Green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization

Abstract

This invention relates to the field of building material proportion optimization technology, and discloses an intelligent green building material proportion optimization and evaluation system based on solid waste resource utilization. The system includes modules for feature construction, mapping prediction, quantitative evaluation, loss reconstruction, collaborative optimization, and solution output. The system acquires raw physicochemical test data of solid waste, constructs solid waste physicochemical feature vectors and building material proportion parameter vectors, inputs them into a multi-task learning network, and outputs predicted values ​​of mechanical properties and life-cycle carbon emissions in parallel. Based on the predicted values, it calculates carbon efficiency indicators and comprehensive quantitative evaluation scores. Using the carbon efficiency indicators, it calculates dynamic penalty weights, reconstructs an adaptive total loss function, and trains the network. It updates the building material proportion parameter vectors based on the evaluation scores and performs closed-loop iteration based on feedback. When the set convergence conditions are met, it outputs the globally optimal building material proportion parameter vector and the result dataset. This invention achieves automated iterative optimization of building material proportions.

Description

Technical Field

[0001] This invention relates to the field of building material proportioning optimization technology, specifically to an intelligent proportioning optimization and evaluation system for green building materials based on solid waste resource utilization. Background Technology

[0002] With the development of the construction industry and the increasing environmental protection requirements, the resource utilization of industrial solid waste in the production of green building materials has become an industry trend. Due to the complex and fluctuating physicochemical properties of solid waste raw materials, traditional building material formulation design usually relies on manual experience and physical adaptation. This method consumes a lot of time and material costs, and it is difficult to accurately find the optimal solution among multiple process parameters.

[0003] In recent years, some technical solutions have begun to incorporate data-driven algorithms for predicting building material performance and calculating mix proportions. However, existing mix proportion optimization methods mostly model single mechanical performance indicators, failing to incorporate life-cycle carbon emissions and solid waste utilization rates into a unified collaborative calculation framework. This severs the connection between the macroscopic mechanical properties of building materials and their environmental impact, resulting in calculated mix proportions that struggle to balance engineering safety and green, low-carbon requirements. Furthermore, conventional algorithms often lack effective constraints on actual physical states during iterative optimization, leading to output mix proportion parameters that exceed tolerance limits for the total absolute volume of materials per unit volume, thus deviating from engineering realities. In addition, existing models lack dynamic penalty feedback mechanisms for high carbon emissions or low solid waste utilization rates during the training phase. The system cannot adaptively adjust towards lower carbon efficiency during backpropagation of the computational graph gradient, making it difficult to efficiently and accurately output globally optimal mix proportions that meet comprehensive evaluation criteria. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent proportioning optimization and evaluation system for green building materials based on solid waste resource utilization. This system solves the problems that existing building material proportioning designs struggle to balance mechanical performance and life-cycle carbon emissions within a unified computational framework, and that the lack of physical volume constraints and low-carbon-efficiency penalty mechanisms during iterative optimization leads to proportioning schemes that are easily detached from engineering realities and have low automated optimization efficiency.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent proportioning optimization and evaluation system for green building materials based on solid waste resource utilization, characterized in that it includes: The feature construction module is used to obtain the original physicochemical test data of solid waste raw materials and construct solid waste physicochemical feature vectors and building material ratio parameter vectors. The mapping prediction module is used to input the solid waste physicochemical feature vector and the building material ratio parameter vector into a multi-task learning network, and output the mechanical performance prediction value and the life cycle carbon emission prediction value. The quantitative evaluation module is used to calculate the carbon efficiency index and the comprehensive quantitative evaluation score based on the predicted mechanical performance value and the predicted carbon emissions over the entire life cycle. The loss reconstruction module is used to calculate dynamic penalty weights based on the carbon efficiency index, reconstruct the adaptive total loss function using the dynamic penalty weights, and train the multi-task learning network. The collaborative optimization module is used to update the building material ratio parameter vector based on the comprehensive quantitative evaluation score, and feed the updated building material ratio parameter vector back to the mapping prediction module for a new round of calculation. The solution output module is used to stop the new round of calculation when the comprehensive quantitative evaluation score meets the set convergence condition, and output the globally optimal building material ratio parameter vector and result dataset.

[0006] Preferably, the feature construction module is specifically used for: Obtain chemical composition and physical morphology data from the original physicochemical test data of solid waste raw materials; Intrinsic feature parameters are extracted from the chemical composition data and the physical morphology data to form an initial solid waste physicochemical feature vector; The derived feature parameters are calculated using the intrinsic feature parameters, and then appended to the initial solid waste physicochemical feature vector for range normalization, outputting the standardized solid waste physicochemical feature vector.

[0007] Preferably, the feature construction module is further used for: Extract the building material proportioning parameters to form the building material proportioning parameter vector; The building material mixing parameters include the dosage of various solid wastes, the dosage of silicate cement clinker, the water-cement ratio, and the dosage of chemical activators. The feature construction module defines hard constraint boundaries for the building material proportioning parameter vector; The hard constraint boundary includes equality constraints and inequality constraints; The equation constraint limits the sum of the mass percentages of all dry powder material parameters in the building material proportioning parameter vector; The inequality constraint limits the lower and upper limits of the process values ​​for each dimension of the building material proportioning parameter vector.

[0008] Preferably, the mapping prediction module is specifically used for: The solid waste physicochemical feature vector and the building material proportioning parameter vector are concatenated along the dimensions to generate a multi-source feature fusion vector. The multi-source feature fusion vector is input into the shared hidden layer inside the multi-task learning network for nonlinear feature extraction, and the hidden layer feature vector is output. At the end of the shared hidden layer, a mechanical performance prediction branch and a carbon emission prediction branch are connected; The hidden layer feature vectors are input into the mechanical performance prediction branch and the carbon emission prediction branch respectively, and the mechanical performance prediction value and the life cycle carbon emission prediction value are output in parallel.

[0009] Preferably, the quantitative evaluation module is specifically used for: The predicted mechanical properties are linearly weighted and summed to generate an equivalent mechanical property scalar. The carbon efficiency index is constructed by performing a division operation on the equivalent mechanical performance scalar and the predicted carbon emissions over the entire life cycle. Extract the amount of various solid wastes from the building material proportioning parameter vector and sum them to generate the solid waste utilization rate; The carbon efficiency index and the preset target carbon efficiency threshold are subjected to dimensionless processing to generate a relative carbon efficiency ratio. The relative carbon efficiency ratio and the solid waste utilization rate are weighted and summed to calculate the comprehensive quantitative evaluation score.

[0010] Preferably, the loss reconstruction module is specifically used for: Determine whether the relative carbon efficiency ratio is less than one; When the relative carbon efficiency ratio is determined to be less than one, a dynamic penalty weight with a value greater than one is generated based on the degree to which the relative carbon efficiency ratio deviates from one. Obtain accurate label values ​​for mechanical properties and carbon emissions; Calculate the mechanical performance prediction loss based on the actual labeled mechanical performance value and the predicted mechanical performance value; Calculate the life-cycle carbon emission prediction loss based on the actual carbon emission label value and the life-cycle carbon emission prediction value; The basic multi-objective collaborative loss function is constructed by weighted summation of the mechanical performance prediction loss and the full life cycle carbon emission prediction loss. The dynamic penalty weights are multiplied and fused with the basic multi-objective collaborative loss function to reconstruct the adaptive total loss function, and the multi-task learning network is trained using the adaptive total loss function.

[0011] Preferably, the collaborative optimization module is specifically used for: Freeze the network layer weight parameters and bias vectors in the multi-task learning network; The comprehensive quantitative evaluation score is converted into a negative value to construct an optimization objective function; The chain rule of the network computation graph is used to calculate the gradient vector of the building material ratio parameter with respect to the building material ratio parameter vector of the optimization objective function; The building material proportioning parameter gradient vector and the preset optimization iteration step size are used to perform numerical update calculations on the building material proportioning parameter vector to obtain the updated building material proportioning parameter vector; The updated building material ratio parameter vector is fed back to the mapping prediction module for a new round of calculation.

