Deep learning-based method for optimizing proportioning of recycled aggregate concrete

Abstract

This invention discloses a deep learning-based method for optimizing the mix proportion of construction waste recycled aggregate concrete, belonging to the field of data processing technology. The method includes: deploying a mix proportion optimization framework to determine the target concrete performance and an initial mix proportion scheme; when the initial mix proportion scheme is an empty set, the mix proportion optimization framework adopts a first optimization mode, and through iterative decision-making until the target concrete performance is met, to obtain the target mix proportion scheme; if the initial mix proportion scheme is a non-empty set, the mix proportion optimization framework adopts a second optimization mode to obtain the target mix proportion scheme. This method addresses the technical problems in existing technologies, such as the lack of precise quantitative parameters for evaluating the performance impact, blind iterative optimization, slow convergence, and poor generalization adaptability of the optimized scheme. It achieves efficient and interpretable generation of mix proportion schemes under multi-objective constraints, improves the adaptability of the mix proportion to fluctuations in recycled aggregates from different sources, reduces the number of trial mixes, and improves the utilization rate of recycled aggregates.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning. Background Technology

[0002] In the field of construction solid waste resource utilization, recycled aggregate concrete faces prominent difficulties in mix design due to the characteristics of old mortar adhering to the aggregate surface and the development of internal microcracks. These difficulties include large component fluctuations, strong nonlinear performance response, and mutual constraints among multiple objectives.

[0003] Traditional proportioning methods mainly rely on empirical formulas and extensive experimental matching, which is not only time-consuming and costly, but also difficult to adapt to the performance uncertainties caused by the diverse sources of recycled aggregates. In recent years, some studies have attempted to use genetic algorithms, response surface methodology, or neural networks for proportioning optimization. However, existing optimization strategies are mostly one-time global searches or fixed rule iterations, lacking differentiated initiation and convergence mechanisms, resulting in low search efficiency or getting stuck in local optima.

[0004] In summary, existing technologies lack precise methods for evaluating the impact of parameters on performance, and iterative optimization is often blind and slow to converge, resulting in poor generalization and adaptability of optimization schemes. How to systematically solve the problem of efficient and interpretable generation of proportioning schemes under multi-objective constraints remains an unsolved technical issue. Summary of the Invention

[0005] This application provides a deep learning-based method for optimizing the mix proportion of recycled aggregate concrete from construction waste. This method addresses the technical problems in existing technologies, such as the lack of precise quantitative parameters for evaluating the performance impact, blind iterative optimization, slow convergence, and poor generalization adaptability of the optimized solution.

[0006] In view of the above problems, this application provides a method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning.

[0007] This application provides a deep learning-based method for optimizing the mix proportion of construction waste recycled aggregate concrete. The method includes: deploying a mix proportion optimization framework, wherein the mix proportion optimization framework consists of a mix proportion decision-maker and a mix proportion judge; determining the target concrete performance and an initial mix proportion scheme; when the initial mix proportion scheme is an empty set, the mix proportion optimization framework adopts a first optimization mode, inputting the target concrete performance into the mix proportion decision-maker, making a mix proportion scheme decision based on the diffusion convergence of random noise, and transferring the decision to the mix proportion judge for execution, performing performance evaluation based on operator regression relationship and the contribution of mix proportion parameters, and making iterative decisions until the target concrete performance is met, thereby obtaining the target mix proportion scheme; if the initial mix proportion scheme is a non-empty set, the mix proportion optimization framework adopts a second optimization mode, inputting the target concrete performance and the initial mix proportion scheme into the mix proportion judge, performing performance evaluation and optimization decision based on the mix proportion decision-maker, and making iterative decisions until the target concrete performance is met, thereby obtaining the target mix proportion scheme.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: The method for optimizing the mix proportion of recycled aggregate concrete based on deep learning provided in this application deploys a mix proportion optimization framework to determine the target concrete performance and an initial mix proportion scheme. When the initial mix proportion scheme is an empty set, the mix proportion optimization framework adopts a first optimization mode, inputting the target concrete performance into the mix proportion decision-maker. The mix proportion scheme decision is performed based on the diffusion convergence of random noise, and then transferred to the mix proportion judge for execution. Performance evaluation is performed based on operator regression relationships and the contribution of mix proportion parameters. The process is iterative until the target concrete performance is met, thus obtaining the target mix proportion scheme. If the initial mix proportion scheme is a non-empty set, the mix proportion optimization framework... The optimization framework adopts a second optimization mode, inputting the target concrete performance and the initial mix proportion scheme into the mix proportion judge, performing performance evaluation and optimization decision based on the mix proportion decision-maker, and obtaining the target mix proportion scheme through in-loop decision-making until the target concrete performance is met. This is used to solve the technical problems in the prior art, such as the lack of precise quantitative parameter evaluation methods for performance impact, blind iterative optimization, slow convergence, and poor generalization adaptability of the optimized scheme. It realizes the efficient and interpretable generation of mix proportion schemes under multi-objective constraints, improves the adaptability of mix proportions to fluctuations in recycled aggregates from different sources, reduces the number of trial mixes, and improves the utilization rate of recycled aggregates. Attached Figure Description

[0009] Figure 1 This application provides a schematic diagram of the process for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning.

[0010] Figure 2This application provides a schematic diagram of the deployment process of the mix proportion optimization framework in the deep learning-based method for optimizing the mix proportion of recycled aggregate concrete from construction waste. Detailed Implementation

[0011] This application provides a deep learning-based method for optimizing the mix proportion of construction waste recycled aggregate concrete to address the technical problems in existing technologies, such as the lack of precise quantitative parameters for evaluating performance impact, blind iterative optimization, slow convergence, and poor generalization adaptability of the optimized solution.

