicted compressive strength - Measured compressive strength of solid material| / Measured compressive strength of solid material × 100% When the prediction error rate exceeds 8%, the mix proportion parameters and measured values ​​of the actual compressive strength of the concrete with high mineral admixture content to be tested will be added to the sample database as new sample data.

[0018] Preferably, the update method of the prediction model is determined based on the amount of new sample data n added to the sample database; When the amount of new sample data n < 30 sets, the model parameters of the current prediction model are used as the initial parameters, and a limited number of rounds of model updates are performed on the new sample data. When the amount of new sample data n≥30 sets, the prediction model is completely retrained using the sample database including the new sample data. The prediction performance of the updated prediction model is validated.

[0019] The present invention has the following beneficial effects: 1. In the model construction process, this invention addresses the issues of varying activity among different mineral admixtures in high-volume mineral admixture concrete and the coupling effect between the water-cement ratio and the mineral admixture substitution rate. It corrects the dosage of mineral admixtures based on their activity index and constructs an interaction feature between the water-cement ratio and the total mineral admixture substitution rate. This allows the model input features to further characterize the activity differences among different mineral admixtures and the impact of the combined changes in water-cement ratio and substitution rate on compressive strength. Simultaneously, a physical consistency constraint based on the relationship between water-cement ratio and compressive strength is introduced during model training. When the model prediction shows a trend of abnormally increasing predicted compressive strength with an increasing water-cement ratio, the training sample weights are adjusted according to the physical constraint penalty value, thereby limiting predictions that do not conform to the basic material laws of concrete. Thus, the material characteristic characterization at the model input end and the physical law constraints at the model training end complement each other, which is beneficial to improving the accuracy and physical rationality of compressive strength prediction in high-volume mineral admixture concrete. Ablation experiments show that when dedicated feature engineering and physical consistency constraints are set together, the model's prediction accuracy and physical consistency are superior to models without corresponding technical measures.

[0020] 2. In the model validation process, this invention determines the prediction uncertainty interval of the test sample based on the prediction error distribution of the prediction model on the training samples, and uses the width of the prediction uncertainty interval as the basis for evaluating prediction reliability. This identifies test samples with large prediction uncertainties and provides a basis for determining whether further physical compressive strength testing is necessary. For samples whose prediction errors exceed a preset threshold as determined by physical testing, their mix proportion parameters and measured compressive strength are added to the database as new sample data. A limited number of gradient updates or complete retraining is selected based on the amount of new sample data, allowing the prediction model to be corrected using the new measured data. This forms a validation feedback loop of prediction, reliability evaluation, physical testing, data feedback, and model update, which helps improve the adaptability of the model in practical applications and the reliability of prediction results, and reduces the necessity of complete retraining when the amount of new sample data is small. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the construction and verification process of the prediction model for the compressive strength of concrete with high mineral admixture content according to the present invention.

[0022] Figure 2 This is a schematic diagram illustrating the implementation process of physical consistency constraints during the training of the compressive strength prediction model of this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] This invention provides a method for constructing and verifying a model for the compressive strength of concrete with high mineral admixture content. For concrete with high mineral admixture content and a mineral admixture replacement rate of not less than 30%, input features reflecting the effective cementitious contribution of different mineral admixtures are constructed based on their activity characteristics, and interactive features between the water-cement ratio and the mineral admixture replacement rate are also constructed. Using XGBoost as the basic model, physical consistency constraints based on the relationship between the water-cement ratio and compressive strength are set during model training, and particle swarm optimization algorithm is used to optimize the model hyperparameters, thereby constructing a predictive model for the compressive strength of concrete with high mineral admixture content.

[0025] After the model is built, the model's predictive performance and physical consistency are verified through a test set, and the actual prediction results of the model are further verified through prediction of concrete samples to be tested and solid compressive strength tests.

[0026] As one application scenario, the method described in this invention can be used for mix design and compressive strength prediction of high-volume mineral admixture concrete in the construction of transportation infrastructure such as highways and bridges in cold and humid regions. During the construction of transportation projects in such areas, a compressive strength prediction model can be constructed and verified using concrete mix proportion parameters and corresponding compressive strength data, providing a predictive basis for mix design and material performance evaluation of high-volume mineral admixture concrete in the construction of roads, bridges, and other transportation infrastructure.

[0027] Example 1: Construction and Verification of Compressive Strength Model for High-Dosage Fly Ash Concrete I. Prediction Model Construction S1. Acquire and process sample data of high-volume fly ash concrete. Mix proportion parameters and corresponding compressive strength data of high-volume fly ash concrete were collected, resulting in 235 sets of original sample data.

