Inplanatable tunnel water gushing grouting material component optimization method and system

Abstract

The invention provides an interpretable tunnel water gushing grouting material component optimization method and system, and relates to the technical field of tunnel grouting material optimizing.The method comprises the steps that engineering condition data of actual tunnel grouting are obtained and preprocessed; inputting the preprocessed engineering condition data into the integrated model, and outputting the predicted target material performance by extracting the association relationship between the engineering condition and the material performance; inputting the target material performance into the self-adaptive generative adversarial network model, firstly generating material components through a generator in combination with random noise, then inputting the generated material components and the target material performance into a discriminator, outputting a probability value of whether the material components conform to the target material performance or not, and introducing an interpretability method to determine whether the material components conform to the target material performance or not. And taking the output of the discriminator as an explanation target, quantifying the influence of the input material component on the output judgment probability of the discriminator, dynamically adjusting the key feature weight of the discriminator, and carrying out reverse prediction to obtain the optimal material component.

Description

Technical Field

[0001] The present disclosure relates to the technical field of tunnel grouting material optimization, and particularly to an interpretable optimization method and system for the components of tunnel water inrush grouting materials. Background Technique

[0002] The statements in this part merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] In tunnel engineering construction, the water inrush problem is one of the common geological disasters. Especially in areas with complex geological conditions and abundant groundwater, water inrush will not only damage the tunnel structure but also seriously affect the construction safety and project progress. To address this problem, grouting treatment methods are usually adopted, and specific materials are injected to reinforce the formation and block the water source, thereby achieving the purpose of stabilizing the tunnel.

[0004] Traditional grouting design methods usually rely on construction experience, laboratory test results, and engineers' subjective judgments. Although the above methods can solve the tunnel water inrush disaster to a certain extent, in the performance design stage of grouting materials, the existing design methods mostly rely on empirical judgments and cannot systematically and scientifically consider the complex interactions between multiple engineering conditions, nor fully consider the influence of on-site actual conditions on the grouting effect; in the material research and development design stage, a large number of cumbersome manual tests are usually required, which not only have a long cycle and high cost but also are difficult to achieve intelligent and efficient optimization of material components.

[0005] In recent years, with the development of machine learning and artificial intelligence technologies, data-driven optimization methods have gradually become an important way to solve this problem. In the performance design stage of grouting materials, machine learning models can deeply mine and analyze historical data to accurately predict the target performance of grouting materials; in the material research and development design stage, using machine learning models to train laboratory test data can achieve intelligent design of material components for the target performance.

[0006] Machine learning has great potential in the optimization design of grouting materials. Commonly used existing machine learning algorithms include: regression models, support vector machines, random forests, artificial neural networks, etc.; however, these algorithms all have certain limitations, resulting in their inability to be widely applied in actual engineering design. For example, regression models are difficult to capture the complex non-linear relationship between the performance of cement-based materials and their components, are prone to overfitting, underfitting, and multicollinearity problems, and are extremely sensitive to outliers; support vector machines have weak multi-objective optimization capabilities and are difficult to consider multiple material performance indicators simultaneously; random forest models are difficult to capture high-dimensional complex relationships, have poor extrapolation capabilities, and cannot generate new material design schemes beyond the training data distribution, with limited generalization capabilities; the training of artificial neural networks depends on a large amount of high-quality data and also has the problem of poor extrapolation capabilities. Summary of the Invention

[0007] To solve the above problems, the present disclosure proposes an interpretable optimization method and system for the components of tunnel water inrush grouting materials. The components of the grouting materials are optimized in a data-driven manner. A machine learning inverse model for reverse inferring the material components based on the material properties is constructed using a generative adversarial network, and an interpretability method is introduced for optimization to achieve the adaptive adjustment of the generative adversarial network. The complex non-linear mapping relationship between the material components and their properties is approximated through adversarial learning, effectively solving the problems of poor interpretability and insufficient generalization ability of traditional machine learning methods, and having a higher degree of automation and reliability.

