Concrete mix proportion intelligent generation and online compliance checking method and system

Abstract

This invention discloses a method and system for intelligent generation and online compliance verification of concrete mix proportions, relating to the fields of intelligent concrete engineering and industrial internet technology. The method includes training a multi-objective prediction model using a joint loss function containing a supervision error term and an engineering physical constraint penalty term, guiding the model output to tend towards the engineering rule boundary set; inputting the sample to be predicted into the model to obtain the original prediction result; verifying the original prediction result; if it does not meet the engineering rule boundary set, generating a pattern vector, and matching a deterministic repair operator sequence from a repair strategy mapping table based on the pattern vector and key process parameters, thereby obtaining the final mix proportion that meets the constraints through repair; if the repair fails, outputting a safe alternative mix proportion or reverting to a historical effective mix proportion, and issuing an alarm. This invention, through the synergy of soft constraint training and rule-based repair, ensures that the output meets 100% of the engineering constraints while maintaining accuracy.

Description

Technical Field

[0001] This invention belongs to the field of intelligent concrete engineering and industrial internet technology. Specifically, it relates to a method and system for intelligent generation and online compliance verification of concrete mix proportions, which is particularly suitable for mix proportion design, dynamic adjustment and quality traceability in the production process of mixing plants. Background Technology

[0002] Concrete mix design is a core element determining the workability and service performance of concrete. With the widespread application of high-performance concrete and the increasing complexity of construction conditions, mix design is gradually shifting from traditional empirical methods to data-driven and intelligent prediction methods. Existing data-driven mix prediction methods typically employ regression models or multi-objective machine learning models to jointly predict key engineering parameters such as unit water consumption, water-cement ratio, and cementitious material dosage. Their optimization objectives are mostly focused on minimizing prediction errors or improving statistical fitting accuracy.

[0003] However, in practical engineering applications, mix proportion prediction results not only need to be statistically reasonable, but also must meet a series of rigid constraints imposed by the physical laws of concrete engineering. These constraints include at least: the engineering allowable range of material usage (such as upper and lower limits of water usage, safe range of water-cement ratio), the volume balance equation (the sum of the volumes of all components equals the unit volume), and feasibility constraints related to workability (such as slump, pumpability requirements). In the absence of effective physical constraint mechanisms, prediction models are prone to outputting mix proportion schemes that fit well at the numerical level but are completely unfeasible in terms of engineering physics. For example, the prediction results may lead to negative aggregate volume (violating volume conservation) or water-cement ratios exceeding the workable range. Such schemes cannot be directly put into production, and in engineering, adjustments or trial mixes can usually only be made retrospectively based on human experience. This not only weakens the engineering application value of the prediction model, but also increases the volatility and uncertainty of the design and construction process.

[0004] To address the aforementioned issues, academia and industry have primarily pursued improvements along two technical routes: one is to introduce physical constraint penalty terms during the model training phase to guide the model output toward the engineering feasible region in the form of soft constraints; the other is to perform post-processing corrections after the prediction results are output, forcing the satisfaction of engineering constraints through pruning, recalculation, or optimization.

[0005] However, while training with soft constraints alone can reduce the overall probability of default, it is difficult to guarantee that all samples will meet the engineering constraints during the inference phase, especially for out-of-distribution or boundary samples, where the risk of default still exists. Relying solely on post-processing corrections, while achieving mandatory compliance, has the following inherent drawbacks: First, general constraint optimization solvers (such as sequential quadratic programming) have high computational complexity and uncontrollable iteration processes, making it difficult to meet the real-time requirements of millisecond-level response in online production at mixing plants; second, the correction process lacks deterministic rules, and the correction magnitude may be too large, severely disturbing the original prediction results and affecting the accurate judgment of engineering performance; third, the correction process is usually not auditable, failing to meet the requirements of end-to-end traceability for engineering quality management.

[0006] Therefore, in concrete mix design engineering applications, there is an urgent need for a technical solution that can guide the model output towards the engineering feasible region through soft constraints during the model training stage, while performing regular, interpretable, and limited-range deterministic corrections on individual non-compliant results during the inference stage, and ensuring that the correction process is controllable in real time and traceable throughout the entire process, so as to simultaneously guarantee prediction accuracy, engineering feasibility, online real-time performance, and quality management compliance. Summary of the Invention

[0007] To address the aforementioned deficiencies in existing technologies, the present invention aims to provide a method and system for intelligent generation and online compliance verification of concrete mix proportions. This system solves the problems in existing technologies, such as the difficulty in ensuring that the output during the inference stage of soft-constraint training always meets engineering physical constraints, and the uncontrollable correction magnitude, poor real-time performance, and unexplainable and untraceable correction process of individual post-processing corrections. This system achieves synergistic optimization of prediction accuracy and engineering feasibility, and meets the engineering requirements of real-time performance and quality traceability for online production at mixing plants.

[0008] This invention solves the above-mentioned technical problems through the following technical solution: a method for intelligent generation and online compliance verification of concrete mix proportions, comprising:

[0009] The acquired concrete engineering sample data is preprocessed to construct the input feature vector and training dataset;

[0010] A multi-objective prediction model is constructed, and the model is trained based on the training dataset using a joint loss function that includes a supervision error term and an engineering physical constraint penalty term; wherein, the engineering physical constraint penalty term is used to guide the model output to tend toward the set of engineering rule boundaries;

[0011] Input the sample to be predicted into the trained multi-objective prediction model to obtain the original prediction result;

[0012] The original prediction results are verified. If the original prediction results do not meet the engineering rule boundary set, a pattern vector is generated to represent the default type. Based on the pattern vector and the pre-configured key process parameters, the corresponding deterministic repair operator sequence is matched from the repair strategy mapping table.

[0013] Within a preset time delay threshold and under the condition that the number of repair iterations does not exceed the maximum number of iterations, the deterministic repair operator sequence is called from the repair operator library to repair the original prediction result, so as to obtain the final mix ratio that satisfies the boundary set of engineering rules;

[0014] If the final mix ratio cannot be obtained within the preset delay threshold, or if the repair iteration count reaches the maximum number of iterations but still fails to converge, the pre-configured safe alternative mix ratio will be output or the historical effective mix ratio will be rolled back, and an alarm message will be generated.

[0015] This invention addresses the technical problems in existing technologies, such as the impracticality of purely data-driven model outputs, the difficulty in guaranteeing 100% compliance through soft-constraint training, and the excessive and uncontrollable perturbations from post-processing corrections. It provides a two-stage collaborative technical solution: "soft-constraint training + rule-based repair." By introducing an engineering physical constraint penalty term during the training phase, the model output is guided to generally converge towards the engineering rule boundary set, reducing the probability of default from the outset. During the inference phase, a deterministic operator sequence is used to perform limited-amplitude repairs on a small number of defaulting samples, ensuring that the final output fully satisfies the engineering constraints. This synergistic effect avoids the problem of excessive correction amplitude caused by relying solely on post-processing repair, and overcomes the deficiency that soft-constraint training alone cannot guarantee complete compliance during the inference phase, achieving a synergistic optimization of prediction accuracy and engineering feasibility.

[0016] During the model training phase, a joint loss function including a supervision error term and an engineering physical constraint penalty term is employed. By penalizing predictions that violate engineering rule boundaries, the model gradually learns an output distribution that conforms to the physical laws of concrete engineering during training. This technique makes the model output statistically tend towards the feasible region, significantly reducing the probability of default in the inference phase. Compared with models trained without soft constraints, the model trained by this invention maintains prediction accuracy while significantly reducing the default rate. Compared with schemes that rely solely on post-processing repair, because the model output itself is closer to the feasible region, the proportion of samples requiring repair is reduced, and the repair magnitude is significantly decreased, thereby preserving the engineering performance characteristics reflected in the original prediction results to the greatest extent.

