A ground stress inversion method based on an LSTM-Transformer hybrid deep learning model

Abstract

The present application relates to a kind of based on LSTM-Transformer hybrid deep learning model's ground stress inversion method, comprising: collecting the ground stress measured data of downhole multiple measuring points;The ground stress measured data is preprocessed;Deep learning model is constructed;Wherein, the deep learning model uses the long short-term memory network and the Transformer encoder of fusion;The data after pre-processing is used to train the deep learning model;Using the deep learning model after training, ground stress inversion is carried out.The present application can significantly reduce the calculation cost, improve the accuracy and robustness of inversion result, applicable to the stability analysis and disaster prediction of mine under complex geological conditions.

Description

Technical Field

[0001] This invention relates to the fields of geological engineering and mining engineering technology, and in particular to a geostress inversion method based on an LSTM-Transformer hybrid deep learning model. Background Technology

[0002] In-situ stress is a key factor affecting the stability of underground engineering projects, the safety of mining operations, and the mechanical properties of rock masses. Accurate acquisition of in-situ stress parameters is crucial for mine design, disaster prevention, and safety assessment of underground structures. Current in-situ stress measurements primarily rely on field testing methods such as hydraulic fracturing and acoustic emission monitoring. While these methods offer high accuracy, they suffer from drawbacks such as complex testing processes, long cycles, high costs, and low data coverage. Furthermore, due to the complexity of geological conditions, measured data often suffers from missing data, noise interference, and insufficient sample sizes. In recent years, machine learning and deep learning technologies have been introduced into the field of geostress prediction, with common methods including random forests, support vector regression, convolutional neural networks (CNNs), and long short-term memory networks (LSTMs). However, existing methods generally suffer from the following shortcomings: traditional regression models and some deep learning models struggle to simultaneously capture the temporal and spatial correlations in geostress characteristics, leading to insufficient prediction accuracy. Existing methods mostly model single stress parameters and cannot simultaneously predict multiple geostress indices within a unified framework.

[0003] When faced with datasets from different mining areas and with varying geological conditions, the model's performance deteriorates significantly, and it exhibits poor adaptability to small sample sizes. Parameters such as vertical principal stress may have missing values ​​in actual measurements, and existing methods lack an effective mechanism for combining classification and regression to address this issue. Summary of the Invention

[0004] The purpose of this invention is to propose a geostress inversion method based on an LSTM-Transformer hybrid deep learning model. Addressing the shortcomings of existing geostress prediction methods in feature extraction, data missingness handling, prediction accuracy, and model generalization, this invention provides an intelligent method and system capable of achieving high-precision prediction of multiple parameters, including stress direction, maximum secondary principal stress, minimum secondary principal stress, and vertical principal stress, even under conditions of small sample size, high noise levels, and missing data. This method can simultaneously model the temporal dependence and global correlation of geostress characteristics, improving prediction accuracy and stability, shortening computation time, and reducing reliance on field testing.

[0005] To achieve the above objectives, the present invention provides the following solution: A geostress inversion method based on an LSTM-Transformer hybrid deep learning model includes: Collect measured ground stress data from multiple measuring points downhole; The measured ground stress data are preprocessed; Construct a deep learning model; wherein the deep learning model employs a fusion of a long short-term memory network and a Transformer encoder; The deep learning model is trained using the preprocessed data; The trained deep learning model is used to perform geostress inversion.

[0006] Optionally, the measured geostress data includes: measuring point number, measuring section burial depth, stress direction, rupture pressure, re-tension pressure, instantaneous closure pressure, maximum secondary principal stress, minimum secondary principal stress, and vertical principal stress.

[0007] Optionally, preprocessing the measured geostress data includes: Missing values ​​in the vertical principal stress are marked with a preset special symbol and a binary classification label is generated to indicate whether there are missing values ​​in the current sample, thus separating the original vertical principal stress data into a numerical part and a missing value label part. Several derived features are constructed based on measured geostress data, including pressure ratio, pressure difference, and burial depth pressure ratio. All numerical features are standardized, and the original data samples are expanded.

[0008] Optionally, augmenting the original data sample includes: Gaussian noise enhancement, interpolation enhancement, SMOTE enhancement, and rotation enhancement are applied to the original data samples.

[0009] Optionally, the deep learning model includes: Long Short-Term Memory (LSTM) networks are used to extract temporal dependency information from input features; Transformer encoders are used to capture global feature associations through a multi-head self-attention mechanism; Fully connected layers are used to output prediction results; Regression model: Stress direction, maximum secondary principal stress, and minimum secondary principal stress are all output using regression models; for vertical principal stress, a classification model is first used to determine missing values, and then a regression model is used to predict its value.