[0008] According to some embodiments, the present disclosure adopts the following technical solutions:

[0009] An interpretable optimization method for the components of tunnel water inrush grouting materials, comprising:

[0010] Obtaining and preprocessing the engineering condition data of actual tunnel grouting;

[0011] Inputting the preprocessed engineering condition data into an integrated model, where the integrated model is a weighted average fusion of a statistical basic model and a machine learning basic model using a fusion strategy, and by extracting the correlation relationship between the engineering conditions and the material properties, outputting the predicted target material properties;

[0012] Inputting the target material properties into an adaptive generative adversarial network model. First, the generator of the adaptive generative adversarial network model combines random noise to generate material components, and then the generated material components and the target material properties are input into the discriminator of the adaptive generative adversarial network model, outputting the probability value of whether the material components meet the target material properties. An interpretability method is introduced, using the output of the discriminator as the interpretation target, quantifying the influence of the input material components on the output judgment probability of the discriminator, and dynamically adjusting the key feature weights of the discriminator to optimize the input material components, that is, reverse predicting to obtain the optimal material components.

[0013] According to some embodiments, the present disclosure adopts the following technical solutions:

[0014] An interpretable optimization system for the components of tunnel water inrush grouting materials, comprising:

[0015] A data acquisition module, configured to obtain and preprocess the engineering condition data of actual tunnel grouting;

[0016] A prediction target acquisition module is configured to input the preprocessed engineering condition data into an integration model. The integration model is a weighted average fusion of a statistical basic model and a machine learning basic model by adopting a fusion strategy. By extracting the correlation relationship between engineering conditions and material properties, it outputs the predicted target material properties.

[0017] A material component reverse prediction module is configured to input the target material properties into an adaptive generative adversarial network model. First, the generator of the adaptive generative adversarial network model combines random noise to generate material components. Then, the generated material components and the target material properties are input into the discriminator of the adaptive generative adversarial network model, and it outputs the probability value indicating whether the material components conform to the target material properties. An interpretability method is introduced, and the output of the discriminator is used as the interpretation target to quantify the influence of the input material components on the judgment probability output by the discriminator. By dynamically adjusting the key feature weights of the discriminator, the input material components are optimized, that is, the optimal material components are obtained through reverse prediction.

[0018] According to some embodiments, the present disclosure adopts the following technical solutions:

[0019] A computer program product includes a computer program, and when the computer program is executed by a processor, it implements the described interpretable optimization method for the grouting material components of tunnel water inrush.

[0020] According to some embodiments, the present disclosure adopts the following technical solutions:

[0021] A non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the described interpretable optimization method for the grouting material components of tunnel water inrush is implemented.

[0022] According to some embodiments, the present disclosure adopts the following technical solutions:

[0023] An electronic device includes: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory so that the electronic device implements the described interpretable optimization method for the grouting material components of tunnel water inrush.

[0024] Compared with the prior art, the beneficial effects of the present disclosure are:

[0025] An interpretable optimization method for the components of tunnel water inrush grouting materials of the present disclosure constructs a database of tunnel water inrush grouting treatment materials by using natural language processing technology, makes full use of historical engineering data, constructs an integrated model by integrating a statistical model and a machine learning model, captures complex non-linear relationships by leveraging the advantages of both models, reduces the bias and variance problems that may occur in a single model, and improves the generalization ability and stability of the model.

[0026] An interpretable optimization method for the components of tunnel water inrush grouting materials of the present disclosure optimizes the components of grouting materials in a data-driven manner, constructs a machine learning inverse model for reverse inferring the material components based on the material properties by using a generative adversarial network, and realizes the adaptive adjustment of the discriminator with the aid of a SHAP interpreter. Through the way of adversarial training, it learns and approximates the complex non-linear mapping relationship between the material components and their properties, solves the difficulties that may not be effectively described by traditional linear models or simple regression methods, and has a higher degree of automation and reliability.

[0027] An interpretable optimization method for the components of tunnel water inrush grouting materials of the present disclosure is a whole process from the engineering conditions of tunnel water inrush treatment to the design of the target performance of grouting materials, and then to the optimization design and interpretation of the components of grouting materials. Through intelligent and automated methods, it realizes the targeted design of grouting treatment materials, greatly improves the design efficiency and accuracy, provides an innovative solution for the material design of tunnel engineering grouting treatment, and promotes the intelligent development of engineering design. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of the disclosure. The schematic embodiments and descriptions thereof of the disclosure are used to explain the disclosure and do not constitute an improper limitation of the disclosure.