[0017] A rule-based repair process was designed during the inference phase, consisting of "default type identification - policy mapping - deterministic operator repair":

[0018] Ensuring 100% compliant output: By verifying the original prediction results and performing deterministic repairs on all prediction results that violate the boundary set of engineering rules, we ensure that the final mix design scheme meets 100% of the allowable range and physical constraints of the engineering, eliminating the risk of unfeasible mix designs flowing into the production process.

[0019] The repair process is controllable in real time: By setting a preset time delay threshold and a maximum number of iterations, the repair process is subject to dual constraints, ensuring that the repair operation is completed within the time window allowed by the mixing plant's online production, meeting the real-time requirements of millisecond-level response in industrial settings. Compared with schemes that call general constraint optimization solvers, the deterministic repair operators of this invention are all explicitly recalculated or element-wise pruned, resulting in low computational complexity and predictability, and eliminating the risk of iteration non-convergence or computation timeout.

[0020] The repair results are interpretable and reproducible: a deterministic repair operator sequence is used instead of black-box optimization. The repair path is entirely determined by the default type and key process parameters. The same input will produce the same output result at any time and under any environment, which ensures the reproducibility and interpretability of the repair process and makes it easier for engineers to understand and trust the system output.

[0021] Limited and controllable repair scope: Since the soft constraints during the training phase have already guided the model output to generally tend towards the feasible region, the proportion of samples that need to be repaired during the inference phase is low and the deviation is small. The repair operation only makes limited adjustments to a small number of non-compliant components, avoiding excessive disturbance to the original prediction performance by large corrections, and maintaining prediction accuracy while ensuring compliance.

[0022] A comprehensive anomaly degradation and alarm mechanism has been established: when a final mix proportion that meets the constraints cannot be obtained within a preset delay threshold, or when the number of repair iterations reaches the upper limit and convergence is still not achieved, the system automatically outputs a pre-configured safe alternative mix proportion or reverts to a historically valid mix proportion, and generates an alarm message. This technical measure ensures that the system can still output an implementable mix proportion scheme in extreme cases (such as model failure, data anomalies, repair logic defects, etc.), avoiding production interruption. At the same time, by alerting manual intervention, an automatic repair-based, manual intervention-assisted robust operation mode has been formed, significantly improving the system's engineering practicality and robustness.

[0023] Furthermore, the specific expression for the joint loss function is as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] in, Indicates joint loss; This represents the supervision error term, which measures the error between the model's predicted values ​​and the actual values. This represents the engineering physical constraint penalty term, used to penalize prediction results that violate the boundary set of engineering rules; Represents the weighting coefficient of the engineering physical constraint penalty term; This represents a range constraint penalty, used to penalize predictions that exceed the allowable range of the engineering project. In this situation, and Let represent the lower and upper limits of the engineering allowable limits for the j-th predicted target, respectively; This represents a feasibility constraint penalty, used to punish prediction results that violate engineering physical constraints. In this situation, F represents the original prediction result, and F represents the result compared to the original prediction result. The corresponding set of engineering physical parameters; , This represents the weight coefficient of the corresponding item; n represents the number of samples in the training batch; m represents the number of predicted targets in the original prediction result; This represents the model output value of the j-th predicted target for the i-th sample; K represents the number of engineering physical constraints; This represents the k-th engineering physical constraint function; This represents the predicted result vector for the i-th sample. ; This represents the set of engineering physical parameters corresponding to the i-th sample.

[0029] This invention defines the specific form of the joint loss function and clarifies the mathematical expressions and symbolic meanings of the supervision error term, range constraint penalty, and feasibility constraint penalty. This technical feature brings the following beneficial effects:

[0030] Implementing computable penalties for physical constraints: By transforming abstract engineering physics laws into quantifiable mathematical expressions (the penalty for range constraints is the accumulation of out-of-bounds distances, and the penalty for feasibility constraints is the accumulation of positive values ​​of the constraint function), the model can clearly perceive the "cost" of violating constraints through backpropagation during training, thereby effectively guiding the model output towards the engineering feasible region.

[0031] Differentiate penalty weights for different types of breaches: By setting independent weighting coefficients, engineers can adjust the relative importance of scope breaches and physical constraint breaches according to actual needs. For example, setting... Prioritizing the suppression of infeasible outputs that violate physical constraints enhances the flexibility of the training strategy.

[0032] Efficient computation for batch training: The double summation in the expression (for the number of samples and the number of predicted targets / constraints) enables the penalty term to be seamlessly integrated with the batch training framework, completing the constraint violation assessment for the entire batch in a single forward propagation, thus ensuring training efficiency.

[0033] Furthermore, the weighting coefficients Adaptive updates will be performed using the following strategy:

[0034] In the early stages of training, the weight coefficients... Equal to the maximum weight coefficient ;

[0035] Later in the training process, the weight coefficients equal ,in, This represents the minimum weight coefficient. Indicates the attenuation factor and ;

[0036] The pre-training phase is defined as the phase in which the number of training rounds has not exceeded a preset round threshold, or the phase in which the validation set default rate is not lower than a preset default rate threshold, wherein the validation set default rate is defined as the proportion of samples in the validation set that violate the boundary set of engineering rules.

[0037] The late training phase is defined as the stage where the number of training rounds exceeds a preset round threshold, or the stage where the validation set default rate is lower than a preset default rate threshold and the validation set error converges.

[0038] The weighting coefficient , It is a positive number and satisfies .

[0039] In the early stages of training, the weight coefficients... Equal to the maximum weight coefficient This allows the physical constraint penalty term to dominate the joint loss function, forcing the model to quickly learn and output compliant fit ratios, thus preventing the model from falling into the trap of pure data fitting and ignoring physical laws. In the later stages of training, the weight coefficients... according to By gradually attenuating the error term, the supervisory error term gradually plays a greater role, guiding the model to approximate the real data distribution as closely as possible while satisfying constraints, thus achieving a dynamic balance between prediction accuracy and compliance.

[0040] The division between the early and late stages of training is clearly defined by "round threshold" or "default rate threshold," which avoids subjective judgment, makes the update strategy deterministic and reproducible, and facilitates engineering implementation and cross-scenario migration.

[0041] limited Ensuring that the penalty weight for violating engineering physical constraints (such as volume non-conservation) is no less than that for simple range violations reflects the engineering principle that "physical laws take precedence over numerical ranges".

[0042] Furthermore, the set of engineering rule boundaries consists of both the permissible engineering range and engineering physical constraints, wherein the engineering physical constraints include at least volume balance constraints, and their constraint functions are defined as follows:

[0043] ;

[0044] in, Represents the volume balance constraint function; This represents the prediction result vector for the i-th sample; This represents the set of engineering physical parameters corresponding to the i-th sample; This represents a quantity that characterizes aggregate volume. This represents the set value for the air content of the concrete in the i-th sample; This represents the predicted unit water consumption for the i-th sample; This represents the predicted amount of cementitious material used for the i-th sample; Indicates the density of water; This indicates the density of the cementitious material.

[0045] This invention clarifies the objects of verification and repair by defining the engineering rule boundary set as composed of both "engineering allowable range" and "engineering physical constraints," avoiding confusion or omission of these two concepts. It provides an explicit expression for volume balance constraints, transforming the abstract concept of "volume conservation" into a function that can be directly substituted into numerical calculations, thus providing a concrete basis for penalty term calculation in the training phase and constraint verification in the inference phase. Treating volume balance constraints as mandatory hard constraints reflects the fundamental physical law that the sum of material volumes must equal the unit volume in concrete mix design, ensuring the engineering feasibility of the output mix proportions.

[0046] Furthermore, the pattern vector is generated in the following manner:

[0047] ;

[0048] Where s represents the pattern vector; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This represents the original prediction result vector for the current sample to be predicted. express The j-th component, m represents the original prediction result vector. The number of predicted targets; and Let represent the lower and upper limits of the engineering allowable limits for the j-th predicted target, respectively; Let F represent the k-th engineering physical constraint function, k=1,2,…,K, where K represents the number of engineering physical constraints; F represents the set of engineering physical parameters corresponding to the current sample to be predicted.