[0028] The mix proportion parameters include water content W, cement content C, fly ash content FA, coarse aggregate content, fine aggregate content, water-reducing agent content, and total mineral admixture replacement rate R.

[0029] In this embodiment, the values ​​of each proportion parameter are in the range of: water usage 150–200 kg / m³ 3 Cement usage: 180–350 kg / m³ 3 Fly ash dosage: 90–200 kg / m³ 3 Coarse aggregate dosage: 1000–1200 kg / m³ 3 Fine aggregate dosage: 600–900 kg / m³ 3 Water-reducing agent dosage: 3–12 kg / m³ 3 .

[0030] Fly ash accounts for 30% to 55% of the total cementitious materials, and the total replacement rate R of mineral admixtures is determined according to the following formula: R = FA / (C + FA) × 100%.

[0031] Where FA represents the amount of fly ash and C represents the amount of cement, and in this embodiment, 30% ≤ R ≤ 55%.

[0032] The compressive strength data include 7-day compressive strength and 28-day compressive strength, where the 7-day compressive strength is 12.5–52.0 MPa and the 28-day compressive strength is 30.0–72.0 MPa.

[0033] Outlier identification was performed on 235 sets of original sample data using the box plot method. After removing 18 sets of outlier data, 217 sets of valid sample data were obtained.

[0034] The obtained valid sample data were normalized, and each input parameter was standardized to the [0, 1] interval to obtain a sample database of compressive strength of high-volume fly ash concrete for model construction.

[0035] S2. Constructing model input features for compressive strength prediction Based on the original mix proportion parameters such as water, cement, fly ash, coarse aggregate, fine aggregate, and water-reducing agent, we further constructed the characteristics of total mineral admixture substitution rate, admixture activity index, water-cement ratio-mineral admixture substitution rate interaction, and curing age.

[0036] Among them, the fly ash dosage was adjusted based on the 28-day activity index of fly ash.

[0037] In this embodiment, the 28-day activity index coefficient of fly ash is taken as 0.80, and the fly ash activity correction dosage FA_effective is determined according to the following formula: FA_effective = FA × 0.80.

[0038] The effective cementitious contribution of fly ash in the corresponding concrete system is characterized by the amount of fly ash used for activity correction.

[0039] The fly ash activity index coefficient can be determined based on the 28-day activity index test value of fly ash, and its value ranges from 0.7 to 0.9.

[0040] Furthermore, the water-cement ratio W / B of the concrete was calculated, and the interaction characteristics of the water-cement ratio and the total mineral admixture substitution rate R were constructed based on the water-cement ratio W / B and the total mineral admixture substitution rate R. X_interaction = (W / B) × R.

[0041] Interactive features are used to characterize the coupled effect of changes in water-cement ratio and fly ash substitution rate on the compressive strength of concrete under high fly ash dosage conditions.

[0042] Meanwhile, the curing age is used as an input feature of the model to characterize the differences in the development of concrete compressive strength at different curing ages of 7 days and 28 days.

[0043] The original mix proportion parameters, total mineral admixture replacement rate, fly ash activity index characteristics, water-cement ratio-mineral admixture replacement rate interaction characteristics, and curing age are combined to form the model input features.

[0044] S3. Construct the XGBoost basic prediction model and optimize the model hyperparameters. The Limiting Gradient Boosting (XGBoost) model is selected as the basic machine learning prediction model. The model input features obtained in step S2 are used as inputs, and the compressive strength of the corresponding samples is used as the model output.

[0045] The hyperparameters of the XGBoost model were optimized using the particle swarm optimization algorithm (PSO).

[0046] In this embodiment, the population size of the particle swarm optimization algorithm is 20, and the number of iterations is 20.

[0047] The hyperparameters to be optimized for the XGBoost model and their corresponding search range are: number of iterations 20–120; Gamma value 0.05–0.9; maximum tree depth 3–11; learning rate 0.01–0.3.

[0048] The particle swarm optimization algorithm was used to optimize the parameters within the above search range, and the hyperparameters of the XGBoost model corresponding to this embodiment were obtained as follows: number of cycles 90; Gamma value 0.4; maximum tree depth 6; learning rate 0.08.

[0049] S4. Set physical consistency constraints during model training. To limit the prediction results generated by the pure data-driven model from being inconsistent with the basic strength variation law of concrete, a physical consistency constraint based on the relationship between water-cement ratio and compressive strength variation is set during the model training process.

[0050] The model-predicted compressive strength is denoted as f, and the water-cement ratio is denoted as W / B. By numerically differencing the water-cement ratio at each sample point of the training data, the trend of the model-predicted compressive strength changing with the water-cement ratio is determined.

