Industrial solid waste cementing material ratio intelligent optimization and soil solidification method based on neural network

Abstract

The invention relates to the crossing field of building materials and artificial intelligence, in particular to an industrial solid waste cementing material proportion intelligent optimization and soil solidification method based on a neural network, comprising the following steps: S1, constructing a data set for forming a standardized database; s2, establishing an artificial neural network model of a multi-layer perceptron structure, including an input layer, a hidden layer and an output layer, adding an L2 regular term and an early-stop mechanism in a training process, and performing collaborative optimization by adopting a combined loss function; s3, executing arrangement importance analysis, disorganizing input parameters one by one, and quantifying the influence of each parameter on a model prediction error; s4, matching intelligent recommendation, including forward prediction and reverse optimization; and S5, performing closed-loop verification. A multi-layer perceptron neural network structure is combined with feature sensitivity analysis, the nonlinear prediction precision is remarkably improved, forward and reverse double-path optimization is adopted to replace manual trial and error, the design period is shortened, a closed-loop verification mechanism ensures the engineering reliability and reduces errors, and solid waste replaces a high-carbon material to reduce carbon emission.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building materials and artificial intelligence, in particular to an intelligent optimization method for industrial solid waste cementitious material proportioning based on a neural network and soil solidification. BACKGROUND

[0002] Currently, in the process of building foundation engineering, underground space development and municipal road construction, there is a common problem of handling a large amount of construction waste. At the same time, the output of industrial by-products such as steel slag, slag, and desulfurization gypsum is increasing year by year, and its resource utilization is limited by large fluctuations in physical and chemical properties and imperfect processing technology. At present, traditional solidification agents such as cement or lime are commonly used in engineering to reinforce and treat soil to improve its bearing capacity and durability. However, this method has the following main shortcomings:

[0003] (1) The proportioning design is highly dependent on the manual trial-and-error mode of "test-feedback-adjustment", lacks the ability of rapid prediction or reverse design based on performance targets, has a long design cycle and low efficiency, and is difficult to meet the requirements of rapid construction and diverse adaptability in actual engineering;

[0004] (2) The existing methods are mostly based on traditional modeling methods such as linear regression and grey correlation, which are difficult to represent the highly nonlinear and multivariate coupled mechanical evolution process in the solid waste cementitious material system, have large model prediction errors, and are difficult to realize reliable popularization and application;

[0005] (3) Although some researches in recent years have tried to introduce machine learning models for performance prediction, they mostly stay at the level of "predicting strength from proportioning", and have not formed a reverse optimization design path of "reversing proportioning structure from performance target", and most of the methods lack a closed-loop mechanism combined with engineering verification, which cannot effectively connect the model prediction to the actual performance confirmation;

[0006] (4) Poor solidification adaptability, for soil with high liquid limit, high water content or loose structure, the solidification effect is easily disturbed, and problems such as insufficient strength, dry shrinkage cracks and unstable structure often occur;

[0007] (5) Heavy environmental burden, the solidification agent itself is a high-energy consumption and high-carbon emission product. Especially under the background of the strategic goal of carbon peak and carbon neutralization, traditional high-carbon materials are not conducive to the green transformation of engineering.

[0008] Therefore, there is still a technical gap between the collaborative utilization of solid waste resources and the design of soil solidification, and an intelligent optimization method that is multivariate, multi-soil, multi-component, and interpretable is needed to establish a full-process closed-loop technology system from material parameter input to proportioning output, from predicted performance to structure verification. SUMMARY

[0009] The purpose of the present application is to provide a neural network-based industrial solid waste cementitious material proportioning intelligent optimization and soil solidification method to solve the problems raised in the background art.