[0049] This invention provides a specific method for generating pattern vectors: various default situations (lower limit default, upper limit default, physical constraint default) are encoded into binary vectors through indicator functions, transforming qualitative judgments into quantitative representations, thus achieving digital encoding of default types and facilitating subsequent computer processing and strategy matching. Indicator bits are set for each prediction target and each physical constraint, enabling precise identification of the specific object of the default (e.g., "water consumption below the lower limit" or "volume balance violation"), supporting fine-grained default identification, and providing a decision-making basis for differentiated remediation strategies.

[0050] Furthermore, the pre-configured key process parameters are predetermined based on the engineering requirements of the concrete, and at least one of the water-cement ratio, unit water consumption, and cementitious material consumption is selected as a key parameter that is preferred to remain constant; wherein, the rules for determining the key process parameters include at least one of the following:

[0051] When the concrete strength grade or durability grade is not lower than the corresponding preset grade threshold, the key process parameter is the water-cement ratio, and the water-cement ratio is kept constant in the first place.

[0052] When the concrete slump or pumping height is not lower than the corresponding preset threshold, the key process parameter is the unit water consumption, and the unit water consumption should be kept constant in the first place.

[0053] When the cost of cementitious materials is higher than the preset cost threshold, or the required temperature rise for concrete insulation is not higher than the preset temperature rise threshold, the key process parameters are the amount of cementitious materials or the amount of cementitious materials and the water-cement ratio. Priority is given to keeping the amount of cementitious materials unchanged or keeping both the amount of cementitious materials and the water-cement ratio unchanged.

[0054] This invention provides objective constraints on the rules for determining key process parameters. These rules cover various engineering scenarios, including strength grade, durability grade, slump requirements, pumping height, cementitious material costs, and adiabatic temperature rise. This allows the system to flexibly select and prioritize key parameters based on actual working conditions, adapting to different construction conditions and economic requirements. Because the selection rules for key process parameters are pre-configured and predetermined, the same input feature will yield the same key process parameter selection results when performing repair at different times, thus ensuring the reproducibility of the entire repair path. Selecting different priority parameters for different scenarios (e.g., prioritizing the water-cement ratio for high-strength concrete and prioritizing water content for highly fluid concrete) reflects a deep understanding of the core control indicators of concrete engineering, making the repair results more consistent with engineering practice experience.

[0055] Furthermore, the repair operator library includes at least the following three types of repair operators:

[0056] Range clipping operator: Used to clip predicted target components that are outside the allowable range of the project to the allowable range of the project;

[0057] Volume boundary recalculation operator: When volume balance constraints are violated, it is used to explicitly recalculate at least one target variable through the volume balance equation based on pre-configured key process parameters, so that the repair result satisfies the volume balance constraints.

[0058] Iterative correction operator: Used to gradually approximate the repair result that satisfies all engineering rule boundaries by performing a combination of range trimming and volume boundary recalculation through a finite number of iterations when any of the following conditions exist, and stops when the iteration converges or the maximum number of iterations is reached:

[0059] This also violates both the permitted scope of the project and the physical constraints of the project;

[0060] Violation of multiple different types of engineering physical constraints;

[0061] Multiple predicted targets simultaneously exceeded the engineering allowable range;

[0062] A single default cannot be eliminated in a single repair using the range trimming operator and the volume boundary recalculation operator.

[0063] This invention defines the composition of the repair operator library and the applicable scenarios for iterative correction operators: the repair operation is encapsulated into three types of deterministic repair operators (range pruning, volume boundary recalculation, and iterative correction), making the repair logic clear, the functions decoupled, and easy to maintain and expand (if a new operator type is needed, it only needs to be added to the operator library without affecting the decision logic); all three types of operators are deterministic operations (pruning, explicit recalculation, and a combination of fixed iteration steps), without random factors, and the same input executing the same operator sequence always produces the same output, satisfying the reproducibility requirement; the iterative correction operator gradually approaches the feasible region through a finite number of iterations, and can handle complex default situations that cannot be solved by a single repair (such as multiple constraints being coupled together), while being constrained by the maximum number of iterations to ensure that the repair does not fall into an infinite loop.

[0064] Furthermore, the repair strategy mapping table maps different mode vectors to corresponding deterministic repair operator sequences, and includes at least the following mapping relationships:

[0065] When the pattern vector indicates that there is only a single predicted target within the permitted engineering range and the default exists, it is mapped to the range pruning operator.

[0066] When the pattern vector indicates that there are only engineering-permissible defaults and the number of defaulted predicted targets is multiple, it is mapped to the iterative correction operator;

[0067] When the pattern vector indicates that only the volume balance constraint is violated, it is mapped to the volume boundary recalculation operator;

[0068] When the pattern vector indicates that both the engineering allowable range violation and the volume balance constraint violation exist simultaneously, it is mapped to a combination operator sequence that first performs range trimming and then performs volume boundary recalculation.

[0069] When the pattern vector indicates that multiple different types of engineering physical constraints are violated simultaneously, it is mapped to the iterative correction operator.

[0070] This invention refines the specific mapping relationships of the repair strategy mapping table: based on pattern vectors, different default types are mapped to corresponding operator sequences, forming a closed-loop decision logic of "diagnosis → prescription", making the repair process completely regular and avoiding the complexity of general optimization solutions; "only range default" is further subdivided into "the number of default prediction targets is one" and "the number of default prediction targets is multiple", which are mapped to the range pruning operator (single operation) and the iterative correction operator (requiring multiple iterations) respectively, reflecting the refined idea of ​​taking differentiated treatment for defaults of different complexities.

[0071] Furthermore, the method also includes: sending the final mix proportion to the production control system and writing the structured information in the verification and repair process into the quality traceability database;

[0072] The structured information includes at least: a summary of the input feature vector, the original prediction result, the pattern vector, the selected deterministic repair operator sequence and its version number, key process parameters, the difference before and after repair, the execution delay, the verification result, and the downgrade flag;

[0073] The input feature vector summary is used to uniquely identify the index information of the input feature vector, including the hash value of the feature vector or a combination of key features; the version number is used to identify the version number to which the selected deterministic repair operator sequence belongs, and is used to trace the iteration history of the repair strategy.

[0074] Structured information from the verification and repair process is stored, making each mix design generation process auditable and reproducible, meeting the stringent traceability requirements of engineering quality management. The "input feature vector summary" is explicitly defined as a hash value or a combination of key features, used to uniquely identify input data while saving storage space. The "version number" is clearly defined to identify the iteration history of the operator sequence, making the evolution of the repair strategy traceable and facilitating problem identification and strategy optimization. Recorded information such as the difference before and after repair and execution latency can serve as a basis for model optimization; for example, it can analyze which types of defaults require larger repairs, thereby strengthening the soft constraint guidance during the training phase. When an anomaly degradation is triggered, the degradation flag and the final output (safe alternative or historical version) are also recorded, ensuring that even in the event of system anomalies, there is still evidence to be found, forming a complete chain of evidence for anomaly handling.

[0075] Based on the same concept, the present invention also provides a concrete mix design intelligent generation and online compliance verification system, including a memory, a processor, and a computer program or instructions stored in the memory, wherein the processor executes the computer program or instructions to implement the concrete mix design intelligent generation and online compliance verification method as described above.

[0076] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0077] The soft constraints (engineering physical constraint penalty term) in the training phase guide the model output to tend towards the feasible region, significantly reducing the probability of default. The rule-based repair (deterministic repair operator sequence) in the inference phase makes limited corrections to a few default samples, ensuring that the final match ratio meets 100% of the engineering constraints, while keeping the correction range to a minimum to avoid excessively disturbing the prediction accuracy.

