A ready-mixed concrete compressive strength risk early warning method based on interpretable prediction and a storage medium

Abstract

This invention discloses a method and storage medium for predictive risk warning of compressive strength in ready-mixed concrete based on interpretable prediction, belonging to the field of concrete strength risk warning technology. The method involves acquiring multi-source production data during the ready-mixed concrete production process, performing missing data processing, anomaly identification, data alignment, and key variable labeling to construct a strength risk warning feature library. An interpretable compressive strength prediction model is established based on a gradient boosting tree model, and hyperparameter optimization is performed using an objective function that includes prediction error terms and cross-condition stability terms. The prediction deviation is calculated based on the predicted compressive strength and the target strength or control range, and a risk score is formed by combining historical verification residuals. Graded warning information is generated according to combined thresholds. For batches triggering warnings, warning triggering factors, influence directions, and sensitive ranges are generated based on SHAP and sensitivity analysis, and mapped to verification order and mix proportion adjustment suggestions. Simultaneously, the applicability of the model and warning thresholds is maintained through rolling updates and online calibration.

Description

Technical Field

[0001] This invention belongs to the field of ready-mixed concrete strength risk early warning technology. Specifically, this invention relates to a method and storage medium for early warning of ready-mixed concrete compressive strength risk based on interpretable prediction. Background Technology

[0002] The compressive strength of ready-mixed concrete is a crucial indicator for evaluating the structural safety and final product quality of engineering projects. Current quality control primarily relies on compressive strength testing of test blocks after standard curing periods, which suffers from significant feedback lag and makes it difficult to identify strength deviation risks in a timely manner during the production stage. To predict the strength level of production batches in advance, existing methods utilize mix proportion parameters, metering records, and raw material inspection data to establish strength prediction models.

[0003] Chinese Patent 104991051A provides a method for predicting concrete strength based on a hybrid model. First, strength experiments are conducted on-site using standard concrete strength testing methods for different concrete mix proportions to obtain multiple sets of learning samples: "cement x1, blast furnace slag powder x2, fly ash x3, water x4, water-reducing agent x5, coarse aggregate x6, fine aggregate x7, curing age x8, concrete mix proportion information - concrete strength y". The hybrid model is then trained using an extreme learning machine, artificial neural network, and support vector machine. The optimal model is determined by minimizing the relative error. Based on this, the predicted concrete strength is determined according to the predicted values ​​from various modeling methods.

[0004] However, existing strength prediction methods primarily focus on outputting predicted strength values, often lacking stable risk warning mechanisms tailored to the production stage. Simply obtaining predicted values ​​is insufficient for production quality control; it's also necessary to assess the strength deviation risk of the current batch in conjunction with the target strength or target strength control range, generating early warning information that can be used for on-site verification. Furthermore, in the actual production of ready-mixed concrete, raw material batches, cementitious material systems, mix proportions, production sites, and seasonal environments continuously change. If model training and parameter optimization only focus on the overall prediction error, lacking constraints on error fluctuations under different working conditions, it may lead to increased prediction deviations in some conditions, affecting the stability of risk scores and early warning levels. Summary of the Invention

[0005] The present invention aims to provide a method and storage medium for early warning of compressive strength risk of ready-mixed concrete based on interpretable prediction, so as to achieve the purpose of early identification, stable early warning, cause explanation and rolling calibration of compressive strength risk of ready-mixed concrete.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] This invention provides a method for early warning of compressive strength risk in ready-mixed concrete based on interpretable prediction, comprising the following steps:

[0008] Step 1: Collect multi-source production data during the ready-mixed concrete production process;

[0009] Step 2: Perform missing data processing, anomaly identification, and data alignment on multi-source production data, and label key variables that affect the risk assessment of concrete compressive strength.

[0010] Step 3: Construct a feature library for intensity risk early warning, including the original variables of mix proportion and feed, as well as their derived variables;

[0011] Step 4: Establish an interpretable compressive strength prediction model based on the gradient boosting tree regression model, and output the predicted compressive strength values ​​for the production batch;

[0012] Step 5: Construct an objective function that includes a prediction error term and a cross-condition stability term, and use the particle swarm optimization algorithm to adaptively optimize the hyperparameters of the compressive strength prediction model;

[0013] Step 6: Calculate the predicted deviation based on the predicted compressive strength and the target strength or the target strength control range, and calculate the risk score by combining the historical verification residual standard deviation, recent rolling residual fluctuation or the error index of the working condition group. Generate graded risk warning information according to the combined threshold of the predicted deviation and the risk score.