[0029] Figure 1 is the optimization design flow chart of the components of tunnel water inrush grouting materials in the embodiment of the present disclosure;

[0030] Figure 2 is the structure diagram of the adaptive generative adversarial network in the embodiment of the present disclosure;

[0031] Figure 3 is the schematic diagram of the optimization process of the adaptive generative adversarial network in the embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0032] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0033] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further descriptions of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0034] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they specify the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0035] Example 1

[0036] In an embodiment of the present disclosure, an interpretable optimization method for the components of a tunnel water inrush grouting material is provided, including:

[0037] Step 1: Obtain the engineering condition data of actual tunnel grouting and preprocess it;

[0038] Step 2: Input the preprocessed engineering condition data into an integrated model, where the integrated model is a weighted average fusion of a statistical basic model and a machine learning basic model using a fusion strategy. By extracting the correlation relationship between engineering conditions and material properties, the predicted target material properties are output;

[0039] Step 3: Input the target material properties into an adaptive generative adversarial network model. First, the generator of the adaptive generative adversarial network model combines random noise to generate material components, and then the generated material components and the target material properties are input into the discriminator of the adaptive generative adversarial network model to output the probability value of whether the material components meet the target material properties. An interpretability method is introduced, and the output of the discriminator is used as the interpretation target to quantify the influence of the input material components on the judgment probability output by the discriminator. By dynamically adjusting the key feature weights of the discriminator, the input material components are optimized, that is, the optimal material components are obtained by reverse prediction.

[0040] As an embodiment, an interpretable optimization method for the components of a tunnel water inrush grouting material according to the present disclosure is to realize the autonomous regulation of the performance of the grouting treatment material, clarify the complex effects and mutual influence relationships of each component of the material with the help of intelligent algorithms, and provide an intelligent means for the research and development of the grouting treatment material. The specific implementation process is as follows:

[0041] Step 1: Obtain historical grouting treatment material data, construct a tunnel water inrush grouting treatment material database, and preprocess the data to divide the data set;

[0042] 1) The acquisition of historical grouting treatment material data includes: collecting the engineering conditions and the corresponding material performance parameters of grouting treatment from literature, engineering case materials or on-site materials;

[0043] Among them, the engineering conditions include: formation type, water inflow, and tunnel depth of burial; the performance parameters of the grouting material include: material type, initial setting time, 3-day compressive strength, and viscosity.

[0044] 2) Apply natural language processing techniques to historical materials, including data collection and preprocessing, key information extraction, and then establish a database.

[0045] Specifically, data collection and preprocessing include: collecting literature, engineering case reports, or on-site records, which can be in formats such as PDF or Word documents, web pages, database records, etc. Preprocess the collected data, including converting non-text content (such as PDFs, charts) into text format, removing elements without text information, and cleaning irrelevant information (such as headers, footers, references, etc.) to ensure the simplicity and accuracy of the data for subsequent processing.

[0046] Key information extraction includes named entity recognition (NER) and relation extraction. This disclosure uses NLP to extract a structured database for subsequent machine learning training, and does not involve complex correlation analysis, so only named entity recognition is required.

[0047] Furthermore, named entity recognition (NER) refers to identifying entities with specific meanings in the text.

[0048] The specific steps include: annotating the dataset, model training, automatic recognition, and evaluation and optimization. The process is as follows:

[0049] 1. Annotating the dataset: Label the engineering conditions and material properties included in the text in the preprocessed dataset as specific entity categories. Specifically, set the engineering conditions: formation type, water inflow, tunnel depth of burial, and the performance parameters of the grouting treatment material: material type, initial setting time, 3-day compressive strength, and viscosity as categories, and set their corresponding values as entities.

[0050] When manually annotating, label each entity with its corresponding category to construct an entity annotation dataset.