[0012] Example: Figure 1 As shown, this application provides a deep learning-based method for optimizing the mix proportions of construction waste recycled aggregate concrete, the method comprising: S1: Deploy a ratio optimization framework, wherein the ratio optimization framework consists of a ratio decision-maker and a ratio judgment-maker.

[0013] Furthermore, such as Figure 2 As shown, the deployment of the ratio optimization framework, step S1 of this application includes: Determine the proportioning parameters and multi-objective performance to form a joint data distribution. The proportioning parameters include water-cement ratio, aggregate gradation, admixture ratio, and admixture dosage. The multi-objective performance includes strength, slump, cost, and carbon emissions. Deploy a proportioning decision-maker based on the joint data distribution. Deploy a proportioning judgment device based on the two-way performance contribution determined by operator regression relationship and the contribution of proportioning parameters. Construct a proportioning optimization architecture based on the proportioning decision-maker and the proportioning judgment device.

[0014] As a preferred implementation, the deployment of the mix design optimization framework specifically includes the following steps. First, the mix design parameters and multi-objective performance are determined, and the two are combined to form a joint data distribution. The mix design parameters refer to the controllable independent variables in concrete mix design decisions. As an exemplary rather than restrictive example, they take recycled coarse aggregate, recycled fine aggregate, recycled sand and cement as basic components, and include at least the water-cement ratio, aggregate gradation, admixture ratio, and admixture dosage.

[0015] The multi-objective performance refers to multi-dimensional indicators used to evaluate the quality and application effect of finished concrete products. For example, these include at least strength, slump, cost, and carbon emissions. It should be noted that the joint data distribution is not a simple summary of the original data, but rather a joint probability distribution representation formed by vectorizing and associating the mix proportion parameter space and the multi-objective performance space. This distribution provides a unified data foundation for the subsequent training and inference of the decision-maker and judge.

[0016] Based on this, a mix design decision-maker is deployed according to the joint data distribution. The core function of the mix design decision-maker is to generate or adjust a mix design within the mix design parameter space, using the target concrete performance as a constraint.

[0017] In one feasible implementation, the underlying logic of the allocation decision-maker is as follows: a generation strategy based on random noise diffusion convergence is adopted, that is, initial variables are sampled from the standard normal noise distribution, and noise is gradually removed through a trained reverse diffusion process, and finally mapped to the feasible solution region in the allocation parameter space, thereby generating an allocation scheme that meets the target performance constraints.

[0018] The deployment process of the matching decision-maker includes: determining training samples based on the joint data distribution, specifically including matching parameter samples and corresponding performance labels, using these as supervision signals, and conducting multiple rounds of supervised training until the loss converges. The convergence condition is that the error between the performance of the matching parameters output by the matching decision-maker and the corresponding performance label is less than a preset difference, thus obtaining the constructed matching decision-maker. Optionally, the matching decision-maker takes performance requirements as input, or takes performance requirements and existing feasible matching schemes as input, and outputs an optimized matching scheme that satisfies performance convergence.

[0019] Simultaneously, a matching factor arbiter is deployed based on a two-way performance contribution determination derived from operator regression relationships and matching parameter contribution. The first stage of the determination is based on operator regression relationships, which are explicit mathematical expressions constructed using symbolic primitives, including basic operators, constants, and variable parameters substituted by the matching parameters, through methods such as genetic programming or exhaustive search. These expressions are used to directly calculate the corresponding predicted performance values ​​based on the input matching parameters, thereby providing an interpretable quantitative mapping between matching and performance.

[0020] Simultaneously, the second stage of judgment is implemented based on the contribution of the proportioning parameters. It calculates the marginal contribution value of each proportioning parameter to the overall performance for the current proportioning scheme. This marginal contribution value is associated with a sign representing the direction of positive or negative influence, indicating the adjustment direction for increasing or decreasing the corresponding parameter. Through the synergistic effect of these two stages of judgment, the proportioning judge can not only assess whether the proportioning scheme meets the standards, but also accurately locate the key parameters affecting performance and their adjustment strategies, providing clear feedback for subsequent iterative optimization.

[0021] Finally, a complete mix design optimization architecture is formed based on the mix design decision-maker and the mix design judgment-maker. In this architecture, a closed-loop interaction is established between the mix design decision-maker and the mix design judgment-maker. Specifically, the candidate mix design schemes output by the mix design decision-maker serve as input to the mix design judgment-maker, while the performance evaluation results and mix parameter contribution information fed back by the mix design judgment-maker are fed back as optimization signals to the mix design decision-maker, driving it to perform the next round of scheme adjustment or regeneration. This cycle repeats until the performance evaluation results output by the mix design judgment-maker meet the preset performance requirements of the target concrete.

[0022] In summary, the ratio optimization framework deployed in the above manner has both the ability to generate solutions and the ability to evaluate performance. The two work together to adaptively optimize the ratio while taking into account both efficiency and accuracy.

[0023] Furthermore, the deployment of the proportioning determiner, step S1 of this application includes: A concrete mix design sample is retrieved, which contains multiple sequences based on mix design parameters and measured performance. A first mathematical relation is constructed using symbolic primitives, which include basic operators, constants, and variable parameters, with the variable parameters substituted from the mix design parameters. Based on the concrete mix design sample, the first mathematical relation is validated and optimized to obtain a second mathematical relation set, wherein the difference between the computational performance based on the first mathematical relation and the corresponding measured performance is used for convergence error validation. The second mathematical relation set is then validated and optimized based on the concrete mix design sample through multiple iterations until the error converges, resulting in a performance evaluation relation. This performance evaluation relation is then embedded into a first mix design decision node.