[0012] Preferably, the solution output module is specifically used for: Extract the proportions of each raw material from the updated building material ratio parameter vector, and convert them into the mass of each raw material by combining them with the preset total mass of building materials per cubic meter. The absolute volume is calculated by converting the mass of the individual raw material and the corresponding apparent density of the material. The total absolute volume is generated by adding the absolute volume of all raw materials to the preset volume of building material gas content. Determine whether the total absolute volume is within a preset volume tolerance range; When it is determined that the total absolute volume is not within the preset volume tolerance range, a volume value penalty item is generated; The volume value penalty term is superimposed on the optimization objective function, and the update direction of subsequent calculation rounds is adjusted through error backpropagation.

[0013] Preferably, the solution output module is further used for: Extract the state parameters from two adjacent iterations, and calculate the absolute value of the score change of the comprehensive quantitative evaluation score in two adjacent calculation rounds; Calculate the gradient norm 2 of the building material proportioning parameter vector; When the absolute value of the score change is less than a preset score convergence threshold and the gradient norm is less than a preset gradient convergence threshold, the comprehensive quantitative evaluation score is determined to meet the set convergence condition.

[0014] Preferably, the solution output module is further used for: Generate the result dataset, which includes a green building materials optimal ratio decision panel data package; The solution output module transforms the original predicted indicators associated with the globally optimal building material ratio parameter vector into normalized evaluation values ​​with multiple independent dimensions. A multidimensional performance radar chart feature is constructed using normalized evaluation values ​​from multiple independent dimensions; Calculate the coverage area scalar of the multidimensional performance radar map features; The coverage area scalar, the list of raw material usage per unit volume, and various original prediction indicators are encapsulated according to key-value pairs, and header information containing project batch identifiers and timestamps is attached to generate the final green building materials optimal ratio decision panel data package.

[0015] This invention provides an intelligent system for optimizing and evaluating the proportioning of green building materials based on solid waste resource utilization. It offers the following advantages: 1. This invention constructs a multi-task learning network comprising a shared hidden layer, a mechanical performance prediction branch, and a carbon emission prediction branch. It integrates solid waste physicochemical feature vectors and building material mix design parameter vectors as input, enabling simultaneous prediction of building material mechanical properties and life-cycle carbon emissions. This structure utilizes the shared hidden layer to extract low-level correlation information from multi-source features, reducing parameter redundancy in the independent operation of multiple single-objective prediction models, improving the computational efficiency of multi-dimensional index prediction, and providing fundamental data support for multi-objective mix design evaluation and optimization.

[0016] 2. This invention introduces a dynamic penalty mechanism based on carbon efficiency indicators. In the loss reconstruction module, dynamic penalty weights are generated based on the relative carbon efficiency ratio and fused with the basic multi-objective collaborative loss function to reconstruct an adaptive total loss function. When the relative carbon efficiency ratio of the current ratio parameter is less than one, the system amplifies the loss proportion of that sample during network training by using a dynamic penalty weight greater than one. This mechanism directly intervenes in the update direction of the network layer weight parameters during the model training phase, enabling the system to effectively tilt towards a ratio with high solid waste utilization and low carbon emissions while ensuring the accuracy of mechanical performance prediction.

[0017] 3. This invention employs a collaborative optimization strategy based on computational graph gradients, combined with a physical volume verification mechanism in the solution output module, to achieve closed-loop automated iteration of building material proportions. The system updates the building material proportion parameter vector using the gradient vector obtained through chain-like differentiation, and determines the tolerance range for the calculated total absolute volume of materials per unit area, converting any volume deviations into volume penalty terms that are added to the optimization objective function. This approach not only avoids the time cost and material consumption of traditional manual trial mixing but also ensures the rationality and engineering practicality of the output globally optimal proportion at the physical space level. Attached Figure Description

[0018] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a graph showing the convergence curve of the optimization iteration of this invention; Figure 4 This is a bar chart showing the multidimensional performance of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0020] See attached document Figure 1 , Figure 1 This invention provides an intelligent ratio optimization and evaluation system for green building materials based on solid waste resource utilization, including: a feature construction module, a mapping prediction module, a quantitative evaluation module, a loss reconstruction module, a collaborative optimization module, and a solution output module.

[0021] The feature construction module is used to acquire physicochemical property data of solid waste raw materials, historical building material mix proportion data, and process constraint boundary data. The module transforms the acquired data into solid waste physicochemical feature vectors and building material mix proportion parameter vectors, and constructs a system feature space for subsequent calculations.

[0022] The mapping prediction module is connected to the feature construction module. Based on a deep neural network, the mapping prediction module establishes a mapping relationship between solid waste physicochemical feature vectors, building material proportioning parameter vectors, and building material performance indicators and environmental impact indicators. Given input conditions, the mapping prediction module outputs in parallel predicted values ​​of mechanical properties and life-cycle carbon emissions for the current proportioning parameters.

[0023] The quantitative evaluation module is connected to the mapping prediction module. The quantitative evaluation module receives the predicted mechanical properties and the predicted life-cycle carbon emissions, and calculates the carbon efficiency index. The carbon efficiency index characterizes the mechanical properties converted from a unit of carbon emission equivalent. The quantitative evaluation module, combined with a preset solid waste utilization rate, calculates the comprehensive quantitative evaluation score of the current formulation scheme.

[0024] The loss reconstruction module is connected to the quantization evaluation module. The loss reconstruction module receives the carbon efficiency index output by the quantization evaluation module. Based on the deviation of this carbon efficiency index from the target carbon efficiency threshold, the loss reconstruction module calculates and updates the dynamic penalty weights. These dynamic penalty weights are then applied to the carbon emission penalty term of the basic loss function, generating a dynamic adaptive loss function for the current iteration state. This function drives the parameter training and error backpropagation of the multi-task learning network.

[0025] The collaborative optimization module is connected to the loss reconstruction module. Under the constraint of a fixed solid waste physicochemical characteristic vector, the collaborative optimization module aims to maximize the comprehensive quantitative evaluation score (i.e., minimize the optimization objective function composed of its negative values) and uses an optimization algorithm to update the position of the building material proportion parameter vector. In each iteration, the collaborative optimization module feeds back the updated building material proportion parameter vector to the mapping prediction module, triggering the next round of forward prediction and evaluation feedback calculation, thus forming a closed-loop iterative circuit of the system.

[0026] The solution output module is connected to the collaborative optimization module. When the system determines that the comprehensive evaluation score meets the set convergence condition or the number of iterations reaches the set maximum threshold, the solution output module controls the collaborative optimization module to stop iterating. The solution output module extracts and outputs the globally optimal building material mix ratio parameter vector, and simultaneously generates a result dataset containing the predicted mechanical properties, predicted life-cycle carbon emissions, optimal solid waste blending rate, and final carbon efficiency index corresponding to these mix ratio parameters.

[0027] See attached document Figure 2 , Figure 2 The present invention provides an intelligent ratio optimization and evaluation method for green building materials based on solid waste resource utilization. This method relies on the above-mentioned intelligent ratio optimization and evaluation system. The overall workflow of the system includes the following steps.

[0028] First, the system obtains raw data through the feature construction module and constructs solid waste physicochemical feature vectors and building material ratio parameter vectors.

[0029] Secondly, the system inputs the aforementioned feature vectors and parameter vectors into the mapping prediction module, and outputs the predicted mechanical performance value for the current cycle and the predicted carbon emission value for the entire life cycle through forward calculation.

[0030] Next, the system processes the above prediction results through the quantitative evaluation module to calculate the carbon efficiency index and the comprehensive quantitative evaluation score.

[0031] Then, the system extracts the carbon efficiency index from the previous round through the loss reconstruction module, calculates the dynamic penalty weight in real time based on the index deviation, and then reconstructs the dynamic adaptive loss function required for the current round of network training, and completes the iterative training of the multi-task learning network.

[0032] Subsequently, the system uses a collaborative optimization module to calculate the gradient of the building material mix ratio parameter vector based on the trained and converged network model and the current comprehensive quantitative evaluation score, and then updates the parameters. The updated building material mix ratio parameter vector is then re-entered into the mapping prediction module for a new round of calculations.

[0033] Finally, the system monitors the iteration process through the solution output module, exits the closed loop when the preset termination conditions are met, and outputs the final optimal ratio solution and related indicator data.

