High-permeability stratum shield slurry automatic preparation method based on deep learning model

Abstract

The invention discloses an automatic preparation method of high-permeability stratum shield slurry based on a deep learning model, and belongs to the technical field of high-permeability stratum shield slurry, and the method comprises the following steps: S1, designing slurry parameters by adopting an orthogonal design method, and selecting three factors of effective particle size, specific gravity and viscosity of slurry for research; the effective particle size of the slurry is prepared from pure bentonite slurry with the swelling-water ratio of 1: 12 and particle materials with different particle sizes; s2, carrying out penetration test on the prepared slurry sample in stratums with different permeability coefficients, and recording the change condition of the water filtering amount along with time, the type of a formed mud film and the thickness of the mud film; by establishing the quantitative evaluation index based on the deep learning model, the mud film performance can be accurately and objectively evaluated, and a scientific basis is provided for mud preparation. And in combination with quantitative evaluation indexes and a deep learning model, automatic optimization of the slurry formula is realized, the slurry preparation efficiency is improved, and the labor cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of shield tunneling mud technology in high-permeability strata, specifically to an automatic preparation method for shield tunneling mud in high-permeability strata based on a deep learning model. Background Technology

[0002] In strata with high groundwater pressure and high permeability, tunnel boring machine (TBM) construction faces challenges such as large mud loss and difficulty in forming an effective mud film. This leads to decreased sealing of the mud chamber, inability to maintain mud pressure, and ultimately, serious consequences such as instability at the excavation face. Therefore, accurately evaluating mud quality and preparing mud that meets engineering requirements is crucial.

[0003] Currently, scholars both domestically and internationally mainly focus on the influence of mud properties on mud film formation, and have achieved certain results. For example, studies have found that mud density, sand content, and additives have a significant impact on mud filtration loss; mud film quality evaluation methods include ultimate support pressure, filtration rate, effective stress ratio, and air tightness value.

[0004] However, existing technologies still have the following problems: evaluation methods for mud film quality mainly rely on empirical judgment and comparative experiments, lacking quantitative analysis methods based on artificial intelligence technologies such as deep learning. They are mostly qualitative descriptions or empirical judgments, lacking objective and quantitative indicators, making it difficult to accurately assess mud film performance. Furthermore, they are often specific to certain geological conditions and difficult to apply to different geological environments, lacking universality. Therefore, an automatic mud preparation method for shield tunneling in high-permeability strata based on a deep learning model is proposed. Summary of the Invention

[0005] This invention aims to address at least one of the technical problems existing in the prior art: the lack of quantitative evaluation indicators, the reliance on single methods, and poor applicability. Therefore, one objective of this invention is to propose an automatic method for preparing shield tunneling mud in high-permeability strata based on a deep learning model.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An automatic mud preparation method for shield tunneling in high-permeability formations based on deep learning models includes the following steps:

[0008] S1. The orthogonal design method was used to design the mud parameters. The effective particle size, specific gravity and viscosity of the mud were selected for study. The effective particle size of the mud was selected by using pure bentonite mud with a swelling ratio of 1:12 and supplemented with granular materials of different particle sizes.

[0009] S2. Conduct permeability tests on the prepared mud samples in formations with different permeability coefficients, and record the changes in filtration volume over time, the type of mud film formed, and the thickness of the mud film;

[0010] S3. From the physical parameters and permeability test data of the mud samples, use regression analysis and t-test to find statistically significant features;

[0011] S4. Using a deep learning model based on the Transformer architecture, train the selected features to establish a mapping relationship between mud filtration volume and mud parameters;

[0012] S5. Automatically prepare drilling mud that meets actual requirements based on formation permeability coefficient and model prediction results.

[0013] Furthermore, in step S1, the characteristics of density, funnel viscosity, and effective particle size are used as independent variables, while the filtration volume under the first-level load and the filtration volume under the sixth-level load are used as dependent variables.

[0014] Furthermore, in step S3, there are n observations, p independent variables, and one dependent variable. The basic model of multivariate regression is as follows:

[0015] Y i =β0+β1X 1i +β 2i X 2i +…+β ni X ni +∈ (4)

[0016] Where Yi is the dependent variable of the i-th observation, and X 1i ,X 2i ,…,X ni Let β0, β1, β2, ..., β be the independent variable for the i-th observation. n ∈ represents the regression coefficient, and ∈ represents the error term.