[0024] As a preferred implementation method, firstly, concrete mix design samples are retrieved. These samples are derived from historical experimental data, literature data, or data collected from actual engineering projects, and contain multiple sequences based on mix design parameters and measured performance.

[0025] It should be noted that each sequence corresponds to an independent matching test sample, where the matching parameters serve as the input feature vector and the measured performance serves as the output label vector. Together, they form the training basis for supervised learning.

[0026] For example, a complete sequence can be represented as: {water-cement ratio 0.35, aggregate gradation (30% for 5-10mm, 50% for 10-20mm, 20% for 20-25mm), fly ash content 15%, water-reducing agent content 1.2%; measured 28-day compressive strength 42.3MPa, slump 185mm}, and the above sequences together constitute the mix design sample set.

[0027] Secondly, symbolic primitives are used to construct the first mathematical relation. The symbolic primitives include basic operators, constants, and variable parameters. The basic operators include at least addition, subtraction, multiplication, division, exponentiation, etc. The constants are numerical constants. The variable parameters are substituted by ratio parameters, that is, each ratio parameter corresponds to a variable symbol, which is used as a placeholder in the construction of the mathematical expression and substituted with a specific value during calculation.

[0028] For example, for the water-cement ratio x1, aggregate gradation characteristic value x2, and fly ash content x3, candidate mathematical expressions such as f(x) = a·x1² + b·x2 / (x3 + c) are generated randomly based on the above symbolic primitives through genetic programming or enumeration combination methods, serving as the initial population of the first mathematical relation. The first mathematical relation is essentially a set of candidate expressions covering various possible proportion-performance mapping forms for subsequent screening and optimization.

[0029] Subsequently, based on the concrete mix design samples, the first mathematical relationship is verified and optimized to obtain a second set of mathematical relationships. Specifically, the verification and optimization process involves substituting the mix design parameters of each sample in the concrete mix design samples into each candidate expression in the first mathematical relationship to calculate the corresponding computational performance value; comparing the computational performance value with the measured performance value in the same mix design sample, and using the difference between the two as a measure of convergence error.

[0030] For example, for a candidate expression, if the sum of squares or mean absolute error of the differences between its calculated values ​​and measured values ​​on multiple samples exceeds a preset threshold, it indicates that the expression's fitting accuracy is insufficient, and it needs to be structurally optimized through operations such as selection, crossover, and mutation to generate new candidate expressions; conversely, if the error is within the allowable range, the expression is retained. After the above verification and optimization, all candidate expressions that meet the accuracy requirements are selected, forming the second mathematical relation set.

[0031] Finally, the second mathematical relation set is re-verified and optimized based on the concrete mix design sample. Through multiple iterations until the error converges, the performance evaluation relation is obtained.

[0032] Specifically, each candidate expression in the second mathematical relation set is substituted back into all concrete mix design samples for calculation and error evaluation. Expressions that do not meet the accuracy requirements are further optimized and replaced, while expressions that meet the accuracy requirements are retained and their error values ​​are recorded. After each round of verification and optimization, the current globally optimal expression and its error are recorded. When the change in the optimal error between two adjacent rounds or multiple consecutive rounds is less than the preset convergence threshold, iterative convergence is determined, and the optimization process stops.

[0033] At this point, the converged optimal expression is used as the performance evaluation relation. This performance evaluation relation uses the proportioning parameters as independent variables and the performance index as the dependent variable, and is fully interpretable. This performance evaluation relation is embedded into the first proportioning judgment node, enabling the first proportioning judgment node to quickly calculate the corresponding predicted performance value based on any set of input proportioning parameters, for subsequent performance evaluation and iterative judgment of the proportioning scheme.

[0034] Furthermore, the deployment of the proportioning determiner, step S1 of this application includes: Using the proportioning parameters as input variables and the marginal contribution value to the proportioning performance as output, a second proportioning judgment node is deployed, wherein the marginal contribution value is identified by a sign representing positive and negative; the proportioning judge is deployed using the first proportioning judgment node and the second proportioning judgment node.

[0035] In a preferred embodiment, the second proportioning judgment node uses proportioning parameters as input variables and marginal contribution values ​​to proportioning performance as output variables. The marginal contribution value quantitatively characterizes the degree of influence of a unit change in each proportioning parameter on overall performance, and is associated with a sign indicating the direction of influence: a positive sign indicates a positive correlation between the parameter and performance, and it should be increased; a negative sign indicates a negative correlation, and it should be decreased.

[0036] For example, if the marginal contribution of the water-cement ratio is negative, it indicates that the water-cement ratio needs to be reduced to improve strength; if the marginal contribution of the water-reducing agent dosage is positive, it indicates that a moderate increase will help improve slump.

[0037] In one feasible implementation, the marginal contribution value is calculated as follows: for each proportion parameter in the current proportioning scheme, positive and negative perturbations of a preset step size are applied while keeping other parameters unchanged. The mean of the performance change rate after the perturbation is calculated as the marginal contribution value of that parameter. Before calculation, the input proportioning parameters and output performance indicators are normalized to the range of 0-1 to eliminate dimensional differences.

[0038] In multi-objective scenarios, the marginal contribution values ​​of each individual performance item are summed after being assigned preset weights. The preset weights can be customized to obtain the comprehensive marginal contribution value.

[0039] After the deployment of the first ratio determination node and the second ratio determination node are completed respectively, the two together constitute the ratio determination device.

[0040] The first allocation judgment node is responsible for performance prediction, calculating specific predicted performance values ​​based on the allocation parameters to determine whether the current scheme meets the standards. The second allocation judgment node is responsible for contribution analysis, outputting a signed marginal contribution value vector to identify key control variables and provide adjustment directions. The two nodes work together to ensure that the allocation judge's output includes both a judgment on whether the standards are met and optimization suggestions on how to adjust, providing comprehensive and accurate feedback information for the iterative decision-making of the allocation decision-maker.