[0010] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0011] A neural network-based industrial solid waste cementitious material proportioning intelligent optimization and soil solidification method, comprising the following steps:

[0012] S1. Construct a data set for forming a standardized database to provide a training basis for an artificial neural network model, the data set including input parameters and output parameters, the input parameters including cementitious material components and soil parameters, and the output parameters being unconfined compressive strength;

[0013] S2. Establish a multi-layer perceptron structure artificial neural network model, the artificial neural network model including an input layer, a hidden layer and an output layer, an L2 regularization term and an early-stopping mechanism being added in the training process to avoid overfitting, and a combined loss function being used for collaborative optimization;

[0014] S3. Perform permutation importance analysis by disturbing the input parameters one by one to quantify the influence of each parameter on the model prediction error, and generate a parameter importance ranking;

[0015] S4. Intelligent recommendation of proportioning, including forward prediction and reverse optimization;

[0016] S5. Closed-loop verification.

[0017] Further, the cementitious material components in S1 include steel slag SS, slag GGBS, desulfurization gypsum DG, cement OPC, fly ash FA and cementitious material content BD, the soil parameters include water-soil ratio W / S, soil water content WC, plastic limit PL and liquid limit LL, and the unconfined compressive strength includes 7-day unconfined compressive strength UCS-7 and 28-day unconfined compressive strength UCS-28.

[0018] Further, the input layer in S2 is 10 nodes, the hidden layer is 5 layers of 32 nodes, and the output layer is 2 nodes.

[0019] Further, the model structure in S2 is expressed as follows:

[0020]

[0021] h (1) =σ(W (1) X+b (1) )

[0022] h (2) =σ(W (2)X + b (2) )

[0023] h (3) = σ(W (3) X + b (3) )

[0024] h (4) = σ(W (4) X + b (4) )

[0025] h (5) = σ(W (5) X + b (5) )

[0026]

[0027] wherein, is an n-dimensional real number vector space, indicates that the input is a 10-dimensional real number vector, indicates that the output is a 2-dimensional real number vector; σ is an activation function; W (l) , b (l) are the weight matrix and the bias vector of the lth layer, respectively; is all the trainable parameters of the model.

[0028] Further, the combined loss function in S2 is:

[0029] L total = α·MSE + β·MAE + γ·SmoothL1;

[0030] wherein:

[0031]

[0032]

[0033] α, β, γ are adjustable weight coefficients, yi is the true value of the ith sample, is the model prediction value of the ith sample, n is the total number of samples, x represents the prediction error of each sample.

[0034] Further, the coefficient of determination R 2 is used to evaluate the precision of the model in S2.

[0035]

[0036] wherein: represents the average value of the true value, represents the sum of squared residuals, represents the total sum of squares.

[0037] Further, the model optimizer in S2 selects Adam optimizer, and the maximum number of iterations is 5000 rounds.

[0038] Further, in S4, the forward prediction is inputting cementitious material components and soil parameters, and outputting the predicted value of unconfined compressive strength, and combining the raw material proportion boundary, total admixture range and construction adaptability requirements, the existing proportioning library samples are screened to recommend the optimized proportioning path meeting the performance target and engineering requirements.

[0039] Further, in S4, the reverse optimization is inputting soil parameter target indicators and required unconfined compressive strength, and automatically generating multiple groups of cementitious material component candidate proportioning structures meeting the constraint adjustment.

[0040] Further, the S5 closed-loop verification includes test verification of the recommended cementitious material component proportioning, comparison of the measured unconfined compressive strength with the predicted value and feedback to the artificial neural network model for retraining.

[0041] Compared with the prior art, the beneficial effects of the present application are:

[0042] (1) The industrial solid waste cementitious material proportioning intelligent optimization and soil solidification method based on neural network establishes a nonlinear mapping model based on artificial neural network, realizes automatic recommendation of solid waste cementitious material proportioning through performance index target input, avoids the traditional low-efficiency design path of relying on multiple rounds of "test-feedback-adjustment", shortens the proportioning design cycle, and improves the response speed and engineering efficiency.

[0043] (2) Compared with traditional linear regression, gray analysis and other methods, the present application adopts a multilayer perceptron neural network structure, combines multiple loss function optimization strategies and feature sensitivity analysis, significantly enhances the fitting ability of the model to nonlinear coupling relationship, realizes high-precision prediction and has good generalization ability.