[0034] The system uses a feature construction module to quantify and vectorize the physicochemical characteristics of solid waste raw materials, generating a mathematical feature space that can be read and computed by the algorithm. This process specifically includes the following steps: The feature construction module acquires raw physicochemical test data of solid waste raw materials, specifically including at least one of fly ash, blast furnace slag, steel slag, coal gangue, and metal tailings. The feature construction module receives test records through an external laboratory information management system or on-site testing equipment; the raw physicochemical test data covers the chemical composition and physical morphology data of the solid waste raw materials.

[0035] The feature construction module extracts intrinsic feature parameters from the raw physicochemical test data and constructs an initial solid waste physicochemical feature vector. The solid waste physicochemical feature vector is denoted as... Its expression is In the formula, This represents the total number of dimensions of the feature parameters. Indicates the first Feature values ​​in each dimension, superscript This represents the vector transpose. The aforementioned intrinsic characteristic parameters specifically include the mass fraction, specific surface area, loss on ignition, and fineness of the key oxides. The key oxides include silicon oxide, aluminum oxide, calcium oxide, magnesium oxide, and ferric oxide.

[0036] The feature construction module calculates derived feature parameters based on the extracted intrinsic feature parameters, and appends the derived feature parameters as new elements to the solid waste physicochemical feature vector. The derived characteristic parameters are used to characterize the potential hydration reactivity of solid waste raw materials, specifically including the alkalinity coefficient and the mass coefficient. The system obtains these coefficients through a pre-set chemical calculation model. Taking the alkalinity coefficient as an example, the alkalinity coefficient is denoted as... The calculation formula is as follows: ; in: This indicates the mass fraction of calcium oxide in solid waste raw materials. This indicates the mass fraction of magnesium oxide. This indicates the mass fraction of silicon dioxide. This indicates the mass fraction of aluminum oxide.

[0037] The quality coefficient is denoted as The calculation formula is as follows: ; The feature construction module incorporates the calculated alkalinity coefficient and mass coefficient as derived feature values ​​into the solid waste physicochemical feature vector. middle.

[0038] The feature construction module completes the physical and chemical feature vector of solid waste. Normalization is performed. Due to the differences in numerical magnitudes between different physical quantities, the system uses range normalization to map the numerical values ​​of each feature in the solid waste physicochemical feature vector to the same specific interval. The normalization calculation formula is: ; in: This represents the characteristic value after normalization. Indicates the first Feature values ​​in each dimension This represents the maximum statistical value of this type of feature in the historical dataset. This represents the minimum statistical value of this type of feature in the historical dataset. For data cleaning during the acquisition of raw physicochemical test data, those skilled in the art can use the Laida criterion to remove outliers and the mean interpolation method to handle missing values. These data cleaning steps are well-known techniques in the field and will not be elaborated upon here. After the above steps, the feature construction module outputs a standardized solid waste physicochemical feature vector, which serves as the direct input data for the subsequent mapping prediction module and collaborative optimization module.

[0039] The system defines the building material proportioning parameter space and sets process boundary constraints through the feature construction module, providing an optimization space and decision boundaries for subsequent optimization algorithms. This process specifically includes the following steps: The feature construction module extracts and defines building material proportioning parameters based on the preparation process of the target green building material. The extracted proportioning parameters are then grouped into a vector, denoted as the building material proportioning parameter vector, whose expression is: .in, This represents the total number of dimensions of the proportioning parameters. Indicates the first The ratio values ​​of each dimension, superscript This represents the vector transpose. The specific building material mix design parameters include the dosage of various solid wastes, the dosage of silicate cement clinker, the water-cement ratio, and the dosage of chemical activators. The dosage of each type of solid waste refers to the percentage by mass of different types of solid waste raw materials in the total cementitious material system. The dosage of chemical activators refers to the percentage by mass of the activator relative to the total mass of the cementitious material system; the activator can be water glass, sodium hydroxide, or sodium sulfate. The water-cement ratio is the ratio of the total mass of mixing water to the total mass of the total cementitious material system, which includes various solid waste raw materials and silicate cement clinker.

[0040] The feature construction module is a vector of building material proportioning parameters. Establish hard constraints on mass conservation and process operability. These hard constraints include equality constraints and inequality constraints. Equality constraints limit the total mass percentage of all dried powder materials in the formulation; the formula for calculating this is: ; in: This represents the total number of dimensions of the dry powder material parameters, and satisfies the following conditions: , Indicates the first The feature construction module generates a vector of building material proportioning parameters. (The text also mentions numerical values ​​for each dimension.) The front of the middle Each dimension is set as a parameter for dry powder materials, specifically including the dosage of various solid wastes and the dosage of silicate cement clinker.

[0041] Inequality constraints are used to limit the allowable fluctuation range of each proportioning parameter in actual engineering production, in order to prevent the generation of proportioning schemes that are physically impossible to achieve. Their expression is: ; in: Indicates the first The lower limit of the process ratio values ​​for each dimension. Indicates the first The upper and lower limits of the process parameters for each dimension of the ratio. These upper and lower limits are directly read by the feature construction module based on empirical data from the historical database or relevant industry standards.

[0042] The feature construction module receives and records the target performance constraints and environmental impact boundary constraints for system optimization. The feature construction module also obtains externally set target mechanical performance benchmark values ​​and life-cycle carbon emission limits. The target mechanical performance benchmark value is denoted as... Its specific characteristics include the 28-day compressive strength benchmark value and the 28-day flexural strength benchmark value. The life-cycle carbon emission limit is denoted as... This limit is used to define the maximum permissible carbon dioxide emission equivalent per unit volume of green building materials throughout their lifecycle, from raw material acquisition and processing to final molding. The feature construction module packages and transmits the set equality constraints, inequality constraints, target mechanical performance benchmark values, and full lifecycle carbon emission limits to the collaborative optimization module as the basic boundary conditions for determining whether the iteratively generated mix design is compliant.

[0043] For the mixing process and curing operation in the trial mixing of building materials, those skilled in the art can use a standard cement mortar mixer to prepare specimens and cure them in a standard constant temperature and humidity curing chamber. The specimen preparation and curing process is a well-known technology in the field and will not be described in detail here.

[0044] The system constructs and calculates a forward mapping network between components, performance, and carbon emissions through a mapping prediction module. This process specifically includes the following steps: The mapping and prediction module acquires the standardized solid waste physicochemical feature vector and building material ratio parameter vector, and concatenates and fuses them to construct the input layer data of the network model. The mapping and prediction module concatenates the solid waste physicochemical feature vector and the building material ratio parameter vector along their dimensions to generate a multi-source feature fusion vector. This multi-source feature fusion vector is denoted as... Its expression is .in, This represents the transpose of the solid waste physicochemical characteristic vector. This represents the transpose of the building material proportioning parameter vector. The multi-source feature fusion vector covers the fundamental input dimensions that influence the hydration reaction process and final macroscopic performance of the solid waste building material system.

[0045] The mapping prediction module inputs the multi-source feature fusion vector into the shared hidden layer for nonlinear feature extraction. The shared hidden layer consists of multiple cascaded fully connected network layers, used to uncover the implicit coupling relationship between the physicochemical properties of solid waste and the proportioning parameters. The feature transfer and calculation process of a layer is described by the following matrix formula: ; in: Indicates the first Layers share the output feature vector of the hidden layer; Indicates the first Layers share the output feature vector of the hidden layer, when hour, ; Indicates the first Layers share the weight coefficient matrix of the hidden layer; Indicates the first Layers share the bias vectors of the hidden layers; This represents a non-linear activation function. The mapping prediction module uses a modified linear unit function as the concrete implementation of this non-linear activation function in the shared hidden layer to avoid the gradient vanishing problem in deep network computation.

[0046] The mapping prediction module connects multiple independent task output branches at the end of the shared hidden layer, constructing the branch topology of the multi-task learning network. The mapping prediction module assigns the output feature vector of the last layer of the shared hidden layer to the mechanical performance prediction branch and the carbon emission prediction branch. The mechanical performance prediction branch calculates and outputs the predicted mechanical performance values ​​for the current mix design parameters. The carbon emission prediction branch calculates and outputs the predicted life-cycle carbon emissions values ​​for the current mix design parameters.

[0047] The mapping prediction module sets up independent fully connected layers in each task output branch for forward computation, obtaining the final output values ​​for each prediction dimension. The output calculation formula for the mechanical performance prediction branch is as follows: ; in: Indicates the predicted mechanical properties; Let represent the output feature vector of the last shared hidden layer, and satisfy . , Indicates the total number of shared hidden layers; This represents a specific weight matrix for the mechanical performance prediction branch; This represents a specific bias vector for the mechanical performance prediction branch.