[0017] Furthermore, the independent samples t-test for two groups of samples is calculated using the following formula:

[0018]

[0019] in: and These are the means of the two groups of samples, and s is the pooled standard deviation, calculated as follows:

[0020]

[0021] Where s1 and s2 are the standard deviations of the two groups of samples, and n1 and n2 are the sample sizes of the two groups of samples.

[0022] Furthermore, in step S4, during the neural network parameter initialization phase, the weights between each layer of the Transformer are initialized, and the number of input layers, encoder layers, decoder layers, and output layers is selected to determine each parameter;

[0023] During the Transformer model training process, the preprocessed training data sample set is input into the Transformer for training. The model output results are obtained and continuously compared and analyzed with the set ideal values. The network parameters are repeatedly optimized until the specified requirements are met, at which point training stops.

[0024] Furthermore, in step S5, after the model training is completed, the preprocessed test data sample set is input into the trained Transformer model for testing to obtain the corresponding prediction results. The prediction results are then compared and analyzed with the actual values ​​to predict the first-level load filtration volume and the sixth-level load filtration volume, thereby evaluating the matching between the mud and the formation.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] This invention establishes quantitative evaluation indicators based on deep learning models, enabling accurate and objective assessment of mud film performance and providing a scientific basis for mud formulation. Furthermore, by combining quantitative evaluation indicators and deep learning models, it achieves automatic optimization of mud formulations, improving mud formulation efficiency and reducing labor costs.

[0027] This invention employs the Transformer architecture, which can effectively process mud effluent volume data under different formation conditions, exhibits good generalization ability, and is applicable to various high-permeability formation environments. Based on extensive experimental data, a relationship model between mud effluent volume and mud parameters is established through regression analysis and Transformer model training. Furthermore, the Transformer architecture is introduced into the field of mud quality assessment, leveraging its powerful feature extraction and sequence modeling capabilities to improve the accuracy and efficiency of mud effluent volume prediction. Attached Figure Description

[0028] Figure 1 This is a particle distribution curve of mud slurry according to the present invention;

[0029] Figure 2 This is a graph showing the change in filtration rate of mud in the No. 1 formation during the present invention.

[0030] Figure 3 This is a graph showing the change in filtration rate of mud in the No. 2 formation during the present invention.

[0031] Figure 4 This is a diagram of the surface of the formation blocked by mud particles, as described in this invention.

[0032] Figure 5 The diagram shows the mud film formed by the other six types of mud in the present invention in Formation No. 2;

[0033] Figure 6This is a graph showing the change in filtration volume of mud in the No. 3 formation during the permeability test of the present invention.

[0034] Figure 7 This is a diagram showing the mud film formed by the remaining mud slurry of the present invention in Formation No. 3;

[0035] Figure 8 These are diagrams of the mud films formed by mud slurries No. 6 to No. 9 of the present invention;

[0036] Figure 9 This is a graph showing the change in filtration volume of mud in the No. 5 formation of the present invention.

[0037] Figure 10 This is a diagram of the mud film formed by mud slurries No. 8 and No. 9 of the present invention;

[0038] Figure 11 This is a diagram showing the Transformer fitting results for the training set, test set, and validation set of this invention.

[0039] Figure 12 This is the histogram of error distribution of the Transformer test set in this invention;

[0040] Figure 13 This is a diagram of the Transformer architecture of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1

[0043] Orthogonal design was used to design the mud parameters, focusing on three factors: effective particle size, specific gravity, and viscosity. The effective particle size was determined by preparing pure bentonite mud (base mud) with a swelling-to-water ratio of 1:12, supplemented with granular materials of different particle sizes. The values ​​of each factor are shown in Table 1, the basic properties of the mud used in the experiment are shown in Table 2, and the particle distribution curves are attached. Figure 1 As shown.