[0041] S2: Determine the target concrete performance and initial mix design.

[0042] S3: When the initial mix design is an empty set, the mix design optimization framework adopts the first optimization mode, inputs the target concrete performance into the mix design decision-maker, makes a mix design decision based on the diffusion convergence of random noise, and transfers it to the mix design judge for execution. The performance evaluation is solved based on the operator regression relationship and the contribution of the mix design parameters. The decision is made through iterative processes until the target concrete performance is met, and the target mix design is obtained. S4: If the initial mix design is a non-empty set, the mix design optimization framework adopts the second optimization mode, inputs the target concrete performance and the initial mix design into the mix design maker, performs performance evaluation and optimization decision based on the mix design maker, and makes decisions within the loop until the target concrete performance is met, thereby obtaining the target mix design.

[0043] As a preferred implementation, after completing the deployment of the mix design optimization framework, the method described in this application first determines the target concrete performance and the initial mix design. The target concrete performance refers to the multi-dimensional performance index thresholds required by the project; for example, it may include constraints such as a strength grade not lower than C30 and a slump controlled within the range of 160mm to 200mm. The initial mix design refers to the prior mix design data existing before formal optimization, which may originate from engineering experience, historical mix design, or preliminary schemes manually entered by operators.

[0044] In the technical solution of this application, it is necessary to identify and judge the initial proportioning scheme to determine whether it is an empty set, so as to serve as the basis for subsequent decision branches for selecting different optimization modes.

[0045] As one of the key technical features of this application, the mix design optimization framework adaptively switches to the corresponding optimization mode based on the presence or absence of an initial mix design scheme. Specifically, when the initial mix design scheme is determined to be an empty set, the mix design optimization framework activates the first optimization mode. In this mode, the mix design decision-maker is first activated, and the target concrete performance is input as a constraint condition. The mix design decision-maker responds to this input by making a mix design scheme decision using a diffusion convergence strategy based on random noise. Specifically, the mix design decision-maker samples a random noise tensor from a standard normal distribution as an initial variable, and performs a denoising process step by step through a trained inverse diffusion network. This denoising process is guided by the target concrete performance condition, so that the variable gradually converges to the feasible solution region in the mix design parameter space during the step-by-step denoising process. Finally, a set of specific mix design parameter values ​​is output through decoding mapping as the first mix design scheme.

[0046] For example, if the target strength is C30, the noise reduction process will guide the generated result to tend towards a combination of proportions that meet the strength level, such as a water-cement ratio of 0.45 and a fly ash content of 18%.

[0047] Next, the first mix design is transferred to the mix design decision unit for performance evaluation. The mix design decision unit first calculates the predicted performance value corresponding to the design based on the operator regression relationship embedded in its first mix design decision node, substituting the values ​​of each mix design parameter in the first mix design into the performance evaluation relationship. This predicted performance value includes indicators such as strength, slump, cost, and carbon emissions. It then determines whether the predicted performance value meets all the constraint thresholds for the target concrete performance. If it does, the first mix design is directly output as the target mix design, and the process ends.

[0048] If the conditions are not met, the ratio determination device further activates its second ratio determination node. According to the aforementioned training logic, for each ratio parameter in the first ratio scheme, positive and negative perturbations with a preset step size are applied while keeping other ratio parameters unchanged. The average value of the performance change rate after the perturbation is used as the marginal contribution value of the ratio parameter.

[0049] Preferably, before calculation, the input ratio parameters and output performance indicators are normalized to eliminate the dimensional differences between different ratio parameters. For example, the normalization boundary is set to the range of 0-1, and each input ratio parameter and output performance indicator is first scaled to this range according to the corresponding ratio.

[0050] Each marginal contribution value is associated with a symbol representing the direction of positive or negative influence, indicating the sensitivity of the parameter to the target performance and the direction of adjustment. Subsequently, the array of marginal contribution values ​​is filtered using a preset threshold. Marginal contribution values ​​with absolute values ​​greater than the preset threshold are marked as optimization variables, indicating that they are highly sensitive parameters in subsequent iterations and should be actively adjusted. Marginal contribution values ​​with absolute values ​​not greater than the preset threshold are marked as optimization parameters, indicating that they are low-sensitivity parameters and should be kept fixed in subsequent iterations to reduce the dimensionality of the optimization space and improve convergence efficiency. The preset threshold can be customized based on experience.

[0051] Subsequently, based on the above division results, a first feedback condition is generated. This feedback condition includes the names of the optimization variables to be adjusted and their adjustment direction indicators (positive increase or negative decrease), as well as a list of optimization quantities to be kept unchanged. The first feedback condition is imported into the mix proportioning decision-maker, driving it to perform a new round of scheme adjustment or regeneration. It then flows back to the mix proportioning judge for evaluation. This decision-judgment-feedback process is repeated until the performance evaluation result output by the mix proportioning judge meets all the constraints of the target concrete performance. At this point, the current mix proportion scheme is determined as the final target mix proportion scheme.

[0052] As another parallel technical path of this application, when the initial mix design is determined to be a non-empty set, the mix design optimization framework initiates a second optimization mode. In this mode, the mix design judge is first activated, and the target concrete performance and the initial mix design are simultaneously input into it. The first mix design judgment node of the mix design judge performs a performance evaluation on the initial mix design based on operator regression, calculates its predicted performance values, and determines the amount and direction of deviation from the target concrete performance. Simultaneously, the second mix design judgment node of the mix design judge solves for the contribution of mix design parameters, calculates the marginal contribution value of each mix design parameter in the initial mix design, filters and divides the optimization variables and optimization quantities using a preset threshold, and generates a second feedback condition. The second feedback condition includes the performance deviation information of the initial mix design and the adjustment direction and priority of each mix design parameter.