[0048] The output calculation formula for the carbon emission prediction branch is as follows: ; in: This represents the projected carbon emissions over the entire life cycle; This represents a specific weight matrix for the carbon emission prediction branch; This represents a specific bias vector for the carbon emission prediction branch. Through the parallel branching structure described above, the mapping prediction module simultaneously outputs predicted mechanical performance and life-cycle carbon emission values ​​during a single forward propagation, providing foundational data for subsequent multi-dimensional evaluation and optimization.

[0049] For the model weight initialization method and backpropagation parameter update mechanism in the construction process of multi-task learning network, those skilled in the art can use the Xavier initialization algorithm to assign weights and use the error backpropagation algorithm combined with historical samples to train parameters. The model initialization and training process is a well-known technology in this field and will not be described in detail here.

[0050] The system employs a parallel prediction mechanism for mechanical performance and life-cycle carbon emissions through a mapping prediction module. This process specifically includes the following steps: The mapping prediction module outputs predicted mechanical properties through the mechanical property prediction branch. These predicted mechanical properties are denoted as... The predicted mechanical properties are a column vector containing multiple mechanical indices. Specifically, the expression for the predicted mechanical properties is: .in, This represents the predicted compressive strength value after 28 days. This indicates the predicted flexural strength value after 28 days, indicated by the superscript. This represents the transpose of a vector. The mapping prediction module adjusts the type of output mechanical index by adding or removing dimensions from this vector according to the requirements of the target application scenario.

[0051] The mapping prediction module outputs the life-cycle carbon emission prediction value through the carbon emission prediction branch. The life-cycle carbon emission prediction value is denoted as... The historical training labels for this branch network are constructed based on a defined carbon footprint accounting boundary. The accounting boundary is set as the production stage from the acquisition of solid waste and raw materials to the mixing and molding of building materials. The formula for calculating the true carbon emission label value of the historical mixing scheme is: ; in: This represents the actual carbon emission label value for the corresponding historical mix design. This indicates the total number of different types of raw materials actually consumed. Indicates the first The actual quality of the raw materials consumed; Indicates the first Carbon emission equivalent factor corresponding to each raw material; This represents the fixed carbon emissions generated during the mixing and stirring process.

[0052] The mapping prediction module calculates the above-mentioned first step based on the building material mix ratio parameter vector and the preset total mass of the building material trial mix, using mass conversion relationships. The actual quality of raw materials consumed.

[0053] The mapping prediction module establishes a carbon emission prediction branch and uses a deep network to fit the nonlinear relationship between the input features and environmental impacts. This carbon emission prediction branch, through forward propagation, directly outputs a life-cycle carbon emission prediction value that approximates the actual carbon emission label value. The mapping prediction module simultaneously acquires and outputs the mechanical performance prediction value and the life-cycle carbon emission prediction value through parallel computation.

[0054] The mapping prediction module encapsulates the predicted mechanical performance values ​​and the predicted carbon emissions over the entire life cycle generated by the forward calculation in the current cycle into a specific data format. This encapsulated data is then transmitted to the quantitative evaluation module as the input for subsequent calculations of the quantitative evaluation score and penalty weights.

[0055] For the selection of various carbon emission equivalent factors and the calibration of basic data for building material raw materials, those skilled in the art can consult and use the life cycle inventory database of relevant industries or the national greenhouse gas emission factor database. The extraction and verification of basic carbon emission factors are well-known technologies in this field and will not be elaborated here.

[0056] The system performs mathematical construction and evaluation calculations of carbon efficiency indicators through a quantitative evaluation module. This process specifically includes the following steps: The quantitative evaluation module receives the predicted mechanical properties and life-cycle carbon emissions from the mapping prediction module. The predicted mechanical properties are column vectors containing multiple mechanical indicators, while the predicted life-cycle carbon emissions are scalar data. The quantitative evaluation module performs dimensionality checks and non-negativity checks on the input data to eliminate abnormal forward prediction data.

[0057] The quantitative evaluation module performs dimensionality reduction and order reduction calculations on the above predicted mechanical properties to obtain an equivalent mechanical property scalar. Since the predicted mechanical properties include both the predicted 28-day compressive strength and the predicted 28-day flexural strength, the quantitative evaluation module uses a linear weighted summation method to integrate them into a single comprehensive strength characterization quantity. The formula for calculating the equivalent mechanical property scalar is: ; in: Represents an equivalent mechanical property scalar; This represents the predicted compressive strength value after 28 days. This represents the predicted flexural strength value after 28 days. This represents the compressive strength weighting coefficient; This represents the flexural strength weighting coefficient. The sum of the two weighting coefficients mentioned above is limited to 1. In specific implementation scenarios, since building materials mainly bear compressive loads in actual engineering projects, the quantitative evaluation module sets the compressive strength weighting coefficient to a real number between 0.8 and 1.0, and the flexural strength weighting coefficient to a real number between 0 and 0.2.

[0058] The quantitative evaluation module constructs and outputs a carbon efficiency index based on the extracted equivalent mechanical performance scalar and the predicted value of carbon emissions throughout the entire life cycle, through a division operation. The mathematical formula for calculating the carbon efficiency index is: ; in: Indicates carbon efficiency indicators; Represents an equivalent mechanical property scalar; This represents the projected carbon emissions over the entire life cycle. The carbon efficiency index is used to characterize the equivalent mechanical performance scalar value corresponding to a unit of carbon emission equivalent.

[0059] The quantitative evaluation module retrieves the system's preset target carbon efficiency threshold. The target carbon efficiency threshold is denoted as... The quantitative evaluation module calculates baseline carbon efficiency data based on historical database data of the actual performance and carbon emissions of traditional ordinary Portland cement building materials of the same strength grade. This baseline carbon efficiency data is then multiplied by a preset enhancement coefficient to generate the aforementioned target carbon efficiency threshold. The preset enhancement coefficient is a real number greater than 1.0, specifically ranging from 1.1 to 1.3. The quantitative evaluation module compares and stores the calculated carbon efficiency index with the target carbon efficiency threshold, providing a parameter basis for subsequent calculations of the comprehensive quantitative evaluation score and the dynamic adaptive loss function.

[0060] For the mathematical solution of the weight coefficients in the above steps, those skilled in the art can use the analytic hierarchy process or the entropy weight method to assign subject and object weights. The specific calculation process for determining the weights is a well-known technology in this field and will not be elaborated here.

[0061] The system calculates a multi-dimensional comprehensive evaluation score for the utilization rate of integrated solid waste through a quantitative evaluation module. This process specifically includes the following steps: The quantitative evaluation module extracts the amount of various solid wastes from the building material proportioning parameter vector and calculates the solid waste utilization rate of the current proportioning scheme. The solid waste utilization rate characterizes the mass proportion of solid waste in the target green building materials. The formula for calculating the solid waste utilization rate is: ; in: Indicates the utilization rate of solid waste; This indicates the total number of types of solid waste raw materials; Indicates the first The quantitative evaluation module assigns values ​​to the various solid waste raw materials in the total cementitious material system using index matching. The quantitative evaluation module sums the extracted percentage values ​​of various types of solid waste to obtain a scalar value of the overall solid waste utilization rate.

[0062] The quantitative evaluation module extracts the carbon efficiency index calculated above and compares it with the system's preset target carbon efficiency threshold, then performs dimensionless processing. The module divides the carbon efficiency index by the target carbon efficiency threshold to obtain the relative carbon efficiency ratio. This relative carbon efficiency ratio is used to measure the degree to which the current formulation scheme meets the preset benchmark in terms of carbon efficiency. Its calculation formula is as follows: ; in: This indicates the relative carbon efficiency ratio. Indicates carbon efficiency indicators; This represents the target carbon efficiency threshold.

