Ground stress field intelligent inversion method based on deep learning algorithm

Abstract

The invention relates to the technical field of crustal stress field inversion, in particular to a crustal stress field intelligent inversion method based on a deep learning algorithm. Comprising the following steps: S1, establishing a three-dimensional geomechanical model; s2, setting a plurality of undetermined influence factors, and respectively acting on the three-dimensional geomechanical model; s3, setting a stress component for each undetermined influence factor, and obtaining a training sample and a verification sample by taking the undetermined influence factor as an independent variable and the stress component as a dependent variable; s4, a CNN-LSTM-Attention deep learning model is constructed; s5, training the deep learning model based on the training sample, taking the stress component as input, taking the combined working condition as output, and obtaining a crustal stress inversion model after training; s6, based on the verification sample, performing prediction precision verification on the crustal stress inversion model, and when the prediction precision verification is not qualified, iterating the steps S1 to S5; and reliable and large-range ground stress field distribution characteristics are obtained.

Description

Technical Field

[0001] This invention relates to the field of geostress field inversion technology, and in particular to an intelligent geostress field inversion method based on deep learning algorithms. Background Technology

[0002] In-situ stress data is a crucial parameter for analyzing the stability of rock masses in underground engineering projects and for engineering design. It is a natural stress that is widely present in the strata before construction. The in-situ stress field is a major factor inducing various geological disasters in underground engineering projects. Therefore, accurately understanding the distribution patterns of the in-situ stress field in the engineering area is of great significance for revealing the stress and displacement field distribution of the surrounding rock during the construction of underground caverns and for studying the stability of the surrounding rock.

[0003] The inversion of rock mass stress fields aims to accurately obtain the stress state of underground rock masses and provide a primary basis for the stability analysis and evaluation of surrounding rock during the construction period of underground engineering projects, as well as for engineering optimization design. Currently, many methods are commonly used to obtain stress information, such as in-situ borehole measurements. However, applying these methods to underground engineering faces several unresolved issues. For example, in-situ core sampling technology is underdeveloped and costly, making large-scale borehole measurements in the engineering area impossible, thus limiting the acquisition of substantial measured stress data. Furthermore, the widespread distribution of geological structures in underground engineering areas results in significant localization of borehole measurement data, failing to reflect the overall initial stress field distribution characteristics of the engineering area. Existing stress field inversion methods do not consider all factors comprehensively, leading to low reliability and limited scope of the obtained stress field distribution characteristics.

[0004] Therefore, there is an urgent need to provide an intelligent inversion method for geostress field based on deep learning algorithms, which can obtain reliable and wide-ranging geostress field distribution characteristics compared with existing technologies. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing an intelligent inversion method for geostress field based on deep learning algorithms.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for intelligent inversion of geostress field based on deep learning algorithm includes the following steps: S1. Establish a three-dimensional geomechanical model based on the topographic and geological survey data of the project area; S2. Set multiple displacement boundary conditions, and set each displacement boundary condition as an undetermined influencing factor. In FLAC3D, apply the set multiple displacement boundary conditions to the three-dimensional geomechanical model. S3. For each undetermined influencing factor, set corresponding stress measurement points to obtain the stress components at each stress measurement point. With the undetermined influencing factor as the independent variable and the stress components as the dependent variable, construct a multivariate linear equation. Eliminate insignificant undetermined influencing factors to obtain a stepwise regression equation with significant regression effect. Obtain multiple sets of combined working conditions of displacement boundary conditions and gravitational acceleration, as well as the corresponding stress components calculated by the stepwise regression equation. Use some combined working conditions and their corresponding stress components as training samples, and use the remaining combined working conditions and their corresponding stress components as validation samples. S4. Construct a CNN-LSTM-Attention deep learning model; S5. Based on the training samples, the CNN-LSTM-Attention deep learning model is trained. During the training process, the stress components are used as inputs and the combined working conditions are used as outputs. After training, the ground stress inversion model is obtained. S6. Based on the validation samples, verify the prediction accuracy of the geostress inversion model. If the prediction accuracy verification fails, iterate through steps S1-S5.