[0014] Step 7: For production batches that trigger risk warnings, generate explanatory outputs based on the Shapley additive interpretation method (SHAP) and sensitivity analysis, including warning triggering factors, direction of influence, and sensitive range. Map the explanatory outputs to suggestions for checking the order and adjusting the mix ratio.

[0015] Step 8: Continuously collect measured intensity results of new production batches, and continuously update and calibrate model parameters, risk scores, and early warning thresholds online.

[0016] In step one, the collected multi-source production data includes raw material inspection data, mixing ratio parameters, metering and feeding records, mixing parameters, slump at the outlet and adjustment records, transportation time and waiting time, and environmental temperature and humidity data; and a unique identifier is established for each production batch to establish the correlation between the multi-source production data.

[0017] In step two, missing data is handled by deletion, statistical imputation, or group imputation; anomaly identification is performed by combining engineering constraint rules and statistical rules, and anomaly marker fields are retained; data alignment includes mapping raw material inspection indicators to corresponding production batches by batch or expiration date, and aggregating environmental temperature and humidity into statistical quantities of mean, extreme values, and fluctuation range by production window.

[0018] In step three, the intensity risk early warning feature library includes original variables and derived variables. The original variables include the raw material ledger field and the metering and feeding field; the derived variables include the feeding deviation statistics, cementitious material combination characteristics, water-cement ratio consistency characteristics, raw material batch change characteristics, and historical rolling characteristics.

[0019] In step four, the strength risk warning feature library is used as input and the compressive strength of concrete production batches is used as output. Cross-validation or time-rolling validation is used to evaluate the model performance, and the mean, standard deviation and quantile of the validation residuals are statistically analyzed. The model version, training data range, feature definition and model information required for interpretable analysis are saved.

[0020] In step five, the particle swarm optimization algorithm is used to search and update the model's hyperparameters; the candidate parameter set is evaluated using the objective function and iteratively updated; after iterative convergence, the optimal parameter set is determined, and the optimized interpretable compressive strength prediction model is retrained. The objective function includes a prediction error term and a cross-condition stability term. The prediction error term is the root mean square error, mean absolute error, or a weighted combination thereof on the validation set; the cross-condition stability term is constructed based on the error index of the condition grouping.

[0021] The objective function (J) is:

[0022] J = a×RMSE + b×MAE + λ×Stab;

[0023] Where RMSE is the root mean square error on the validation set, MAE is the mean absolute error on the validation set, Stab is the cross-condition stability term, a is the weighting coefficient of RMSE, b is the weighting coefficient of MAE, and λ is the stability weighting coefficient. The cross-condition stability term is constructed based on one or more of the following: root mean square error, mean absolute error, group error variance, group error range, or maximum group error under different condition groups.

[0024] In step six, the predicted deviation is calculated based on the predicted compressive strength and the target strength or the target strength control range; the risk score is calculated based on one or more of the absolute value of the predicted deviation, the standard deviation of historical verification residuals, the recent rolling residual fluctuations, and the error index of the working condition group; and graded early warning information is generated according to the combined threshold of the predicted deviation and the risk score.

[0025] In step seven, global and local interpretation results are calculated based on the Shapley additive interpretation method (SHAP) to identify key factors that contribute significantly to the risk warning of the current batch; the influence direction and sensitive range of the key factors are obtained by combining sensitivity analysis to form interpretation output; the interpretation output is mapped into a list of actionable disposal suggestions and priorities, including suggestions for checking the order and adjusting the mixing ratio.

[0026] The present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, performs the above-described method.

[0027] The technical effects of this invention are as follows:

[0028] (1) This invention enables interpretable output of compressive strength risk warning results. Based on the interpretable compressive strength prediction model and SHAP analysis, the key factors that cause the current batch to trigger risk warning can be identified, and the direction of influence and sensitive range can be output, so that on-site personnel can clearly identify the source of risk.

[0029] (2) This invention can identify the risk of compressive strength deviation in advance during the production stage. By calculating the predicted deviation between the predicted compressive strength and the target strength or the target strength control range, and combining the historical verification residual standard deviation, recent rolling residual fluctuation and the error index of the working condition group, a risk score is formed, which can generate graded risk warning information before the standard maintenance strength result is recovered.

[0030] (3) This invention realizes risk classification early warning and rapid response based on combined thresholds. By establishing a classification early warning mechanism through a combination of predicted deviation and risk score thresholds, it can distinguish between normal predicted fluctuations and abnormal deviations that require intervention, and output the risk level, deviation magnitude, triggering rules, verification suggestions and verification order in the early warning information, providing a clear investigation path for on-site personnel.