[0063] The quantitative evaluation module constructs a multi-dimensional comprehensive quantitative evaluation function based on the relative carbon efficiency ratio and solid waste utilization rate. The module calculates the comprehensive quantitative evaluation score of the current formulation scheme using a linear weighted average. The mathematical formula for calculating the comprehensive quantitative evaluation score is as follows: ; in: This represents the comprehensive quantitative evaluation score; This represents the weighting coefficient for carbon efficiency evaluation; This represents the weighting coefficient for solid waste utilization evaluation. The sum of the two evaluation weighting coefficients mentioned above is limited to 1. In the implementation scenario, the system sets the weighting coefficients according to the specific engineering carbon reduction requirements or solid waste disposal indicators. When the project focuses on high performance and low carbon emissions, the quantitative evaluation module sets the carbon efficiency evaluation weighting coefficient to a real number between 0.6 and 0.8; when the project focuses on maximizing the disposal of large quantities of solid waste, the quantitative evaluation module sets the solid waste utilization evaluation weighting coefficient to a real number between 0.6 and 0.8.

[0064] The quantitative evaluation module stores and outputs the calculated comprehensive quantitative evaluation score. This comprehensive quantitative evaluation score, as one of the core indicators for determining whether the system's optimization iteration meets the convergence conditions, is synchronously transmitted to the loss reconstruction module and the solution output module.

[0065] For the normalization mapping and dimensionless method of multiple indicators in the above evaluation system, those skilled in the art can use the max-min normalization or Z-score normalization method for data processing. The specific steps of the dimensionless data processing are well known in the art and will not be described in detail here.

[0066] The system constructs a basic multi-objective collaborative loss function through a loss reconstruction module, which is used to calculate the basic prediction error of the multi-task learning network during the forward propagation process. This process specifically includes the following steps: The loss reconstruction module acquires the predicted mechanical properties and life-cycle carbon emissions from the mapping prediction module, and simultaneously extracts corresponding real test data from the historical database as label values. The real test data includes the actual label values ​​for mechanical properties and the actual label values ​​for carbon emissions of the corresponding historical mix design. The actual label value for mechanical properties is denoted as... .

[0067] The loss reconstruction module calculates the prediction error of the mechanical performance prediction branch, generating the mechanical performance prediction loss. The module uses the mean square error function to calculate the sum of squares of the differences between the predicted mechanical performance values ​​and the actual labeled mechanical performance values. The formula for calculating the mechanical performance prediction loss is: ; in: Indicates the predicted loss of mechanical properties; Indicates the predicted mechanical properties; This represents the true labeled value of the mechanical properties; This represents the L2 norm operation of a vector.

[0068] The loss reconstruction module calculates the prediction error of the carbon emission prediction branch, generating the life-cycle carbon emission prediction loss. The module generates the prediction error for this branch by calculating the squared scalar difference between the life-cycle carbon emission prediction value and the actual carbon emission label value of the corresponding historical mix design. The formula for calculating the life-cycle carbon emission prediction loss is: ; in: This represents the predicted loss of carbon emissions over the entire life cycle; This represents the projected carbon emissions over the entire life cycle; This represents the true carbon emission label value for the corresponding historical mix design.

[0069] The loss reconstruction module weights and sums the predicted mechanical performance loss and the predicted life-cycle carbon emission loss to construct a basic multi-objective collaborative loss function. The formula for calculating the basic multi-objective collaborative loss function is as follows: ; in: Represents the basic multi-objective collaborative loss function; Indicates the initial weight of mechanical property loss; Indicates the initial weight of carbon emission losses; The observation noise parameter represents the branch of mechanical property prediction; This represents the observation noise parameter for the carbon emission prediction branch.

[0070] In a specific implementation, to balance task branches with different dimensions, the loss reconstruction module uses a homoscedastic uncertainty weighted algorithm to dynamically assign values ​​to the initial weights. Specifically, the loss reconstruction module defines the initial weight of mechanical performance loss as... The initial weight of carbon emission loss is defined as and will and These parameters serve as learnable network parameters at the end of a multi-task learning network. During error backpropagation, the network adaptively updates these parameters based on the gradient variance between the mechanical performance prediction branch and the carbon emission prediction branch.

[0071] Regarding the regularization term processing in the above loss function calculation process, those skilled in the art can add an L1 regularization term or an L2 regularization term to the end of the basic multi-objective collaborative loss function to reduce the complexity of the network layer parameters. The specific calculation process of the regularization constraint is a well-known technology in the field and will not be described in detail here.

[0072] The system performs dynamic update calculations of the reverse driving and penalty weights based on evaluation feedback through the loss reconstruction module. This process specifically includes the following steps: The loss reconstruction module obtains the relative carbon efficiency ratio output by the quantitative evaluation module. The relative carbon efficiency ratio represents the carbon efficiency level under forward prediction of the current historical training samples. Based on the magnitude of this value, the loss reconstruction module determines whether the currently input historical training samples belong to the category of samples that have not met the engineering low-carbon benchmark.

[0073] The loss reconstruction module calculates dynamic penalty weights based on the relative carbon efficiency ratio. The module employs a one-sided penalty mechanism, increasing the dynamic penalty weight for samples with a relative carbon efficiency ratio less than 1. This amplifies the corresponding predicted loss to improve the network's fitting accuracy for low-carbon-efficiency boundary samples. The formula for calculating the dynamic penalty weights is: ; in: Indicates dynamic penalty weight; This represents the preset penalty amplification factor; This represents the function that takes the maximum value. This represents the relative carbon efficiency ratio. The preset penalty amplification coefficient is assigned by the loss reconstruction module based on the convergence state of the network training, and is specifically set to a real number between 0.5 and 2.0. When the relative carbon efficiency ratio is greater than or equal to 1, the dynamic penalty weight is 1; when the relative carbon efficiency ratio is less than 1, the dynamic penalty weight will be greater than 1, and the greater the deviation from 1, the larger the weight value.

[0074] The loss reconstruction module multiplies and fuses the calculated dynamic penalty weights with the basic multi-objective collaborative loss function to reconstruct an adaptive total loss function. The formula for calculating the adaptive total loss function is: ; in: Represents the adaptive total loss function; This represents the basic multi-objective collaborative loss function. The loss reconstruction module obtains the final error value for the current forward propagation round through the above calculations.

[0075] The loss reconstruction module performs backpropagation calculations for the multi-task learning network based on the final error value output by the adaptive total loss function. During backpropagation, the module treats the dynamic penalty weights as constant scalars, truncating their gradient backpropagation paths. The module calculates the gradient values ​​of the adaptive total loss function with respect to the network parameters in the shared hidden layers and the output branches of each task, and uses the optimizer to iteratively update the weight coefficient matrix and bias vector of each fully connected network layer based on these gradient values.

[0076] For the gradient descent calculation method in the above-mentioned network parameter iterative update process, those skilled in the art can use the Adam optimization algorithm or the stochastic gradient descent algorithm with momentum term to control the network learning rate and update parameters. The specific calculation process of the optimization algorithm is a well-known technology in the field and will not be described in detail here.

[0077] The system executes a gradient descent-based optimization strategy for the matching parameter space through the scheme output module. This process specifically includes the following steps: The solution output module freezes all network layer weight parameters and bias vectors in the trained multi-task learning network. The solution output module obtains the solid waste physicochemical feature vector of the current target project batch and randomly generates an initial building material ratio parameter vector within a preset feasible ratio region, denoted as... The aforementioned solid waste physicochemical feature vectors remain constant inputs during the current optimization iteration process.

[0078] The solution output module concatenates the solid waste physicochemical feature vector with the building material ratio parameter vector of the current iteration, inputs it into a multi-task learning network for forward propagation, and calculates the comprehensive quantitative evaluation score corresponding to the current ratio through the mapping prediction module and the quantitative evaluation module. The solution output module transforms the ratio optimization task into a problem of maximizing the comprehensive quantitative evaluation score. To adapt to the automatic differentiation and minimization mechanism of the deep learning computing framework, the solution output module constructs an optimization objective function that is the negative value of the comprehensive quantitative evaluation score. The formula for calculating the optimization objective function is: ; in: This represents the objective function for optimization. This represents the comprehensive quantitative evaluation score generated by forward calculation based on the input solid waste physicochemical feature vector and building material proportioning parameter vector; Represents the physicochemical characteristic vector of solid waste; This represents a vector of building material proportioning parameters.