[0007] Furthermore, the method for verifying the qualification of the geostress inversion model in step S6 is as follows: substitute the stress components in the verification sample into the model, invert the corresponding combined working conditions, calculate the first evaluation index and the second evaluation index, set the first evaluation threshold and the second evaluation threshold, and if the first evaluation index is less than the first evaluation threshold and / or the second evaluation index is greater than the second evaluation threshold, it is considered unqualified; otherwise, it is considered qualified.

[0008] Furthermore, the first evaluation indicator is the coefficient of determination, and the second evaluation indicator is the mean absolute percentage error.

[0009] Furthermore, the coefficient of determination is calculated using the following formula: ; In the above formula, The coefficient of determination is represented by the coefficient of determination. This represents the stress component in the i-th verification sample. This represents the stress components output by the geostress inversion model for the i-th validation sample. This represents the average value of the stress components in all verification samples.

[0010] Furthermore, the mean absolute percentage error is specifically calculated using the following formula: ; In the above formula, This represents the mean absolute percentage error. This represents the total number of validation samples.

[0011] Furthermore, the CNN-LSTM-Attention deep learning model includes a CNN network structure, a long short-term memory neural network, and an attention mechanism. The CNN network structure, the long short-term memory neural network, and the attention mechanism process the input data in sequence to obtain the output results.

[0012] Furthermore, the CNN grid structure is represented as: ; In the above formula, This represents the activation function. Represents convolutional feature mapping, Represents the weight matrix. This represents the convolution operation. Indicates the input item. This indicates the offset.

[0013] Furthermore, the Long Short-Term Memory (LSTM) neural network is represented as: ; ; ; ; ; ; In the above formula, Represents the Gate of Oblivion This represents the sgmoid function. The weight matrix represents the forget gate. This indicates the hidden state in the previous moment. Indicates the current input time. This represents the bias vector of the forget gate. Indicates the input gate. This represents the weight matrix of the input gate. This represents the bias vector of the input gate. Indicates the current state of the candidate cell. Indicates the use of generating The weight matrix, Indicates the use of generating The bias vector, This represents the updated state of the memory cell at time t. Indicates update The state of the memory unit at any given moment. Indicates the output gate. This represents the weight matrix of the output gate. This represents the bias vector of the output gate. Indicates the hidden state in the next moment. This indicates that elements are accumulated one by one.

[0014] Furthermore, the attention mechanism can be represented as: ; ; ; In the above formula, This represents the attention probability distribution value obtained from the LSTM layer output at time t. , Both represent weight coefficients used for linearly combining the outputs of the LSTM layer, and b represents the bias parameter. This represents the attention-standardized weights. This represents the attention probability distribution value obtained from the LSTM layer output at time n, where n ranges from 1 to t. This represents the output of the attention mechanism layer at time t; The input to the attention mechanism is the hidden state obtained in the Long Short-Term Memory neural network for the next time step.

[0015] Furthermore, in step S3, the specific method for eliminating insignificant undetermined influencing factors is as follows: for the multiple linear equation, each undetermined influencing factor is introduced one by one using stepwise regression, and the significance of different undetermined influencing factors in the regression equation is tested. When the significance level of the undetermined influencing factor is not greater than 0.05 and / or the VIF value of the collinearity index is not greater than 10, the undetermined influencing factor is considered insignificant and is eliminated.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention can deeply explore the nonlinear characteristic relationship between displacement boundary conditions and geostress values ​​at measuring points, resulting in simulation effects that are closer to measured values. It boasts high computational efficiency and a simple modeling method, allowing the inversion calculation results to be applied to subsequent engineering construction analysis. This invention considers various constituent factors involved in the formation of the initial geostress field of the rock mass, such as topography and lithology, while also taking into account the complexity of geological conditions in different engineering areas and the degree of influence of tectonic movements on geostress. It selects appropriate scientific inversion analysis methods to obtain reliable and wide-ranging geostress field distribution characteristics. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are not all embodiments of the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0019] like Figure 1 As shown, this invention provides an intelligent inversion method for geostress field based on deep learning algorithms, comprising the following steps: S1. Based on the topographic and geological survey data of the project area, establish a three-dimensional geomechanical model.

[0020] In order to make the created three-dimensional geomechanical model conform to the actual engineering, the boundary of the model is constructed according to certain principles: first, the selected model calculation range should include the main engineering area; second, the selected model boundary line should be clear and unambiguous, and should not be too complex.

