A deep learning-based grouting parameter selection optimization method and system

Abstract

The application discloses a grouting parameter selection optimization method and system based on deep learning, belongs to the technical field of tunnel grouting, and extracts features from pretreated geological condition and construction condition data, reduces dimensions of the extracted features, and obtains low-dimensional feature data; the low-dimensional feature data is preliminarily evaluated for importance to obtain feature importance scores; features with a feature importance score higher than a set threshold value are input into a pre-trained global optimization model to obtain a grouting effect score; and the grouting effect score and feature data corresponding to the grouting effect are input into a pre-trained local fine-tuning model to obtain a fine-tuning value of the grouting parameter. The double structure of the global optimization model and the local fine-tuning model is designed for global optimization and local fine-tuning of the grouting parameter respectively, and the hierarchical model design enables the system to better adapt to changes under complex construction conditions, and improves the stability and reliability of the overall grouting effect.

Description

Technical Field

[0001] This invention belongs to the field of tunnel grouting technology, and in particular relates to a method and system for selecting and optimizing grouting parameters based on deep learning. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] With the rapid development of underground engineering, tunnel construction faces increasingly complex geological conditions, especially in weak surrounding rock and water-rich strata, where the risks of disasters such as soil erosion and landslides increase significantly. Grouting technology, as an important engineering method, is widely used in reinforcing surrounding rock, controlling groundwater, and preventing stratum deformation. However, the grouting effect is affected by various factors, including geological conditions, grouting material properties, and grouting parameters. This complexity makes it difficult to determine the grouting parameters, thus affecting construction results and costs.

[0004] Traditional grouting parameter design relies heavily on the experience of construction personnel and field tests, lacking scientific rigor and systematic approach. Due to the uncertainty and complexity of geological conditions, grouting parameters are difficult to standardize in different engineering environments, often requiring continuous adjustments, which increases construction time and costs. Furthermore, geological conditions may change at any time during construction, such as instability of the surrounding rock structure and fluctuations in groundwater velocity and flow rate; these factors directly affect the grouting effect. Therefore, how to achieve real-time optimization and intelligent control of grouting parameters has become an important research direction in current tunnel construction technology.

[0005] In recent years, with the development of big data and artificial intelligence technologies, deep learning-based intelligent optimization methods have provided new solutions for tunnel grouting technology. By building and training deep learning models, key features and patterns can be automatically extracted from massive amounts of historical data, effectively improving the scientific nature of grouting parameter design. Although such intelligent optimization methods have provided new opportunities for tunnel grouting technology, some problems still exist. For example, when the model is applied in a resource-constrained field environment, its real-time performance and computational efficiency face certain challenges, which may make it difficult to meet the needs of rapid response construction. Existing models lag in responding to changes in the construction environment, making it difficult to optimize grouting parameters in real time, resulting in insufficient adaptability of the solution and the need for frequent manual intervention. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, this invention provides a deep learning-based method and system for selecting and optimizing grouting parameters. It adopts a dual structure of "global optimization model + local fine-tuning model", which designs for global optimization and local fine-tuning of grouting parameters respectively. This hierarchical model design enables the system to better adapt to changes under complex construction conditions, thereby improving the stability and reliability of the overall grouting effect.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] In a first aspect, the present invention provides a method for optimizing grouting parameters based on deep learning, comprising:

[0009] Acquire and preprocess current geological and construction condition data;

[0010] Feature extraction is performed on the preprocessed geological and construction condition data, and the extracted features are then dimensionality reduced to obtain low-dimensional feature data. A preliminary importance assessment is performed on the low-dimensional feature data to obtain feature importance scores.

[0011] Features with importance scores higher than a set threshold are input into a pre-trained global optimization model to obtain the grouting effect score;

[0012] The grouting effect score and the corresponding feature data are input into a pre-trained local fine-tuning model to obtain the fine-tuned values ​​of the grouting parameters.

[0013] In a further technical solution, the construction condition data includes the characteristics of the grouting material, grouting parameters, and grouting effect.

[0014] A further technical solution yields the following low-dimensional feature data:

[0015] Relevant features were extracted from the preprocessed geological and construction condition data and then normalized.

[0016] The correlation between each feature and the grouting effect index was calculated using the Pearson correlation coefficient, and features related to the grouting effect index were screened.

[0017] An improved linear discriminant analysis was applied to reduce the dimensionality of the selected features, resulting in low-dimensional feature data.

[0018] A further technical solution is that the improved linear discriminant analysis introduces a feature weight matrix and a class contribution weight matrix on the basis of linear discriminant analysis, specifically expressed as follows:

[0019]

[0020] Where ζ(w) represents the weighted objective function, w T W is the feature projection direction vector. c S is the class weight matrix. b W is the inter-class scatter matrix. f Let S be the feature weight matrix, w be the corresponding eigenvector, and S be the feature weight matrix. w It is the intra-class scatter matrix.