[0044] (3) The present application introduces proportioning screening, test verification and retraining mechanism on the basis of model prediction, constructs a whole-process closed-loop system from data input, proportioning output to experimental feedback and learning, ensures the operability and stability of the model output proportioning in actual engineering, and improves the practicability and popularization potential of the method.

[0045] (4) By introducing liquid limit, plastic limit, moisture content and other original soil indicators as model input variables, a cementitious material proportioning scheme suitable for different foundation soils (especially high liquid limit, high moisture content and loose structure soil) is constructed, which effectively improves the mechanical properties and volume stability of the solidified body, and reduces the risk of dry shrinkage crack and structure instability.

[0046] (5) The invention makes full use of industrial by-products such as steel slag, slag, desulfurization gypsum, etc. as the main cementitious material, and realizes partial or complete replacement of high-carbon materials such as cement and lime through proportioning optimization, significantly reducing carbon dioxide emissions and energy consumption, and meeting the development trend of green building materials under the background of carbon peak and carbon neutralization. BRIEF DESCRIPTION OF DRAWINGS

[0047] Fig. 1 The flowchart of the present application is shown in the figure;

[0048] Fig. 2 The neural network model structure of the multi-layer perceptron structure of the present application is shown in the figure;

[0049] Fig. 3 The model precision evaluation is shown in the figure. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0051] Please refer to Figs. 1-3 The present application provides an embodiment: a neural network-based industrial solid waste cementitious material proportioning intelligent optimization and soil solidification method, comprising the following steps:

[0052] S1. Construct a data set for forming a standardized database, adopt a combination of "experimental data + literature data", and construct a data set for model training and verification, the data set includes input parameters and output parameters, the input parameters include cementitious material components and soil parameters, the cementitious material components include steel slag SS, slag GGBS, desulfurization gypsum DG, cement OPC, fly ash FA, and cementitious material dosage BD, the soil parameters include water-soil ratio W / S, soil water content WC, plastic limit PL and liquid limit LL, and the output parameters are unconfined compressive strength, 7-day unconfined compressive strength UCS-7 and 28-day unconfined compressive strength UCS-28, all original data are processed by Min-Max normalization before inputting into the model to form a standardized database, and provide a training basis for an artificial neural network model;

[0053] S2. Establish an artificial neural network model of multi-layer perceptron structure (MLP), the artificial neural network model (ANN model) includes 10-node input layer, 5-layer 32-node hidden layer and 2-node output layer, in order to ensure the stability and generalization ability of the model, L2 regular term and early-stopping mechanism are added in the training process to avoid overfitting, the model optimizer selects Adam optimizer, self-learning rate adjustment, the maximum number of iterations is 5000 rounds, at the same time, the combined loss function is used for collaborative optimization during the model training process, R 2 is selected to evaluate the model accuracy, provide intuitive accuracy explanation, and measure the explanation ability of the model to the variance of the target variable.

[0054] The model structure is expressed as follows:

[0055]

[0056] h (1) =σ(W (1) X+b (1) )

[0057] h (2) =σ(W (2) X+b (2) )

[0058] h (3) =σ(W (3) X+b (3) )

[0059] h (4) =σ(W (4) X+b (4) )

[0060] h (5) =σ(W (5) X+b (5) )

[0061]

[0062] Wherein, is an n-dimensional real number vector space, indicates that the input is a 10-dimensional real number vector, indicates that the output is a 2-dimensional real number vector; σ is an activation function; W (l) , b (l) are the weight matrix and bias vector of the lth layer respectively; is all the trainable parameters of the model.

[0063] The combined loss function is:

[0064] L total =α·MSE+β·MAe+γ·SmoothL1;

[0065]

[0066] In the formula, α, β, and γ are adjustable weight coefficients, and yi is the true value of the i-th sample. Let α be the model prediction value for the i-th sample, and n be the total number of samples. In this embodiment, α, β, and γ are set to 0.4, 0.3, and 0.3, respectively.