[0010] Optionally, residual connections and layer normalization are provided between the long short-term memory network and the Transformer encoder.

[0011] Optionally, training the deep learning model includes: A multi-task joint training approach is adopted, and a comprehensive loss function is defined to collaboratively optimize various geostress parameters.

[0012] Optionally, a multi-task joint training approach can be adopted, defining a comprehensive loss function to collaboratively optimize multiple geostress parameters, including: To address the strong coupling relationship among multiple geostress parameters, the maximum horizontal principal stress, minimum horizontal principal stress, and vertical principal stress are incorporated into the same network framework for collaborative training. Mechanical constraint loss is introduced to achieve coupled inversion of multiple stress parameters. The multi-task joint training method is as follows: under the same neural network framework, three prediction branches, namely the maximum horizontal principal stress, the minimum horizontal principal stress, and the vertical principal stress, are constructed simultaneously. Feature learning is carried out uniformly through a shared feature extraction layer, and each task branch outputs the corresponding stress parameter prediction value. Furthermore, a comprehensive loss function composed of regression errors of each sub-task and physical consistency constraints of geostress is constructed. The network parameters are uniformly optimized through joint backpropagation, thereby achieving synergistic inversion and accuracy improvement of multiple geostress parameters.

[0013] Optionally, using the trained deep learning model to perform ground stress inversion includes: The standardized input features are fed into the trained deep learning model to obtain the predicted values ​​of stress direction, maximum secondary principal stress, minimum secondary principal stress and vertical principal stress, respectively. The prediction of the vertical principal stress employs a two-step method: The classification sub-model is used to determine whether it is the preset special symbol; If the value is numerical, a regression sub-model is used for prediction, and the results are combined with the classification results to obtain the final predicted value of the vertical principal stress.

[0014] Optionally, training the deep learning model further includes: An early stopping mechanism is used during model training; The model uses mean square error and coefficient of determination as evaluation indicators when predicting stress direction, maximum secondary principal stress and minimum secondary principal stress; and uses accuracy, confusion matrix and classification report as evaluation indicators when predicting vertical principal stress classification.

[0015] The beneficial effects of this invention are as follows: 1. The fusion of LSTM and Transformer can simultaneously capture the temporal dependence and global features of geostress data, thereby improving prediction accuracy.

[0016] 2. Under the same framework, the joint prediction of stress direction, maximum secondary principal stress, minimum secondary principal stress and vertical principal stress can be achieved, thereby improving modeling efficiency.

[0017] 3. The vertical principal stress adopts a two-step prediction strategy that combines classification and regression, which effectively solves the problem of irregular missing values.

[0018] 4. Before training the data, use algorithms to expand the sample data, making the sample data more diverse and improving the model's prediction performance.

[0019] 5. It can still maintain high prediction accuracy even with small sample sizes, noise interference, and variable geological conditions.

[0020] 6. Combine the adaptive optimization algorithm (Adam) to dynamically adjust the learning rate, thereby improving the training convergence speed and stability. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a diagram of the LSTM-Transformer architecture according to an embodiment of the present invention; Figure 2 This is a diagram of the Transformer framework according to an embodiment of the present invention; Figure 3 This is a diagram of the LSTM framework according to an embodiment of the present invention; Figure 4 The loss function curve is shown in the embodiment of the present invention; Figure 5 This is an example diagram of the data prediction results in an embodiment of the present invention; Figure 6 This is a flowchart illustrating a geostress inversion method based on an LSTM-Transformer hybrid deep learning model, according to an embodiment of the present invention. Detailed Implementation

[0023] 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.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] like Figure 6 As shown, this embodiment proposes a geostress inversion method based on an LSTM-Transformer hybrid deep learning model, including: Collect measured ground stress data from multiple measuring points downhole; The measured ground stress data are preprocessed; Construct a deep learning model; wherein the deep learning model employs a fusion of a long short-term memory network and a Transformer encoder; The deep learning model is trained using the preprocessed data; The trained deep learning model is used to perform geostress inversion.

[0026] Specifically, in this embodiment, the measured geostress data includes parameters such as: measuring point number, measuring section depth, stress direction, rupture pressure, re-tension pressure, instantaneous closure pressure, maximum secondary principal stress, minimum secondary principal stress, and vertical principal stress. Some vertical principal stress values ​​are missing and are represented by a pre-defined special symbol "\". After data collection, the data is organized into a structured table file and saved in CSV format for subsequent processing and analysis.