[0051] 2. Model training: Select a suitable NER model. Optional models include BERT, RoBERTa, SpaCy, etc. Use the entity annotation dataset to train the model so that the model learns the characteristics of entities in the literature, engineering case materials, and on-site materials to identify entities of the same or similar types in unannotated data.

[0052] 3. Automatic recognition: The trained NER model can automatically recognize different entities in new text data.

[0053] 4. Evaluation and Optimization: Evaluate the performance of the NER model using metrics such as accuracy, recall, and F1-score. Improve the model's accuracy by adjusting model parameters or increasing training data. The recognition ability of the model can also be optimized by introducing a domain vocabulary library.

[0054] 3) Establish a database, structure the recognized information, and store it in the form of a table to associate different engineering conditions and material properties.

[0055] Furthermore, clean and standardize the data, clean the noise and errors in the data, such as removing irrelevant information or filling in missing values; unify the units and standardize various attribute values to ensure data consistency, and finally obtain the tunnel water inrush grouting treatment material database.

[0056] 4) Preprocess the tunnel water inrush grouting treatment material database, including numericalization of discrete features, normalization, class balancing, etc.

[0057] Among them, for numericalization of discrete features, encoding methods such as one-hot encoding, label encoding, and embedding encoding can be selected for conversion. For normalization methods, Min-Max normalization, Z-score normalization, logarithmic normalization, etc. can be selected. For class balancing, undersampling and oversampling methods can be selected for processing.

[0058] 5) The division of the dataset can be divided into two parts. Shuffle the data, select 70% as the training set for model training; 30% as the test set for model evaluation.

[0059] Step 2: Construct a statistical basic model and a machine learning basic model, apply a fusion strategy to construct an integrated model, and establish a relatively accurate relationship between engineering conditions and material properties, including: input the engineering conditions in the tunnel water inrush grouting treatment material database into the integrated model for analysis and feature extraction, and output the target material properties under this engineering condition.

[0060] Specifically, the integrated model is composed of the integration and fusion of a statistical basic model and a machine learning basic model. Among them, the statistical basic model can analyze the linear and non-linear relationships between engineering conditions and material properties, and the machine learning basic model can extract features of complex non-linear relationships. Combining traditional engineering experience with modern machine learning technology improves the generalization ability and robustness of the model.

[0061] After training and testing the integrated model, input the current engineering condition into the integrated model, and the relatively accurate target material properties required under this condition can be obtained.

[0062] Furthermore, the training process of the integrated model of the present disclosure is as follows:

[0063] 1. Build a statistical basic model, including exploratory data analysis, model selection, parameter fitting, and testing and evaluation.

[0064] Among them, exploratory data analysis includes:

[0065] 1) Check the data characteristics to see if there is a strong linear or non-linear relationship between the independent variable engineering conditions and each dependent variable material property;

[0066] 2) Variable correlation analysis, using scatter plots, correlation coefficients (such as Pearson, Spearman, etc.) to analyze the correlation between independent variables and dependent variables;

[0067] 3) Variable distribution check: Check the distribution pattern of the data to see if it satisfies the normality assumption;

[0068] 4) Multicollinearity detection: Use methods such as variance inflation factor (VIF) to detect collinearity problems. If the VIF value is too high, variables with high correlation need to be deleted, or regularization methods such as ridge regression and Lasso regression are used.

[0069] Furthermore, model selection includes: Select a model according to the data characteristics. For linear relationships, ordinary linear regression, generalized linear regression, etc. can be selected; for non-linear relationships, polynomial regression can be selected; for multicollinearity, regularization methods such as ridge regression and Lasso regression can be selected.

[0070] Furthermore, parameter fitting includes: Divide the data into a training set and a test set; Use methods such as least squares method and maximum likelihood estimation to fit the model to obtain model parameters; Evaluate the goodness of fit (R 2 ) and residual distribution of the model to ensure that the assumptions of the model are satisfied with its data characteristics.

[0071] Furthermore, testing and evaluation include: Check the independence and normality of the residuals. If the residuals show a pattern, there may be missing non-linearity in the model; Use t-test or F-test to judge the significance of independent variables to dependent variables; Through cross-validation or resampling, test the stability of the model under different data distributions.

[0072] Finally, a statistical basic model is obtained.