[0021] Further technical solutions yield the following feature importance scores:

[0022] The gradient boosting decision tree algorithm is used to perform a preliminary importance assessment of the low-dimensional feature data. An XGBoost model is then constructed and trained to evaluate the importance of the features and output a feature importance score.

[0023] A weighted feature importance assessment mechanism is introduced, combining engineering experience and goal-oriented weight adjustments to make feature importance scores more targeted and interpretable; specifically including:

[0024] First, a set of engineering experience weights is introduced to adjust the feature importance score, expressed as:

[0025] I i ′=I i ·W e,i

[0026] Among them, I i Score the feature importance of the i-th feature; W e,i Weighted by engineering experience; I i ′ represents the corrected feature importance score;

[0027] Then, target relevance correction is introduced, and feature importance is corrected a second time in combination with the optimization target to obtain the feature importance score after the second correction. After normalization, the final feature importance score is obtained.

[0028] A further technical solution is that the global optimization model is constructed using a multilayer perceptron, including an input layer, several hidden layers, a dropout layer, and an output layer, with a dropout layer added after each hidden layer of the multilayer perceptron.

[0029] A further technical solution involves constructing the local fine-tuning model using a lightweight convolutional neural network. This model receives the grouting effect score and the corresponding feature data, sets an optimization objective to maximize the grouting effect score, and expresses the objective function as follows:

[0030]

[0031] Where N is the number of training samples, S local S is the local score obtained after model optimization.global A global score is given for the grouting effect;

[0032] Then, the local fine-tuning model adjusts the input features, which are then passed through the hidden layers of a multi-layer fully connected network. Finally, the fine-tuned values ​​of each grouting parameter are output, as follows:

[0033] ΔP inj =W out,1 ·h (L) +b out,1 ΔT inj =W out,2 ·h (L) +b out,2

[0034] Wherein, ΔP inj With ΔT inj W represents the fine-tuning values ​​for grouting pressure and grouting time, respectively. out,i h represents the output layer weight matrix. (L) b represents the final output of the hidden layer. out,i This is the bias value for the output layer.

[0035] Secondly, the present invention provides a grouting parameter selection and optimization system based on deep learning, comprising:

[0036] The data acquisition module is configured to acquire and preprocess current geological and construction condition data.

[0037] The feature filtering module is configured to: extract features from the preprocessed geological and construction condition data, reduce the dimensionality of the extracted features to obtain low-dimensional feature data, and perform a preliminary importance assessment on the low-dimensional feature data to obtain a feature importance score.

[0038] The global optimization module is configured to input features whose feature importance scores are higher than a set threshold into a pre-trained global optimization model to obtain a grouting effect score.

[0039] The local adjustment module is configured to input the grouting effect score and the feature data corresponding to the grouting effect into a pre-trained local fine-tuning model to obtain the fine-tuned values ​​of the grouting parameters.

[0040] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the deep learning-based grouting parameter selection optimization method described in the first aspect.

[0041] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the deep learning-based grouting parameter selection optimization method described in the first aspect.

[0042] The above one or more technical solutions have the following beneficial effects:

[0043] This invention employs a dual structure of "global optimization model + local fine-tuning model," designed for both global optimization and local fine-tuning of grouting parameters. The global optimization model focuses on the grouting parameters with significant impact, achieving a globally optimal solution through continuous variation. The local fine-tuning model, based on the global optimization model, identifies potential deficiencies in the current grouting scheme and further optimizes the parameters output by the global optimization model. Through the combined effect of the global optimization model and the local fine-tuning model, a set of grouting parameters, including grouting pressure and grouting speed, is selected to achieve the best grouting effect under current geological and construction conditions. This hierarchical model design allows the system to better adapt to changes under complex construction conditions, improving the stability and reliability of the overall grouting effect.

[0044] This invention emphasizes the coupling effect between different grouting parameters, achieving comprehensive prediction of grouting effects through multi-parameter coupling analysis. Unlike traditional linear superposition models, this system can capture complex nonlinear relationships between parameters, thereby improving the accuracy of grouting effect prediction. This innovative application enables engineers to more comprehensively consider the mutual influence between parameters when designing grouting processes, optimizing grouting schemes and improving construction safety and efficiency.

[0045] This invention employs a gradient boosting decision tree algorithm for feature importance evaluation and introduces a weighted feature importance evaluation mechanism to effectively identify the features that have the greatest impact on grouting results. This process not only helps optimize model feature input, improving training efficiency and prediction performance, but also avoids computational complexity caused by feature redundancy. Feature dimensionality reduction further improves the model's computational efficiency, enabling the system to provide high-quality prediction results in a shorter time and improving the real-time response of both the global optimization model and the local fine-tuning model. Attached Figure Description

[0046] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0047] Figure 1 This is a flowchart of the grouting parameter selection and optimization method according to an embodiment of the present invention;

[0048] Figure 2 This is a structural diagram of the global optimization model in an embodiment of the present invention. Detailed Implementation

[0049] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0051] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0052] Example 1

[0053] like Figure 1 As shown in the figure, this embodiment discloses a deep learning-based method for optimizing grouting parameters. The method includes the following steps:

[0054] S1: Obtain and preprocess current geological and construction condition data.