[0031] (4) This invention improves the stability of risk warning under different production conditions. By introducing a cross-condition stability term into the objective function, the model not only pursues the minimum overall prediction error, but also constrains the error fluctuation under different raw material systems, mix proportions, sites, seasonal windows or intensity levels, thereby reducing the risk of prediction inaccuracy and unstable warning caused by changes in operating conditions.

[0032] (5) This invention has the ability to continuously update and calibrate online. By continuously collecting the measured intensity results of new production batches, the model parameters, risk scores and warning thresholds are iteratively updated, so that the risk warning mechanism can adapt to changes in raw material batches, production conditions and environmental conditions, and ensure the continuous effectiveness and engineering availability of the warning function. Attached Figure Description

[0033] This manual includes the following figures, which illustrate the following:

[0034] Figure 1 This is a flowchart of a method for early warning of compressive strength risk of ready-mixed concrete based on interpretable prediction and a storage medium according to the present invention;

[0035] Figure 2 This invention provides the cross-condition stability constraint optimization process and objective function convergence curve for the interpretable prediction model.

[0036] Figure 3 This is an explanatory diagram of the risk warning triggering factors based on SHAP in this invention;

[0037] Figure 4 This is a sample diagram showing the output of compressive strength risk warning information and verification and handling suggestions in this invention. Detailed Implementation

[0038] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, in order to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention, and to facilitate its implementation.

[0039] This invention provides a method for early warning of compressive strength risk in ready-mixed concrete based on interpretable prediction, comprising the following steps:

[0040] Step 1: Collect multi-source production data during the ready-mixed concrete production process;

[0041] Step 2: Perform missing data processing, anomaly identification, and data alignment on multi-source production data, and label key variables that affect the risk assessment of concrete compressive strength.

[0042] Step 3: Construct a feature library for intensity risk early warning, including the original variables of mix proportion and feed, as well as their derived variables;

[0043] Step 4: Establish an interpretable compressive strength prediction model based on the gradient boosting tree regression model, and output the predicted compressive strength values ​​for the production batch;

[0044] Step 5: Construct an objective function that includes a prediction error term and a cross-condition stability term, and use the particle swarm optimization algorithm to adaptively optimize the hyperparameters of the compressive strength prediction model;

[0045] Step 6: Calculate the predicted deviation based on the predicted compressive strength and the target strength or the target strength control range, and calculate the risk score by combining the historical verification residual standard deviation, recent rolling residual fluctuation or the error index of the working condition group. Generate graded risk warning information according to the combined threshold of the predicted deviation and the risk score.

[0046] Step 7: For production batches that trigger risk warnings, generate explanatory outputs based on the Shapley additive interpretation method (SHAP) and sensitivity analysis, including warning triggering factors, direction of influence, and sensitive range. Map the explanatory outputs to suggestions for checking the order and adjusting the mix ratio.

[0047] Step 8: Continuously collect measured intensity results of new production batches, and continuously update and calibrate model parameters, risk scores, and early warning thresholds online.

[0048] In step one, the collected multi-source production data includes raw material inspection data, mixing ratio parameters, metering and feeding records, mixing parameters, slump at the outlet and adjustment records, transportation time and waiting time, and environmental temperature and humidity data; and a unique identifier is established for each production batch to establish the correlation between the multi-source production data.

[0049] In step two, missing data is handled by deletion, statistical imputation, or group imputation; anomaly identification is performed by combining engineering constraint rules and statistical rules, and anomaly marker fields are retained; data alignment includes mapping raw material inspection indicators to corresponding production batches by batch or expiration date, and aggregating environmental temperature and humidity into statistical quantities of mean, extreme values, and fluctuation range by production window.

[0050] In step three, the intensity risk early warning feature library includes original variables and derived variables. The original variables include raw material ledger fields and metering input fields; the derived variables include input deviation statistics, cementitious material combination characteristics, water-cement ratio consistency characteristics, raw material batch variation characteristics, and historical rolling characteristics.

[0051] In step four, the strength risk warning feature library is used as input and the compressive strength of concrete production batches is used as output. Cross-validation or time-rolling validation is used to evaluate the model performance, and the mean, standard deviation and quantile of the validation residuals are statistically analyzed. The model version, training data range, feature definition and model information required for interpretable analysis are saved.