[0044]

[0045] Table 1. Values ​​of each factor

[0046] Test No. A B C <![CDATA[Density (g / cm 3 )]]> Funnel viscosity (s) <![CDATA[Effective particle size d 85 (μm)]]> 1 A1 B1 C1 1.1 49 78 2 A1 B2 C2 1.2 87 82 3 A1 B3 C3 1.3 158 85 4 A2 B1 C2 1.1 75 155 5 A2 B2 C3 1.2 139 162 6 A2 B3 C1 1.3 46 170 7 A3 B1 C3 1.1 129 260 8 A3 B2 C1 1.2 39 270 9 A3 B3 C2 1.3 77 282

[0047] Table 2 Basic property parameters of mud (where the effective particle size of mud is the particle size at 85% of the cumulative particle size curve of mud).

[0048] Regarding the selection of the formation used in the experiment:

[0049] The stratum used was highly permeable sand. All sand used in the experiment was sand with a single particle size range, which was obtained by sieving natural sand. The particle size range and permeability coefficient of the bottom layer particles are shown in Table 3.

[0050] Method for determining the formation permeability coefficient: Sand and soil are strictly layered and placed into a test container according to a controlled dry density of 1.5 g / cm3, and a constant head test is conducted to obtain the coefficient.

[0051] Stratigraphic sequence 1 2 3 4 5 Particle size range (mm) 0.075～0.25 0.25～0.5 0.5～1 1～2 2～5 Formation permeability coefficient (cm / s) 0.012 0.051 0.22 0.81 3.3 The stratigraphic representative grain size D15 (mm) 0.11 0.32 0.64 1.22 2.69 Average formation pore size D0 (μm) 20.9 60.8 121.6 232 511

[0052] Table 3. Grain size range and permeability coefficient of five strata

[0053] Regarding the test equipment and methods:

[0054] Step 1: Conduct mud permeation film formation test using the test device. Before the test begins, a medium-coarse sand layer with a height of 5cm and a particle size of 2-5mm is placed at the bottom of the permeation column as a filter layer, followed by a sand layer with a height of 27cm.

[0055] Step 2: Use the forward saturation method, which involves first injecting water into the permeation column and then loading the filter layer and formation, to saturate the filter layer and formation, and control the dry density of the formation to 1.56 g / cm3;

[0056] Step 3: Then, inject mud to a certain height, seal the flange, apply the required air pressure to the permeation column through an air compressor and pressure stabilizing valve, open the connecting valve at the bottom of the permeation column, and start the permeation test.

[0057] Step 4: Apply mud pressure in a graded loading manner. There are six levels of mud pressure. After the filtration rate stabilizes at each level, the next level is added. The six levels are 0.05MPa, 0.1MPa, 0.15MPa, 0.2MPa, 0.25MPa and 0.3MPa. The purpose of the first level of mud pressure is to determine whether the mud can form a film in this formation. The purpose of the next five levels of pressure is to evaluate the quality of the mud film formed in this formation.

[0058] Nine types of mud as shown in Table 2 were used to conduct mud filtration tests in five types of formations as shown in Table 3, for a total of 45 sets of tests. The changes in filtration volume over time, the type of mud film formed, and the thickness of the mud film were recorded during the tests.

[0059] Regarding the experimental results and analysis:

[0060] For formation No. 1 with a permeability coefficient of 0.012 cm / s:

[0061] Nine types of drilling mud, as shown in Table 2, were used in permeability tests in Formation No. 1, which has a permeability coefficient of 0.012 cm / s. The curves showing the change in filtration rate per unit area over time for each of the nine types of drilling mud in this formation are attached. Figure 2 As shown.

[0062] The infiltration tests of the nine types of drilling mud in Formation No. 1 showed very similar processes and results: after each load level was applied, the filtration volume was small, and the filtration volume stabilized rapidly; the total filtration volume under six load levels was small; ultimately, all nine types of drilling mud formed a thin layer of solid particles on the formation surface, such as... Figure 3 As shown. By Figure 3 It can be seen that the mud with the best adaptability to this stratum is mud No. 3.