[0055] In this embodiment, various historical data related to tunnel grouting are collected through construction logs, geological reports, and references. These include geological and construction condition data for different projects. The construction condition data includes data on grouting material characteristics, grouting parameters, and grouting effects. The grouting effect is quantified through grouting effect indicators to establish a comprehensive database.

[0056] The collected raw data, namely geological and construction condition data, are cleaned and preprocessed: outliers are processed using statistical or identification methods. For example, for data where the compressive strength deviates significantly from the reasonable range, it is analyzed whether the anomaly is caused by testing errors or other reasons, and a decision is made on whether to remove or correct it; for missing values, interpolation or methods to delete missing data are used to handle null values ​​in the data.

[0057] After data preprocessing, the data is categorized and stored according to its characteristics: multi-dimensional classification is performed based on the geological conditions of different projects (including rock strata distribution, lithology, groundwater conditions, etc.), grouting material characteristics (including cement grout ratio, viscosity, setting time, etc.), construction parameters (including grouting pressure, grouting speed, pipeline layout, etc.), and grouting effects (including permeability improvement rate, water resistance, construction stability, etc.). After classification, data with similar characteristics are aggregated into their respective independent modules.

[0058] Structured data (including geological parameters, numerical characteristics of grouting materials, etc.) is stored in a relational database (SQL database) to ensure data integrity, standardization, and retrieval; while unstructured data (including project logs, construction records, on-site images or videos, etc.) is managed using a non-relational database (MongoDB) to facilitate the storage of information in different formats.

[0059] S2: Extract features from the preprocessed geological and construction condition data, and reduce the dimensionality of the extracted features to obtain low-dimensional feature data.

[0060] The process of obtaining low-dimensional feature data involves: extracting relevant features from preprocessed geological and construction condition data and normalizing them; calculating the correlation between each feature and grouting effect indicators using the Pearson correlation coefficient, and selecting features related to grouting effect indicators; and applying improved linear discriminant analysis to reduce the dimensionality of the selected features to obtain low-dimensional feature data. Specifically:

[0061] S201: Relevant features are extracted from the preprocessed geological and construction condition data. These features cover key factors affecting the grouting effect, such as grouting parameters, geological environment, grouting material properties, and construction conditions. The extracted features are divided into training and testing sets at a ratio of 70% and 30% to ensure the model's generalization ability.

[0062] The key features closely related to grouting effect include: geological conditions such as lithology, rock layer thickness, rock layer fracture rate, groundwater level, and geological pressure; grouting material characteristics such as grout ratio, grout viscosity, setting time, and compressive strength; grouting parameters such as grouting pressure, grouting speed, grouting diffusion radius, and construction time; and grouting effect indicators such as improvement rate, water resistance performance, and construction stability.

[0063] Statistical methods were used for feature extraction, statistical analysis was performed on the known data distribution, the Pearson correlation coefficient was calculated, and parameters with significant influence were screened. The specific steps are as follows:

[0064] First, the preprocessed data of different dimensions are normalized to scale the feature values ​​to the [0,1] interval, eliminating dimensional differences. Then, the Pearson correlation coefficient r is used to calculate the correlation between each feature and the grouting effect index, and features that are significantly correlated with the grouting effect index are selected. In this embodiment, when |r|>0.5, the feature is considered to be significantly correlated.

[0065] The specific formula for the Pearson correlation coefficient is as follows:

[0066]

[0067] Where, x i Let y be the eigenvalue of the i-th feature. i Here, i represents the ith feature, and n represents the total number of feature values. This represents the average value of characteristic x. It is represented by the average value of the grouting effect index y.

[0068] The specific formula for the above data normalization process is as follows:

[0069]

[0070] Where x is the original feature value, x′ is the normalized feature value, and min(x) and max(x) are the minimum and maximum values ​​of the feature, respectively.

[0071] S202: Dimensionality reduction is performed using an improved Linear Discriminant Analysis (LDA) technique. Specifically, the mean and covariance of the feature data for each category are calculated to construct between-class and within-class scatter matrices. The optimal projection direction is then obtained by solving the generalized eigenvalue problem. Projecting the original high-dimensional data onto this direction generates low-dimensional feature data, which serves as the primary input for subsequent model training.