[0078] A deterministic repair operator sequence is used to replace the general constraint optimization solver. All repair operations are explicitly recalculated or element-wise pruned, resulting in low computational complexity and predictability. Through the dual constraints of preset time delay threshold and maximum number of iterations, the repair process is ensured to be completed within the time window allowed by the online production of the mixing plant, meeting the real-time requirements of industrial sites.

[0079] A complete anomaly degradation mechanism is set up so that when a compliant mix ratio cannot be obtained within a specified time, a pre-configured safe alternative mix ratio is automatically output or the system reverts to a historically valid version, and an alarm message is generated to avoid production interruption due to system anomalies.

[0080] The repair path is entirely determined by the type of breach and key process parameters. The same input always yields the same output, ensuring reproducibility. The information such as input feature summary, original prediction results, pattern vector, repair operator sequence and version number, repair difference, and execution delay are stored in a structured manner to form a complete quality traceability evidence chain, meeting the requirements of engineering supervision and quality management. Attached Figure Description

[0081] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0082] Figure 1 This is a flowchart of the method for intelligent generation and online compliance verification of concrete mix proportions in this embodiment of the invention;

[0083] Figure 2 This is a schematic diagram illustrating the guiding effect of soft constraints on the joint loss function in an embodiment of the present invention;

[0084] Figure 3 This is a single-point example diagram of the rule correction and repair process in an embodiment of the present invention. Detailed Implementation

[0085] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0086] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0087] Example 1

[0088] Reference Figure 1 The method for intelligent generation and online compliance verification of concrete mix proportions provided in this embodiment includes the following steps:

[0089] Step S1: Preprocess the acquired concrete engineering sample data to construct the input feature vector and training dataset.

[0090] Concrete engineering sample data are collected from historical databases, laboratory test records, or the production control system of the batching plant. The collected concrete engineering sample data includes at least two types of information: feature data as input to the model and expected output data as the model's supervised learning objective.

[0091] Input feature data should include at least: raw material type (such as cement type, fly ash type, mineral powder grade, etc.), raw material density calibration values ​​(such as cement density, fly ash density, mineral powder density, etc.), concrete air content setting value, ambient temperature, target slump grade, target strength grade, and other engineering parameters affecting mix design. In addition, other relevant operating condition data, such as production date, batch number, and raw material origin, can also be included for subsequent traceability.

[0092] The expected output data consists of concrete mix proportion parameters used in actual production or testing, corresponding to the input feature data. These parameters include at least continuous engineering parameters with engineering-physical coupling relationships, such as unit water consumption, water-cement ratio, and cementitious material dosage. These expected output data are actual mix proportions directly obtained from historical records, serving as labels for supervised learning to train the model and predict mix proportions in actual production.

[0093] The collected raw data is preprocessed, including the following sub-steps:

[0094] Data cleaning: Remove samples that are obviously abnormal or incomplete, such as samples that are missing key input features or expected outputs, or samples whose values ​​exceed the physical possibility range (such as negative unit water consumption or water-glue ratio greater than 1).

[0095] Missing value imputation: For a small number of missing input features, imputation is performed using mean imputation, interpolation, or prediction methods based on similar samples. For the desired output, if a sample is missing, it is directly removed to ensure the accuracy of the training labels.

[0096] Normalization: To eliminate the influence of different units on model training, the input features are normalized, scaling each feature value to a uniform range (such as [0,1] or standardized to a mean of 0 and a variance of 1).

[0097] All preprocessed input features are combined into a fixed-dimensional vector, resulting in the input feature vector. Each sample's input feature vector and its corresponding expected output vector Y (composed of unit water consumption, water-to-binder ratio, cementitious material dosage, etc.) form a training sample. The set of all training samples constitutes the training dataset.

[0098] Step S2: Construct a multi-objective prediction model and train it using a joint loss function that includes a supervision error term and an engineering physical constraint penalty term based on the training dataset.

[0099] The core of step S2 lies in integrating the physical laws of concrete engineering into the model training process through a joint loss function. This allows the model to learn and output mix proportion parameters that conform to the allowable range and physical constraints of engineering while fitting historical data. The specific implementation method is as follows:

[0100] First, a multi-objective prediction model is constructed to simultaneously predict multiple combination ratio parameters. The input to the multi-objective prediction model is the input feature vector constructed in step S1, and the output is the predicted combination ratio vector. Among them, the predicted mix ratio vector It includes at least three continuous engineering parameters: unit water consumption, water-cement ratio, and amount of cementitious material. The model structure can adopt a fully connected neural network, gradient boosting tree, or multi-task learning architecture. The specific choice is determined according to the data scale and complexity, but the model must be differentiable to facilitate backpropagation training.

[0101] Secondly, based on the training dataset constructed in step S1, the model is trained using a joint loss function. The joint loss function is a weighted sum of a supervision error term and an engineering physical constraint penalty term:

[0102] (1)

[0103] in: Indicates joint loss; This represents the supervision error term, used to measure the error between the model's predicted values ​​and the true values ​​(i.e., the expected output vector of the samples). It can be expressed as mean squared error (MSE) or mean absolute error (MAE), for example... , where n represents the number of samples in the training batch. This represents the predicted result vector (i.e., the predicted combination ratio vector or the original predicted result) for the i-th sample. This represents the expected output vector (i.e., the true combination ratio vector) corresponding to the i-th sample. This represents the L2 norm. This represents the engineering physical constraint penalty term, which is used to penalize prediction results that violate the boundary set of engineering rules, guiding the model output towards the feasible region; This represents the weighting coefficient of the engineering physical constraint penalty term, used to balance the importance of the monitoring error term and the engineering physical constraint penalty term.

[0104] The engineering physical constraint penalty term is further composed of a weighted average of the scope constraint penalty and the feasibility constraint penalty:

[0105] (2)

[0106] in: , This represents the preset positive weight coefficient, and satisfies... Prioritize suppressing infeasible outputs that violate engineering physics constraints (such as volume imbalance); Indicates range constraint penalties; This indicates a feasibility constraint penalty.

[0107] Range constraint penalty is used to penalize prediction results that exceed the allowable range of the project. The specific definition of this situation is:

[0108] (3)

[0109] Where: m represents the number of predicted targets in the original prediction result (i.e., the predicted mix ratio vector output by the model); This represents the model output value of the j-th prediction target (such as unit water consumption, water-cement ratio, etc.) for the i-th sample; and Let represent the lower and upper limits of the engineering allowable limits for the j-th predicted target, respectively.

[0110] The range constraint penalty calculates and sums the distances of each out-of-bounds component to the boundary, gradually adjusting the model's output to within the engineering-allowed range during training.

[0111] Feasibility constraint penalties are used to punish prediction results that violate engineering physical constraints. In this case, engineering physical constraints include at least volume equilibrium constraints, which are specifically defined as:

[0112] (4)

[0113] Where: K represents the number of engineering physical constraints; This represents the k-th engineering physical constraint function; This represents the predicted result vector for the i-th sample. ; This represents the set of engineering physical parameters corresponding to the i-th sample (such as the density of water, the density of cementitious materials, the air content of concrete, etc.).

[0114] Taking volume equilibrium constraints as an example, its constraint function is defined as:

[0115] (5)

[0116] in, Represents the volume balance constraint function; F represents the original prediction result, and F represents the result compared to the original prediction result. The corresponding set of engineering physical parameters; This represents a quantity that characterizes aggregate volume. This represents the set value for the air content of the concrete in the i-th sample; This represents the predicted unit water consumption for the i-th sample; This represents the predicted amount of cementitious material used for the i-th sample; Indicates the density of water; This represents the density of the cementitious material. When the prediction result leads to aggregate volume... When it is negative, This will produce a positive punishment.

[0117] Weighting coefficient , , Configure parameters for the project, including weighting coefficients. An adaptive update strategy is adopted to balance compliance learning in the early stages of training with accuracy in the later stages.