[0053] Subsequently, the initial mix design and the second feedback condition are input into the mix design decision-maker. Starting with the initial mix design and guided by the adjustment instructions in the second feedback condition, the mix design decision-maker generates an updated mix design based on the same underlying logic. This updated design is then passed to the mix design judge for performance evaluation and contribution calculation, generating new feedback conditions which are then input into the decision-maker again. This forms a closed-loop decision-making cycle between the mix design decision-maker and the mix design judge, iterating repeatedly until the mix design judge confirms that the current design meets all the constraints of the target concrete performance. At this point, the design is output as the target mix design.

[0054] It should be noted that, unlike the first optimization mode where the proportioning decision-maker is generated globally starting from random noise, the second optimization mode makes local adjustments based on the existing proportioning scheme. Its search space is smaller and its convergence speed is faster, thus making full use of prior information.

[0055] Furthermore, the steps in this application also include: Identify the initial proportioning scheme. If the initial proportioning scheme is an empty set, the proportioning optimization framework adopts the first optimization mode to obtain the target proportioning scheme. If the initial proportioning scheme is a non-empty set, the proportioning optimization framework adopts the second optimization mode to obtain the target proportioning scheme.

[0056] As a preferred embodiment, after the mix design optimization framework is deployed and the target concrete performance and initial mix design scheme are determined, this application first identifies and judges the initial mix design scheme, and adaptively selects the corresponding optimization execution path based on whether it is an empty set.

[0057] Specifically, when the initial proportioning scheme is identified as an empty set, it indicates that there is currently no prior knowledge of the proportioning, and the proportioning optimization framework initiates the first optimization mode. In the first optimization mode, the system lacks a reference starting proportion, so the proportioning decision maker needs to generate a scheme without prior guidance, and gradually approach the target performance through subsequent iterative decisions, finally obtaining the target proportioning scheme.

[0058] Conversely, when the initial proportioning scheme is identified as a non-empty set, it indicates the existence of existing prior proportioning data, and the proportioning optimization framework initiates a second optimization mode. In this second optimization mode, the prior information carried by the initial proportioning scheme is fully utilized. Using this scheme as the starting point for optimization, the target proportioning scheme is ultimately obtained through subsequent local searches and iterative adjustments. Compared to the first optimization mode, this mode has a smaller search space and a faster convergence speed, achieving effective reuse of existing knowledge.

[0059] Through the branch identification and mode switching mechanism based on whether the initial proportion scheme is empty or not, the proportion optimization framework can maintain optimal optimization efficiency and convergence performance under different prior information conditions.

[0060] Furthermore, the ratio optimization framework adopts a first optimization mode to obtain the target ratio scheme. Step S3 of this application includes: If the initial mix design is an empty set, the mix design optimization framework first activates the mix design decision-maker, and generates a first mix design by performing diffusion convergence based on random noise, with the target concrete performance as a constraint. The first mix design is then imported into the mix design judge, and the mix design parameters in the first mix design are used as variable parameters. Based on the first mix design judgment node, the first mix design performance is calculated. If the first mix design performance meets the target concrete performance, the first mix design is taken as the target mix design.

[0061] In a preferred implementation, under the first optimization mode, the mix design optimization framework first activates the mix design decision-maker. The mix design decision-maker is configured to output a reasonable combination of mix design parameters under specific constraints. Specifically, the target concrete performance is input into the mix design decision-maker as a set of constraints, including but not limited to multi-dimensional indicators such as strength grade, slump range, cost ceiling, and carbon emission limits. The mix design decision-maker uses this target performance as a guiding condition and performs a diffusion-convergence process based on random noise to determine the mix design scheme.

[0062] Specifically, the diffusion convergence based on random noise refers to the process whereby the mix design decision maker samples a random noise tensor from a standard normal distribution as an initial variable. This variable does not possess any semantic information in the mix design parameter space. Subsequently, a denoising process is executed step by step through a trained reverse diffusion mechanism. Each denoising step is guided by the target concrete performance as a condition, so that the variable gradually converges to the feasible solution region in the mix design parameter space.

[0063] For example, when the target strength is C30 and the slump requirement is 180±20mm, the reverse diffusion process gradually shapes random noise into specific values ​​for proportioning parameters such as water-cement ratio, aggregate gradation, admixture ratio, and additive dosage to meet the strength and slump constraints during iterative denoising. Finally, the proportioning decision-maker decodes and maps the converged variables to output a set of specific proportioning parameter values, and this output is recorded as the first proportioning scheme.

[0064] Next, the first proportioning scheme is imported into the proportioning decision-maker for performance evaluation. The proportioning decision-maker is pre-deployed with a first proportioning decision-making node, which embeds a performance evaluation relationship trained by operator regression, using proportioning parameters as independent variables and various performance indicators as dependent variables. Each proportioning parameter in the first proportioning scheme is substituted into the performance evaluation relationship as a variable to calculate the predicted performance value corresponding to the scheme, and this predicted performance value is recorded as the first proportioning performance.

[0065] For example, if the first formulation includes a water-cement ratio of 0.44, a fly ash content of 16%, and a water-reducing agent content of 1.1%, then these values ​​can be substituted into the corresponding evaluation relationships for strength, slump, cost, and carbon emissions to obtain the corresponding predicted values ​​for compressive strength, slump, unit cost, and carbon emissions.

[0066] Finally, the performance of the first mix design is compared with the performance of the target concrete item by item. If all the indicators in the performance of the first mix design meet all the constraint thresholds set for the performance of the target concrete, the scheme is deemed qualified, and the first mix design is directly output as the target mix design, and the process terminates.