[0052] In step five, the particle swarm optimization algorithm is used to search and update the model's hyperparameters; the candidate parameter set is evaluated using the objective function and iteratively updated; after iterative convergence, the optimal parameter set is determined, and the optimized interpretable compressive strength prediction model is retrained. The objective function includes a prediction error term and a cross-condition stability term. The prediction error term is the root mean square error, mean absolute error, or a weighted combination thereof on the validation set; the cross-condition stability term is constructed based on the error index of the condition grouping.

[0053] The objective function (J) is:

[0054] J = a×RMSE + b×MAE + λ×Stab;

[0055] Where RMSE is the root mean square error on the validation set, MAE is the mean absolute error on the validation set, Stab is the cross-condition stability term, a is the weighting coefficient of RMSE, b is the weighting coefficient of MAE, and λ is the stability weighting coefficient.

[0056] In step six, the predicted deviation is calculated based on the predicted compressive strength and the target strength or the target strength control range; the risk score is calculated based on one or more of the absolute value of the predicted deviation, the standard deviation of historical verification residuals, the recent rolling residual fluctuations, and the error index of the working condition group; and graded early warning information is generated according to the combined threshold of the predicted deviation and the risk score.

[0057] In step seven, global and local interpretation results are calculated based on the Shapley additive interpretation method (SHAP) to identify key factors that contribute significantly to the risk warning of the current batch; the influence direction and sensitive range of the key factors are obtained by combining sensitivity analysis to form interpretation output; the interpretation output is mapped into a list of actionable disposal suggestions and priorities, including suggestions for checking the order and adjusting the mixing ratio.

[0058] In step eight, the measured intensity results of new production batches are continuously collected, and the model parameters, risk scores, and early warning thresholds are continuously updated and calibrated online.

[0059] The present invention provides a storage medium having a computer program stored thereon, which, when executed by a processor, performs the above-described method.

[0060] The present invention will now be described in detail.

[0061] This invention is deployed within the data environment of ready-mixed concrete production enterprises. Data sources include raw material inspection and material ledger data, mix proportion and metering data, and are associated with corresponding 28-day standard curing compressive strength test results. This invention establishes a unique batch identifier based on the production batch, linking the batch identifier to the raw material batch ledger, mix proportion number, metering records, and strength results to form training samples and online prediction samples.

[0062] Data such as raw material inspection data, mix proportion parameters, metering and feeding records, mixing parameters, slump at discharge and adjustment records, transportation time and waiting time, and environmental temperature and humidity are collected for each production batch. A unique identifier is established for each production batch, and a correlation is established between the production batch and the mix proportion, feeding, discharge, transportation, and strength results. Among them, the multi-source production data includes at least the fields related to compressive strength risk warning, including cement, cement type, mineral powder, fly ash, water-cement ratio, water, sand ratio, natural sand, dry manufactured sand, medium and large crushed stone, admixtures, and water-reducing agent dosage.

[0063] In this embodiment, for each production batch, the following data are collected from the raw material ledger and inspection records: cement type, mineral powder, fly ash, natural sand, dry manufactured sand, medium and large crushed stone. The mix proportions and metering records are also collected for cement, water, water-cement ratio, sand ratio, admixtures, and water-reducing agent dosages. Raw material ledger indicators are mapped to production batches according to raw material batch numbers, with a validity period of 7 to 30 days. Metering records are aligned with batches according to batch number or feeding time, with the alignment window taking 10 minutes before and after the feeding time. Strength results are associated with the standard 28-day compressive strength according to batch number.

[0064] The collected data underwent missing value processing, anomaly identification, standardization, and data alignment to form training and prediction samples. Missing values ​​were handled through deletion, statistical imputation, or group imputation. Anomaly identification employed a combination of engineering constraint rules and statistical rules, retaining anomaly marker fields. Data alignment utilized time or batch alignment, including mapping raw material inspection indicators to corresponding production batches by batch or expiration date, and aggregating environmental temperature and humidity into statistical quantities such as mean, extreme values, and fluctuation ranges by production window. Key variables affecting strength risk assessment were labeled according to their controllability, traceability, and operating condition attributes. Controllable variables included at least the water-cement ratio, water, sand ratio, and the dosage of admixtures and water-reducing agents; traceable variables included at least the raw material batch information corresponding to the largest component of crushed stone among cement type, mineral powder, fly ash, natural sand, dry manufactured sand, and crushed stone; and operating condition variables included at least the raw material system, mix proportion type, site, seasonal window, or strength grade.