[0118] In this embodiment, the weighting coefficient The value range is [0.1, 10], and the maximum weight coefficient is... The minimum weighting coefficient is 10. The value is 0.1. Early training phase: [The value is missing from the original text.] This strengthens the role of the engineering physical constraint penalty term, enabling the model to quickly learn and output fit ratios that conform to the rule boundaries. The early training phase is defined as the stage where the number of training epochs does not exceed a preset epoch threshold (e.g., 50), or the validation set is not lower than a preset default rate threshold, where the validation set default rate is the proportion of samples in the validation set that violate the engineering rule boundaries. In this embodiment, the preset default rate threshold is 5%.

[0119] Later stages of training: according to Gradually decaying, among which Indicates the attenuation factor and (For example (0.95). The later stage of training is defined as the stage where the number of training epochs exceeds the preset epoch threshold, or the stage where the validation set default rate is lower than the preset default rate threshold and the validation set error converges. At this time, the physical constraint penalty weight is gradually reduced so that the model can more accurately fit the real data distribution while satisfying the physical constraints.

[0120] Weighting coefficient , It is a positive number and satisfies This prioritizes suppressing infeasible outputs that violate engineering physics (such as volume imbalance). Specific values ​​can be set based on engineering experience, for example... It is 1.0. It is 0.5.

[0121] In each training batch, perform the following steps:

[0122] Forward propagation: Inputting the feature vectors from the batch into the multi-objective prediction model to obtain the original prediction results. ;

[0123] Calculate the supervision error term Calculate the mean square error or mean absolute error based on the predicted and actual values;

[0124] Calculation range constraint penalty : Traverse all samples and the predicted target, and accumulate the out-of-bounds distances;

[0125] Calculate feasibility constraint penalties : Traverse all samples and engineering physical constraints, and accumulate the positive values ​​of the constraint function.

[0126] Calculate the joint loss and its gradient with respect to the model parameters, then update the model parameters.

[0127] Iteration: Repeat the above steps until the loss on the training dataset converges or the preset maximum number of training rounds is reached.

[0128] During training, the default rate and prediction error are periodically evaluated on the validation set, and the weight coefficients are dynamically adjusted according to an adaptive update strategy. After training, a multi-objective prediction model that meets statistical accuracy and generally tends towards the engineering feasible region is obtained, laying the foundation for rule correction and repair in the subsequent online inference stage.

[0129] Figure 2 The process by which soft constraints guide the joint loss function is illustrated. The unconstrained optimum corresponds to the state determined solely by the supervision error term L. sup The point where the minimum value is obtained may be outside the feasible region of the project. The feasible region of the project is defined by both the allowable scope of the project and the physical constraints of the project. Figure 2 The area is represented by shaded regions. By introducing an engineering physical constraint penalty term, a "penalty wall" is formed at the boundary of the engineering feasible region, which pulls the optimal point (L minimum) of the joint loss function into the feasible region, thereby guiding the model output to tend towards the set of engineering rule boundaries, thus reducing the overall deviation between the prediction results and the set of engineering rule boundaries.

[0130] Step S3: Input the sample to be predicted into the trained multi-objective prediction model to obtain the original prediction result.

[0131] Step S4: Verify the original prediction results to determine whether they meet the engineering rule boundary set; if not, generate a pattern vector to represent the default type and proceed to step S5; if they meet, use the original prediction results directly as the final mix ratio.

[0132] The core of step S4 is the original prediction result obtained in step S3. The process involves conducting an engineering compliance check, determining whether all constraints are met based on a predefined set of engineering rule boundaries, and deciding whether to output directly or proceed to the repair process based on the result. The specific implementation method is as follows:

[0133] Read the set of engineering rule boundaries under the current working condition. This set consists of two parts:

[0134] Project scope: Where m is the quantity of the predicted target (such as unit water consumption, water-cement ratio, amount of cementitious material, etc.). and Let represent the lower and upper limits of the engineering allowable limits for the j-th predicted target, respectively.

[0135] Engineering physical constraints: Where K is the number of engineering physical constraints. Let F represent the k-th engineering physical constraint function, and let F represent the result compared to the original prediction. The corresponding set of engineering physical parameters (for volume equilibrium constraints, the set of engineering physical parameters includes the density of water, the density of cementitious materials, the air content of concrete, etc.).

[0136] The original prediction results for the current sample to be predicted The following two types of checks will be performed respectively:

[0137] Engineering allowable range verification: For each predicted target j, check... Does it meet the requirements? Record two types of breach of contract: if If the lower limit is breached, it is recorded as 1; otherwise, it is recorded as 0. If the upper limit is breached, it is recorded as 1; otherwise, it is recorded as 0.

[0138] Engineering physical constraint verification: For each engineering physical constraint k, calculate the constraint function value. Check if it meets the requirements. If yes, then it is determined that the physical constraints of the project have not been violated, and it is recorded as 0; if no (i.e. If the condition is true, it is determined that the physical constraints of the project are violated, and it is recorded as 1.

[0139] All verification results are combined into a pattern vector to fully describe the default type of the current prediction result. The pattern vector is constructed as follows:

[0140] (6)

[0141] Where s represents the pattern vector; This indicates an indicator function that takes the value 1 when the condition within the parentheses is true, and 0 otherwise.

[0142] If all elements of the pattern vector are 0, it indicates that the original prediction result fully satisfies the engineering rule boundary set and there is no breach of contract. In this case, the original prediction result is directly used as the final mix proportion, and the process proceeds to step S8 (output and traceability) for distribution and recording.

[0143] If the pattern vector contains non-zero elements, it indicates that the original prediction result violates at least one engineering rule boundary (engineering allowable range or engineering physical constraint), and rule correction and repair are required. At this time, the pattern vector, the original prediction result of the current sample, the set of engineering physical parameters, and other data are passed to step S5 (rule correction and repair) to enter the repair process.

[0144] Regardless of whether the repair process is initiated, the verification results (including pattern vectors and various verification flags) must be temporarily stored in memory or cache for later writing to the quality traceability database. If the original prediction results are output directly, the difference before and after repair is 0, and the execution delay is recorded as the verification time. If the repair process is initiated, complete repair process information will be recorded subsequently.

[0145] Step S4 provides a precise decision-making basis for subsequent rule-based repair by explicitly encoding the default type, ensuring that the repair operation can select the repair operator sequence in a targeted manner, which reflects the core idea of ​​"rule-based repair" in this invention.

[0146] Step S5: Based on the mode vector and pre-configured key process parameters, obtain the corresponding deterministic repair operator sequence from the repair strategy mapping table.

[0147] The core of step S5 is to determine a deterministic repair operator sequence applicable to the current default situation based on the pattern vector generated in step S4 and the pre-configured key process parameters, through a repair strategy mapping table, thus providing a clear, rule-based path for subsequent repair execution. The specific implementation is as follows:

[0148] The system reads pre-configured key process parameters from the system configuration. These parameters are predetermined based on the engineering requirements of the concrete, and at least one parameter is selected from the water-cement ratio, unit water consumption, and cementitious material dosage as the priority to remain constant. The rules for determining key process parameters include at least one of the following:

[0149] If the current concrete strength grade or durability grade is not lower than the corresponding preset grade threshold (e.g., strength grade ≥ C40 or durability grade ≥ a certain standard), then the key process parameter is the water-cement ratio, that is, the water-cement ratio is kept constant.

[0150] If the concrete slump requirement or pumping height is not lower than the corresponding preset threshold (e.g., slump ≥ 180 mm or pumping height ≥ 50 m), then the key process parameter is taken as the unit water consumption, that is, the unit water consumption is kept constant in priority.

[0151] If the cost of cementitious materials is higher than the preset cost threshold, or the required temperature rise of concrete insulation is not higher than the preset temperature rise threshold (e.g., temperature rise ≤ 35℃), then the key process parameter is the amount of cementitious materials, or the key process parameters are the amount of cementitious materials and the water-cement ratio. That is, the amount of cementitious materials is kept constant or the amount of cementitious materials and the water-cement ratio are kept constant at the same time.