[0051] When ∂f / ∂(W / B)>0, it indicates that the model predicts that the compressive strength increases with the increase of the water-to-glue ratio, which is inconsistent with the preset law that the compressive strength decreases with the increase of the water-to-glue ratio. The physical constraint penalty value for the corresponding training samples is set as follows: Lphysics,i=[∂f / ∂(W / B)] 2 ; When ∂f / ∂(W / B)≤0, set: Lphysics,i=0 The physical penalty term is not a built-in regularization term in the XGBoost tree splitting stage, but plays a constraint role in the outer iterative loop. By quantifying the physical bias after each round of training, it is transformed into a dynamic adjustment of the sample weights. The physical mechanism constraint is indirectly introduced by using the sample weight interface natively supported by XGBoost. Through multiple rounds of iteration, prediction results that violate physical laws are gradually suppressed, thereby integrating the physical penalty effect into the complete training process.

[0052] In the process of numerical difference of water-cement ratio, it is not only necessary to modify the W / B eigenvalue in the feature matrix, but also to keep the parameters of cementitious materials, aggregates, mineral admixtures, water-reducing agents, etc. unchanged, and only adjust the water consumption to achieve a small change in W / B. The interactive features are updated simultaneously and the physical boundary of water consumption is checked, so as to ensure the consistency of mix proportion parameters in the difference calculation process.

[0053] The physical constraint penalty term does not directly participate in tree structure optimization as a built-in regularization term in the XGBoost tree splitting stage. Instead, it is set up as an outer-layer iterative process outside the XGBoost model, indirectly affecting model training by adjusting the weights of training samples.

[0054] After obtaining the physical constraint penalty value Lphysics for the i-th training sample, the sample weights for the next round of XGBoost training are dynamically updated using the following formula: ，i In the formula, Wi is the weight of the i-th training sample to be normalized. =1.0 is the baseline sample weight. The physical penalty coefficient is 0.05.

[0055] When the i-th training sample does not trigger the physical constraints =0, therefore, its sample weights to be normalized are ； When training samples trigger physical consistency constraints If the value is greater than 0, the weight of the sample to be normalized increases with the degree of physical violation, thereby increasing the influence of the sample in the next round of XGBoost training.

[0056] The sample weights to be normalized obtained according to the above formula are subjected to amplitude limiting to avoid samples with large physical violations from having excessively large sample weights. After amplitude limiting, all sample weights are normalized so that the sum of all normalized sample weights is equal to the total number of training samples, while maintaining the relative weight relationship between different training samples.

[0057] The normalized sample weights are input into the model through the sample weight interface provided by XGBoost for the next round of XGBoost training.

[0058] The outer iteration adopts a dual termination condition: if the preset maximum number of iterations is reached, or if the maximum physical penalty value of all samples in multiple consecutive iterations is lower than the preset convergence threshold, the outer iteration stops and the final prediction model is output.

[0059] S5. Train the compressive strength prediction model using training samples. The 217 valid samples obtained in step S1 are randomly divided into a training set and a test set at a ratio of 75% and 25%, respectively. The training set includes 163 samples, and the test set includes 54 samples.

[0060] The 163 sets of training samples were used to construct the corresponding model input features according to step S2, and the XGBoost model established and optimized by parameters in step S3 was trained using the training samples.

[0061] During the model training process, physical consistency constraints are applied according to step S4. After training is completed, a prediction model for the compressive strength of high-volume fly ash concrete is obtained.

[0062] II. Model Validation Process S6. Verify the model's predictive performance using test samples. Input the 54 test samples that were not involved in model training in step S5 into the constructed compressive strength prediction model, and calculate the coefficient of determination R based on the model prediction values ​​and the corresponding measured compressive strength of the samples. 2 Mean absolute error (MAE) and mean square error (MSE).

[0063] In this embodiment, the model test result is: R 2=0.922; MAE=2.81 MPa; MSE=12.37 MPa 2 .

[0064] When the model satisfies R 2 ≥0.90, MAE≤3.5 MPa and MSE≤15 MPa 2 At that time, it was determined that the model met the preset compressive strength prediction performance requirements.

[0065] The compressive strength prediction model constructed in this embodiment meets the above performance requirements.

[0066] S7. Verify the physical consistency of the model's prediction results. While keeping the model input features, model hyperparameters and training data the same, a control model without physical consistency constraints is constructed, and the prediction results of the control model and the model in this embodiment are compared.