[0065] In embodiments of this invention, the collected data undergoes standardization and unit conversion. Mass units are standardized to kg, strength units to MPa, and admixture-related fields to kg or mass percentage, while retaining the original unit fields for traceability. Missing values ​​are filled using deletion, statistical imputation, or grouping imputation; batches with more than 20% missing key fields are excluded from the training set. Outlier identification employs a combination of engineering constraint rules and statistical rules, retaining anomaly marker fields. Engineering constraint rules include at least a water-cement ratio of 0.25 to 0.70, a sand ratio of 30% to 50%, an admixture dosage of 0% to 3%, and a water-reducing agent dosage of 0% to 3%; statistical rules employ IQR or 3σ methods. After batch alignment, training and prediction samples are formed, and key variables are labeled with adjustable, traceable, and operational attributes.

[0066] A strength risk early warning feature library is constructed, comprising original variables and derived variables of mix proportion and material input. Original variables include raw material ledger fields and metered material input fields. Derived variables include at least: material input deviation statistics, calculated as the difference and relative deviation between actual material input and the set mix proportion; cementitious material combination characteristics, such as the proportion of mineral powder and fly ash in cementitious materials; water-cement ratio consistency characteristics, calculated based on the amount of water and cementitious materials when the water-cement ratio field is missing, and verified against the recorded water-cement ratio, with deviations exceeding 0.02 marked as abnormal; raw material batch variation characteristics, used to characterize changes in cement type, admixtures, or aggregate sources; and historical rolling characteristics, calculating the mean and variance of strength, key variables, and residuals for the last N batches grouped by the same mix proportion, cement type, or working condition, where N ranges from 30 to 100.

[0067] Operating condition grouping is used to improve the stability of interpretable predictive models under different production conditions. Operating condition grouping can be determined by one or more of the following: raw material system, mix proportion type, site, seasonal window, and strength grade. For example, operating condition groups can be formed according to cement type, mineral powder or fly ash content range, strength grade, production site, and month or quarter. By grouping operating conditions, the validation error under different operating conditions is statistically analyzed to constrain the error fluctuation of the model under different production conditions.

[0068] A gradient boosting tree regression model was constructed on the training set as an interpretable compressive strength prediction model, with XGBoost regression being the preferred option. Using the intensity risk warning feature library as input and the standard maintenance 28-day compressive strength corresponding to each batch as output, 5-fold cross-validation or time-rolling validation was employed to evaluate model performance, with an early cessation number of batches set to 50 to avoid overfitting. The prediction error sequence was calculated from the validation results, and the mean, standard deviation, and 90th and 95th quantiles of the residuals were statistically analyzed in the validation set as error references for subsequent risk scoring and warning threshold setting. After training, the model version, training data time range, feature list, feature calculation methods, and model structure information required for interpretable analysis were saved. The intensity age field was also recorded as 28 days to ensure consistency between the training label and the recovery intensity results.

[0069] The particle swarm optimization (PSO) algorithm is used to search and update the key hyperparameters of XGBoost. These hyperparameters include at least the learning rate, maximum tree depth, subsampling ratio, column sampling ratio, minimum leaf weight, regularization term, and number of trees. In this embodiment, the hyperparameter search range is set as follows: learning rate 0.01 to 0.30, maximum tree depth 3 to 10, subsampling ratio 0.60 to 1.00, column sampling ratio 0.60 to 1.00, minimum leaf weight 1 to 10, L1 regularization coefficient 0 to 5, L2 regularization coefficient 0 to 10, and number of trees 100 to 1500. The particle swarm size is set to 20 to 40, the number of iterations is set to 30 to 80, and an early stopping condition is set to terminate when the objective function improvement is less than 1e-4 after 10 consecutive iterations.

[0070] Construct an objective function J for optimization, including at least a prediction error term and a cross-condition stability term. The prediction error term is the RMSE, MAE, or a weighted combination thereof on the validation set. The cross-condition stability term is constructed based on the error index of the condition grouping; the stability term can be the grouped RMSE variance, grouped MAE variance, maximum grouping error, or grouping error range. By simultaneously minimizing the prediction error and the cross-condition stability term, the applicability of the interpretable prediction model under different raw material systems, mix proportions, sites, and seasonal windows is improved.

[0071] The prediction error term is a weighted combination of the validation set RMSE and MAE:

[0072] J = 0.7×RMSE + 0.3×MAE + λ×Stab

[0073] The units for RMSE and MAE are MPa. The cross-condition stability term Stab is constructed based on the error grouping of conditions, which includes raw material system, site, or mix design type. λ is the stability weighting coefficient, which can range from 0.2 to 0.6, used to balance the overall error with cross-condition error fluctuations. Stab can be the variance or maximum value of the grouped RMSE.