[0079] The solution output module performs backpropagation and differentiation of the error with respect to the network input. Utilizing the chain rule of the network computation graph, the module calculates the partial derivative of the objective function with respect to the building material proportion parameter vector, obtaining the iterative gradient information. The formula for calculating the gradient vector of the building material proportion parameters is: ; in: This represents the gradient vector of building material mix proportion parameters. Based on the calculated gradient vector, the solution output module performs numerical updates on the building material mix proportion parameter vector. The update calculation formula is: ; in: Indicates the first The building material proportioning parameter vector generated in the next iteration; Indicates the first The building material proportioning parameter vector generated in the next iteration; This indicates the preset optimization iteration step size; Indicates the first The gradient vector of the building material proportioning parameters obtained from the previous iteration.

[0080] The solution output module applies engineering physical constraints to the updated building material mix proportion parameter vector after each iteration. Since the dosage of each raw material in the building material mix has physical boundaries, the solution output module sets upper and lower limit constraint vectors for the mix proportion parameters and uses a truncation function to restrict mix proportion values ​​exceeding the boundaries to the feasible region boundary. The implementation formula is as follows: ; in: This represents the vector of building material mix proportion parameters after boundary constraint processing; This represents the preset lower limit vector of the proportioning parameters; This represents the upper limit vector of the preset proportioning parameters; and This indicates an extremum constraint operation performed on an element-by-element basis.

[0081] When dealing with mix proportion parameters subject to constant total proportion constraints (e.g., the total content of various cementitious materials is limited to 100%), the scheme output module performs normalization calculations to proportionally scale the constrained values. The normalization calculation formula is as follows: ; in: Represents the normalized th The item is subject to the constraint of the proportioning parameter value; Indicates the truncated number after processing. The item is subject to the constraint of the proportioning parameter value; This represents the total number of dimensions of the dry powder material parameters subject to the summation constraint. The solution output module uses the above calculation results to replace the corresponding elements in the building material proportioning parameter vector.

[0082] The solution output module determines whether the current optimization iteration process meets the stopping condition. The stopping condition is set as follows: the change in the comprehensive quantitative evaluation score between two consecutive iterations is less than a preset convergence threshold, or the number of iterations reaches the maximum optimization round limit set by the system. When the stopping condition is met, the solution output module ends the optimization loop and marks the current building material ratio parameter vector as the candidate optimal solution.

[0083] For the non-gradient global optimization alternatives in the above-mentioned ratio parameter space optimization process, those skilled in the art can use particle swarm optimization algorithm, differential evolution algorithm or genetic algorithm to construct fitness function and perform population iterative search. The specific calculation process and parameter configuration of the heuristic optimization algorithm are well known in the art and will not be described in detail here.

[0084] The system performs boundary checks, truncation processing, and iterative convergence determination calculations through the scheme output module. This process specifically includes the following steps: The solution output module obtains the building material proportioning parameter vector after each iteration update and extracts its feature dimension data. To ensure the physical feasibility of the proportioning scheme, the solution output module introduces a physical boundary verification mechanism based on the absolute volume method. The solution output module calculates the mass of each raw material based on the proportions of various raw materials in the building material proportioning parameter vector, combined with the preset total mass per cubic meter of building materials, and retrieves the corresponding apparent density to calculate the total absolute volume of the proportioning scheme. The conversion formula for the mass of each raw material is based on the water-cement ratio and activator dosage ratio in the proportioning parameter vector, combined with the total mass per cubic meter of building materials to construct a mass balance equation. First, the total mass of the cementitious material is calculated and separated, and then multiplied by the corresponding percentage of solid waste and clinker dosage to obtain the mass of each individual raw material. In the specific calculation logic, the quality of a single raw material The conversion process is as follows: Let the preset total mass of building materials per cubic meter be... The water-to-binder ratio parameter is: The chemical activator dosage parameter is: The total mass of cementitious materials The calculation formula is: Subsequently, the system will calculate the total mass of the cementitious material. Multiply by the corresponding first element in the building material proportioning parameter vector The specific mass of a dry powder material can be obtained by determining its percentage content. At the same time, the system will measure the total mass of cementitious materials. Multiply by the water-cement ratio parameter respectively With chemical activator dosage parameters This yields the corresponding quality of the mixing water and the chemical activator, thus providing a comprehensive understanding of the target... Individual quality data for each raw material.

[0085] The solution output module determines whether the total absolute volume satisfies the set unit volume conservation constraint. The formula for calculating unit volume conservation is: ; in: Represents the total absolute volume; Indicates the total number of raw materials; Indicates the first The quality of the raw materials; Indicates the first The apparent density of the raw materials; This indicates the preset gas content volume of the building materials. The solution output module makes the following judgments. Whether it is within the preset volume tolerance range (this range can be set to [0.98, 1.02] cubic meters).

[0086] Simultaneously, the solution output module retrieves the pre-set target mechanical performance benchmark value. and life-cycle carbon emission limits Verify whether the predicted mechanical performance value for the current cycle and the predicted carbon emissions over the entire life cycle exceed the above-mentioned limit boundaries.

[0087] If the calculated total absolute volume is outside the tolerance range, the solution output module uses a penalty function method, adding a volume numerical penalty term to the above optimization objective function. The specific formula for calculating this volume numerical penalty term is as follows: ; in, Indicates a numerical penalty term; This represents the penalty coefficient, which takes a value greater than 10. 4 real numbers; This represents the volume tolerance threshold. The solution output module adds this numerical penalty term to the optimization objective function, and adjusts the update direction of the building material proportion parameter vector through error backpropagation.

[0088] The solution output module performs optimization convergence determination calculations. It extracts state parameters from two adjacent iterations and employs a dual-determination logic combining the objective function difference and the gradient vector norm to monitor the iteration state. The specific calculation formula for the dual-determination logic is as follows: ; ; in: This represents the absolute value of the score change; Indicates the first The comprehensive quantitative evaluation score generated in the next iteration; Indicates the first The comprehensive quantitative evaluation score generated in the next iteration; Represents the gradient 2 norm; L2 norm operation for vectors; Indicates the first The gradient vector of building material proportioning parameters obtained from the next iteration.

[0089] Solution output module judgment Whether the score is less than the preset convergence threshold is determined simultaneously. Is it less than the preset gradient convergence threshold? In specific implementation scenarios, in order to ensure both the system's optimization accuracy and computational convergence efficiency, the solution output module sets the preset score convergence threshold to 10. -5 Up to 10 -3 Real numbers between (for example, a value of 1.0 × 10) -4 ), and set the preset gradient convergence threshold to 10. -4 Up to 10 -2 Real numbers between (for example, a value of 1.0 × 10) -3 When both of the above conditions are met simultaneously, or when the current optimization iteration round reaches the preset maximum iteration limit, the solution output module determines that the current optimization process has converged and stops iterating. The solution output module uses the building material ratio parameter vector generated in the last iteration as the candidate optimal building material ratio solution.

[0090] The solution output module outputs the candidate optimal building material mix ratio schemes. It multiplies the various proportional parameters in the optimal building material mix ratio scheme with the preset total mass of building materials per unit volume to generate a list of raw material usage per unit volume. The solution output module then encapsulates this usage list, the corresponding predicted mechanical properties, the predicted life-cycle carbon emissions, and the comprehensive quantitative evaluation score into a data format and stores it in a local database or outputs it to an external human-computer interaction terminal.

[0091] For the above-mentioned methods for measuring the apparent density of various raw materials in the absolute volume method and the standard determination process for gas content, those skilled in the art can refer to the relevant national building materials testing standards or industry specifications. The specific testing methods and parameter acquisition processes are well-known technologies in this field and will not be elaborated here.

[0092] The system generates and outputs data for the optimal mix design decision panel for green building materials through the solution output module. This process specifically includes the following steps: The solution output module obtains the list of raw material usage per unit volume locked after iterative optimization and convergence, along with the corresponding predicted mechanical properties, predicted life-cycle carbon emissions, and comprehensive quantitative evaluation scores. The solution output module normalizes this data, mapping it to the standard value range of [0,1], and extracts it into performance evaluation node data with multiple independent dimensions. Mathematically, this performance evaluation node data is represented by the normalized evaluation values ​​for the corresponding dimensions.