[0067] The results show that in the control model without physical consistency constraints, approximately 27.7% of the test samples showed that the predicted results did not conform to the law of decreasing compressive strength as the water-cement ratio increased.

[0068] After setting the physical consistency constraint in step S4, the model in this embodiment did not produce any prediction results that violated the above-mentioned preset change rules in the test samples, and the physical consistency compliance rate was 100%.

[0069] This verifies that the physical consistency constraint can reduce the number of prediction results generated by the model that are inconsistent with the preset water-cement ratio-compressive strength variation law.

[0070] S8. Perform interpretability analysis on the model. The SHAP method is used to calculate the contribution of each input feature of the model to the compressive strength prediction result, and the key features affecting the compressive strength prediction result are determined based on the contribution of each input feature.

[0071] In this embodiment, the total replacement rate of mineral admixtures, water-cement ratio, and curing age are three input characteristics that significantly affect the compressive strength of high-volume fly ash concrete, contributing 28.6%, 24.1%, and 18.3%, respectively.

[0072] Further univariate sensitivity analysis was conducted.

[0073] Keep the other model input parameters, except for the mineral admixture replacement rate, at their corresponding average values. Gradually increase the mineral admixture replacement rate from 30% to 55% in increments of 5%, and record the predicted 28-day compressive strength output by the model.

[0074] Within the data range corresponding to this embodiment, for every 5% increase in the mineral admixture replacement rate, the model-predicted 28-day compressive strength decreases by an average of approximately 3.2 MPa.

[0075] S9. Verify the actual prediction results of the model using concrete samples to be tested. A set of high-volume fly ash concrete samples were selected for testing, with the following mix proportions: water 165 kg / m³. 3 ; Cement 240 kg / m 3 ; fly ash 120 kg / m³ 3 Coarse aggregate 1120 kg / m³ 3 Fine aggregate 780 kg / m³ 3 Water-reducing agent 5.2 kg / m³ 3 .

[0076] The fly ash replacement rate in the tested concrete was 33.3%.

[0077] The mix proportion parameters of the concrete to be tested are processed according to the data processing method in step S1 and the input feature construction method in step S2, and the obtained model input features are input into the compressive strength prediction model constructed in step S5.

[0078] Model output: 7-day compressive strength prediction is 28.6 MPa; 28-day compressive strength prediction is 52.3 MPa.

[0079] Simultaneously, the prediction uncertainty interval for the test sample is determined based on the prediction error distribution of the model on the training samples. The prediction error for the training samples is defined as the difference between the measured compressive strength and the corresponding predicted compressive strength of the training sample.

[0080] The prediction errors of the training samples are sorted by size, and the prediction uncertainty interval is determined based on the 5th and 95th percentiles of the prediction error distribution.

[0081] In this embodiment: The uncertainty range for the 7-day compressive strength prediction is [27.1 MPa, 30.1 MPa]; The uncertainty range for the 28-day compressive strength prediction is [50.8 MPa, 53.8 MPa].

[0082] The training samples are input into the trained prediction model, and the prediction error for each group of training samples is calculated. All prediction errors are sorted from smallest to largest, and the 5th and 95th quantiles of the error distribution are calculated. For the test sample, the predicted value of the compressive strength point output by the model is obtained. The 5th and 95th quantiles of the error are added to this predicted value to obtain the lower and upper limits of the prediction uncertainty interval. The width of the uncertainty interval is calculated. When the interval width is greater than 8 MPa, the prediction reliability is deemed insufficient, and a physical compressive strength test is required. In the example, the predicted value of the 7-day compressive strength point of the test sample is 28.6 MPa, the 5th quantile of the training sample error is -1.5 MPa, and the 95th quantile is +1.5 MPa. The calculated uncertainty interval is [27.1 MPa, 30.1 MPa], with an interval width of 3.0 MPa, which is less than the 8 MPa threshold, and no physical test verification is required.

[0083] Both prediction uncertainty intervals have a width of 3.0 MPa, which is less than the preset threshold of 8 MPa.

[0084] To further verify the actual prediction results of the model, three sets of standard test blocks of 150 mm × 150 mm × 150 mm were made according to the above concrete mix proportions and cured at a temperature of 20 ± 2℃ and a relative humidity of not less than 95%.

[0085] After solid compressive strength test: The measured compressive strength after 7 days is 27.9 MPa; The measured compressive strength after 28 days is 51.6 MPa.

[0086] The prediction error rate is calculated using the following formula: Prediction error rate = |Measured value - Predicted value| / Measured value × 100%.

[0087] Among them, the prediction error rate for 7-day compressive strength was 2.5%; and the prediction error rate for 28-day compressive strength was 1.4%.