[0072] This application introduces a feature weight matrix W on the basis of the original LDA. f When constructing the inter-class and intra-class scatter matrices, each feature is assigned a weight based on its importance score. Simultaneously, a class contribution weight matrix W is added. c The discrimination direction is adjusted according to the importance of different geological environment categories. The specific formula is:

[0073]

[0074] Where ζ(w) represents the weighted objective function, represents the optimization objective of w, and w T W is the feature projection direction vector; c W is the class weight matrix, used to adjust the influence of different classes in the inter-class scatter matrix. c =diag(α1,α2,...,α)c ), α z S represents the importance weight of project category z; b W is the inter-class scatter matrix, used to measure the degree of dispersion between the means of different classes. f W is the feature weight matrix, used to adjust the contributions of the inter-class and intra-class scatter matrices. f =diag(w1,w2,...,w d ), w i S represents the i-th eigenvector; w is the corresponding eigenvector; S w This is an intra-class scatter matrix, used to measure the degree of dispersion of data within the same category.

[0075] This application improves LDA by introducing a feature-weighted and class contribution evaluation mechanism, combined with specific prior knowledge from engineering, to adjust the scatter matrix. The optimized generalized eigenvalue problem enhances class discrimination ability, resulting in higher model training accuracy and stronger generalization ability.

[0076] S3: Perform a preliminary importance assessment on the low-dimensional feature data to obtain a feature importance score.

[0077] In this embodiment, the gradient boosting decision tree algorithm is used to perform a preliminary importance assessment of the extracted features, and an XGBoost model is constructed and trained to evaluate the importance of the features and output a feature importance score.

[0078] 1) The XGBoost model is trained using a pre-defined training set. During training, the model continuously adjusts the structure and parameters of the decision tree based on the training data to minimize the loss between the predicted and actual values. The algorithm iterates gradually, generating a new decision tree with the goal of fitting the residual between the model's previous prediction and the actual value. Through multiple iterations, the model's predictive ability gradually improves.

[0079] 2) Set the hyperparameters of the XGBoost model. For example, the learning rate: a small learning rate may lead to excessively long training time, while a large learning rate may prevent the model from converging to the optimal value. To ensure the model's convergence speed without causing the model to skip the optimal solution during training, the learning rate is set to 0.1.

[0080] 3) Utilize metrics from the trained XGBoost model to obtain feature importance scores. For example, frequency metrics indicate the number of times a feature is used as a splitting node across all decision trees. If a feature frequently appears in decision tree splits, it indicates that it has a high influence in the model. Through the evaluation of these metrics, the model can quantify the importance of each feature to the final prediction result and output a feature importance score, providing a preliminary basis for subsequent steps to select key feature metrics for training.

[0081] 4) Building upon XGBoost's initial feature importance assessment, a weighted feature importance assessment mechanism is introduced. This mechanism, incorporating engineering experience and goal-oriented weight adjustments, makes feature importance scoring more targeted and interpretable. Specifically, the following improvement steps are included:

[0082] First, when initially assessing the importance of features, a set of engineering experience weights W is introduced. e The feature importance score is adjusted. Weights are derived from expert experience, historical project analysis, or the priority of specific indicators. The specific formula is:

[0083] I i ′=I i ·W e,i

[0084] Among them, I i Assign a feature importance score to the i-th feature (the initial importance score is obtained from the XGBoost model); W e,i Weighted by engineering experience; I i ′ represents the corrected feature importance score.

[0085] Then, a target relevance correction is introduced, and the feature importance is further corrected based on the optimization objective. For example, if the objective is to predict the grouting diffusion radius or the sealing effect, features highly correlated with diffusion and fluid properties need to have their weights increased. The specific formula is as follows:

[0086] I i "=I i ′·ρ i

[0087] Among them, I i ′ represents the feature importance score after one correction; I i "This is the feature importance score after secondary correction; ρ" i To assign correlation weights between features and the target variable, the Pearson correlation coefficient corr(X) is calculated. i We get: ,Y) Where x i Let be the feature value of the i-th feature, U be the target variable, and o be the total number of features.

[0088] Finally, after normalization, the final feature importance score is obtained, providing a basis for subsequent key feature selection. The specific formula is:

[0089]

[0090] Among them, I final,i This is a normalized feature importance score, which facilitates subsequent feature selection.

[0091] This application improves upon the original XGBoost model's evaluation of initial feature importance by introducing engineering experience weights and target relevance weights, making the model more closely aligned with actual grouting application needs. The revised feature scoring avoids misjudgments caused by uneven sample distribution or data noise, thus improving the stability of feature selection.

[0092] S4: Input the features whose feature importance scores are higher than the set threshold into the pre-trained global optimization model (large model) to obtain the grouting effect score.