[0067] Wherein, MSE (mean squared error) is used to penalize samples with large biases, making the model focus more on prediction accuracy, and yi is the true value of the i-th sample. Here, n represents the model prediction value for the i-th sample, and n represents the total number of samples.

[0068] MAE (mean absolute error) reduces the interference of outliers on the model and improves the model's robustness.

[0069] SmoothL1 loss combines gradient continuity with outlier resistance, maintaining gradient stability in small error ranges and avoiding gradient explosion in large error ranges, thus balancing accuracy and stability.

[0070] The accuracy of the model is evaluated using the coefficient of determination R. 2 :

[0071]

[0072] In the formula: The average value representing the true value Represents the sum of squared residuals, This represents the total sum of squares.

[0073] After training, the model accuracy is as follows: Fig. 3 As shown, the final accuracy is summarized as follows: Average R 2 The (UCS7+UCS28) ratio is 0.9505, and the average (combined loss) is 0.0837, indicating that the model has high accuracy and reliability in predicting the compressive strength of solidified soil. This result further verifies the significant linear correlation between the input and output parameters, indicating that factors such as the content of different cementitious materials, the amount of cementitious materials used, the water-to-soil ratio, soil moisture content, soil plastic limit, and soil liquid limit have a significant impact on the unconfined compressive strength of solidified waste soil.

[0074] S3. Perform importance analysis by shuffling the input parameters one by one, quantifying the impact of each parameter on the model prediction error (such as MSE), and generating a parameter importance ranking. This process forms a parameter weight reference sequence, which helps to identify the main control factors and provides a basis for subsequent ratio adjustment and sensitivity analysis.

[0075] S4. Intelligent ratio recommendation, including forward prediction and reverse optimization;

[0076] In S4, the forward prediction is inputting the cementitious material components and soil parameters, and outputting the UCS-7 and UCS-28 prediction values, and combining the raw material proportion boundary, total admixture range and construction adaptability requirements, the existing proportion library samples are screened, and the optimized proportion path meeting the performance target and engineering requirements is recommended.

[0077] In S4, the reverse optimization is inputting the soil parameter target indicators (such as W / S, WC, PL, LL) and the required unconfined compressive strength (such as UCS-7≥2MPa), and the system can automatically generate multiple groups of cementitious material component candidate proportion structures meeting the constraint adjustment through the ANN model and high-throughput search strategy, which is suitable for subsequent sample preparation and performance verification.

[0078] S5. Closed loop verification, including test verification of the recommended cementitious material component proportion, comparison of the measured unconfined compressive strength with the predicted value and feedback to the artificial neural network model for retraining. Through this method, a closed loop optimization path of "model prediction-proportion generation-performance verification-result feedback" can be constructed, which has high efficiency, wide adaptability and good engineering implementability, and is especially suitable for rapid decision-making and green building material selection design under complex foundation conditions.

[0079] To further verify the applicability of the neural network proportion optimization method proposed in the application in actual engineering, three types of typical soil are selected as research objects, which are high liquid limit clay, medium sand and high water content mucky soil. After testing the basic physical indicators of each type of soil, the obtained soil indicators are used as model input parameters, the cementitious material proportion recommended by the method is used, and indoor strength test is carried out to verify the reliability and engineering adaptability of the recommended formula.

[0080] Example 1

[0081] To verify the applicability of the method in high plastic soft soil, a typical alluvial clay is selected as the research object. This kind of clay is grayish brown in color, mainly composed of clay particles and silt particles, with a liquid limit LL of 55.0%, a plastic limit PL of 23.0%, a soil water content WC of 36.0%, and characteristics of strong compressibility, large dry shrinkage and unstable structure.