[0027] Furthermore, the preprocessing of the measured geostress data includes: Missing values ​​in the vertical principal stress are marked with a preset special symbol and a binary classification label is generated to indicate whether there are missing values ​​in the current sample, thus separating the original vertical principal stress data into a numerical part and a missing value label part. Several derived features are constructed based on measured geostress data, including pressure ratio, pressure difference, and burial depth pressure ratio. All numerical features are standardized, and the original data samples are expanded.

[0028] In the data preprocessing stage, the vertical principal stress field is first identified by a missing value flag, converting it into a binary label to indicate whether a value is missing in the current sample. Then, the original vertical principal stress data is separated into a numerical part and a missing value flag part, with missing values ​​converted to NaN for special handling during model training. Before training, the original data samples are augmented using an algorithm to enhance the sample data and improve the model's prediction performance. To enhance the model's feature representation ability, multiple derived features are constructed from the original data. All continuous features are standardized to have a mean of zero and a variance of one, thereby improving the stability and convergence speed of model training.

[0029] Specifically, in this embodiment, data preprocessing includes: Missing values ​​in the vertical principal stress (vertical principal stress refers to the normal stress component acting on the rock mass unit along the gravity direction in the original stress field of the strata, usually generated by the self-weight of the overlying strata) are marked with a special symbol ("\") and a binary classification label (whether it is a missing value) is generated. (Due to the influence of factors such as field testing conditions, packer sealing failure, and borehole wall collapse, some hydraulic fracturing measurement points cannot effectively obtain vertical principal stress data, and in actual data tables, missing values ​​are usually represented by the special symbol "\". Therefore, this invention proposes a two-step prediction method of "vertical principal stress missing discrimination + numerical inversion prediction", that is: first, the vertical principal stress is judged to be missing through a classification model; if it is a numerical sample, quantitative prediction is then performed through a regression model.) Missing values ​​are replaced with NaN, and derived features are constructed, including pressure ratio, pressure difference, and burial depth pressure ratio, etc. All numerical features are standardized (the Z-score used in this invention). Standardization methods (which are mature existing data preprocessing techniques in this field) improve the stability of model training. Gaussian noise, interpolation, the SMOTE algorithm, and rotation augmentation are combined to augment the original data samples, enhancing the model's predictive performance.

[0030] Furthermore, the expansion of the original data sample includes: Gaussian noise enhancement, interpolation enhancement, SMOTE enhancement, and rotation enhancement are applied to the original data samples.

[0031] Specifically, in this embodiment, data augmentation includes Gaussian noise augmentation, interpolation augmentation, SMOTE augmentation, and rotation augmentation. To ensure numerical stability, the input feature matrix X is first normalized using Z-score on the training set. : The following augmentation operations are performed on the training samples in the normalized space: 1. Gaussian noise: Add noise to sample x in the j-th dimension. ,in Let j be the standard deviation of the j-th dimension. 1. Noise intensity. 2. Interpolation (Mixup): Randomly select two samples. , With random coefficients generate The corresponding labels are processed using the same linear interpolation (regression task) or selected based on probability (classification task). 3. SMOTE: For minority class samples Randomly select a nearest neighbor from its k nearest neighbors. ,according to , Synthesize new samples, inherit labels 4. Rotation Enhancement: Apply scale-simulated rotational perturbation to the sample. , It contains a small amount of Gaussian noise. After enhancement, for Implement anti-standardization The augmented samples are then written back to the dataset for subsequent training. The above augmentation methods are all known techniques in the art, but in this embodiment, they are used in conjunction with LSTM-Transformer to support robust inversion tasks with small samples, noisy interference, and missing values.

[0032] More specifically, this embodiment also incorporates GAN (Generative Adversarial Network) for data augmentation to improve the model's robustness in situations with insufficient data. (A GAN consists of a generator G and a discriminator D. The generator G generates new synthetic geostress samples, taking random noise z and a conditional vector c (burial depth, pressure features, missing data markers, etc.) as inputs and outputting synthetic geostress samples.) This is used to learn the mapping relationship from random noise and geological conditions to geostress samples. The discriminator D determines whether a sample is real data; it takes a real sample x as input and generates a sample... The condition vector c outputs the probability or score that the sample is a "real sample," used to constrain the generator to produce samples consistent with the real data distribution. The basic mathematical model for GAN training is the classic adversarial objective function. The complete process of GAN data augmentation is as follows: original real data → data preprocessing and normalization → construction of conditional vector c → training GAN → generation of synthetic geostress data → physical constraints and anomaly screening → merging of synthetic data and real data → for training prediction models. Generative Adversarial Networks (GANs) belong to existing deep learning data augmentation techniques.