[0093] S401: Constructing the Dataset: Divide the low-dimensional feature data into a training set, a validation set, and a test set. In this embodiment, the dataset division ratio is set to 70:15:15 to ensure that the training set provides sufficient data samples, the validation set is used for model tuning, and the test set is used for final performance evaluation. The input of the globally optimized model includes grouting parameter features with significant impact, i.e., features whose feature importance scores are higher than a set threshold, such as grouting pressure, grouting speed, and grout type. The output is the grouting effect. In the training and validation sets, data features closely related to the grouting effect, i.e., features whose feature importance scores are higher than a set threshold, are selected and standardized to improve the efficiency and stability of model training.

[0094] The aforementioned grouting parameter features with significant impact were selected based on feature importance scores, and the final importance score I for each feature was obtained using an improved XGBoost algorithm. final,i Features with scores higher than a set threshold T are selected as input values ​​for the global optimization model. Specifically:

[0095] I final,i >T

[0096] Where T is the set filtering threshold, and in this embodiment, T = 0.5.

[0097] S402: Construct a global optimization model. For example... Figure 2 As shown, the global optimization model is constructed using a multilayer perceptron (MLP), comprising an input layer, several hidden layers, a dropout layer, and an output layer. A dropout layer is added after each hidden layer of the MLP. In the specific network configuration, the hidden layers use the ReLU activation function to enhance the model's nonlinear modeling ability and prevent the gradient vanishing problem.

[0098] The ReLU function is defined as follows:

[0099] R(x R ) = max(0, x R )

[0100] For negative input numbers, the output is 0; for positive input numbers, the output is equal to the input.

[0101] To address the risk of overfitting in MLP models when there are many input features, this application adds a neural network architecture with a Dropout layer after each hidden layer. By randomly discarding some neuron features, the network's dependence during training is reduced, enhancing the model's robustness and preventing overfitting. The specific Dropout formula is as follows:

[0102]

[0103] in, This is the output after the Dropout layer, representing the output features after discarding; h g These are the original output features of neurons; z g ∈{0,1} is a Bernoulli random variable representing whether the neuron's feature g is preserved. This value is determined by the preservation probability p. P(z) g =1)=p, meaning the output of the neuron feature is retained with probability p and discarded with probability 1-p. In this embodiment, p=0.5, meaning that in each training iteration, there is a 50% probability of discarding a neuron feature.

[0104] This application addresses the global optimization model MLP by incorporating a Dropout layer into the hidden layers of a neural network architecture. This prevents overfitting by randomly discarding some neuron features in each iteration, while simultaneously improving training efficiency. For the locally fine-tuned model TinyCNN, it introduces diverse kernel sizes and residual connections to ensure that the convolutional layers can fully capture features at different scales, avoiding the vanishing gradient problem. Both models improve performance on the original model framework, enhancing the efficiency and accuracy of model training.

[0105] S403: Training the Global Optimization Model. The constructed global optimization model is trained using the training set data. Stochastic gradient descent (SGD) is used to optimize the network parameters to minimize prediction error. During training, k-fold cross-validation is used to evaluate the model's generalization ability, ensuring good performance on unseen data. Mean squared error (MSE) is selected as the evaluation metric. MSE is the average of the squared differences between predicted and actual values, measuring the average degree of prediction error and ensuring the reliability of the prediction results.

[0106] The formula for calculating the mean square error is:

[0107]

[0108] Among them, G i This is the actual value. denoted as the predicted value, and D as the total number of data points.

[0109] To further enhance the predictive ability of the model, a coupling relationship model among the parameters is established by analyzing the interaction relationships between different parameters, as detailed below:

[0110] First, the importance of parameter interactions is quantified based on SHAP values, and the importance of interactive features is evaluated to obtain the degree of influence of the interaction between features on the model output.

[0111]

[0112] Where, x i Let S be the specific value of the feature i being evaluated; S∈F / i is any subset of the feature set after removing feature i; |S| is the number of features contained in the feature subset S; |F| is the total number of features; v(S∪{i}) is the predicted value of the model when using the feature subset S and feature i; v(S) is the predicted value of the model when using only the feature subset S.

[0113] Then, based on the analysis results, a coupling relationship model was selected, as follows:

[0114]

[0115] Where Y is the grouting effect index; β0 is the bias constant; X i X is the i-th parameter; i X j It is a second-order interactive feature; β ij For variable X i and X j The interaction between them affects Y; ε is the random error. β i Let X represent the regression coefficient of the i-th parameter, m represent the total number of parameters, and X represent the regression coefficient of the i-th parameter. j This is represented as the j-th parameter.

[0116] Ultimately, the output coupling model reveals the interactive effects between different parameters, identifying the parameters that contribute the most or have the strongest impact on the grouting effect. This provides a reference for optimizing the model through local fine-tuning, improving the model's predictive ability and the optimization effect of grouting parameters. The coupling model refers to the coupling between parameters; for example, grouting pressure and grouting rate parameters influence each other, jointly affecting the grouting effect. Based on the output of the coupling model, it identifies which parameters contribute the most or have the strongest impact on the grouting effect, thus guiding the local fine-tuning of the model. In practical engineering, real-time monitoring data can be input into the coupling model, and the model updates to output parameters with strong interactions and optimized values, guiding the dynamic adjustment of construction parameters.