[0074] Stab = max(RMSEg) − min(RMSEg)

[0075] Where RMSEg represents the verification error of group g. The raw material system group is g1, the site group is g2, and the mix proportion type group is g3. When all three types of working conditions exist simultaneously, instead of selecting only one group, a set of group RMSEs is calculated for each of g1, g2, and g3, and the worst result is taken from the union to form the composite Stab.

[0076] After optimization, record the optimal parameter set, objective function value, and error of each group for subsequent public disclosure and auditing.

[0077] like Figure 2As shown, it illustrates the cross-condition stability constraint optimization process and objective function convergence curve of the interpretable prediction model.

[0078] The optimized prediction model is applied to the new production batch samples to output the predicted compressive strength fc, in MPa. The prediction deviation Δ is calculated based on the target strength; when the target strength (ft) is used as the baseline,

[0079] Δ = fc − ft

[0080] The risk score R can be constructed by combining the rolling statistics of historical validation residuals and standardized using the standard deviation σe of the validation residuals:

[0081] R = |Δ| / σe

[0082] σe is updated on a rolling basis according to the intensity results of the most recent N batches of recovered samples, where N is set to 200 to 1000.

[0083] The tiered early warning threshold can be determined using a combination of deviation amount and risk score rules, for example:

[0084] Level 1 warning: |Δ| ≥ 4 MPa or R ≥ 2.5

[0085] Level II warning: 2 MPa ≤ |Δ| < 4 MPa or 1.5 ≤ R < 2.5

[0086] Level 3 warning: 1 MPa ≤ |Δ| < 2 MPa or 1.0 ≤ R < 1.5

[0087] No warning: |Δ| < 1 MPa and R < 1.0

[0088] The aforementioned thresholds are jointly determined by the historical validation residual distribution and the enterprise's process control requirements: The standard deviation σe and quantiles of the residual e, such as P90 and P95, are calculated on the validation set or the most recent N batches of samples, where R is set according to the stratified intervals of |Δ| / σe; the Δ threshold adopts the minimum executable level of the enterprise's commonly used intensity control bandwidth and is aligned with the magnitude of σe for initial setting. The thresholds are initialized based on the validation set results at the initial stage of deployment and periodically reviewed based on the error distribution using a rolling window. Adjustments are made as needed according to changes in operating conditions, and records are kept.

[0089] The early warning information should include at least the batch identifier, predicted intensity fc, target intensity ft or control range, deviation Δ, risk score R, warning level, and generation time, and output verification suggestions and verification order. Verification suggestions are generated based on model input features and should include at least the feeding records for water-cement ratio, water, cement, sand ratio, admixtures, and water-reducing agent dosages, as well as the raw material ledger information and corresponding batch numbers for cement type, mineral powder, fly ash, natural sand, dry manufactured sand, and medium and large crushed stone. The verification order is determined according to the principle of strength sensitivity and rapid on-site verification: first, verify the water-cement ratio and water usage; second, verify the admixture and water-reducing agent dosages; third, verify the sand ratio and cement usage; and finally, verify the batch information of the raw material system, including changes in cement type, admixture type, and main aggregate source batches.

[0090] like Figure 3 The risk warning trigger factor interpretation diagram based on SHAP shows that for batches that trigger warnings or require analysis, the local interpretation results are calculated based on SHAP, outputting the top K factors that contribute the most to the risk warning for that batch, where K ranges from 6 to 10, and outputting the magnitude and positive / negative direction of the SHAP value of each factor. The SHAP interpretation results are used to identify the key trigger factors that cause the prediction deviation or the warning level to rise for that batch, rather than just for ranking the importance of ordinary features.

[0091] Sensitivity analysis was conducted on key factors using a univariate perturbation approach. While keeping other variables constant, the key variables were taken from their historical 5% to 95th percentile ranges. The predicted intensity change curve was calculated, and the variable change that caused the predicted intensity change to reach 1 MPa was output as the sensitivity index.

[0092] Combination Figure 4 The output template for the compressive strength risk warning information and verification and handling suggestions is provided. The handling suggestions are output in list form, including at least the suggested item, suggested scope, execution order, and traceability field. For adjustable variables, suggestions for adjusting and optimizing the mix proportion are generated; for traceable variables, suggestions for raw material batch verification are generated; for operating condition variables, suggestions for operating condition review or threshold calibration are generated. The direction of the suggestions is determined based on the predicted deviation direction of the current batch, the SHAP contribution direction of key factors, sensitivity analysis results, and on-site process constraints; the execution order is determined comprehensively based on the sensitivity index and the adjustability of the variable. Generally, the smaller the sensitivity index, the more significant the change in the predicted strength can be caused by a small change in the variable, and the higher the priority; however, for variables limited by workability, safety, or process conditions, their execution priority should be adjusted in conjunction with on-site constraints.