[0088] Both prediction error rates do not exceed 8%, and the measured values ​​of the corresponding solid compressive strength are all within the corresponding prediction uncertainty range. This verifies that the model constructed in this embodiment has the corresponding compressive strength prediction capability for the high-volume fly ash concrete to be tested.

[0089] S10. Dynamically update the model based on the newly added valid data. In the actual application of the model, new mix proportions and solid compressive strength data of high-volume mineral admixture concrete are obtained, and the new data that meets the requirements are added to the sample database.

[0090] The model update method is determined based on the amount of newly added valid data, n.

[0091] When the amount of newly added valid data n < 30 sets, the current model parameters are used as the initial parameters, and the model parameters are updated in a limited number of rounds on the newly added valid data. When the amount of newly added valid data n ≥ 30 sets, the complete model is retrained using the database after adding the newly added valid data.

[0092] After six months of application of the model in this embodiment, 45 new sets of data on high-volume slag-fly ash composite concrete were added.

[0093] Since the amount of new data n = 45 ≥ 30, a complete retraining is triggered.

[0094] After retraining, the model test set R 2 The error rate was reduced to 4.8% for the prediction of high-volume slag-fly ash composite concrete, with an MAE of 0.935 and a value of 2.52 MPa.

[0095] Twelve more sets of valid data were subsequently added.

[0096] Since the amount of new data n = 12 < 30, the model parameters are updated three times based on the new data, using the current model parameters as the initial parameters.

[0097] This model update took approximately 2 minutes, which is 15% of the time required for a complete retraining. The updated model R... 2 It increased from 0.935 to 0.941.

[0098] Using the above method, a choice can be made between updating model parameters in a limited number of rounds and retraining the entire model, depending on the amount of new valid data.

[0099] Example 2

[0100] This embodiment takes high-volume granulated blast furnace slag concrete as the object and constructs and verifies the compressive strength prediction model according to the same basic model construction and verification ideas as in Embodiment 1.

[0101] I. Model Building Process S1. Acquire and process sample data of high-volume granulated blast furnace slag concrete. Mix proportion parameters and corresponding compressive strength data of high-volume granulated blast furnace slag concrete were collected, and a total of 198 sets of original sample data were obtained.

[0102] The values ​​for each mix proportion parameter are as follows: water 155–195 kg / m³ 3 Cement 200-320 kg / m³ 3 Granulated blast furnace slag 90-180 kg / m³ 3 Coarse aggregate 1050~1180 kg / m³ 3Fine aggregate 620–880 kg / m³ 3 Water-reducing agent: 3.5–10.5 kg / m³ 3 .

[0103] In the samples used for model building, the granulated blast furnace slag replacement rate was 30%–60%.

[0104] The compressive strength data corresponding to the samples include 7-day compressive strength and 28-day compressive strength, where: 7-day compressive strength is 18.0–48.0 MPa; 28-day compressive strength is 35.0–68.0 MPa.

[0105] The box plot method was used to identify and remove 12 abnormal samples, resulting in 186 valid sample data.

[0106] S2. Input features for constructing a high-volume granulated blast furnace slag concrete model. The dosage of granulated blast furnace slag was adjusted based on the 28-day activity index of the granulated blast furnace slag.

[0107] In this embodiment, the activity index coefficient of granulated blast furnace slag is taken as 0.95. The activity correction amount of granulated blast furnace slag is obtained by multiplying the amount of granulated blast furnace slag by 0.95, and the activity index feature characterizing the effective cementing contribution of granulated blast furnace slag is constructed with the activity correction amount.

[0108] Further interaction features were constructed based on the water-binder ratio (W / B) and the granulated blast furnace slag replacement rate (R): X_interaction = (W / B) × R.

[0109] Meanwhile, the maintenance age is used as an input feature of the model.

[0110] The original mix proportion parameters, granulated blast furnace slag replacement rate, granulated blast furnace slag activity index characteristics, water-binder ratio-mineral admixture replacement rate interaction characteristics, and curing age are combined to form the model input characteristics.

[0111] S3. Establish the XGBoost basic prediction model and optimize its parameters. XGBoost was used as the basic prediction model, and the hyperparameters of the XGBoost model were optimized using the particle swarm optimization algorithm.

[0112] After particle swarm optimization, the XGBoost model parameters obtained in this embodiment are: number of cycles 85; Gamma value 0.35; maximum tree depth 7; learning rate 0.09.

[0113] S4. Set physical consistency constraints for the model Following the method described in step S4 of Example 1, physical consistency constraints based on the relationship between water-cement ratio and compressive strength variation are set during model training.