[0117] The parameters output by the global optimization model are used as input. Bayesian optimization techniques are employed to optimize and adjust the hyperparameters of the global optimization model to maximize the grouting effect, thereby achieving higher grouting stability and economy. The specific steps are as follows:

[0118] 1) Clearly define the hyperparameters to be optimized and their value ranges. For example, the learning rate range is set to [0.01, 0.1] to control the learning step size of the model; the number of network layers is set to [2, 4] to control the depth of the neural network, etc. A Gaussian process (GP) is used as the surrogate model, and the objective function is defined as:

[0119] f(θ) = -MSE val (θ)

[0120] Where θ represents different combinations of hyperparameters, MSE val This represents the mean squared error evaluated on the validation set.

[0121] 2) Define an acquisition function to select combinations from the hyperparameter space, and evaluate these combinations using a surrogate model. The acquisition function is defined as follows:

[0122] EI(A)=E[max(f(A)-f best )]

[0123] Among them, f best This represents the current best performance. f(A) represents the performance value predicted by the surrogate model at point A, which is a random variable; EI(A) represents the expected improvement value, which represents the potential expected improvement value of point A in the hyperparameter space, and E represents the expected value. The improvement value of point A is a weighted average under its probability distribution.

[0124] 3) Update the surrogate model GP to incorporate the new evaluation results in order to improve the accuracy of the surrogate model.

[0125] 4) Repeat steps 2 and 3 until the preset number of iterations is reached. Output the current optimal hyperparameter combination and apply it to the final training of the model to ensure the best predictive ability.

[0126] After completing model training and hyperparameter tuning, the validation set data is input into the globally optimized model for prediction to evaluate its accuracy in predicting grouting effects. Specifically, a scatter plot of predicted and actual values ​​is generated to observe the overall data distribution and the model's fit. The scatter points are checked to see if they are closely distributed around the ideal 45-degree diagonal. If so, the prediction performance is satisfactory, the model parameters are finalized, and the globally optimized model is saved. Otherwise, there is a prediction bias, and retraining should be iterated.

[0127] S5: Input the grouting effect score and the feature data corresponding to the grouting effect into the pre-trained local fine-tuning model (small model) to obtain the fine-tuning value of the grouting parameters.

[0128] The grouting effect corresponds one-to-one with the feature data. Different feature data inputs result in different grouting effects. The feature data corresponding to the grouting effect can be expressed as: a certain grouting effect caused by a certain part of the feature values. That is, the feature data is a subset of features whose feature importance scores are higher than a set threshold. It is a part of the feature data because not all feature data needs to be fine-tuned.

[0129] The input to the local fine-tuning model consists of two parts: the grouting effect score output by the global optimization model and the feature data set corresponding to the grouting effect score, which is the data set that is input into the global optimization model to obtain the corresponding score.

[0130] S501: Constructing the dataset: Divide the input features into a training set and a validation set.

[0131] S502: Constructing the Local Fine-Tuning Model: The local fine-tuning model is constructed using a lightweight convolutional neural network (TinyCNN), consisting of 2-3 convolutional layers and 1-2 fully connected layers. In the convolutional layers, small-sized convolutional kernels are used to capture local correlations of the input features. The pooling layers employ max pooling to further extract features and reduce the dimensionality of the feature maps. Finally, the prediction result is output through a fully connected layer.

[0132] To ensure that the convolutional layers can fully capture features at different scales and avoid the vanishing gradient problem, this application diversifies the convolutional kernel size and introduces residual connections, as follows:

[0133] First convolutional layer: using a 3×3 convolutional kernel, stride of 1, and 32 output channels;

[0134] The second convolutional layer uses a 5×5 convolutional kernel with a stride of 1 and 64 output channels.

[0135] The third convolutional layer uses a 3×3 kernel with a stride of 1 and 128 output channels.

[0136] The specific formula for residual connection is:

[0137]

[0138] Where J is the output signal of the convolutional layer; W i These are the convolution kernel parameters; This is the output of the residual module, representing the result after a series of convolution operations; K is the output signal. In this way, the gradient can be directly backpropagated through skip connections, thus solving the gradient vanishing problem.

[0139] S503: Training the Locally Fine-Tuned Model: The Adam optimization algorithm is used in conjunction with mini-batch gradient descent for training. During training, the loss value is gradually reduced to improve prediction accuracy. k-fold cross-validation is employed to evaluate the performance and stability of the locally fine-tuned model. Mean Squared Error (MSE) is chosen as the loss function to quantify the model's error under different grouting environments for evaluation, as detailed in S403 above.