[0063] For formation No. 2 with a permeability coefficient of 0.051 cm / s:

[0064] Nine types of drilling mud, as shown in Table 4.1-2, were used to conduct infiltration tests in Formation No. 2, which has a permeability coefficient of 0.051 cm / s. All particles in this formation had a particle size between 0.25 and 0.5 mm (dry density of the formation was 1.5 g / cm³), and the formation permeability coefficient was 0.051 cm / s. The curves showing the change in filtration rate per unit area over time for the nine types of drilling mud in this formation are attached. Figure 3 As shown.

[0065] For the infiltration tests of mud types 7, 8, and 9:

[0066] During the mud filtration test, muds No. 7, 8, and 9 produced a significant amount of filtration water even under the first-level load, with some mud seeping into the formation. Subsequently, as the mud pressure increased, the amount of filtration water gradually decreased, and the stabilization rate was relatively fast. Ultimately, some mud particles seeped into the formation, while a large amount of fine sand particles from the mud blocked the formation surface, as shown in the attached image. Figure 4 As shown, this indicates that the mud particle size distribution is poor and its adaptability to the formation is poor.

[0067] For the remaining 6 types of mud percolation tests:

[0068] The test procedures and results for the other six types of mud were the same as those for Formation No. 2, all of which formed a thin layer of solid particles on the formation surface, as shown in the attached figure. Figure 5 As shown.

[0069] As the effective particle size of the mud increases, the latter six types of mud can only form mud cake-type mud films on the formation surface. The filtration rate and the permeability coefficient of the mud film during the formation process both decrease. When the particle size of the solid phase particles in the mud is too large, the filtration rate and the permeability coefficient of the mud film during the formation process both increase. Among the nine types of mud, mud No. 3 forms the best quality mud film, that is, the filtration loss of the mud during the formation process and the permeability coefficient of the mud film after formation are both the smallest. In other words, mud No. 3 is the best match with formation No. 2.

[0070] For formation No. 3 with a permeability coefficient of 0.22 cm / s:

[0071] Nine types of drilling mud, as shown in Table 4.1-2, were used in permeability tests in Formation No. 3, which has a permeability coefficient of 0.22 cm / s. The curves showing the change in filtration rate per unit area over time for each of the nine types of drilling mud in this formation are attached. Figure 6 As shown in Table 4.2-7, the total amount of water filtered per unit area under Level 1 and Level 6 loads during the test is shown in Table 4.2-7.

[0072] For mud types 1, 7, 8, and 9:

[0073] During the mud filtration test of muds No. 7, 8, and 9 in stratum No. 3, the permeation was the same as that in stratum No. 2. A layer of fine sand accumulated on the surface of the stratum, and the permeation of this stratum was greater, with all the mud permeating. Mud No. 1 produced a large amount of filtration water as a first-level load was applied. Subsequently, each level of load application caused a large amount of filtration water. Under the last level of load, the mud permeated completely, and the filtered water was turbid with no mud film formation.

[0074] For the remaining mud percolation tests:

[0075] The experimental process and results of muds No. 2, 3, 4, 5, and 6 show that under pressure, some particles in the mud can penetrate into the formation, while others can accumulate on the formation surface, forming... Figure 7 The mud film shown.

[0076] For formation No. 4 with a permeability coefficient of 0.81 cm / s:

[0077] Nine types of mud, as shown in Table 5, were used to conduct mud filtration tests in Formation No. 4, which has a permeability coefficient of 0.81 cm / s. The total amount of water filtered per unit area under Level 1 and Level 6 loads during the tests is shown in Table 4.2-10.

[0078] For mud percolation tests No. 1 to No. 5:

[0079] The test process and results of mud filtration tests No. 1 to No. 5 were similar to those of mud No. 1. The difference was that the mud was completely filtered out under the first load, and the five types of mud failed to form any type of mud film in the No. 4 formation.

[0080] For mud percolation tests No. 6-9:

[0081] The test procedures and results of the infiltration tests for mud slurries 6-9 were similar to those for mud slurry 1. Under pressure, some particles in the mud slurry could penetrate into the formation, while others could remain on the formation surface, forming a mud film. The cross-sectional diagram of mud slurry 9 in formation 4 is shown below. Figure 8 As shown in the figure, the mud particles tightly seal the formation pores. Therefore, the No. 9 mud has the smallest water filtration volume in this formation, meaning that the No. 9 mud is the most compatible with the No. 4 formation.