[0067] For example, if the target requirements are strength ≥ 30 MPa, slump 160-200 mm, cost ≤ 350 yuan / m³, and carbon emissions ≤ 280 kg CO2 / m³, and the calculated performance of the first proportion is 32.5 MPa, 185 mm, 325 yuan / m³, and 265 kg CO2 / m³, all falling within the threshold range, then the first proportion scheme is confirmed to be the target proportion scheme. If any indicator does not meet the requirements, the process proceeds to the subsequent feedback optimization process, where the proportion judge further calls the second proportion judgment node to generate feedback conditions and sends them back to the proportion decision-maker, driving iterative loops until the target is met.

[0068] Furthermore, step S3 of this application includes: If the first mix proportion performance does not meet the target concrete performance, the first mix proportion performance is imported into the second mix proportion judgment node to determine the marginal contribution value array of each mix proportion parameter; the marginal contribution value array is judged by a preset threshold, and the mix proportion parameter corresponding to the marginal contribution value greater than the preset threshold is used as the optimization variable, otherwise it is used as the optimization quantity; based on the optimization variable and the optimization quantity, a first feedback condition is generated and imported into the mix proportion decision-maker, and the target mix proportion scheme is determined through iterative optimization until the target concrete performance is met.

[0069] In a preferred implementation, under the first optimization mode, when the first mix proportion performance does not meet the target concrete performance, a feedback optimization process is initiated. Specifically, the mix proportion determiner imports the first mix proportion performance into a second mix proportion determination node. The second mix proportion determination node is configured to take mix proportion parameters based on the mix proportion scheme as input and output the marginal contribution value of each parameter to the overall performance. The marginal contribution value quantitatively characterizes the degree of influence of a unit change in each mix proportion parameter on the overall performance and is associated with a symbol representing the direction of positive or negative influence.

[0070] Specifically, in response to the import of the first proportion performance, the second proportion judgment node calculates the marginal contribution value corresponding to each proportion parameter in the current first proportion scheme, and combines the marginal contribution values ​​of all proportion parameters to form a marginal contribution value array. For example, if the current proportion scheme includes four parameters: water-cement ratio, aggregate gradation characteristic value, fly ash content, and water-reducing agent content, then the marginal contribution value array can be represented as a vector form of [-0.35, 0.12, 0.08, 0.42].

[0071] Subsequently, the array of marginal contribution values ​​is filtered and judged using a preset threshold. The preset threshold is a pre-defined absolute value threshold for contribution, which is used to distinguish the sensitivity of the allocation parameters to the overall performance. Specifically, the absolute value of the marginal contribution value of each allocation parameter is compared with the preset threshold. The allocation parameters corresponding to marginal contribution values ​​with absolute values ​​greater than the corresponding preset threshold are marked as optimization variables, indicating that the parameter has a significant impact on the overall performance and should be actively adjusted in subsequent iterations. The allocation parameters corresponding to marginal contribution values ​​with absolute values ​​not greater than the preset threshold are marked as optimization variables, indicating that the parameter has a weak impact on the overall performance and should be kept fixed in subsequent iterations to reduce the dimensionality of the optimization space and accelerate convergence.

[0072] Finally, a first feedback condition is generated based on the division results of the optimized variables and optimized quantities. The first feedback condition includes at least the following information: the names of the optimized variables to be adjusted and their adjustment direction indicators, determined by the sign of their marginal contribution values ​​(positive indicating increase, negative indicating decrease), and a list of optimized quantities to be kept unchanged. The first feedback condition and the first mix design are imported into the mix design decision-maker, driving it to perform directional adjustments only on the optimized variables in the first mix design in the next iteration, using the first feedback condition as a priori guide, while fixing the optimized quantities, thus generating an updated mix design. This updated design is then passed to the mix design decision-maker for performance evaluation and contribution calculation, repeating the evaluation-division-feedback process. This iterative optimization continues until the mix design decision-maker confirms that the current design meets all the constraints of the target concrete performance. At this point, the current mix design is determined as the final target mix design.

[0073] Furthermore, the ratio optimization framework adopts a second optimization mode to obtain the target ratio scheme. Step S4 of this application includes: If the initial mix design is a non-empty set, the mix design optimization framework first activates the mix design decision-maker to obtain the second feedback condition; then the initial mix design and the second feedback condition are input into the mix design decision-maker, and iterative optimization is performed until the target concrete performance is met, thus obtaining the target mix design.

[0074] In a preferred implementation, in the second optimization mode, when the initial proportioning scheme is identified as a non-empty set, the proportioning optimization framework first activates the proportioning diagnostic unit. Unlike the first optimization mode where the decision-maker generates the scheme first, the second optimization mode fully utilizes existing prior information and prioritizes the diagnostic evaluation of the initial scheme through the diagnostic unit.

[0075] Specifically, the target concrete performance and the initial mix design are input into the mix design decision-maker. The first mix design decision-making node of the mix design decision-maker calculates the initial mix performance based on the embedded performance evaluation relationship, using the mix design parameters from the initial mix design as variables. Simultaneously, the second mix design decision-making node calculates an array of marginal contribution values ​​of each mix design parameter to the overall performance, and uses a preset threshold to filter and classify optimization variables and optimization quantities. The mix design decision-maker integrates the above two decision results to generate a second feedback condition. The second feedback condition includes at least the deviations between the current initial mix design performance and the target concrete performance, the marginal contribution values ​​of each mix design parameter and their sign direction, and the list of optimization variables and optimization quantities.

[0076] Subsequently, the initial proportioning scheme and the second feedback condition are input into the proportioning decision-maker. The proportioning decision-maker takes the initial proportioning scheme as the starting point for optimization, and unlike the global generation that starts from random noise in the first optimization mode, it performs local search optimization under the guidance of the second feedback condition.