[0140] Bayesian optimization techniques are used to fine-tune the hyperparameters of the local fine-tuning model, as shown in S4.

[0141] In actual grouting projects, based on the current environmental monitoring data and the output of the global optimization model, namely the grouting effect score and the corresponding feature data set, the data is input into the local fine-tuning model. The local fine-tuning model outputs the fine-tuning values ​​of parameters such as grouting pressure and grouting time.

[0142] After receiving the grouting effect score and corresponding feature data (such as geological features and construction conditions) from the global optimization model, the local fine-tuning model identifies potential deficiencies in the current grouting scheme based on the score or effect data, sets an optimization objective, and maximizes the grouting effect score. The objective function is expressed as:

[0143]

[0144] Where N is the number of training samples, S local S is the local score obtained after model optimization. global The overall grouting effect is scored.

[0145] Then, the local fine-tuning model adjusts the input features, such as optimizing grouting pressure and grouting time. Specifically:

[0146] Model mapping input data X input The fine-tuned grouting parameters are expressed as follows:

[0147] ΔP inj ,ΔT inj =f(X) input )

[0148] Wherein, ΔP inj ΔT is the fine-tuning value for the grouting pressure. inj This is a fine-tuning value for the grouting time. The hidden layer uses a multi-layer fully connected network, represented as:

[0149] h (l) =σ(W (l) ·h (l-1) +b (l) )

[0150] Among them, h (l) W is the output of the l-th hidden unit. (l) b (l) These are the weight matrix and the bias vector, respectively, and σ() is the activation function.

[0151] Finally, the fine-tuning values ​​for each grouting parameter are output, as follows:

[0152] ΔP inj =W out,1 ·h (L) +b out,1 ΔT inj =W out,2 ·h (L) +b out,2

[0153] Wherein, ΔP inj With ΔT inj W represents the fine-tuning values ​​for grouting pressure and grouting time, respectively. out,i h represents the output layer weight matrix. (L) b represents the final output of the hidden layer. out,i This is the bias value for the output layer.

[0154] These fine-tuning parameters are further adjusted based on the global optimization model to cope with different geological and construction conditions. In this way, the grouting effect is ensured to reach its optimal level under the coupling effect of parameters, enabling real-time dynamic fine-tuning of parameters to adapt to the changing needs of the grouting environment.

[0155] Real-time application and dynamic adjustment of grouting parameter selection optimization methods: Data on current geological and construction conditions is acquired, preprocessed, and then processed through feature extraction, dimensionality reduction, and feature importance scoring before being input into the global optimization model, which outputs a grouting effect score. The grouting effect score and the corresponding feature data are then input into the local fine-tuning model to adjust the grouting parameters. The global optimization model has been trained on a large amount of historical data, covering grouting cases under various geological and construction scenarios. It provides a relatively macroscopic optimal parameter combination as a basis for the current situation. Based on this, the local fine-tuning model further optimizes the parameters output by the global optimization model. Through the collaborative work of the global optimization model and the local fine-tuning model, the deep learning model can generate the optimal grouting parameters suitable for the current conditions in real time.

[0156] When geological conditions change during construction (such as variations in stratum hardness or groundwater level) or other factors affect the grouting effect, this invention can promptly acquire new data, re-analyze and recalculate, and quickly adjust grouting parameters to ensure the stability and reliability of the grouting effect. For example, if the actual grouting effect deviates from the expected result during construction, the model can be recalculated and subsequent grouting parameters adjusted based on real-time monitoring data to adapt to the changing situation, ensuring the smooth progress of the entire grouting project and achieving the expected results.

[0157] Example 2

[0158] This embodiment discloses a grouting parameter selection and optimization system based on deep learning, including:

[0159] The data acquisition module is configured to acquire and preprocess current geological and construction condition data.

[0160] The feature filtering module is configured to: extract features from the preprocessed geological and construction condition data, reduce the dimensionality of the extracted features to obtain low-dimensional feature data, and perform a preliminary importance assessment on the low-dimensional feature data to obtain a feature importance score.

[0161] The global optimization module is configured to input features whose feature importance scores are higher than a set threshold into a pre-trained global optimization model to obtain a grouting effect score.

[0162] The local adjustment module is configured to input the grouting effect score and the feature importance score that are higher than a set threshold into a pre-trained local fine-tuning model to obtain the fine-tuned values ​​of the grouting parameters.