[0152] The repair strategy mapping table is a predefined rule table that maps pattern vectors (combined with the implicit effects of key process parameters) to corresponding deterministic repair operator sequences. The repair strategy mapping table contains at least the mapping relationships shown in Table 1.

[0153] Table 1. Examples of mapping relationships in the repair strategy mapping table.

[0154]

[0155] As shown in Table 1, each record in the repair strategy mapping table contains the following fields:

[0156] Description of breach type: A natural language summary of the breach situation;

[0157] Pattern vector matching conditions: logical combinations of values ​​taken at specific positions in the pattern vector (e.g., "only range default and the number of default targets is 1" corresponds to only one range default flag being 1 and all physical constraint flags being 0 in the pattern vector).

[0158] Key process parameter application conditions: Some mappings may need to be combined with the values ​​of key process parameters (for example, the volume boundary recalculation operator needs to select a specific calculation formula according to the key process parameters, such as formula (5));

[0159] The corresponding deterministic repair operator sequence is an ordered list of operators, such as [range pruning operator], [volume boundary recalculation operator], [range pruning operator, volume boundary recalculation operator], [iterative correction operator], etc.

[0160] Based on the value of the pattern vector and the current key process parameters, the first record that satisfies all matching conditions is found in the repair strategy mapping table, thus obtaining the corresponding deterministic repair operator sequence. The specific matching logic is as follows:

[0161] First, analyze the position and number of non-zero elements in the pattern vector to determine the type of default:

[0162] Range default flag: Check the first m bits (lower limit default) and the next m bits (upper limit default). If any bit is 1, a range default exists; count the number r of predicted default targets.

[0163] Physical constraint default flag: Check the last K bits. If any bit is 1, there is a physical constraint default; count the number p of defaulted physical constraints.

[0164] The repair strategy mapping table is usually organized from simplest to most complex in terms of default complexity, and the matching is performed in the following order (or by using priority rules):

[0165] Only range defaults and the number of default targets is single: if p=0 and r=1, then map to the range clipping operator.

[0166] Only for range-based defaults and multiple default targets: If p=0 and r≥2, then map to the iterative correction operator.

[0167] Only volume balance constraint violation: If p=1 and the violation is exactly a volume balance constraint, and r=0, then the volume boundary recalculation operator is mapped. In this case, the specific recalculation method needs to be selected according to the key process parameters (such as recalculating the unit water consumption while keeping the water-cement ratio constant, or recalculating the amount of cementitious material while keeping the unit water consumption constant, etc.).

[0168] Both range default and volume balance constraint default exist: If both range default (r≥1) and volume balance constraint default exist simultaneously ( If the flag is 1), it is mapped to the sequence of combined operators [range clipping operator, volume boundary recalculation operator], that is, range clipping is performed first, and then volume boundary recalculation is performed.

[0169] Multiple different types of engineering physical constraints are violated simultaneously: if p≥2 (i.e. at least two different physical constraints are violated), then it is mapped to the iterative correction operator.

[0170] It should be noted that the above matching rules are exemplary and can be supplemented or adjusted based on engineering experience in actual applications. The repair strategy mapping table supports flexible configuration to adapt to the repair strategy requirements in different scenarios.

[0171] Upon successful matching, a deterministic repair operator sequence is obtained. This sequence may contain one or more operators, and the execution order of the operators is strictly defined by the sequence. Simultaneously, the version number of the currently selected operator sequence (e.g., "v1.0") is recorded. This version number is associated with the versions of the repair strategy mapping table and the repair operator library, and is used to track the iteration history of the repair strategy during subsequent quality tracing.

[0172] If a matching failure occurs due to an abnormal pattern vector or incomplete coverage of the repair mapping table, the system should trigger the default processing mechanism: a preset general iterative correction operator can be selected as a backup solution, and an alarm message should be generated to indicate that the repair strategy mapping table needs to be updated.

[0173] Step S6: Within the preset time delay threshold and under the condition that the number of repair iterations does not exceed the maximum number of iterations, call the deterministic repair operator sequence from the repair operator library to repair the original prediction result and obtain the final mix ratio that satisfies the boundary set of engineering rules.

[0174] The core of step S6 is to perform the repair operation under strict real-time constraints based on the deterministic repair operator sequence obtained in step S5, ensuring that the final output mix ratio meets the engineering rule boundary set. The specific implementation method is as follows:

[0175] Before entering the repair process, the system records the current timestamp and sets the following parameters:

[0176] Preset delay threshold Δt: The upper limit of the total online inference-verification-repair delay (e.g., 200ms) given by the engineering configuration, used to ensure that the repair process does not exceed the time window allowed by the production cycle of the mixing plant.

[0177] Maximum number of iterations T max The maximum number of iterations allowed for the iterative correction operator (e.g., T) max =10), to prevent infinite loops.

[0178] Convergence threshold (e.g., 0.1 kg / m³) 3 ): This is a precision parameter used to determine whether the iteration has converged. The iteration will terminate prematurely if the change in the predicted vector between two iterations is less than this convergence threshold.

[0179] Meanwhile, the following data are obtained from steps S4 and S5: original prediction results, mode vector, deterministic repair operator sequence, key process parameters, and set of engineering physical parameters.

[0180] The repair operator library predefines three types of deterministic repair operators, whose specific implementations are as follows:

[0181] Range clipping operator: Used to handle simple numerical out-of-bounds problems. For the original prediction results... Each component in To constrain it within the allowable range of the project, that is The clipping operation is performed element-wise and does not change the values ​​of other components. The range clipping operator is a one-time operation with no iteration process.

[0182] Volume boundary recalculation operator: Used to correct prediction results that violate volume equilibrium constraints. Depending on the key process parameters, the recalculation methods can be categorized into the following typical cases:

[0183] Scenario 1: The key process parameter is the water-to-binder ratio (preferably keep the water-to-binder ratio unchanged).

[0184] Assuming the water-cement ratio in the original prediction results Deemed reliable, unit water consumption needs to be recalculated. and the amount of cementitious materials This ensures that volume balance is achieved.

[0185] (7)

[0186] Solving for:

[0187] (8)

[0188] If there are other components in the original prediction result (such as sand ratio), they remain unchanged.

[0189] Scenario 2: The key process parameter is the unit water consumption (prioritizing keeping the unit water consumption constant).

[0190] Keep Recalculate the amount of cementitious material to achieve volume balance:

[0191] (9)

[0192] At this point, the water-to-glue ratio becomes... .

[0193] Scenario 3: The key process parameters are the amount of cementitious material or the amount of cementitious material and the water-cement ratio.

[0194] Prioritize keeping the amount of cementitious material constant and recalculate the unit water consumption:

[0195] (10)

[0196] If the water-cement ratio and the amount of cementitious material are kept constant, the aggregate or other parameters need to be readjusted according to the volume balance equation. However, in such cases, the volume balance is usually automatically satisfied (because the water-cement ratio and cementitious material are fixed, so the water consumption is also fixed). If it is still not satisfied, it may trigger iterative correction.

[0197] The volume boundary recalculation operator is also a deterministic one-time operation, which directly yields a new mix proportion after execution.

[0198] When the default situation is complex (such as multiple constraints defaulting simultaneously, multiple prediction targets going out of range, or a single default not being eliminated by a single operator), an iterative correction operator is required. Its execution flow is as follows:

[0199] Initialization: Set the current solution Original prediction results The iteration counter t=0.

[0200] Enter the iteration loop: for t=1,2,…,T max :

[0201] Range clipping: for Y (t−1) Apply the range trimming operator to obtain temporary results, ensuring that all components are within the allowable range of the project.

[0202] Volume boundary recalculation: Apply the volume boundary recalculation operator to the provisional results based on the current critical process parameters to obtain Y. (t) .