[0114] In this embodiment, the physical constraint penalty coefficient λ is set to 0.05.

[0115] When the model predicts that the compressive strength increases with the increase of the water-cement ratio, a corresponding physical constraint penalty term is added to the total loss function of the model training; when the model prediction results conform to the preset water-cement ratio-compressive strength change law, the physical constraint penalty term is zero.

[0116] S5. Training a predictive model for the compressive strength of high-volume granulated blast furnace slag concrete. The 186 valid samples were divided into training and test sets at a ratio of 75% and 25%, respectively, with 140 sets in the training set and 46 sets in the test set.

[0117] Using 140 sets of training samples, the model input features were constructed according to step S2, and the XGBoost model established in step S3 was trained. During the training process, physical consistency constraints were applied according to step S4 to obtain a prediction model for the compressive strength of high-volume granulated blast furnace slag concrete.

[0118] II. Model Validation Process S6. Verify model prediction performance and physical consistency. The 46 test samples were input into the compressive strength prediction model obtained in step S5 to verify the model's prediction performance.

[0119] The test result is: R 2 =0.915; MAE=2.62 MPa; MSE=11.84 MPa 2 .

[0120] The model satisfies R 2 ≥0.90, MAE≤3.5 MPa and MSE≤15 MPa 2 The preset model performance requirements.

[0121] Meanwhile, the physical consistency rate of the test samples was 100%, thus confirming that the model constructed in this embodiment met the predetermined prediction performance and physical consistency requirements.

[0122] S7. Validate the actual prediction results of the model using the test samples. A set of high-volume granulated blast furnace slag concrete samples were selected for testing, with the following mix proportions: water 170 kg / m³. 3 ; Cement 260 kg / m 3 Granulated blast furnace slag 130 kg / m³ 3 Coarse aggregate 1100 kg / m³ 3 Fine aggregate 760 kg / m³ 3 Water-reducing agent 6.0 kg / m³3 .

[0123] Based on the above mix proportion parameters, the granulated blast furnace slag replacement rate of the sample to be tested is 33.3%.

[0124] The sample to be tested is processed in accordance with the methods described in steps S1 and S2, and the resulting model input features are input into the compressive strength prediction model constructed in step S5.

[0125] Model output: 7-day compressive strength prediction is 32.5 MPa; 28-day compressive strength prediction is 55.8 MPa.

[0126] The uncertainty range for the 7-day compressive strength prediction is [31.0 MPa, 34.0 MPa]; the uncertainty range for the 28-day compressive strength prediction is [54.0 MPa, 57.6 MPa].

[0127] Further solid compressive strength tests were conducted, and the measured value of the 7-day compressive strength was 31.8 MPa, with a prediction error rate of 2.2%; the measured value of the 28-day compressive strength was 54.9 MPa, with a prediction error rate of 1.6%.

[0128] Both prediction error rates do not exceed 8%, and the measured values ​​of the solid compressive strength are both within the corresponding prediction uncertainty range.

[0129] This verifies that the compressive strength prediction model constructed using the model construction method described in this invention for high-volume granulated blast furnace slag concrete has the corresponding compressive strength prediction capability.

[0130] Comparative experiment To further verify the technical effect of the model construction method of the present invention, under the same data conditions, the compressive strength prediction model constructed by the present invention was compared with the traditional Bolomey formula, the unoptimized XGBoost model, the GA-XGBoost model and the GridSearchCV-XGBoost model.

[0131] The test results are as follows:

[0132] To further analyze the technical effects of dedicated feature construction at the model input end and physical consistency constraints at the model training end, ablation experiments were conducted under identical conditions. The results are as follows:

[0133] The results above show that, without setting specific features and physical consistency constraints, model R... 2The strength was 0.883, the MAE was 3.42 MPa, and the physical consistency compliance rate was 65.8%.

[0134] When setting dedicated features without setting physical consistency constraints, model R 2 The MAE decreased to 2.96 MPa, which is higher than the previous value. This indicates that constructing model input features based on the coupling relationship between the activity of mineral admixtures and the water-cement ratio and the substitution rate of mineral admixtures is beneficial to improving the model's ability to characterize the relationship between the compressive strength of concrete with high mineral admixture content.

[0135] When physical consistency constraints are set without setting specific features, the physical consistency achievement rate of the model increases to 89.5%, indicating that penalizing prediction trends that violate the preset water-cement ratio-compressive strength change relationship during model training is beneficial to improving the physical consistency of the model prediction results.