[0093] Based on the aforementioned performance evaluation node data, the solution output module constructs a multi-dimensional performance radar chart feature. To characterize the overall balance of the current optimal mix across multiple dimensions including mechanics, environment, and solid waste disposal, the solution output module calculates the coverage area scalar of the multi-dimensional performance radar chart. The formula for calculating the coverage area scalar is: ; in: Indicates the coverage area scalar; This represents the total number of dimensions for performance evaluation nodes; Indicates the first Normalized evaluation values ​​for each dimension; Indicates the first Normalized evaluation values ​​for each dimension, and the system is limited. .

[0094] The solution output module combines the calculated coverage area scalar, unit raw material usage list, and various original forecast indicators to generate a green building materials optimal ratio decision panel data package. In a specific implementation, the solution output module uses JSON data serialization format to encapsulate the above data into a key-value pair structure, and appends header information containing project batch identifiers and timestamps to the beginning of the data package for subsequent identification and traceability in the production control system.

[0095] The solution output module sends the generated green building material optimal ratio decision panel data package to an external terminal device via a network communication interface. The external terminal device receives and parses the decision panel data package, converting it into a visual operation panel or into specific formula execution instructions, which are then sent to the programmable logic controller at the building material production and mixing station for material mixing.

[0096] Regarding the network communication protocols and security verification mechanisms in the above-mentioned data packet transmission process, those skilled in the art can use the MQTT protocol or the HTTP protocol combined with the TLS encryption layer for data transmission. The specific operation process of the underlying network communication and encryption handshake is a well-known technology in this field and will not be described in detail here.

[0097] Specific application examples: To further illustrate the operation of this invention, we take the mix design of a batch of C40 strength grade green building materials (prepared by blending fly ash and blast furnace slag) as an example.

[0098] Data preparation and feature construction: The system receives raw physicochemical test data of solid waste through a feature construction module. In this embodiment, the chemical composition of fly ash is: silicon dioxide 50.2%, aluminum oxide 26.5%, calcium oxide 4.8%, and magnesium oxide 1.5%; the chemical composition of blast furnace slag is: silicon dioxide 31.0%, aluminum oxide 14.2%, calcium oxide 38.5%, and magnesium oxide 8.5%. The feature construction module extracts the above mass fractions to calculate derived feature parameters. Taking blast furnace slag as an example, its alkalinity coefficient is calculated to be 1.04, and its mass coefficient is 1.99. The system combines the above parameters with physical morphology data such as specific surface area, performs range normalization processing, and generates a standardized solid waste physicochemical feature vector.

[0099] Initial inequality constraints are set for the building material mix design parameters: the lower limit for fly ash content is 10%, and the upper limit is 40%; the lower limit for blast furnace slag content is 20%, and the upper limit is 60%; the lower limit for water-cement ratio is 0.28, and the upper limit is 0.45. Equality constraints are set such that the total content of fly ash, slag, and silicate cement clinker is 100%.

[0100] Model optimization iteration: The mapping prediction module concatenates the solid waste physicochemical feature vector with a randomly initialized building material mix ratio parameter vector and inputs it into the multi-task learning network. In the initial iteration phase, the network's forward propagation output shows a predicted 28-day compressive strength of 36.5 MPa and a predicted life-cycle carbon emission of 285 kg / m³. 3 .

[0101] The quantitative evaluation module calculated the carbon efficiency index of this initial ratio to be 0.128 MPa / (kg / m³). 3 The carbon efficiency threshold is lower than the preset target carbon efficiency threshold of 0.150 MPa / (kg / m³). 3 The relative carbon efficiency ratio is calculated to be 0.85. Meanwhile, the initial solid waste utilization rate is 45%. The quantitative evaluation module calculates the comprehensive quantitative evaluation score with a carbon efficiency weight of 0.6 and a solid waste utilization weight of 0.4.

[0102] Since the relative carbon efficiency ratio is less than 1, the loss reconstruction module calculates a dynamic penalty weight of 1.15 and amplifies the carbon emission prediction loss during the backpropagation of the current network error, causing the gradient update to tilt towards the low-carbon direction. The collaborative optimization module uses the chain rule of the computation graph to calculate the gradient vector for the building material proportion parameters and updates the admixture and water-cement ratio values ​​with an iteration step size of 0.01. During the iteration process, the scheme output module performs a total absolute volume check on the proportion parameters after each update. In the 45th iteration, the total absolute volume calculated due to the combination of water-cement ratio and cementitious material admixture reaches 1.025 cubic meters, exceeding the preset tolerance range of [0.98, 1.02]. The scheme output module triggers a volume value penalty term and adds it to the optimization objective function, forcing the system to revert the water-cement ratio parameter in the 46th iteration.

[0103] Optimal solution output: When the iteration reaches the 182nd iteration, the absolute value of the change in the comprehensive quantitative evaluation score between two adjacent iterations drops below 0.0001, and the gradient L2 norm meets the convergence threshold. The system stops iterating and outputs the globally optimal building material mix ratio parameter vector. The corresponding converted unit mix ratio is: fly ash content 28%, blast furnace slag content 45%, silicate cement clinker content 27%, water-cement ratio 0.33, and chemical activator content 1.5%.

[0104] At this point, the predicted 28-day compressive strength is 42.6 MPa, and the predicted life-cycle carbon emissions have decreased to 198 kg / m³. 3 The solid waste utilization rate reached 73%, and the total absolute volume was verified to be 1.005 cubic meters, meeting all engineering constraints.

[0105] Experimental verification and effect comparison: To verify the effectiveness of this system, three sets of comparative experiments were designed. The test objective for all comparative groups was set as the formulation of green building materials that meet the 28-day compressive strength requirement of 40 MPa, and the calculations were performed under a unified raw material database and carbon emission equivalent factor standard.

[0106] Comparison scheme settings: Control Group 1: The traditional manual trial mixing method was used. The testers consulted the specifications and conducted three rounds of physical sample mixing and strength testing to screen out the formula that met the strength requirements.

[0107] Control Group 2: A conventional single-objective machine learning optimization method was used. A single-objective random forest algorithm without dynamic carbon efficiency penalty mechanism and volume constraint was used, with maximizing compressive strength as the optimization objective for mix proportion prediction.

[0108] Experimental group: The intelligent proportioning optimization and evaluation system for green building materials based on solid waste resource utilization provided by this invention was adopted.

[0109] Comparative data analysis: After system operation and physical testing verification, the optimal ratio data of each group were extracted and compared. The results are shown in the table below.

[0110] Solution Group Optimization time / trial fitting cycle 28-day measured compressive strength (MPa) Solid waste utilization rate (%) <![CDATA[Carbon emission accounting value (kg / m 3 )]]> Absolute volume deviation Control group 1 35 days (including maintenance) 41.5 45.0 295 qualified Control group 2 0.5 hours 45.2 30.0 330 Out of tolerance (+3.5%) experimental group 0.8 hours 42.1 73.0 198 Qualified (+0.5%) Data analysis and accompanying graphs demonstrate that: See attached document Figure 3 , Figure 3 The iterative convergence characteristics of the system in the automated optimization process are demonstrated. The curves show that as the number of iterations increases, the comprehensive quantitative evaluation score (solid line on the left Y-axis), reflecting the overall quality of the ratio, rises rapidly in the early stage (approximately 0-50 iterations), accompanied by a significant and smooth decay of the gradient norm 2 (dashed line on the right Y-axis). Around the 180th iteration, the system's comprehensive evaluation score stabilizes in the high range around 0.95, while the gradient norm 2 approaches zero and no local oscillations or divergence occur. This directly verifies that the collaborative optimization strategy based on computational graph gradients adopted in this system can efficiently and stably converge to the global optimum.

[0111] Combined with appendix Figure 4 The multidimensional performance bar chart features allow for a direct comparison of the normalized performance of each group of schemes in four independent dimensions: mechanical properties, solid waste utilization rate, degree of low carbonization, and volume stability. Although control group 1 (light gray bar) ensured that the volume and strength were qualified (corresponding to the high volume stability and mechanical performance indicators in the figure), the trial mixing cycle was extremely long, the solid waste utilization rate remained at the conventional level, and the carbon emissions were high (corresponding to the lower solid waste utilization rate and degree of low carbonization scale in the figure).