[0163] Example 3

[0164] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0165] Example 4

[0166] The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0167] The steps and methods involved in the apparatuses of Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0168] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0169] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0170] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A deep learning-based grouting parameter selection optimization method, characterized in that, The method comprises the following steps: obtaining and preprocessing current geological condition and construction condition data; extracting features from the preprocessed geological condition and construction condition data, and performing dimensionality reduction on the extracted features to obtain low-dimensional feature data; performing preliminary importance evaluation on the low-dimensional feature data to obtain feature importance scores; inputting features with a feature importance score higher than a set threshold into a pre-trained global optimization model to obtain a grouting effect score; the global optimization model is built using a multilayer perceptron, including an input layer, a plurality of hidden layers, a Dropout layer and an output layer, and a Dropout layer is added after each hidden layer of the multilayer perceptron; inputting the grouting effect score and the feature data corresponding to the grouting effect into a pre-trained local fine-tuning model to obtain fine-tuning values of grouting parameters; the local fine-tuning model is built using a lightweight convolutional neural network, receives the grouting effect score and the feature data corresponding to the grouting effect, sets an optimization goal to maximize the grouting effect score, and the objective function is represented as: wherein, is the number of training samples, is the local score obtained after model optimization, is the global score of grouting effect; Then the local fine-tuning model adjusts the input features and passes them through a plurality of hidden layers of a multilayer fully connected network to finally output fine-tuning values of the grouting parameters, represented as: ， wherein, with respectively represent fine adjustment values of the grouting pressure and the grouting time, represents an output layer weight matrix, represents the final output of the hidden layer, is a bias quantity of the output layer.

2. The grouting parameter selection and optimization method based on deep learning according to claim 1, characterized in that, The construction condition data includes grouting material properties, grouting parameters and grouting effects.

3. The deep learning-based grouting parameter selection and optimization method according to claim 1, wherein, The low-dimensional feature data is obtained by: extracting relevant features from the preprocessed geological condition and construction condition data and performing normalization processing; calculating the correlation of each feature with the grouting effect index using the Pearson correlation coefficient, and screening features related to the grouting effect index; applying improved linear discriminant analysis to perform dimensionality reduction processing on the screened features to obtain low-dimensional feature data; The improved linear discriminant analysis introduces a feature weight matrix and a class contribution weight matrix on the basis of linear discriminant analysis, and is specifically represented as: wherein, is represented as a weighted objective function, is a feature projection direction vector, is a class weight matrix, is an inter-class scatter matrix, is a feature weight matrix, is a corresponding feature vector, is an intra-class scatter matrix.

4. The deep learning-based grouting parameter selection and optimization method according to claim 1, wherein, The feature importance score is obtained by: using a gradient boosting decision tree algorithm to preliminarily evaluate the importance of the low-dimensional feature data, constructing and training an XGBoost model to evaluate the importance of the features, and outputting the feature importance score; introducing a weighted feature importance evaluation mechanism, combining engineering experience and target-oriented weight adjustment to make the feature importance score targeted and interpretable; specifically including: First, introduce a set of engineering experience weights to adjust the feature importance score, represented as: wherein, is a feature importance score for the th feature; is an engineering experience weight; is a revised feature importance score; Then, introduce target correlation correction, combine the optimization goal to perform secondary correction on the feature importance, obtain the secondary corrected feature importance score, and obtain the final feature importance score after normalization.

5. A deep learning-based grouting parameter selection and optimization system, characterized in that, The method comprises the following steps: a data acquisition module configured to obtain and preprocess current geological condition and construction condition data; a feature screening module configured to extract features from the preprocessed geological condition and construction condition data, and perform dimensionality reduction on the extracted features to obtain low-dimensional feature data; perform preliminary importance evaluation on the low-dimensional feature data to obtain feature importance scores; The global optimization module is configured to input a feature with a feature importance score higher than a set threshold to a pre-trained global optimization model to obtain a grouting effect score; the global optimization model is constructed by using a multilayer perception mechanism and includes an input layer, a plurality of hidden layers, a Dropout layer, and an output layer, and the Dropout layer is added after each hidden layer of the multilayer perception mechanism; The local adjustment module is configured to input the grouting effect score and feature data corresponding to the grouting effect to a pre-trained local fine-tuning model to obtain a fine-tuning value of a grouting parameter; The local fine-tuning model is constructed by using a lightweight convolutional neural network, receives the grouting effect score and the feature data corresponding to the grouting effect, sets an optimization goal to maximize the grouting effect score, and a target function is represented as: wherein, is the number of training samples, is the local score obtained after model optimization, is the global score of grouting effect; Then the local fine-tuning model adjusts the input feature and passes through a plurality of hidden layers of a multilayer full connection network to finally output the fine-tuning value of each grouting parameter, which is represented as: ， wherein, represents the output layer weight matrix, respectively represent fine adjustment values of the grouting pressure and the grouting time, represents the output layer weight matrix, represents the final output of the hidden layer, is a bias quantity of the output layer.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps in the grouting parameter selection and optimization method based on deep learning in any one of claims 1-4.

7. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps in the grouting parameter selection and optimization method based on deep learning in any one of claims 1-4.

Images