[0077] Specifically, the proportioning decision-maker updates the corresponding parameter values ​​in the initial proportioning scheme according to the optimization variables marked in the second feedback conditions and their adjustment direction indications, and keeps the parameters marked as optimization quantities unchanged.

[0078] For example, if the second feedback condition indicates that the water-cement ratio is an optimized variable with a negative marginal contribution value, and the water-reducing agent dosage is an optimized variable with a positive marginal contribution value, then the mix proportioning decision-maker reduces the water-cement ratio and increases the water-reducing agent dosage based on the initial scheme, while keeping other non-sensitive parameters unchanged, generating an updated mix proportioning scheme. The updated scheme is then passed to the mix proportioning judge for performance evaluation and contribution calculation, generating a new second feedback condition, which is then input into the mix proportioning decision-maker again. This forms a closed-loop decision-making cycle between the mix proportioning decision-maker and the mix proportioning judge, iterating repeatedly until the mix proportioning judge confirms that the current scheme meets all the constraints of the target concrete performance. At this point, the current scheme is output as the target mix proportioning scheme.

[0079] By adopting the aforementioned local optimization strategy that starts with a priori schemes and is guided by decision-maker feedback, the second optimization mode achieves effective reuse of existing knowledge and has a smaller search space and faster convergence speed compared to the first optimization mode.

[0080] Furthermore, after obtaining the target formulation, this application also includes the following steps: According to the target formulation scheme, small-batch process testing is carried out to collect test data; based on the test data, performance evaluation is performed to obtain test performance, wherein the performance generalization is measured by the evaluation variance, quantiles and deviation distribution of each test sample.

[0081] If the test performance does not meet the target concrete performance, test feedback conditions are generated. The target ratio scheme and the test feedback conditions are input into the ratio decision-maker for ratio optimization.

[0082] As a preferred implementation, after obtaining the target formulation scheme through the aforementioned iterative optimization, this application further introduces a process verification step to test the feasibility and robustness of the scheme under actual production conditions.

[0083] Specifically, based on the values ​​of each proportion parameter in the target mix design, concrete trial mixing and curing are carried out according to a preset small-batch process testing procedure to prepare multiple test samples. Actual performance data are collected for each test sample according to standard testing methods. The test data includes at least indicators such as compressive strength and slump for each sample, thereby forming an actual performance dataset.

[0084] Subsequently, the target formulation scheme is evaluated based on the test data to obtain the test performance. This test performance is derived from actual physical experiments and better reflects the performance of the target formulation scheme under real-world process conditions.

[0085] To comprehensively measure the performance consistency and generalization ability of the target formulation across different test samples, this application uses the evaluation variance, quantiles, and deviation distribution of each test sample to measure performance generalization. The evaluation variance characterizes the dispersion of performance values ​​for each test sample; a smaller variance indicates higher stability of the formulation across different batches or specimens. The quantiles characterize the distribution characteristics of performance values; for example, the 25%, 50%, and 75% quantiles can be calculated to assess the central tendency and dispersion range of each sample's performance. The deviation distribution characterizes the degree of deviation of each test sample's performance value from the target performance threshold and its distribution pattern, including the direction of the deviation (positive or negative bias) and its magnitude.

[0086] By using the above multi-dimensional generalization metrics, it is possible to comprehensively determine whether the target formulation scheme has sufficient consistency and reliability under actual process conditions.

[0087] In one feasible implementation, if the test performance meets all the constraint thresholds of the target concrete performance and the generalization metrics all meet the requirements, such as the variance being less than the preset tolerance value and the overall deviation distribution being within the target threshold range, then the target mix design is confirmed to be effective and can be put into formal production application.

[0088] If the tested performance does not meet the target concrete performance, it indicates that although the mix design meets the requirements in the virtual optimization stage, there is a performance deviation under actual process conditions. In this case, test feedback conditions are generated. The test feedback conditions include at least the deviation between the measured performance values ​​of each test sample and the target performance, the direction of the deviation, and information on the performance differences of each mix design parameter under the actual test conditions.

[0089] Subsequently, the target proportion scheme and the test feedback conditions are input into the proportion decision-maker for proportion optimization.

[0090] Specifically, the proportioning decision-maker takes the target proportioning scheme as the starting point for optimization and the test feedback conditions as prior guidance to perform targeted fine-tuning of the proportioning parameters, generating a corrected proportioning scheme. This corrected scheme can be verified again through small-batch process testing until the test performance meets the target requirements. Through the closed-loop feedback mechanism combining digital optimization and practical verification, this application effectively improves the success rate and reliability of the proportioning scheme from theory to engineering implementation.

[0091] The deep learning-based method for optimizing the mix proportion of recycled aggregate concrete from construction waste provided in this application has the following technical effects: A dataset is constructed by integrating multi-dimensional parameters and performance indicators such as aggregate gradation and carbon emissions, and a mix proportion decision-maker and a two-node mix proportion judge are built accordingly. A single framework is compatible with two business scenarios: generating mix proportions from scratch and iteratively optimizing existing mix proportions. When there is no initial solution, candidate mix proportions are generated through diffusion convergence. Existing solutions are first evaluated for defects and then targeted for correction, with feedback conditions generated based on parameter contribution for iterative optimization. This targeted narrowing of the optimization scope accelerates the convergence speed of multi-objective mix proportions.

[0092] Based on the operator regression performance formula that converges to the sample iterative fitting error, the positive and negative marginal contributions of each mix proportion parameter are simultaneously calculated. This accurately predicts concrete strength, cost, and other indicators, identifies key adjustable parameters, and reduces ineffective iterations.