[0082] For formation No. 5 with a permeability coefficient of 3.3 cm / s:

[0083] The curves showing the change in filtration rate per unit area over time for the nine types of drilling mud in Formation No. 5 are attached. Figure 9 As shown.

[0084] For mud percolation tests No. 1 to No. 7:

[0085] The test process and results of mud filtration tests No. 1 to No. 7 were similar to those of mud No. 1. The difference was that mud No. 1, 2, 3, 4, 5 and 7 were completely filtered out under the first load, while mud No. 6 was completely filtered out under the second load. None of the seven types of mud formed any type of mud film in the No. 5 formation.

[0086] For the infiltration tests of mud No. 8 and No. 9:

[0087] The test procedures and results of mud slurry infiltration tests No. 8 and No. 9 were similar to those of mud slurry No. 1. Under pressure, some particles in the mud could penetrate into the formation, while others could remain on the formation surface, forming... Figure 10 The mud film shown.

[0088] Analysis of factors affecting mud quality:

[0089] Density (g / cm³), funnel viscosity (s), and effective particle size d85 (μm) were used as independent variables, while the filtration rate under level 1 load (m³ / m²) and level 6 load (m³ / m²) were used as dependent variables. Regression analysis and t-tests were used to identify statistically significant features, and then a backpropagation (BP) neural network was used to fit the selected statistical features. The percentages of the training set, validation set, and test set were 70%, 15%, and 15%, respectively.

[0090] Specifically, for multivariate regression and T-test:

[0091] Multivariate regression is used to predict the relationship between one or more continuous dependent variables and one or more independent variables. Given n observations, p independent variables, and one dependent variable, the basic model of multivariate regression is as follows:

[0092] Y i =β0+β1X 1i +β 2i X 2i +…+β ni X ni +∈ (1)

[0093] Where Yi is the dependent variable of the i-th observation, and X 1i ,X 2i ,…,X ni Let β0, β1, β2, ..., β be the independent variable for the i-th observation. n ∈ represents the regression coefficient, and ∈ represents the error term.

[0094] The formula for calculating the independent samples t-test for two groups of samples is as follows:

[0095]

[0096] in: and These are the means of the two groups of samples, respectively. s is the pooled standard deviation, calculated as follows:

[0097]

[0098] Where s1 and s2 are the standard deviations of the two groups of samples, and n1 and n2 are the sample sizes of the two groups of samples.

[0099] For the Transformer architecture:

[0100] The Transformer architecture is used to replace the traditional BP neural network model to predict the amount of mud filtration.

[0101] By introducing the Transformer's self-attention mechanism into a traditional MLP model, the model's ability to capture input features is enhanced, thereby improving prediction accuracy. This approach leverages the Transformer's strengths in processing sequential data while maintaining the simplicity and efficiency of the MLP model.

[0102] The Transformer model consists of an encoder and a decoder. The encoder is composed of multiple identical encoder layers stacked together, each containing two sub-layers: a multi-head self-attention mechanism and a feedforward neural network. The decoder is also composed of multiple identical decoder layers stacked together, each containing three sub-layers: a multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network.

[0103] In the encoder structure, input embedding transforms the input sequence into a fixed-dimensional embedding vector, while positional encoding adds positional information, enabling the model to capture the sequence order. The multi-head self-attention mechanism extracts global features by calculating the self-attention weights at each position in the input sequence. The feedforward neural network applies two linear transformations and a ReLU activation function to further process the extracted features.

[0104] The input sequence consists of various parameters of the mud, including the effective particle size (x1), the mud funnel viscosity (x2), the mud filtration loss after seepage stabilization under level 6 loading (x3), density (x4), and funnel viscosity (x5). These parameters form the input vector:

[0105] X = [x1, x2, x3, x4, x5]

[0106] Transform the input vector into a fixed-dimensional embedding vector:

[0107] E = Embedding(X)

[0108] Adding positional information to the embedding vectors enables the model to capture the sequence's order information:

[0109] E pos =E + PositionalEncoding(E)

[0110] The encoder layer incorporates a multi-head self-attention mechanism and a feedforward neural network. There are four encoder layers, and the output of each encoder layer is represented as follows:

[0111] H i =EncoderLayer i (H i-1 )

[0112] Among them, H 0 =E pos , 0≤i≤4.