[0073] 2. Build a machine learning basic model. The available models include: random forest, support vector machine, KNN, and neural network.

[0074] Specifically, for the independent variable engineering conditions and the dependent variable material properties, test the above models, and then select the best-performing model through cross-validation.

[0075] Among them, five-fold cross-validation or ten-fold cross-validation is used for cross-validation, and the performance indicators (such as accuracy, recall rate, mean square error, etc.) of each fold are recorded to compare the advantages and disadvantages of the models. For the selected model, grid search or random search is used to adjust the hyperparameters of the model, and the error distribution, precision, recall rate, and feature importance of the model are checked. Finally, a machine learning basic model is obtained.

[0076] 3. Apply a fusion strategy to integrate the statistical basic model and the machine learning basic model. The selected appropriate fusion strategies include weighted average, Boosting, Stacking, etc. Combine the prediction results of the statistical basic model and the machine learning basic model to improve the overall prediction performance, that is, establish a relatively accurate relationship between engineering conditions and material properties. Optionally, the present disclosure uses the weighted average method for model fusion. The specific process is as follows:

[0077] 1) For each input sample (engineering condition) x i and the corresponding output (material property) y i , the predicted output of the statistical basic model is The predicted output of the machine learning basic model is Calculate the error of each model, preferably the mean square error:

[0078]

[0079] where N is the number of samples.

[0080] 2) Find the optimal weights by minimizing the weighted average prediction error. The weighted average prediction error is:

[0081]

[0082] where is the predicted output of the i-th sample of the statistical basic model and the machine learning basic model, N is the number of samples, ω stat , ω ml are the weights of the statistical basic model and the machine learning basic model respectively.

[0083] 3) Use the least squares method to minimize the weighted average prediction error: Take the partial derivatives of MSE with respect to ω stat , ω ml respectively:

[0084]

[0085] Set the partial derivatives to 0 and solve for the weights:

[0086]

[0087] The final predicted value can be obtained by performing weighted averaging on the models:

[0088]

[0089] Among them, is the predicted output of the statistical base model, is the predicted output of the machine learning base model.

[0090] Furthermore, the training of the integrated model should train the base models separately. For the statistical base model, use the selected regression method to fit and test the engineering conditions and material properties; for the machine learning base model, similarly, use the engineering conditions as the input and the material properties as the output to train and test the selected machine learning method.

[0091] Finally, a trained integrated model is obtained.

[0092] Furthermore, by inputting the engineering conditions into the integrated model, the target material properties under these engineering conditions can be output.

[0093] Step 3: By establishing a database of grouting material components and properties, construct a machine learning inverse model for inversely inferring material components based on material properties. Input the target material properties output by the integrated model into this inverse model to obtain the predicted material components.

[0094] Step 1) Establish a database of grouting material components and properties;

[0095] Specifically, the establishment of the database of grouting material components and properties can be carried out through laboratory test data collection and data augmentation means. Collect the material properties of various grouting materials under different components through experiments. Specifically, in the grouting treatment of tunnel water inrush, the influence of cement and admixtures in cement-based materials on material properties is the most direct and important;

[0096] The material components include: water-cement ratio, cement type, water reducer type and dosage, retarder type and dosage, early strength agent type and dosage; the material properties are consistent with the output of the integrated model, that is, material type, initial setting time, 3-day compressive strength, viscosity.

[0097] After augmentation through data augmentation means, preprocess the data, and finally establish a database of grouting material components and properties. Data augmentation means can enhance the number of data samples and enrich data diversity when the amount of experimental data is insufficient, effectively improving the generalization ability of the model. Optional means include: interpolation method, variational autoencoder, SMOTE, etc.

[0098] Furthermore, the data preprocessing includes one-hot encoding and numerical processing of the type and addition amount of water reducing agent, retarder, and early strength agent, and standardizing and normalizing the data of each material component.

[0099] Step 2) The prediction of the material components is realized through a machine learning inverse model. A machine learning inverse model for inversely inferring the material components based on the material properties is constructed and trained.