[0093] After optimization, small-batch trial runs were conducted. The generalization ability of the solution was evaluated by variance and deviation distribution. Measured data that did not meet the standards were fed back for re-optimization. The algorithm calculation and on-site process were integrated to improve the reliability of the solution implementation.

[0094] In summary, it can quickly and adaptively output recycled aggregate concrete mix proportions that take into account various performance indicators, thereby improving the utilization rate of recycled aggregates.

[0095] Through the foregoing detailed description of the deep learning-based method for optimizing the mix proportion of recycled aggregate concrete from construction waste, those skilled in the art can clearly understand the deep learning-based method for optimizing the mix proportion of recycled aggregate concrete from construction waste in this embodiment. As for the apparatus disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section description.

[0096] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning, characterized in that, The method includes: A ratio optimization framework is deployed, wherein the ratio optimization framework consists of a ratio decision-maker and a ratio judgment-maker; Determine the target concrete performance and initial mix design; When the initial mix design is an empty set, the mix design optimization framework adopts the first optimization mode, inputs the target concrete performance into the mix design decision-maker, makes a mix design decision based on the diffusion convergence of random noise, and then transfers it to the mix design judge for execution. The performance evaluation is solved based on the operator regression relationship and the contribution of the mix design parameters. The decision is made through iterative processes until the target concrete performance is met, and the target mix design is obtained. If the initial mix design is a non-empty set, the mix design optimization framework adopts the second optimization mode, inputs the target concrete performance and the initial mix design into the mix design maker, performs performance evaluation and optimization decision based on the mix design maker, and makes decisions within the loop until the target concrete performance is met, thus obtaining the target mix design.

2. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 1, characterized in that, Deployment of a ratio optimization framework, including: Determine the proportioning parameters and multi-objective performance to form a joint data distribution, wherein the proportioning parameters include water-cement ratio, aggregate gradation, admixture ratio and additive dosage, and the multi-objective performance includes strength, slump, cost and carbon emissions; Deploy the allocation decision-maker based on the joint data distribution; A matching decision maker is deployed based on the two-way performance contribution determination derived from operator regression relationships and matching parameter contribution. The ratio decision-maker and the ratio judge form a ratio optimization architecture.

3. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 2, characterized in that, Deploy the proportioning determiner, including: Retrieve concrete mix design samples, wherein the concrete mix design samples contain multiple sequences based on mix design parameters and measured performance; The first mathematical relation is constructed using symbolic primitives, wherein the symbolic primitives include basic operators, constants and variable parameters, and the variable parameters are substituted by the ratio parameters; Based on the concrete mix design sample, the first mathematical relationship is verified and optimized to obtain a second set of mathematical relationships, wherein the convergence error is verified by the difference between the computational performance based on the first mathematical relationship and the corresponding measured performance. The second mathematical relation set is validated and optimized based on the concrete mix design sample. Through multiple iterations until the error converges, the performance evaluation relation is obtained, and the performance evaluation relation is embedded into the first mix design decision node.

4. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 3, characterized in that, Deploy the proportioning determiner, including: Using the proportioning parameters as input variables and the marginal contribution value to the proportioning performance as output, a second proportioning judgment node is deployed, wherein the marginal contribution value is identified by a sign representing positive or negative. The ratio determination device is deployed using the first ratio determination node and the second ratio determination node.

5. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 4, characterized in that, Identify the initial proportioning scheme; if the initial proportioning scheme is an empty set, the proportioning optimization framework adopts the first optimization mode to obtain the target proportioning scheme. If the initial proportioning scheme is a non-empty set, the proportioning optimization framework adopts the second optimization mode to obtain the target proportioning scheme.

6. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 5, characterized in that, The ratio optimization framework adopts the first optimization mode to obtain the target ratio scheme, including: If the initial mix design is an empty set, the mix design optimization framework first activates the mix design decision-maker, and generates the first mix design by performing diffusion convergence based on random noise, with the target concrete performance as a constraint. The first proportioning scheme is imported into the proportioning judge, and the proportioning parameters in the first proportioning scheme are used as variable parameters. Based on the first proportioning judge node, the first proportioning performance is calculated. If the performance of the first mix proportion meets the target concrete performance, the first mix proportion shall be used as the target mix proportion.

7. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 6, characterized in that, If the first mix proportion performance does not meet the target concrete performance, the first mix proportion performance is imported into the second mix proportion judgment node to determine the marginal contribution value array of each mix proportion parameter; The marginal contribution value array is judged by a preset threshold. The matching parameter corresponding to the marginal contribution value that is greater than the preset threshold is used as the optimization variable, and otherwise it is used as the optimization quantifier. Based on the optimization variables and optimization quantities, a first feedback condition is generated and imported into the mix proportion decision-maker. Through iterative optimization, the target concrete performance is met, and the target mix proportion scheme is determined.

8. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 5, characterized in that, The ratio optimization framework adopts a second optimization mode to obtain the target ratio scheme, including: If the initial proportioning scheme is a non-empty set, the proportioning optimization framework first activates the proportioning judge to obtain the second feedback condition; The initial mix design and the second feedback condition are input into the mix design decision-maker, and iterative optimization is performed until the target concrete performance is met, thus obtaining the target mix design.

9. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 1, characterized in that, After obtaining the target formulation, the following is included: Based on the target formulation scheme, conduct small-batch process testing and collect test data; Based on the test data, a performance evaluation is performed to obtain the test performance, wherein the performance generalization is measured by the evaluation variance, quantiles, and bias distribution of each test sample.

10. The method for optimizing the mix proportion of construction waste recycled aggregate concrete based on deep learning as described in claim 9, characterized in that, If the test performance does not meet the target concrete performance, a test feedback condition is generated. The target ratio scheme and the test feedback conditions are input into the ratio decision-maker for ratio optimization.

Images