[0021] The model construction process should adhere to the basic principles of geometric simulation and constitutive simulation, fully considering the influence of topographic and geological structures within the engineering area, such as faults and joints, on the initial geostress field of the rock mass, as well as the lithological range of each stratum. A topographic model should be constructed in Rhino based on the contour lines of the engineering area, while a three-dimensional geomechanical model containing the main cavern structure should be constructed in Civil 3D.

[0022] S2. Based on the geological conditions of the engineering area, six displacement boundary conditions, including rock mass self-weight, compression, and shear motion, are considered as undetermined influencing factors. The three-dimensional geomechanical model is imported into HyperMesh for discretization to obtain a mesh model for calculation. The mesh model is then imported into FLAC3D. Given the lithological parameters of each stratum, a set constitutive model is selected for numerical simulation. In FLAC3D, the selected displacement boundary conditions and the set number of calculation steps are applied to the three-dimensional geomechanical model to achieve the set unit displacement load. The self-weight stress is then calculated using the gravitational acceleration given in the three-dimensional geomechanical model.

[0023] S3. For each undetermined influencing factor, stress measurement points are set up to obtain the stress components of the corresponding stress measurement points. Using different undetermined influencing factors as independent variables and the stress components of their corresponding stress measurement points as dependent variables, a multiple linear equation is constructed. Each undetermined influencing factor is introduced one by one using a stepwise regression method, and the significance of different undetermined influencing factors in the regression equation is tested. When the significance level of an undetermined influencing factor is not greater than 0.05 and / or the VIF value of the collinearity index is not greater than 10, the undetermined influencing factor is considered insignificant. Insignificant undetermined influencing factors are removed to obtain a stepwise regression equation with significant regression effect. Multiple sets of combined working conditions of displacement boundary conditions and gravitational acceleration are obtained, and the stress components corresponding to each set of combined working conditions are calculated through the regression equation. Some combined working conditions and their corresponding stress components are used as training samples, and the remaining combined working conditions and their corresponding stress components are used as validation samples.

[0024] S4. Construct a CNN-LSTM-Attention deep learning model, including a CNN network structure, a long short-term memory neural network, and an attention mechanism. The CNN network structure, the long short-term memory neural network, and the attention mechanism process the input data in sequence to obtain the output results.

[0025] (1) The CNN network structure is a deep feedback-forward neural network. Its core part is to extract local patterns and features of the data through convolutional layers, and then to perform dimensionality reduction on the output of the convolutional layers through pooling layers. This reduces the number of parameters and thus reduces the computation speed, while retaining important information of the data. Finally, the retained important features are integrated through fully connected layers for the final classification or regression task. The convolutional feature mapping of the CNN network structure is obtained by the following formula: ; In the above formula, This represents the activation function. Represents convolutional feature mapping, Represents the weight matrix. This represents the convolution operation. Indicates the input item. This indicates the offset.

[0026] (2) Long Short-Term Memory Neural Network (LSTM network model) is a special type of Recurrent Neural Network (RNN) model. It has the same input and output as RNNs, but LSTM networks can capture sequence features, thus achieving a more accurate surrogate model. LSTM network models can extract the temporal characteristics of data and unfold them along the time dimension. By introducing memory cells and a gate mechanism that controls the flow of information by changing parameters, it can effectively solve the problems of gradient vanishing and gradient explosion that traditional RNNs encounter during long sequence training. The core of the LSTM neural network is the memory cell, which can learn and store information during model training. The memory cell controls the inflow and outflow of information by using input gates, forget gates, and output gates. Its model network structure is represented by the following formula: ; ; ; ; ; ; In the above formula, Represents the Gate of Oblivion This represents the sgmoid function. The weight matrix represents the forget gate. This indicates the hidden state in the previous moment. Indicates the current input time. This represents the bias vector of the forget gate. Indicates the input gate. This represents the weight matrix of the input gate. This represents the bias vector of the input gate. Indicates the current state of the candidate cell. Indicates the use of generating The weight matrix, Indicates the use of generating The bias vector, This represents the updated state of the memory cell at time t. Indicates update The state of the memory unit at any given moment. Indicates the output gate. This represents the weight matrix of the output gate. This represents the bias vector of the output gate. Indicates the hidden state in the next moment. This indicates that elements are accumulated one by one.