[0093] It is recommended to first use an initial range obtained from historical statistics, and then perform a simple verification using sensitivity analysis results: retain the original range for stable responses, reduce the priority of responses that are too weak, and appropriately narrow the range for responses that are too strong. The final range must also meet the constraints of the on-site process. For example, after a batch triggers an early warning, the local interpretation results show that cement, mineral powder, and water-cement ratio are the top three influencing factors, with negative SHAP values ​​for cement and mineral powder, and positive SHAP values ​​for water-cement ratio. Sensitivity analysis shows that a 1.2% change in water-cement ratio can cause a 1MPa change in predicted strength, a 1.5% change in cement can cause a 1MPa change in predicted strength, and a 2% change in mineral powder can cause a 1MPa change. Therefore, water-cement ratio has the highest sensitivity, followed by cement and mineral powder. If the water-cement ratio is not suitable for priority adjustment due to workability constraints, then cement measurement and cement batch information should be checked first, and cement should be the priority adjustment item for subsequent batches with similar working conditions; mineral powder should be the second priority adjustment item; and water-cement ratio should be the adjustment item after workability verification. The current batch that triggered the early warning should be output with priority for verification, and the mix proportion adjustment is recommended for subsequent batches with similar working conditions.

[0094] This invention continuously collects new batches of data and intensity results using a rolling window. An update is triggered when the verification error exceeds a threshold for M consecutive batches, the operating condition distribution changes, the error fluctuation of the operating condition group increases, or the error changes due to raw material batch switching. This triggers data write-back and retraining. If necessary, the optimal parameter set is updated again and version iteration is completed. Simultaneously, the model version, feature definitions, risk scoring parameters, and warning thresholds are updated and recorded. Online calibration prioritizes the most recent window data to maintain the applicability of the model and warning mechanism under the current operating conditions.

[0095] The beneficial effects of the present invention are described in detail below.

[0096] This invention enables interpretable output of compressive strength risk warning results. Based on an interpretable compressive strength prediction model and SHAP analysis, it can identify key factors that trigger risk warnings for the current batch and output the direction of influence and sensitive range, enabling on-site personnel to clearly identify the source of risk.

[0097] This invention can identify the risk of compressive strength deviation in advance during the production stage. By calculating the predicted deviation between the predicted compressive strength and the target strength or the target strength control range, and combining it with the standard deviation of historical verification residuals, recent rolling residual fluctuations, and the error index of the corresponding working conditions, a risk score is formed, which can generate graded risk warning information before the standard maintenance strength results are recovered.

[0098] This invention realizes risk classification early warning and rapid response based on combined thresholds. By establishing a classified early warning mechanism through a combination of predicted deviation and risk score thresholds, it can distinguish between normal predicted fluctuations and abnormal deviations requiring intervention. The early warning information outputs the risk level, deviation magnitude, trigger rules, verification suggestions, and verification order, providing on-site personnel with a clear investigation path.

[0099] This invention improves the stability of risk warnings under different production conditions. By introducing a cross-condition stability term into the objective function, the model not only aims to minimize the overall prediction error but also constrains error fluctuations under different raw material systems, mix proportions, sites, seasonal windows, or intensity levels, thereby reducing the risk of prediction inaccuracies and unstable warnings caused by changes in operating conditions.

[0100] This invention features rolling updates and online calibration capabilities. By continuously collecting measured intensity results from new production batches, the model parameters, risk scores, and early warning thresholds are iteratively updated, enabling the risk early warning mechanism to adapt to changes in raw material batches, production conditions, and environmental conditions, thus ensuring the continuous effectiveness and engineering usability of the early warning function.