[0112] The optimization algorithm for control group 2 (medium gray column) overemphasized the improvement of strength index (corresponding to a mechanical property score of 1.0 in the figure), causing the system to automatically reduce the solid waste content and increase the proportion of cement clinker, resulting in a surge in carbon emissions to 330 kg / m³. 3 (This corresponds to the extremely low solid waste utilization rate and low carbonization level shown in the figure). At the same time, due to the lack of closed-loop verification of absolute volume, the output ratio showed a volume deviation of 3.5% in actual engineering conversion, and could not be directly applied to the production of the mixing plant (corresponding to the volume stability index in the figure dropping sharply to 0.05).

[0113] The experimental group of this invention (dark gray columns), through the synergistic effect of a multi-task learning network and an adaptive total loss function, increased the solid waste utilization rate to 73% and reduced carbon emissions by approximately 32.8% compared to control group 1 (corresponding to the significant increase in the height of the columns representing solid waste utilization rate and low carbonization level in the figure). The volume penalty function mechanism ensured that the absolute volume deviation of the output ratio was controlled within 0.5% (corresponding to the excellent volume stability in the figure, on par with control group 1), demonstrating that this system, while ensuring engineering practicality, achieved automated comprehensive optimization of low carbon emissions and high solid waste disposal, presenting an optimal overall balance across four dimensions.

Claims

1. A green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization, characterized in that, include: The feature construction module is used to obtain the original physicochemical test data of solid waste raw materials and construct solid waste physicochemical feature vectors and building material ratio parameter vectors. The mapping prediction module is used to input the solid waste physicochemical feature vector and the building material ratio parameter vector into a multi-task learning network, and output the mechanical performance prediction value and the life cycle carbon emission prediction value. The quantitative evaluation module is used to calculate the carbon efficiency index and the comprehensive quantitative evaluation score based on the predicted mechanical performance value and the predicted carbon emissions over the entire life cycle. The loss reconstruction module is used to calculate dynamic penalty weights based on the carbon efficiency index, reconstruct the adaptive total loss function using the dynamic penalty weights, and train the multi-task learning network. The collaborative optimization module is used to update the building material ratio parameter vector based on the comprehensive quantitative evaluation score, and feed the updated building material ratio parameter vector back to the mapping prediction module for a new round of calculation. The solution output module is used to stop the new round of calculation when the comprehensive quantitative evaluation score meets the set convergence condition, and output the globally optimal building material ratio parameter vector and result dataset.

2. The green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization according to claim 1, characterized in that, The feature construction module is specifically used for: Obtain chemical composition and physical morphology data from the original physicochemical test data of solid waste raw materials; Intrinsic feature parameters are extracted from the chemical composition data and the physical morphology data to form an initial solid waste physicochemical feature vector; The derived feature parameters are calculated using the intrinsic feature parameters, and then appended to the initial solid waste physicochemical feature vector for range normalization, outputting the standardized solid waste physicochemical feature vector.

3. The intelligent proportioning optimization and evaluation system for green building materials based on solid waste resource utilization according to claim 1, characterized in that, The feature construction module is also used for: Extract the building material proportioning parameters to form the building material proportioning parameter vector; The building material mixing parameters include the dosage of various solid wastes, the dosage of silicate cement clinker, the water-cement ratio, and the dosage of chemical activators. The feature construction module defines hard constraint boundaries for the building material proportioning parameter vector; The hard constraint boundary includes equality constraints and inequality constraints; The equation constraint limits the sum of the mass percentages of all dry powder material parameters in the building material proportioning parameter vector; The inequality constraint limits the lower and upper limits of the process values ​​for each dimension of the building material proportioning parameter vector.

4. The intelligent proportioning optimization and evaluation system for green building materials based on solid waste resource utilization according to claim 1, characterized in that, The mapping prediction module is specifically used for: The solid waste physicochemical feature vector and the building material proportioning parameter vector are concatenated along the dimensions to generate a multi-source feature fusion vector. The multi-source feature fusion vector is input into the shared hidden layer inside the multi-task learning network for nonlinear feature extraction, and the hidden layer feature vector is output. At the end of the shared hidden layer, a mechanical performance prediction branch and a carbon emission prediction branch are connected; The hidden layer feature vectors are input into the mechanical performance prediction branch and the carbon emission prediction branch respectively, and the mechanical performance prediction value and the life cycle carbon emission prediction value are output in parallel.

5. The green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization according to claim 1, characterized in that, The quantitative evaluation module is specifically used for: The predicted mechanical properties are linearly weighted and summed to generate an equivalent mechanical property scalar. The carbon efficiency index is constructed by performing a division operation on the equivalent mechanical performance scalar and the predicted carbon emissions over the entire life cycle. Extract the amount of various solid wastes from the building material proportioning parameter vector and sum them to generate the solid waste utilization rate; The carbon efficiency index and the preset target carbon efficiency threshold are subjected to dimensionless processing to generate a relative carbon efficiency ratio. The relative carbon efficiency ratio and the solid waste utilization rate are weighted and summed to calculate the comprehensive quantitative evaluation score.

6. The green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization according to claim 5, characterized in that, The loss reconstruction module is specifically used for: Determine whether the relative carbon efficiency ratio is less than one; When the relative carbon efficiency ratio is determined to be less than one, a dynamic penalty weight with a value greater than one is generated based on the degree to which the relative carbon efficiency ratio deviates from one. Obtain accurate label values ​​for mechanical properties and carbon emissions; Calculate the mechanical performance prediction loss based on the actual labeled mechanical performance value and the predicted mechanical performance value; Calculate the life-cycle carbon emission prediction loss based on the actual carbon emission label value and the life-cycle carbon emission prediction value; The basic multi-objective collaborative loss function is constructed by weighted summation of the mechanical performance prediction loss and the full life cycle carbon emission prediction loss. The dynamic penalty weights are multiplied and fused with the basic multi-objective collaborative loss function to reconstruct the adaptive total loss function, and the multi-task learning network is trained using the adaptive total loss function.

7. The green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization according to claim 1, characterized in that, The collaborative optimization module is specifically used for: Freeze the network layer weight parameters and bias vectors in the multi-task learning network; The comprehensive quantitative evaluation score is converted into a negative value to construct an optimization objective function; The chain rule of the network computation graph is used to calculate the gradient vector of the building material ratio parameter with respect to the building material ratio parameter vector of the optimization objective function; The building material proportioning parameter gradient vector and the preset optimization iteration step size are used to perform numerical update calculations on the building material proportioning parameter vector to obtain the updated building material proportioning parameter vector; The updated building material ratio parameter vector is fed back to the mapping prediction module for a new round of calculation.

8. The green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization according to claim 7, characterized in that, The solution output module is specifically used for: Extract the proportions of each raw material from the updated building material ratio parameter vector, and convert them into the mass of each raw material by combining them with the preset total mass of building materials per cubic meter. The absolute volume is calculated by converting the mass of the individual raw material and the corresponding apparent density of the material. The total absolute volume is generated by adding the absolute volume of all raw materials to the preset volume of building material gas content. Determine whether the total absolute volume is within a preset volume tolerance range; When it is determined that the total absolute volume is not within the preset volume tolerance range, a volume value penalty item is generated; The volume value penalty term is superimposed on the optimization objective function, and the update direction of subsequent calculation rounds is adjusted through error backpropagation. 9.The green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization according to claim 1, characterized in that, The solution output module is also used for: Extract the state parameters from two adjacent iterations, and calculate the absolute value of the score change of the comprehensive quantitative evaluation score in two adjacent calculation rounds; Calculate the gradient norm 2 of the building material proportioning parameter vector; When the absolute value of the score change is less than a preset score convergence threshold and the gradient norm is less than a preset gradient convergence threshold, the comprehensive quantitative evaluation score is determined to meet the set convergence condition. 10.The green building material intelligent proportioning optimization and evaluation system based on solid waste resource utilization according to claim 1, characterized in that, The solution output module is also used for: Generate the result dataset, which includes a green building materials optimal ratio decision panel data package; The solution output module transforms the original predicted indicators associated with the globally optimal building material ratio parameter vector into normalized evaluation values ​​with multiple independent dimensions. A multidimensional performance radar chart feature is constructed using normalized evaluation values ​​from multiple independent dimensions; Calculate the coverage area scalar of the multidimensional performance radar map features; The coverage area scalar, the list of raw material usage per unit volume, and various original prediction indicators are encapsulated according to key-value pairs, and header information containing project batch identifiers and timestamps is attached to generate the final green building materials optimal ratio decision panel data package.

Images