[0033] Furthermore, the deep learning model includes: Long Short-Term Memory (LSTM) networks are used to extract temporal dependency information from input features; Transformer encoders are used to capture global feature associations through a multi-head self-attention mechanism; Fully connected layers are used to output prediction results; Regression Model: Stress direction, maximum secondary principal stress, and minimum secondary principal stress are all output using regression models; for vertical principal stress, a classification model is first used to identify missing values, and then a regression model is used to predict its value. (Data inversion is performed using regression. Vertical principal stress is missing in actual measurements, so it is first classified to filter out missing values, and then regression is used to invert and predict the value.)

[0034] Specifically, in this embodiment, the geostress inversion model is constructed as follows: Addressing the complex nonlinear mapping relationship between hydraulic fracturing monitoring data and geostress parameters, this embodiment constructs a geostress inversion model based on an LSTM-Transformer deep neural network. Using multi-source feature parameters as input, it jointly models and predicts the rupture direction, maximum secondary principal stress, minimum secondary principal stress, and vertical principal stress. The vertical principal stress employs a two-stage inversion strategy of "classification first, regression later," achieving intelligent identification and accurate inversion of missing and valid values. LSTM networks, Transformer attention mechanisms, MSE loss functions, and Adam optimization algorithms are all existing deep learning modeling techniques. Building upon these, this invention addresses the coexistence of multi-parameter regression and missing vertical principal stress in geostress inversion by constructing a multi-objective geostress inversion model based on LSTM-Transformer. It also proposes a two-stage intelligent modeling strategy of "classification + regression" for vertical principal stress, achieving unified modeling of missing value identification and numerical inversion, thereby significantly improving the stability and engineering applicability of geostress inversion. The LSTM-Transformer architecture is as follows: Figure 1 As shown.

[0035] A deep learning model integrating a Long Short-Term Memory (LSTM) network and a Transformer encoder is defined. The LSTM part is used to extract the temporal dependency information of the input features, while the Transformer part captures global feature associations through a multi-head self-attention mechanism. The fully connected layer outputs the prediction results. The stress direction, maximum secondary principal stress, and minimum secondary principal stress are all output using regression models. The vertical principal stress is first identified as missing using a classification model, and then its value is predicted using a regression model. The Transformer framework is as follows: Figure 2 As shown, the LSTM framework is as follows Figure 3 As shown.

[0036] Furthermore, residual connections and layer normalization are set between the Long Short-Term Memory network and the Transformer encoder.

[0037] Specifically, in this embodiment, residual connections and layer normalization are set between the LSTM and the Transformer encoder to improve training stability and convergence speed. (Residual connections are a technique used to alleviate the vanishing gradient problem during deep neural network training. The core idea is to directly pass the input to the output through cross-layer connections, allowing the network to learn the residual between the input and output. Specific process: superposition of input and output: assuming the input is...) The output after the LSTM layer is LSTM ( The output of the residual connection is: Here, + indicates element-wise addition. Layer normalization is a technique used to stabilize neural network training. By normalizing the output of each layer, the distribution of the input data remains stable. Unlike batch normalization, layer normalization normalizes all features of a single sample, making it suitable for small batches or dynamic sequence data. Specific process: ① Calculate the mean and variance: For the input... Calculate the mean and variance: , ② Normalization: ,in It is a very small constant used for numerical stability. ③ Scaling and translation: ,in and These are learnable parameters.

[0038] In this embodiment, the depth of the LSTM layers is increased to enhance the complexity of time series modeling and more accurately capture the evolution characteristics of geostress. (The depth of the LSTM layers refers to the number of stacked LSTM layers. Increasing the depth enhances the model's ability to abstract time series data, thereby more accurately capturing the complex temporal evolution characteristics in geostress data.)

[0039] Furthermore, model training and evaluation include: A multi-task joint training approach is adopted, and the Adam optimizer is used to adjust the model parameters (the Adam optimizer is an existing technology widely used in deep network training). An early stopping mechanism is combined to avoid overfitting. The model performance is comprehensively evaluated using metrics such as mean squared error (MSE), coefficient of determination (R²), accuracy, confusion matrix, and classification report.

[0040] Mean Square Error (MSE): Coefficient of determination R²: The above evaluation indicators all belong to existing technologies.

[0041] The entire training and evaluation process can be described as follows: (1) Input the augmented training set and validation set ; (2) Input LSTM layer to extract temporal features ; (3) Input Transformer encoder to extract global dependencies ; (4) Input fully connected layer maps to regression values ​​or classification probabilities ; (5) Calculate the loss function based on the task type; (6) Use the Adam optimizer to iteratively update the network parameters, and combine it with an early stopping mechanism to prevent overfitting.