[0100] Specifically, the machine learning inverse model adopts an adaptive generative adversarial network model. The adaptive generative adversarial network model includes a generative adversarial network model and an interpreter. The generative adversarial network model includes a generator and a discriminator. The generator takes the target material properties as the conditional variable input, combines with random noise to generate the material components, and selects a convolutional neural network structure. The generation process of the generator is as follows:

[0101]

[0102] Among them, G is the generator, c is the random noise, y is the conditional variable, that is, the material properties, is the output of the generator, that is, the material components.

[0103] The task of the discriminator is to receive the material properties and the generated samples output by the generator as inputs, and output the probability value of whether the material components conform to the target material properties, that is, to judge whether the generated material components are "real", and selects a multi-layer perceptron (MLP) structure. The generation process of the discriminator is as follows:

[0104] D(z) = σ(ω T z + b)

[0105] Among them, ω and b are the hyperparameters of the discriminator, ω is the feature weight, b is the feature bias, z is the input feature, and should take the generated samples output by the generator. σ is the Sigmoid activation function, and D(z) represents the probability that the data z is real data.

[0106] The generator loss and the discriminator adversarial loss are defined respectively:

[0107] The calculation formula for the generator loss is:

[0108] L G = -log[D(y, G(c, y))]

[0109] Among them, G(c, y) is the generated sample output by the generator, and D(y, G(c, y)) is the judgment probability of the discriminator for the generated sample.

[0110] The adversarial loss of the discriminator:

[0111] L D = Lreal +L fake = -log(D(y, z real )) - log(1 - D(y, z fake ))

[0112] The adversarial loss of the discriminator is the sum of the losses of real samples and generated samples, where y and z real are the material properties and their corresponding real material components in the grouting material component and performance database, i.e., real samples; z fake is the material component output by the generator, i.e., the generated sample; D(y, z real ) and D(y, z fake ) are the probability outputs of the discriminator for real samples and generated samples.

[0113] Furthermore, the model training is to alternately train the generator and the discriminator using the preprocessed data. The specific process is as follows:

[0114] 1) Train the discriminator: Randomly select a batch of real samples y and z from the training set real and input them into the discriminator to calculate the loss of real samples; Use the generator to randomly generate a batch of initial material components y fake and the corresponding material properties z fake , and use them as fake samples to input into the discriminator to calculate the loss of generated samples; Calculate the gradient through backpropagation and update its parameters to minimize the total loss.

[0115] 2) Train the generator: Input random noise and material properties into the generator. The generator generates the corresponding material components through its neural network, i.e., the generated samples. Use the discriminator to evaluate these generated samples, calculate the gradient through backpropagation, and update the generator parameters to maximize the comprehensive loss of the generator.

[0116] 3) Alternately repeat the process of training the discriminator and the generator. Each training step will optimize their respective parameters through backpropagation and gradient update.

[0117] Furthermore, the reverse prediction is to input the target material properties output by the integrated model and random noise into the trained generative adversarial network, and the predicted material components can be output by the generator, and the component ratio can better conform to the input target performance.

[0118] In step 3), an interpreter is introduced after the generative adversarial network. The interpreter applies the SHAP (Shapley Additive exPlanations) interpretability method to quantitatively explain which generated samples (i.e., material components and corresponding material properties) in the discriminator contribute to generating "more real" samples, so as to optimize the process of predicting material components;

[0119] Specifically, the adaptive generative adversarial network dynamically adjusts the key feature weights of the discriminator with the help of the SHAP interpreter, and then optimizes the reverse prediction output. The specific steps are as follows:

[0120] 1) Take the output scoring probability of the discriminator as the interpretation target, and use SHAP to quantify the impact of the input generated sample on the discriminator output judgment probability D(y, z).

[0121] For the Shapley value of the generated sample, the calculation formula is:

[0122]

[0123] where M is the set of all input features, that is, the set composed of all features of the generated sample z fake ; |M| represents the total number of features; i represents the i-th sample; S is any subset obtained by removing the i-th sample from the set M, which can be an empty set; |S| represents the size of the subset S; f(S) represents the output of the discriminator when only the features in the subset S are used as input, that is, D(y, z fake ).