[0082] According to the requirements of the site engineering, the target strength (UCS-7≥1.5MPa, UCS-28≥2.5MPa) is set, the water-soil ratio W / S is fixed at 0.12, and the above soil indicators are input into the artificial neural network model described in the application, and the system automatically generates three groups of candidate proportion paths meeting the performance constraints. The candidate proportions and their model predicted strengths are shown in the following table:

[0083]

[0084] The selected ratio 3 is comprehensively considered for experimental verification. The glue joint material according to the ratio 3 is prepared and cured with the original clay according to the test design to form a cubic test piece (50*50*50mm), which is cured for 7 days and 28 days under standard curing conditions (20±2℃, RH≥95%) and the unconfined compressive strength is tested. The measured results are UCS-7=1.62MPa, UCS-28=2.74MPa, and the deviations from the predicted values are 16.4%, 5.8% respectively, which are controlled within ±20%, indicating that the predicted values and the measured values have high consistency.

[0085] Example 2

[0086] In this embodiment, the typical building site backfilling medium sand is selected as the verification object. The particle size of this kind of soil body is concentrated between 0.25-0.5mm, the structure is loose, the liquid limit LL is 24.0%, the plastic limit PL is 13.5%, the soil water content WC is 16.0%, and the overall strength is low, which needs to be improved by cementation.

[0087] According to the target strength setting (UCS-7≥2.0MPa, UCS-28≥5.5MPa), the water-soil ratio W / S is fixed at 0.12, the soil body index is input into the artificial neural network model constructed by the application, and the system automatically generates multiple groups of ratio results. Three representative candidate ratios are as follows:

[0088]

[0089] Considering the selected ratio 1, the test is carried out. The recommended ratio is used to prepare the test sample, and the strength is tested after standard curing. The measured UCS-7 is 3.85MPa, and the UCS-28 is 5.68MPa, which deviates from the predicted value by 9.6%, 3.9% respectively, indicating that the model prediction accuracy is good.

[0090] Example 3

[0091] In this embodiment, the high water content sludge in urban foundation pit is selected as the object, which is in a flow-plastic state, contains organic matter and fine particle construction waste, the liquid limit LL is 42.8%, the plastic limit PL is 18.4%, and the soil water content WC is as high as 68.3%. This kind of soil body has high fluidity, low strength and difficult construction.

[0092] According to the engineering requirement, the target strength (UCS-7≥1.2MPa, UCS-28≥3.5MPa) is set, the water-soil ratio W / S is fixed at 0.12, the soil body index is input into the neural network model for automatic ratio recommendation, and the system generates the following 3 groups of candidate schemes that meet the performance constraints:

[0093]

[0094] The selected ratio 2 is used for subsequent sample preparation and verification. After curing, the test piece is detected, the 7-day strength is 2.19 MPa, the 28-day strength is 3.64 MPa, the prediction error is 7.2% and 3.1% respectively, and both are within the engineering tolerance range.

[0095] The beneficial effects of the present application include:

[0096] 1. The present application takes the proportion of main solid waste components such as slag, steel slag and desulfurization gypsum as input parameters, and also integrates soil moisture content WC, water-soil ratio W / S, plastic limit PL, liquid limit LL and other soil factors representing the original state of foundation, to form a high-dimensional feature space containing 10 input dimensions. At the same time, the model output includes two important indicators of 7-day and 28-day unconfined compressive strength, realizing the direct mapping from the conventional input parameters to the solidification effect.

[0097] 2. The present application innovatively introduces the permutation importance analysis method to sort and identify the influence of each input variable on the model prediction output, helping users understand the model behavior and assisting actual proportioning decision-making, and providing a clear direction for subsequent design optimization.

[0098] 3. On the basis of model prediction, by limiting the proportioning boundary and total amount range of materials such as steel slag, slag and gypsum, a number of recommended proportioning paths with engineering feasibility are screened out, reducing the workload of blind test and repeated adjustment.

[0099] 4. The present application verifies the actual mechanical properties of the model recommended proportioning through experiments, compares the model output with the experimental results, and builds a simple closed-loop mechanism of prediction-verification-feedback, improving the usability of the model in engineering practice.

[0100] 5. The method does not depend on complex material analysis means, the input information is easy to obtain, and is suitable for soil solidification research and engineering site material design work under general laboratory conditions, and has good adaptability and practical popularization prospect.