[0027] (3) Attention mechanisms are widely used in neural network algorithms. As an advanced neural network technology, it can simulate the human ability to selectively focus on information, thereby directly capturing the relationship between any two parts of input data. This breaks the limitations of traditional recurrent neural networks and convolutional neural networks in processing long sequences of data, which are confined to local receptive fields or temporal order. It allows for the analysis of the relationships between input sequence data during data processing, assigning corresponding importance weights to the features of different data. It can be applied to LSTM layers. After processing the data, a series of hidden states in the output can be weighted to determine the correlation between each hidden state and the final prediction result, thus determining the importance of the input data and focusing on the information most important to the result. Attention mechanisms have significant advantages in extracting global dependencies and enhancing the expressive power of neural network models. Specifically, attention mechanisms are expressed as follows: ; ;

[0028] In the above formula, Let represent the attention probability distribution value obtained from the output of the LSTM layer at time t. It is a scalar and can represent the importance of the hidden state output by the LSTM layer at time t to the overall attention weight allocation. , Both represent weight coefficients used to linearly combine the outputs of the LSTM layer to calculate the attention score; b represents the bias parameter, which can be used to adjust the baseline value of the attention score, thereby helping the model learn the attention distribution. This represents the attention-standardized weights. This represents the attention probability distribution value obtained from the LSTM layer output at time n, where n ranges from 1 to t. This represents the output of the attention mechanism layer at time t. The input to the attention mechanism is the hidden state obtained in the Long Short-Term Memory neural network for the next time step.

[0029] S5. Input the training samples into the CNN-LSTM-Attention deep learning model, use the stress components in each training sample as input to the deep learning model, and use the combined working conditions as output to train the deep learning model to obtain the ground stress inversion model.

[0030] S6. Using the validation samples, the prediction accuracy of the geostress inversion model obtained in step S5 is validated, and a geostress inversion model with higher prediction accuracy is finally obtained. Specifically, the stress components in the validation samples are substituted into the model to invert the corresponding combined working conditions, thereby judging the prediction accuracy of the geostress inversion model. For the geostress inversion model with higher prediction accuracy, the geostress field distribution in the engineering area is obtained.

[0031] Calculate the first evaluation index and the second evaluation index, and set the first evaluation threshold and the second evaluation threshold. If the first evaluation index is less than the first evaluation threshold and / or the second evaluation index is greater than the second evaluation threshold, it is considered unqualified, and steps S1-S5 are iterated. Otherwise, it is considered qualified, indicating that the geostress inversion model has high accuracy.

[0032] The primary evaluation indicator is the coefficient of determination, which is calculated using the following formula: ; In the above formula, The coefficient of determination is represented by the coefficient of determination. This represents the stress component in the i-th verification sample. This represents the stress components output by the geostress inversion model for the i-th validation sample. This represents the average value of the stress components in all verification samples.

[0033] The second evaluation indicator is the mean absolute percentage error, which is calculated using the following formula: ; In the above formula, This represents the mean absolute percentage error. This represents the total number of validation samples.

[0034] This invention can deeply explore the nonlinear characteristic relationship between displacement boundary conditions and geostress values ​​at measuring points, resulting in simulation effects that are closer to measured values. It boasts high computational efficiency and a simple modeling method, allowing the inversion calculation results to be applied to subsequent engineering construction analysis. This invention considers various constituent factors involved in the formation of the initial geostress field of the rock mass, such as topography and lithology, while also taking into account the complexity of geological conditions in different engineering areas and the degree of influence of tectonic movements on geostress. It selects appropriate scientific inversion analysis methods to obtain reliable and wide-ranging geostress field distribution characteristics.