[0124] 2) According to the magnitude and sign of the SHAP value, identify which material components contribute the most to the authenticity or performance of the discriminator output. The larger the SHAP value, the greater the contribution of the feature to the discriminator output (i.e., the probability of sample authenticity). A positive value indicates that the feature makes the sample more likely to be judged as real, while a negative value indicates that the feature makes the sample more likely to be judged as false. Specifically, set a threshold, and regard the features exceeding this threshold as key features.

[0125] 3) According to the key features identified by the SHAP value, construct an adaptive discriminator so that it assigns higher weights to the key features when judging true and false samples. Specifically, expand the weights of the identified key features by 5 times.

[0126] 4) Retrain the reverse model and perform reverse prediction again, and recalculate the SHAP value of the input generated sample for the discriminator, and iteratively update the weights. When the change in feature weights is less than 1% or the number of iterations reaches a predetermined threshold, stop the iterative process and save the current weights of the discriminator.

[0127] 5) Output the samples generated by the current generator, that is, optimize the material components with higher quality and interpretability.

[0128] Example 2

[0129] In an embodiment of the present disclosure, an interpretable optimization system for the grouting material components of tunnel water inrush is provided, including:

[0130] A data acquisition module, configured to acquire the engineering condition data of actual tunnel grouting and perform preprocessing;

[0131] A prediction target acquisition module, configured to input the preprocessed engineering condition data into an integrated model, where the integrated model is a weighted average fusion of a statistical basic model and a machine learning basic model using a fusion strategy, and outputs the predicted target material properties by extracting the correlation between the engineering conditions and the material properties;

[0132] A material component inverse prediction module, configured to input the target material properties into an adaptive generative adversarial network model. First, the generator of the adaptive generative adversarial network model combines random noise to generate material components, and then inputs the generated material components and the target material properties into the discriminator of the adaptive generative adversarial network model, and outputs the probability value of whether the material components meet the target material properties. An interpretability method is introduced, and the output of the discriminator is used as the interpretation target to quantify the influence of the input material components on the discrimination probability output by the discriminator. By dynamically adjusting the key feature weights of the discriminator, the input material components are optimized, that is, the optimal material components are obtained by inverse prediction.

[0133] Example 3

[0134] In one embodiment of the present disclosure, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, it implements the interpretable tunnel water inrush grouting material component optimization method described above.

[0135] Example 4

[0136] In one embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, and the non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, it implements the interpretable tunnel water inrush grouting material component optimization method described above.

[0137] Example 5

[0138] In one embodiment of the present disclosure, an electronic device is provided, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device runs, the processor executes the computer program stored in the memory, so that the electronic device executes and implements the interpretable tunnel water inrush grouting material component optimization method described above.

[0139] This disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the disclosure. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general-purpose computers, special-purpose computers, embedded processors, or other programmable data processing devices to generate a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or means for implementing the functions specified in one or more of the blocks.

[0140] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more of the flows Figure 1 one or more of the flows and / or blocks Figure 1 or steps for implementing the functions specified in one or more of the blocks.

[0141] Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, they do not limit the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present disclosure, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. An interpretable method for optimizing the composition of grouting materials for sudden water inrush in tunnels, characterized in that: include: Obtain and preprocess the engineering condition data of actual tunnel grouting; Inputting the preprocessed engineering condition data into an integrated model, wherein the integrated model adopts a fusion strategy to fuse the weighted average of the statistical basic model and the machine learning basic model, extracts the correlation between the engineering conditions and the material properties, and outputs the predicted target material properties; The target material properties are input into the adaptive generative adversarial network model. First, the material components are generated by combining random noise through the generator of the adaptive generative adversarial network model. Then, the generated material components and the target material properties are input into the discriminator of the adaptive generative adversarial network model. The probability value of whether the material component meets the target material properties is output. An interpretability method is introduced, and the output of the discriminator is used as the explanation target. The influence of the input material component on the judgment probability of the discriminator output is quantified. By dynamically adjusting the key feature weights of the discriminator, the input material component is optimized, that is, the optimal material component is obtained by reverse prediction.