[0136] When both dedicated features and physical consistency constraints are set, model R 2 The MAE was reduced to 2.81 MPa, and the physical consistency rate reached 100%. This shows that the present invention can improve the prediction accuracy and physical rationality of the compressive strength prediction model of high-volume mineral admixture concrete by characterizing the material properties at the model input end and constraining the physical laws at the model training end.

[0137] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a predictive model for the compressive strength of road concrete in cold and humid regions, characterized in that, Includes the following steps: S1. Obtain the mix proportion parameters and compressive strength data corresponding to the mix proportion parameters of the high-volume mineral admixture concrete sample. The mix proportion parameters include at least the water content, cement content, mineral admixture content, and total mineral admixture substitution rate, wherein the total mineral admixture substitution rate is not less than 30%. S2. The dosage of mineral admixtures is modified according to the activity index of the mineral admixtures to construct an admixture activity index characteristic that characterizes the effective cementing contribution of mineral admixtures. Based on the water-cement ratio and the total replacement rate of mineral admixtures, an interaction feature of water-cement ratio and mineral admixture replacement rate is constructed, and the mix proportion parameters, the admixture activity index features, and the interaction feature of water-cement ratio and mineral admixture replacement rate are used as model input features; S3. Construct a machine learning basic model for predicting the compressive strength of the high-volume mineral admixture concrete, and set a physical consistency constraint to characterize the relationship between the water-cement ratio and the compressive strength during the training process of the machine learning basic model. When the compressive strength predicted by the model increases with the increase of the water-cement ratio, the physical consistency constraint is used to penalize the model training. S4. Using the model input features as model input and the compressive strength corresponding to the model input features as model output, train the machine learning basic model with the physical consistency constraints to obtain a compressive strength prediction model for high-volume mineral admixture concrete.

2. The method for constructing a prediction model for the compressive strength of road concrete in cold and humid regions according to claim 1, characterized in that: In S1, the mineral admixture includes at least one of fly ash and granulated blast furnace slag; the mix proportion parameters also include the amount of coarse aggregate, the amount of fine aggregate, and the amount of water-reducing agent. The total replacement rate R of the mineral admixtures is determined according to the following formula: R=(FA+GGBS) / (C+FA+GGBS)×100%; Wherein, FA represents the amount of fly ash, GGBS represents the amount of granulated blast furnace slag, and C represents the amount of cement, and 30%≤R≤60%; the compressive strength data includes 7-day compressive strength and 28-day compressive strength; the box plot method is used to identify and remove outliers from the obtained high-volume mineral admixture concrete samples, and the sample data after removing outliers are normalized.

3. The method for constructing a predictive model for the compressive strength of road concrete in cold and humid regions according to claim 1, characterized in that, In S2, the activity index characteristics of the admixture include the amount of fly ash activity correction and / or the amount of granulated blast furnace slag activity correction. The fly ash activity correction dosage is the product of the fly ash dosage and the fly ash activity index coefficient, wherein the fly ash activity index coefficient is 0.7 to 0.9; The amount of granulated blast furnace slag used for activity correction is the product of the amount of granulated blast furnace slag used and the activity index coefficient of granulated blast furnace slag, wherein the activity index coefficient of granulated blast furnace slag is 0.9 to 1.

05. The activity index coefficients of fly ash and granulated blast furnace slag are determined based on the 28-day activity index test values ​​of the corresponding mineral admixtures.

4. The method for constructing a predictive model for the compressive strength of road concrete in cold and humid regions according to claim 1, characterized in that, In S2, the water-cement ratio-mineral admixture substitution rate interaction characteristic is determined according to the following formula: X_interaction = (W / B) × R; Wherein, W / B is the water-cement ratio, and R is the total replacement rate of mineral admixtures; the interaction feature of water-cement ratio-mineral admixture replacement rate is used to characterize the coupling relationship between water-cement ratio and concrete compressive strength under different mineral admixture replacement rates. The model input features also include maintenance age, which includes 7 days and 28 days, and is used as an independent classification input feature.

5. The method for constructing a predictive model for the compressive strength of road concrete in cold and humid regions according to claim 1, characterized in that, In S3, the underlying machine learning model is the XGBoost model; In step S4, the physical consistency constraint is set using an outer-layer iteration method, and the trend of the predicted compressive strength f relative to the water-cement ratio (W / B) is determined by numerical difference: When ∂f / ∂(W / B)>0, it is determined that the model's predicted trend does not conform to the preset law of decreasing compressive strength as the water-cement ratio increases, and the physical constraint penalty value for the corresponding training sample is determined. Lphysics=[∂f / ∂(W / B)] 2 ; When ∂f / ∂(W / B)≤0, Lphysics=0; The sample weights of the corresponding training samples are adjusted according to the physical constraint penalty values ​​of each training sample, and the adjusted sample weights are used as the sample weights of the XGBoost model for the next round of training. Through multiple rounds of outer layer iterations, the trends in the model predictions that do not conform to the stated change rules are constrained.