[0035] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A method for intelligent inversion of geostress field based on deep learning algorithm, characterized in that, Includes the following steps: S1. Establish a three-dimensional geomechanical model based on the topographic and geological survey data of the project area; S2. Set multiple displacement boundary conditions, and set each displacement boundary condition as an undetermined influencing factor. In FLAC3D, apply the set multiple displacement boundary conditions to the three-dimensional geomechanical model. S3. For each undetermined influencing factor, set corresponding stress measurement points to obtain the stress components at each stress measurement point. With the undetermined influencing factor as the independent variable and the stress components as the dependent variable, construct a multivariate linear equation. Eliminate insignificant undetermined influencing factors to obtain a stepwise regression equation with significant regression effect. Obtain multiple sets of combined working conditions of displacement boundary conditions and gravitational acceleration, as well as the corresponding stress components calculated by the stepwise regression equation. Use some combined working conditions and their corresponding stress components as training samples, and use the remaining combined working conditions and their corresponding stress components as validation samples. S4. Construct a CNN-LSTM-Attention deep learning model; S5. Based on the training samples, the CNN-LSTM-Attention deep learning model is trained. During the training process, the stress components are used as inputs and the combined working conditions are used as outputs. After training, the ground stress inversion model is obtained. S6. Based on the validation samples, verify the prediction accuracy of the geostress inversion model. If the prediction accuracy verification fails, iterate through steps S1-S5.

2. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 1, characterized in that, The method for verifying the qualification of the geostress inversion model in step S6 is as follows: Substitute the stress components in the verification sample into the model, invert the corresponding combined working conditions, calculate the first evaluation index and the second evaluation index, set the first evaluation threshold and the second evaluation threshold, and if the first evaluation index is less than the first evaluation threshold and / or the second evaluation index is greater than the second evaluation threshold, it is considered unqualified; otherwise, it is considered qualified.

3. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 2, characterized in that, The first evaluation indicator is the coefficient of determination, and the second evaluation indicator is the mean absolute percentage error.

4. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 3, characterized in that, The coefficient of determination is calculated using the following formula: ； In the above formula, The coefficient of determination is represented by the coefficient of determination. This represents the stress component in the i-th verification sample. This represents the stress components output by the geostress inversion model for the i-th validation sample. This represents the average value of the stress components in all verification samples.

5. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 4, characterized in that, The mean absolute percentage error is calculated using the following formula: ； In the above formula, This represents the mean absolute percentage error. This represents the total number of validation samples.

6. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 1, characterized in that, The CNN-LSTM-Attention deep learning model includes a CNN network structure, a long short-term memory neural network, and an attention mechanism. The CNN network structure, the long short-term memory neural network, and the attention mechanism process the input data in sequence to obtain the output results.

7. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 6, characterized in that, The CNN grid structure is represented as follows: ； In the above formula, This represents the activation function. Represents convolutional feature mapping, Represents the weight matrix. This represents the convolution operation. Indicates the input item. This indicates the offset.

8. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 6, characterized in that, Long Short-Term Memory (LSTM) neural networks are represented as follows: ； ； ； ； ； ； In the above formula, Represents the Gate of Oblivion This represents the sgmoid function. The weight matrix represents the forget gate. This indicates the hidden state in the previous moment. Indicates the current input time. This represents the bias vector of the forget gate. Indicates the input gate. This represents the weight matrix of the input gate. This represents the bias vector of the input gate. Indicates the current state of the candidate cell. Indicates the use of generating The weight matrix, Indicates the use of generating The bias vector, This represents the updated state of the memory cell at time t. Indicates update The state of the memory unit at any given moment. Indicates the output gate. This represents the weight matrix of the output gate. This represents the bias vector of the output gate. Indicates the hidden state in the next moment. This indicates that elements are accumulated one by one.

9. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 8, characterized in that, The attention mechanism is represented as: ； ； ； In the above formula, This represents the attention probability distribution value obtained from the LSTM layer output at time t. , Both represent weight coefficients used for linearly combining the outputs of the LSTM layer, and b represents the bias parameter. This represents the attention-standardized weights. This represents the attention probability distribution value obtained from the LSTM layer output at time n, where n ranges from 1 to t. This represents the output of the attention mechanism layer at time t; The input to the attention mechanism is the hidden state obtained in the Long Short-Term Memory neural network for the next time step.

10. The intelligent inversion method for geostress field based on deep learning algorithm according to claim 1, characterized in that, In step S3, the specific method for eliminating insignificant undetermined influencing factors is as follows: For the multiple linear equation, each undetermined influencing factor is introduced one by one using stepwise regression, and the significance of different undetermined influencing factors in the regression equation is tested. When the significance level of the undetermined influencing factor is not greater than 0.05 and / or the VIF value of the collinearity index is not greater than 10, the undetermined influencing factor is considered insignificant and is eliminated.

Images