[0113] Therefore, the output of the last encoder layer is the encoder output:

[0114] H 4 =EncoderOutput

[0115] The decoder takes a masked version of the target sequence as input, used during training to ensure that the model only sees previous values ​​when generating the current value. Assume the target sequence is Y, and its masked version is Y0. shift The decoder layer includes a multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network. There are four decoder layers, and the output of each decoder layer can be represented as:

[0116] Z j =DecoderLayer j (Z j -1,H 4 )

[0117] Among them, Z 0 =Embedding(Y shifted )+PositionalEncodingY shifted ), 0≤j≤4. The decoder output is transformed into the final prediction result through a linear transformation:

[0118] Output = Linear(Z) 4 )

[0119] Regression Model Results and Validation:

[0120] Before performing multiple regression and t-test, we first performed a batch difference analysis on the quantitative data, as shown in the table below:

[0121]

[0122]

[0123] Table 4. Batch Difference Analysis of Quantitative Data

[0124] Table 4 presents the batch variance analysis results of quantitative data on filtration volume under Level 1 load for different mud film types, formations, and mud densities. The analysis shows that mud film type and formation have a significant impact on filtration volume, while mud density does not show a significant effect. This means that mud permeability varies significantly under different mud film types and formations, but changes in mud density have little impact on filtration volume. Furthermore, the batch variance analysis results for Level 6 load are less than 0.1 different from those for Level 1 load.

[0125] In multiple regression analysis, R 2 The R-value can assess the explanatory power of the model, understand the contribution of explanatory variables to the dependent variable, and thus determine the reliability and applicability of the model. Table 4 shows the regression results for Level 1 loadings; Table 5 shows the regression results for Level 6 loadings, where R... 2 Value: 0.1664; F-statistic: 1.9966; p-value: 0.1135; regression standard error: 0.0047.

[0126]

[0127]

[0128] Table 5 Regression Analysis and T-Test Results for Level 1 Load

[0129] Regression analysis showed that the effective particle size of the mud, the mud funnel viscosity, the mud filtration loss after seepage stabilization under stage 6 load, density, and funnel viscosity all significantly affected the filtration rate under stage 1 load. The model's R² value... 2 The value is 1.9966, demonstrating that the model has good explanatory power. The F-statistic and p-value indicate that the model is significant overall, and the small regression standard error indicates that the model has small prediction error.

[0130]

[0131] Table 6. Regression Analysis and T-Test Results for Level 6 Load

[0132] In Table 6, R 2 Value: 0.0903; F-statistic: 0.9927; p-value: 0.4227; regression standard error: 0.0052.

[0133] Comparative analysis of regression results under Level 1 and Level 6 loads reveals that the effective particle size and funnel viscosity of the mud significantly affect the filtration rate under both loads, but the impact is greater under Level 1 load. This indicates that these factors have a more pronounced effect on mud performance under initial pressure conditions. Under Level 6 load, the filtrate loss after seepage stabilization significantly affects the filtration rate under both loads, but the effect is more significant under Level 1 load. Density and effective particle size d85 have no significant effect under either load. These results provide important references for optimizing mud performance, especially under initial pressure conditions, where the control of the effective particle size and funnel viscosity should be a key focus.

[0134] Regression analysis under level one load showed a high goodness of fit (R0). 2 =0.1664), while regression analysis under level six loading showed a higher goodness of fit (R² = 0.1664). 2 =0.0903), indicating that the model has stronger explanatory power under level 6 loading. Meanwhile, the F-statistic and p-value under level 6 loading show higher model significance and lower regression standard error, further validating the model's reliability and robustness.

[0135] Transformer model results and validation:

[0136] As attached Figure 13 As shown, the Transformer architecture is used to predict mud filtration rate. Input data includes effective mud particle size, mud funnel viscosity, mud filtration loss after seepage stabilization under six levels of loading, density, and funnel viscosity. These data are preprocessed, normalized, and divided into training and testing sets.