2. An interpretable method for optimizing the composition of grouting materials for sudden water inrush in tunnels according to claim 1, characterized in that: The engineering condition data include stratum type, water inflow and tunnel burial depth; the grouting material properties include material type, initial setting time, compressive strength and viscosity; the preprocessing includes discrete feature digitization, normalization and category balance.

3. The interpretable method for optimizing the composition of grouting materials for sudden water inrush in tunnels according to claim 1, characterized in that: The integrated model is integrated and fused by the statistical basic model and the machine learning basic model using a fusion strategy, wherein the statistical basic model performs linear and nonlinear relationship analysis on engineering conditions and material properties, and the machine learning basic model performs feature extraction of complex nonlinear relationships on engineering condition data, and a weighted average is performed on the two to obtain the final predicted target material properties: in, is the prediction output of the statistical base model, It is the prediction output of the basic machine learning model.

4. The interpretable method for optimizing the composition of grouting materials for sudden water inrush in tunnels according to claim 1, characterized in that: Material components include water-cement ratio, cement type, water reducer type and addition amount, retarder type and addition amount, early strength agent type and addition amount. The reverse prediction of material components is achieved through a machine learning inverse model. The machine learning inverse model adopts an adaptive generative adversarial network model. The adaptive generative adversarial network model includes a generator, a discriminator, and an interpretable. The generator accepts the target material properties as conditional variable input and generates material components in combination with random noise.

5. An interpretable method for optimizing the composition of grouting materials for sudden water inrush in tunnels as claimed in claim 4, characterized in that: The discriminator receives the target material properties and the generated samples output by the generator as input, and outputs the probability value of whether the generated samples output by the generator meet the target material properties, that is, it determines whether the generated material components are real. The discriminator selects a multi-layer perceptron structure, and the generation process of the discriminator is: D(z)=σ(ω T z+b) Among them, ω and b are the hyperparameters of the discriminator, z is the input feature, that is, the generated sample output by the generator, σ is the Sigmoid activation function, and D(z) represents the probability that the data z is the real data.

6. The interpretable method for optimizing the composition of grouting materials for sudden water inrush in tunnels according to claim 1, characterized in that: The reverse prediction is to input the target material properties output by the integrated model and the generated random noise into the adaptive generative adversarial network, and output the predicted material components in the generator. The interpretability method in the interpretable is introduced to dynamically adjust the key feature weights of the discriminator, thereby optimizing the reverse prediction output. The Shapley value of the sample generated by the generator is calculated as follows: Among them, M is the set of all input features, that is, the generated sample z fake is a set of all features of the set M; M| represents the total number of features; i represents the i-th sample; S is any subset of the set M after removing the i-th sample, which can be an empty set; |S| represents the size of the subset S; f(S) represents the output of the discriminator when only the features in the subset S are used as input, that is, D(y,z fake ).

7. An interpretable tunnel water inrush grouting material composition optimization system, characterized in that: include: Data acquisition module, used to obtain and pre-process the engineering condition data of actual tunnel grouting; A prediction target acquisition module is used to input the preprocessed engineering condition data into an integrated model, wherein the integrated model adopts a fusion strategy to fuse the weighted average of the statistical basic model and the machine learning basic model, and outputs the predicted target material properties by extracting the correlation between the engineering conditions and the material properties; The material composition reverse prediction module is used to input the target material properties into the adaptive generative adversarial network model. First, the material composition is generated by combining the random noise with the generator of the adaptive generative adversarial network model. Then, the generated material composition and the target material properties are input into the discriminator of the adaptive generative adversarial network model. The probability value of whether the material composition meets the target material properties is output. An interpretability method is introduced, and the output of the discriminator is used as the explanation target. The influence of the input material composition on the judgment probability of the discriminator output is quantified. By dynamically adjusting the key feature weights of the discriminator, the input material composition is optimized, that is, the optimal material composition is obtained by reverse prediction.

8. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, it implements an interpretable method for optimizing the composition of tunnel water inrush grouting materials as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by the processor, an interpretable method for optimizing the composition of tunnel sudden water grouting materials as described in any one of claims 1-6 is implemented.

10. An electronic device, characterized in that: include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes an interpretable method for optimizing the components of tunnel sudden water grouting materials as described in any one of claims 1-6.

Images