6. The method for constructing a predictive model for the compressive strength of road concrete in cold and humid regions according to claim 5, characterized in that, In S3, the particle swarm optimization algorithm is used to optimize the hyperparameters of the XGBoost model within a preset search space. The hyperparameters include the number of cycles, Gamma value, maximum tree depth, and learning rate. The search range for the number of iterations is 20 to 120, the search range for the Gamma value is 0.05 to 0.9, the search range for the maximum tree depth is 3 to 11, and the search range for the learning rate is 0.01 to 0.

3. The particle swarm optimization algorithm has a population size of 20 and an iteration count of 20.

7. A verification method for a predictive model of compressive strength of road concrete in cold and humid regions, wherein the verification method is used to verify the predictive model of compressive strength of high-volume mineral admixture concrete constructed by the construction method as described in claim 1, characterized in that, Includes the following steps: S5. Obtain the compressive strength prediction model of high-volume mineral admixture concrete to be verified and the test samples that have not participated in the training of the prediction model. Input the test samples into the prediction model to obtain the predicted compressive strength value of the test samples. Determine the model prediction performance evaluation index based on the predicted compressive strength value and the measured compressive strength of the corresponding test samples. Determine whether the prediction model meets the preset prediction performance requirements based on the model prediction performance evaluation index. S6. When the prediction model meets the preset prediction performance requirements, obtain the mix proportion parameters of the concrete with high mineral admixture content to be tested, and process the mix proportion parameters according to the data processing method and model input feature construction method used during the training of the prediction model to obtain the model input features of the sample to be tested. The model input features of the sample to be tested are input into the prediction model to obtain the predicted value of compressive strength and the corresponding prediction uncertainty interval. The width of the prediction uncertainty interval is compared with a preset interval width threshold to evaluate the prediction reliability of the predicted value of compressive strength. S7. When the width of the predicted uncertainty interval exceeds the preset interval width threshold, a standard test block is made according to the mix proportion of the concrete with high mineral admixture to be tested, and the compressive strength of the standard test block is tested to obtain the actual compressive strength value of the solid, and the prediction error rate is determined according to the predicted compressive strength value and the actual compressive strength value of the solid. S8. Compare the prediction error rate obtained in step S7 with the preset error rate threshold. When the prediction error rate exceeds the preset error rate threshold, add the mix proportion parameters of the concrete with high mineral admixture content to be tested and the corresponding measured values ​​of solid compressive strength as new sample data to the sample database used to train the prediction model, and update the prediction model based on the supplemented sample database.

8. The verification method for the prediction model of compressive strength of road concrete in cold and humid areas according to claim 7, characterized in that, The prediction uncertainty interval is determined based on the prediction error distribution of the prediction model on the training samples; the prediction error of each training sample is calculated and sorted according to the magnitude of the prediction error, and the prediction uncertainty interval corresponding to the high-volume mineral admixture concrete to be tested is determined based on the quantile of the prediction error distribution. The width of the predicted uncertainty interval is used as the prediction reliability evaluation parameter. When the width of the predicted uncertainty interval exceeds the preset interval width threshold, a physical test verification of the concrete with high mineral admixture content to be tested is triggered.

9. The verification method for the prediction model of compressive strength of road concrete in cold and humid areas according to claim 8, characterized in that, The preset interval width threshold is 8 MPa; The standard test block was tested according to the concrete compressive strength test method to obtain the measured value of the solid compressive strength; The prediction error is represented by a prediction error rate, which is determined according to the following formula: Prediction error rate = |Predicted compressive strength - Measured compressive strength of solid material| / Measured compressive strength of solid material × 100% When the prediction error rate exceeds 8%, the mix proportion parameters and measured values ​​of the actual compressive strength of the concrete with high mineral admixture content to be tested will be added to the sample database as new sample data.

10. The verification method for the prediction model of compressive strength of road concrete in cold and humid areas according to claim 9, characterized in that, The update method of the prediction model is determined based on the amount of new sample data n added to the sample database; When the amount of new sample data n < 30 sets, the model parameters of the current prediction model are used as the initial parameters, and a limited number of rounds of model updates are performed on the new sample data. When the amount of new sample data n≥30 sets, the prediction model is completely retrained using the sample database including the new sample data. The prediction performance of the updated prediction model is validated.

Images