[0101] It is apparent to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, but can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered exemplary and non-limiting, and the scope of the present application is defined by the appended claims rather than the above description, and all changes falling within the meaning and range of equivalent elements of the claims are intended to be encompassed by the present application. Any reference signs in the claims should not be considered as limiting the claims involved.

Claims

1. A method for intelligent optimization of industrial solid waste cementitious material proportioning and soil solidification based on neural networks, characterized in that, Includes the following steps: S1. Construct a dataset to form a standardized database, providing a training basis for artificial neural network models. The dataset includes input parameters and output parameters. The input parameters include cementitious material components and soil parameters. The output parameter is unconfined compressive strength. S2. Establish an artificial neural network model with a multilayer perceptron structure, the artificial neural network model including an input layer, a hidden layer and an output layer, and add an L2 regularization term and an early-stopping mechanism during training to avoid overfitting, and use a combined loss function for collaborative optimization; S3. Perform a ranking importance analysis. By shuffling the input parameters one by one, quantify the impact of each parameter on the model prediction error and generate a parameter importance ranking. S4. Intelligent ratio recommendation, including forward prediction and reverse optimization; S5. Closed-loop verification.

2. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 1, characterized in that, The cementitious material components in S1 include steel slag SS, blast furnace slag GGBS, desulfurized gypsum DG, cement OPC, fly ash FA, and cementitious material admixture BD. The soil parameters include soil-water ratio W / S, soil moisture content WC, plastic limit PL, and liquid limit LL. The unconfined compressive strength includes 7-day unconfined compressive strength UCS-7 and 28-day unconfined compressive strength UCS-28.

3. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 2, characterized in that, In S2, the input layer has 10 nodes, the hidden layer has 5 layers and 32 nodes, and the output layer has 2 nodes.

4. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 3, characterized in that, The model structure in S2 is expressed as follows: h (1) =σ(W (1) X+b (1) ) h (2) =σ(W (2) X+b (2) ) h (3) =σ(W (3) X+b (3) ) h (4) =σ(W (4) X+b (4) ) h (5) =σ(W (5) X+b (5) ) in, Let n be a real vector space. The input is a 10-dimensional real vector. The output is a 2D real vector; σ is the activation function; W (l) b (l) These are the weight matrix and bias vector of the l-th layer, respectively; These are all the trainable parameters of the model.

5. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 1, characterized in that, The combined loss function in S2 is: L total = α·MSE+β·MAE+γ·SmoothL1; in: α, β, and γ are adjustable weight coefficients, and yi is the true value of the i-th sample. Let x be the model prediction value for the i-th sample, n be the total number of samples, and x represent the prediction error for each sample.

6. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 5, characterized in that, The accuracy of the evaluation model in S2 is assessed using the coefficient of determination R. 2 : In the formula: The average value representing the true value Represents the sum of squared residuals, This represents the total sum of squares.

7. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 1, characterized in that, The model optimizer used in S2 is the Adam optimizer, with a maximum of 5000 iterations.

8. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 1, characterized in that, In S4, the positive prediction involves inputting cementitious material components and soil parameters, and then outputting the predicted value of unconfined compressive strength. At the same time, it combines the raw material proportion boundaries, total admixture range, and construction adaptability requirements to screen existing mix design samples and recommend an optimized mix design path that meets performance objectives and engineering requirements.

9. The method for intelligent optimization of industrial solid waste cementitious material ratio and soil solidification based on neural networks according to claim 1, characterized in that, In S4, the reverse optimization involves inputting the target soil parameters and the required unconfined compressive strength, and then automatically generating multiple sets of candidate mix proportions of cementitious material components that meet the constraint adjustment.

10. The method for intelligent optimization of industrial solid waste cementitious material proportioning and soil solidification based on neural networks according to claim 9, characterized in that, The S5 closed-loop verification includes experimental verification of the recommended cementitious material composition ratio, comparing the measured unconfined compressive strength with the predicted value and feeding it back to the artificial neural network model for retraining.

Images