[0203] Convergence judgment and output: Calculate the change between two iterations (using Euclidean distance or maximum component deviation). If this change is less than the preset convergence threshold, convergence is considered achieved, and the loop is terminated early, setting the final mix proportion Y* = Y. (t) If the change is greater than or equal to the preset convergence threshold, then it is determined whether the current iteration number is less than or equal to the maximum iteration number. If so, the loop is terminated and the result of the last iteration is output; otherwise, the iteration continues.

[0204] Figure 3 A single-point example diagram of the rule-based correction and repair process is shown. In the original prediction result, the unit water consumption W is 205.3 kg / m³. 3 Exceeding the maximum allowable limit of 200 kg / m 3 The water-cement ratio (W / B) is 0.38, which is within the allowable range. The mode vector indicates "range default," and the key process parameter is set to prioritize maintaining the water-cement ratio unchanged. Based on the repair strategy mapping table, the deterministic repair operator sequence is matched to the range trimming operator. After executing the range trimming operator, the unit water consumption W is trimmed to 200 kg / m³. 3 With the water-cement ratio W / B remaining constant, the final mix proportion is {W:200.0, W / B:0.38}. This result satisfies the engineering rule boundary set, and compliance is marked with "√".

[0205] During the execution of the repair operator, the system monitors the following two conditions in real time:

[0206] Latency monitoring: After each operator execution, check if the current time exceeds Δt. If it does, immediately terminate the repair and proceed to the abnormal degradation process in step S7.

[0207] Iteration count monitoring: For iterative correction operators, after each iteration, check whether the current iteration count t has reached T. max If convergence is reached but not yet achieved, the iteration is terminated, and the current Y is set to... (t) As a result of the repair (which may still not meet the constraints), the process is then transferred to the abnormal degradation procedure in step S7.

[0208] After the repair is completed, the system should immediately perform a rapid check on the repair output to confirm whether it meets the engineering rule boundary set. If it does, the repair output is used as the final mix ratio, and the process proceeds to step S8 (output and traceability). If it does not meet the requirements (possibly due to non-convergence of iteration or time delay truncation), the repair is considered a failure, and the process proceeds to the abnormal degradation process in step S7.

[0209] Step S7: If the final mix ratio cannot be obtained within the preset delay threshold, or if the repair iteration count reaches the maximum iteration count and still fails to converge, output the pre-configured safe alternative mix ratio or roll back to the historical effective mix ratio, and generate an alarm message.

[0210] An abnormal degradation is triggered when any of the following conditions occur:

[0211] The total time spent during the repair process exceeded Δt;

[0212] The number of iterations reaches T max And the result still does not meet the constraints;

[0213] The repaired output still does not meet the constraints after verification.

[0214] The exception degradation strategy performs one of the following operations in order of priority:

[0215] Output pre-configured safe alternative mix proportions: The system pre-stores a set of engineering-verified conservative mix proportion templates. These templates are based on historical experience or industry standards (e.g., typical mix proportions for low strength grades), have a high safety margin, and can ensure the feasibility of the project. Based on the current working conditions (e.g., strength grade, slump requirements), the system selects the best-matching template from the template library as the final mix proportion.

[0216] Rollback to a historical effective mix proportion: Retrieve the last successfully verified and issued final mix proportion version from the quality traceability database for the current production line or similar operating conditions, and output this as the final mix proportion. This historical version has been verified in actual production and has high reliability.

[0217] If neither of the above is available, a very conservative default mix ratio (such as the baseline mix ratio for C30 level) will be output, and a high-level alarm will be triggered.

[0218] Regardless of the degradation scheme adopted, the system generates an alarm message, which includes: the cause of the anomaly (timeout, non-convergence, verification failure), the original prediction results, current working parameters (strength grade, slump, etc.), the selected degradation scheme, and the corresponding mix proportion. The alarm message can be sent to engineering personnel via the system interface, SMS, or email for manual analysis.

[0219] Step S8: Send the final mix proportion to the production control system and write the structured information from the verification and repair process into the quality traceability database.

[0220] Regardless of whether the repair is successful or a degradation is triggered, the system will temporarily store the complete information of this repair process in a memory buffer, ready to write it to the quality traceability database. The recorded structured information includes at least: a summary of the input feature vector (such as a hash value), the original prediction result, the mode vector, the selected deterministic repair operator sequence and its version number, key process parameters, the difference before and after repair, execution delay (total time), verification result (pass / fail), and degradation flag (if degradation is triggered, the degradation type and the selected scheme will be recorded).

[0221] Through the above steps, this invention achieves rule-based repair under strict real-time constraints, ensuring both the engineering compliance of the final mix proportion and production continuity through the anomaly degradation mechanism, fully demonstrating the engineering practicality of the collaborative design of "soft constraint training + rule-based correction reasoning".

[0222] To verify the effectiveness of the engineering system of this invention, 300 sets of historical production and trial mixing data from a certain mixing plant were selected as a sample for comparison. The prediction targets are the unit water consumption W and the water-cement ratio W / B, with an allowable engineering range of W∈[160,200] kg / m³. 3 W / B ∈ [0.35, 0.55]. Example of engineering physical parameters. =1000kg / m 3 , =2800kg / m 3 The air content of the concrete is 2.5%.

[0223] The comparison methods include:

[0224] Method 1 (baseline): Supervised learning only, without physical constraints;

[0225] Method 2 (Soft Constraints): Incorporate engineering physics constraint penalties during the training phase;

[0226] Method 3 (rule correction only): Execute the repair operator library during the inference phase;

[0227] Method 4 (this invention): Soft constraint training + rule-based corrective reasoning.

[0228] Experimental results:

[0229] Method 1: Predicted MAE = 8.2 kg / m 3 The default rate was 18.7%.

[0230] Method 2: Predicted MAE = 9.1 kg / m 3 The default rate was 8.3%.

[0231] Method 3: Predicted MAE = 8.2 kg / m 3 The default rate was 0%, and the average correction margin was 15.3 kg / m³. 3 .

[0232] Method 4 (Invention): Predicted MAE = 9.0 kg / m 3 The default rate was 0%, and the average correction margin was 6.8 kg / m³. 3 This reduces the performance by 55.6% compared to Method 3, fully demonstrating the synergistic advantages of "soft constraint training + rule-based corrective reasoning" in terms of compliance and performance.

[0233] Example 2

[0234] This invention also provides a system for intelligent generation and online compliance verification of concrete mix proportions. The system includes a memory, a processor, and a computer program or instructions stored in the memory. The processor executes the computer program or instructions to implement the method for intelligent generation and online compliance verification of concrete mix proportions in this invention.

[0235] Although not shown, the system includes a processor that can perform various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) or loaded from a storage portion into random access memory (RAM). The processor can be a multi-core processor or may contain multiple processors. In some embodiments, the processor may include a general-purpose main processor and one or more specialized coprocessors, such as a central processing unit, graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for device operation are also stored in the RAM. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0236] The processor and memory described above are used together to execute programs / instructions stored in the memory. When the program / instructions are executed by the computer, they can implement the methods, steps, or functions described in the above embodiments.