[0101] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution; or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A ready-mixed concrete compressive strength risk early warning method based on interpretable prediction, characterized in that, Includes the following steps: Step 1: Collect multi-source production data during the ready-mixed concrete production process; Step 2: Perform missing data processing, anomaly identification, and data alignment on multi-source production data, and label key variables that affect the risk assessment of concrete compressive strength. Step 3: Construct a feature library for intensity risk early warning, wherein the features include the original variables of mix proportion and feeding and their derived variables; Step 4: Establish an interpretable compressive strength prediction model based on the gradient boosting tree regression model, and output the predicted compressive strength values ​​for the production batch; Step 5: Construct an objective function that includes a prediction error term and a cross-condition stability term, and use a particle swarm optimization algorithm to adaptively optimize the hyperparameters of the compressive strength prediction model based on the objective function; Step 6: Calculate the predicted deviation based on the predicted compressive strength and the target strength or the target strength control range, and calculate the risk score by combining the historical verification residual standard deviation, recent rolling residual fluctuation or the error index of the working condition group. Generate graded risk warning information according to the combined threshold of the predicted deviation and the risk score. Step 7: For production batches that trigger risk warnings, generate explanatory outputs based on the Shapley additive interpretation method (SHAP) and sensitivity analysis, including warning triggering factors, direction of influence, and sensitive range. Map the explanatory outputs to suggestions for checking the order and adjusting the mix ratio. Step 8: Continuously collect measured intensity results of new production batches, and continuously update and calibrate model parameters, risk scores, and early warning thresholds online.

2. The ready-mixed concrete compressive strength risk early-warning method based on interpretable prediction of claim 1, characterized in that: In step one, the collected multi-source production data includes raw material inspection data, mixing ratio parameters, metering and feeding records, mixing parameters, slump at the outlet and adjustment records, transportation time and waiting time, and environmental temperature and humidity data; and a unique identifier is established for each production batch to establish the correlation between the multi-source production data.

3. The method for early warning of compressive strength risk of ready-mixed concrete based on interpretable prediction as described in claim 1, characterized in that: In step two, missing data is handled by deletion, statistical imputation, or group imputation; anomaly identification is performed by combining engineering constraint rules and statistical rules, and anomaly marker fields are retained; data alignment includes mapping raw material inspection indicators to corresponding production batches by batch or expiration date, and aggregating environmental temperature and humidity into statistical quantities of mean, extreme values, and fluctuation range by production window.

4. The ready-mixed concrete compressive strength risk early-warning method based on interpretable prediction of claim 1, wherein: In step three, the intensity risk early warning feature library includes original variables and derived variables. The original variables include the raw material ledger field and the metering and feeding field; the derived variables include the feeding deviation statistics, cementitious material combination characteristics, water-cement ratio consistency characteristics, raw material batch change characteristics, and historical rolling characteristics.

5. The ready-mixed concrete compressive strength risk early-warning method based on interpretable prediction of claim 1, wherein: In step four, the strength risk warning feature library is used as input and the compressive strength of concrete production batches is used as output. Cross-validation or time-rolling validation is used to evaluate the model performance, and the mean, standard deviation and quantile of the validation residuals are statistically analyzed. The model version, training data range, feature definition and model information required for interpretable analysis are saved.

6. The ready-mixed concrete compressive strength risk early-warning method based on interpretable prediction of claim 1, wherein: In step five, the objective function includes a prediction error term and a cross-condition stability term. The prediction error term is the root mean square error, mean absolute error, or a weighted combination thereof on the validation set. The cross-condition stability term is constructed based on the error index of condition grouping.

7. The ready-mixed concrete compressive strength risk early-warning method based on interpretable prediction of claim 6, wherein: The cross-condition stability term is constructed based on one or more of the following: root mean square error, mean absolute error, group error variance, group error range, or maximum group error under different condition groups; the condition groups include one or more of the following: raw material system, mix proportion type, site, seasonal window, and strength grade.

8. The ready-mixed concrete compressive strength risk early-warning method based on interpretable prediction of claim 1, wherein: In step six, the predicted deviation is calculated based on the predicted compressive strength and the target strength or the target strength control range; the risk score is calculated based on one or more of the absolute value of the predicted deviation, the standard deviation of historical verification residuals, the recent rolling residual fluctuations, and the error index of the working condition group; and graded early warning information is generated according to the combined threshold of the predicted deviation and the risk score.

9. The ready-mixed concrete compressive strength risk early-warning method based on interpretable prediction of claim 1, wherein: In step seven, global and local interpretation results are calculated based on the Shapley additive interpretation method (SHAP) to identify key factors that trigger risk warnings for the current batch; combined with sensitivity analysis, the influence direction and sensitive range of the key factors are obtained, forming the interpretation output; Based on the controllable, traceable, or operational attributes of key factors, the explanatory output is mapped into a verification sequence, a list of disposal recommendations, and suggestions for adjusting and optimizing the mix ratio.

10. A storage medium having stored thereon a computer program, characterized in that When the computer program is executed by a processor, it performs the method as described in any one of claims 1 to 9.

Images