[0137] During the neural network parameter initialization phase, the weights of each layer of the Transformer are initialized, and the number of input, encoder, decoder, and output layers is selected. Parameters such as the activation function, number of training iterations, and learning rate are also determined. During Transformer model training, the preprocessed training data sample set is input into the Transformer for training. The model output results are obtained and continuously compared and analyzed with the set ideal values. By repeatedly optimizing various network parameters, training stops when the specified requirements are met.

[0138] After model training is complete, the preprocessed test data sample set is input into the trained Transformer model for testing to obtain the corresponding prediction results. The prediction results are then compared and analyzed with the actual values. This allows for accurate prediction of the first-stage load filtration rate and the sixth-stage load filtration rate, thereby evaluating the matching between the mud and the formation.

[0139] The fitted data in each set fluctuates slightly around the true value until more and more training data is collected, the training results tend to approach the true value, and the accuracy of the obtained model gradually increases. The error histogram of the test set is attached. Figure 12 As shown in the figure, the error distribution is around 0, and the maximum absolute value of the error is only 0.03047, indicating that the prediction results are relatively good.

[0140] All parts not described in this invention are the same as or can be implemented using existing technology. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automatic method for preparing shield tunneling mud in high-permeability strata based on a deep learning model, characterized in that, Includes the following steps: S1. The orthogonal design method was used to design the mud parameters. The effective particle size, specific gravity and viscosity of the mud were selected for study. The effective particle size of the mud was selected by using pure bentonite mud with a swelling ratio of 1:12 and supplemented with granular materials of different particle sizes. S2. Conduct permeability tests on the prepared mud samples in formations with different permeability coefficients, and record the changes in filtration volume over time, the type of mud film formed, and the thickness of the mud film; S3. From the physical parameters and permeability test data of the mud samples, use regression analysis and t-test to find statistically significant features; S4. Using a deep learning model based on the Transformer architecture, train the selected features to establish a mapping relationship between mud filtration volume and mud parameters; S5. Automatically prepare drilling mud that meets actual requirements based on formation permeability coefficient and model prediction results.

2. The automatic preparation method for shield tunneling mud based on a deep learning model in high-permeability strata according to claim 1, characterized in that: In step S1, the characteristics of density, funnel viscosity, and effective particle size are used as independent variables, and the filtration volume under the first-level load and the filtration volume under the sixth-level load are used as dependent variables.

3. The automatic preparation method for shield tunneling mud based on a deep learning model in high-permeability strata according to claim 1, characterized in that: In step S3, there are n observations, p independent variables, and one dependent variable. The basic model of multivariate regression is as follows: Y i =β0+β1X 1i +b 2i X 2i +…+b ni X ni +∈ (1) Where Yi is the dependent variable of the i-th observation, and X 1i ,X 2i ,…,X ni Let β0, β1, β2, ..., β be the independent variable for the i-th observation. n ∈ represents the regression coefficient, and ∈ represents the error term.

4. The automatic preparation method for shield tunneling mud based on a deep learning model according to claim 3, characterized in that: The formula for calculating the independent samples t-test for two groups of samples is as follows: in: and These are the means of the two groups of samples, and s is the pooled standard deviation, calculated as follows: Where s1 and s2 are the standard deviations of the two groups of samples, and n1 and n2 are the sample sizes of the two groups of samples.

5. The automatic preparation method for shield tunneling mud based on a deep learning model according to claim 1, characterized in that: In step S4, during the neural network parameter initialization phase, the weights between each layer of the Transformer are initialized, and the number of input layers, encoder layers, decoder layers, and output layers is selected to determine each parameter. During the Transformer model training process, the preprocessed training data sample set is input into the Transformer for training. The model output results are obtained and continuously compared and analyzed with the set ideal values. The network parameters are repeatedly optimized until the specified requirements are met, at which point training stops.

6. The automatic preparation method for shield tunneling mud based on a deep learning model according to claim 1, characterized in that: In step S5, after the model training is completed, the preprocessed test data sample set is input into the trained Transformer model for testing to obtain the corresponding prediction results. The prediction results are then compared and analyzed with the actual values ​​to predict the first-level load filtration rate and the sixth-level load filtration rate, thereby evaluating the matching between the mud and the formation.

Images