[0237] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent generation and online compliance verification of concrete mix proportions, characterized in that, The method includes: The acquired concrete engineering sample data is preprocessed to construct the input feature vector and training dataset; A multi-objective prediction model is constructed and trained on the training dataset using a joint loss function that includes a supervision error term and an engineering physical constraint penalty term. The engineering physical constraint penalty term guides the model output towards the boundary set of engineering rules. This penalty term is composed of a weighted sum of a range constraint penalty and a feasibility constraint penalty, with the range constraint penalty penalizing prediction results that exceed the permissible engineering range. The specific definition of this situation is: ； in: Indicates range constraint penalty, and represents the lower and upper limits of the engineering allowable limit for the j-th predicted target, respectively; n represents the number of samples in the training batch; m represents the number of predicted targets in the original prediction result; This represents the model output value of the j-th predicted target for the i-th sample; The feasibility constraint penalty is used to penalize prediction results that violate engineering physical constraints. The specific definition of this situation is: ； in, K represents the penalty for feasibility constraints; K represents the number of engineering physical constraints. This represents the k-th engineering physical constraint function; This represents the predicted result vector for the i-th sample. ; This represents the set of engineering physical parameters corresponding to the i-th sample; F represents the original prediction result, and F represents the result compared to the original prediction result. The corresponding set of engineering physical parameters; Input the sample to be predicted into the trained multi-objective prediction model to obtain the original prediction result; The original prediction results are verified. If the original prediction results do not meet the engineering rule boundary set, a pattern vector is generated to represent the default type. Based on the pattern vector and the pre-configured key process parameters, the corresponding deterministic repair operator sequence is matched from the repair strategy mapping table. Within a preset time delay threshold and under the condition that the number of repair iterations does not exceed the maximum number of iterations, the deterministic repair operator sequence is called from the repair operator library to repair the original prediction result, so as to obtain the final mix ratio that satisfies the boundary set of engineering rules; If the final mix ratio cannot be obtained within the preset delay threshold, or if the repair iteration count reaches the maximum number of iterations but still fails to converge, the pre-configured safe alternative mix ratio will be output or the historical effective mix ratio will be rolled back, and an alarm message will be generated.

2. The method for intelligent generation and online compliance verification of concrete mix proportions according to claim 1, characterized in that, The specific expression for the joint loss function is as follows: ； ； in, Indicates joint loss; This represents the supervision error term, which measures the error between the model's predicted values ​​and the actual values. This represents the engineering physical constraint penalty term, used to penalize prediction results that violate the boundary set of engineering rules; Represents the weighting coefficient of the engineering physical constraint penalty term; , This represents the weight coefficient of the corresponding item.

3. The method for intelligent generation and online compliance verification of concrete mix proportions according to claim 2, characterized in that, The weighting coefficient Adaptive updates will be performed using the following strategy: In the early stages of training, the weight coefficients... Equal to the maximum weight coefficient ; Later in the training process, the weight coefficients equal ,in, This represents the minimum weight coefficient. Indicates the attenuation factor and ; The pre-training phase is defined as the phase in which the number of training rounds has not exceeded a preset round threshold, or the phase in which the validation set default rate is not lower than a preset default rate threshold, wherein the validation set default rate is defined as the proportion of samples in the validation set that violate the boundary set of engineering rules. The late training phase is defined as the stage where the number of training rounds exceeds a preset round threshold, or the stage where the validation set default rate is lower than a preset default rate threshold and the validation set error converges. The weighting coefficient , It is a positive number and satisfies .

4. The method for intelligent generation and online compliance verification of concrete mix proportions according to claim 1, characterized in that, The set of engineering rule boundaries consists of both the permissible engineering range and engineering physical constraints. The engineering physical constraints include at least volume balance constraints, and their constraint functions are defined as follows: ； in, Represents the volume balance constraint function; This represents the prediction result vector for the i-th sample; This represents the set of engineering physical parameters corresponding to the i-th sample; This represents a quantity that characterizes aggregate volume. This represents the set value for the air content of the concrete in the i-th sample; This represents the predicted unit water consumption for the i-th sample; This represents the predicted amount of cementitious material used for the i-th sample; Indicates the density of water; This indicates the density of the cementitious material.

5. The method for intelligent generation and online compliance verification of concrete mix proportions according to claim 1, characterized in that, The pattern vector is generated in the following manner: ； Where s represents the pattern vector; This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This represents the original prediction result vector for the current sample to be predicted. express The j-th component, m represents the original prediction result vector. The number of predicted targets; and Let represent the lower and upper limits of the engineering allowable limits for the j-th predicted target, respectively; Let F represent the k-th engineering physical constraint function, k=1,2,…,K, where K represents the number of engineering physical constraints; F represents the set of engineering physical parameters corresponding to the current sample to be predicted.

6. The method for intelligent generation and online compliance verification of concrete mix proportions according to claim 1, characterized in that, The pre-configured key process parameters are predetermined based on the engineering requirements of the concrete, and at least one of the water-cement ratio, unit water consumption, and cementitious material consumption is selected as the key parameter that is preferred to remain unchanged; wherein, the rules for determining the key process parameters include at least one of the following: When the concrete strength grade or durability grade is not lower than the corresponding preset grade threshold, the key process parameter is the water-cement ratio, and the water-cement ratio is kept constant in the first place. When the concrete slump or pumping height is not lower than the corresponding preset threshold, the key process parameter is the unit water consumption, and the unit water consumption should be kept constant in the first place. When the cost of cementitious materials is higher than the preset cost threshold, or the required temperature rise for concrete insulation is not higher than the preset temperature rise threshold, the key process parameters are the amount of cementitious materials or the amount of cementitious materials and the water-cement ratio. Priority is given to keeping the amount of cementitious materials unchanged or keeping both the amount of cementitious materials and the water-cement ratio unchanged.

7. The method for intelligent generation and online compliance verification of concrete mix proportions according to claim 1, characterized in that, The repair operator library includes at least the following three types of repair operators: Range clipping operator: Used to clip predicted target components that are outside the allowable range of the project to the allowable range of the project; Volume boundary recalculation operator: When volume balance constraints are violated, it is used to explicitly recalculate at least one target variable through the volume balance equation based on pre-configured key process parameters, so that the repair result satisfies the volume balance constraints. Iterative correction operator: Used to gradually approximate the repair result that satisfies all engineering rule boundaries by performing a combination of range trimming and volume boundary recalculation through a finite number of iterations when any of the following conditions exist, and stops when the iteration converges or the maximum number of iterations is reached: This also violates both the permitted scope of the project and the physical constraints of the project; Violation of multiple different types of engineering physical constraints; Multiple predicted targets simultaneously exceeded the engineering allowable range; A single default cannot be eliminated in a single repair using the range trimming operator and the volume boundary recalculation operator.

8. The method for intelligent generation and online compliance verification of concrete mix proportions according to claim 7, characterized in that, The repair strategy mapping table maps different mode vectors to corresponding deterministic repair operator sequences, and includes at least the following mapping relationships: When the pattern vector indicates that there is only a single predicted target within the permitted engineering range and the default exists, it is mapped to the range pruning operator. When the pattern vector indicates that there are only engineering-permissible defaults and the number of defaulted predicted targets is multiple, it is mapped to the iterative correction operator; When the pattern vector indicates that only the volume balance constraint is violated, it is mapped to the volume boundary recalculation operator; When the pattern vector indicates that both the engineering allowable range violation and the volume balance constraint violation exist simultaneously, it is mapped to a combination operator sequence that first performs range trimming and then performs volume boundary recalculation. When the pattern vector indicates that multiple different types of engineering physical constraints are violated simultaneously, it is mapped to the iterative correction operator.

9. The method for intelligent generation and online compliance verification of concrete mix proportions according to any one of claims 1 to 8, characterized in that, The method further includes: sending the final mix proportion to the production control system and writing the structured information in the verification and repair process into the quality traceability database; The structured information includes at least: a summary of the input feature vector, the original prediction result, the pattern vector, the selected deterministic repair operator sequence and its version number, key process parameters, the difference before and after repair, the execution delay, the verification result, and the downgrade flag; The input feature vector summary is used to uniquely identify the index information of the input feature vector, including the hash value of the feature vector or a combination of key features; the version number is used to identify the version number to which the selected deterministic repair operator sequence belongs, and is used to trace the iteration history of the repair strategy.

10. A system for intelligent generation and online compliance verification of concrete mix proportions, comprising a memory, a processor, and a computer program or instructions stored in the memory, characterized in that, The processor executes the computer program or instructions to implement the intelligent generation and online compliance verification method for concrete mix proportions as described in any one of claims 1 to 9.

Images