[0042] (7) Calculate regression or classification indices on the test set or validation set for the model after training, and output the comprehensive performance evaluation results.

[0043] More specifically, training deep learning models also includes: A multi-task joint training approach is adopted, and a comprehensive loss function is defined to collaboratively optimize various geostress parameters.

[0044] An early stopping mechanism is used during model training; The model uses mean square error and coefficient of determination as evaluation indicators when predicting stress direction, maximum secondary principal stress and minimum secondary principal stress; and uses accuracy, confusion matrix and classification report as evaluation indicators when predicting vertical principal stress classification.

[0045] In this embodiment, the model training adopts a multi-task joint training approach, defining a comprehensive loss function to collaboratively optimize multiple geostress parameters. (Multi-task joint learning is a common modeling method in existing deep learning. This embodiment, based on this, addresses the strong coupling relationship among multiple geostress parameters by incorporating the maximum horizontal principal stress, minimum horizontal principal stress, and vertical principal stress into the same network framework for collaborative training, and introducing mechanical constraint loss to achieve coupled inversion of multiple stress parameters, thereby improving prediction accuracy and engineering stability. The multi-task joint training approach involves simultaneously constructing three prediction branches for the maximum horizontal principal stress, minimum horizontal principal stress, and vertical principal stress within the same neural network framework. Feature learning is performed uniformly through a shared feature extraction layer, and each task branch outputs its corresponding stress parameter prediction value. Furthermore, a comprehensive loss function composed of the regression error of each sub-task and the physical consistency constraint term of geostress is constructed. The network parameters are uniformly optimized through joint backpropagation, thereby achieving collaborative inversion of multiple geostress parameters and improving accuracy.)

[0046] In this embodiment, the model training adopts a multi-task joint training method. Figure 4 The solid blue line represents the loss function curve for the original training data set, the dashed blue line represents the loss function curve for the original validation data set, the solid green line represents the loss function curve for the augmented data training set, and the dashed green line represents the loss function curve for the augmented data validation set.

[0047] In this embodiment, the output feature importance analysis and attention weight visualization results are provided to improve the interpretability of the prediction.

[0048] Furthermore, the inversion of ground stress using the trained deep learning model includes: The standardized input features are fed into the trained deep learning model to obtain the predicted values ​​of stress direction, maximum secondary principal stress, minimum secondary principal stress and vertical principal stress, respectively. The prediction of the vertical principal stress employs a two-step method: The classification sub-model is used to determine whether it is the preset special symbol; If the value is numerical, a regression sub-model is used for prediction, and the results are combined with the classification results to obtain the final predicted value of the vertical principal stress.

[0049] In the prediction phase, the standardized input features are fed into the trained model to obtain predicted values ​​for stress direction, maximum secondary principal stress, minimum secondary principal stress, and vertical principal stress. For the vertical principal stress prediction, if the classification network determines it is missing, the symbol "\" is output directly; otherwise, the regression prediction value is output, retaining two decimal places of precision. Finally, all prediction results are merged with the original data to generate a new data file containing the predicted columns, which is then output to local storage. Experimental results show that this method can still achieve high-precision prediction of multiple geostress parameters even with limited data sample size and missing values, and significantly outperforms traditional statistical modeling methods in terms of generalization performance and stability. The data prediction results are as follows: Figure 5 As shown.

[0050] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A geostress inversion method based on an LSTM-Transformer hybrid deep learning model, characterized in that, include: Collect measured ground stress data from multiple measuring points downhole; The measured geostress data includes: measuring point number, measuring section burial depth, stress direction, rupture pressure, re-tension pressure, instantaneous closure pressure, maximum secondary principal stress, minimum secondary principal stress, and vertical principal stress. The measured ground stress data are preprocessed; Construct a deep learning model; wherein the deep learning model employs a fusion of a long short-term memory network and a Transformer encoder; The deep learning model is trained using the preprocessed data: A multi-task joint training approach is adopted, and a comprehensive loss function is defined to collaboratively optimize multiple geostress parameters: To address the strong coupling relationship among multiple geostress parameters, the maximum horizontal principal stress, minimum horizontal principal stress, and vertical principal stress are incorporated into the same network framework for collaborative training. Mechanical constraint loss is introduced to achieve coupled inversion of multiple stress parameters. The multi-task joint training method is as follows: under the same neural network framework, three prediction branches, namely the maximum horizontal principal stress, the minimum horizontal principal stress, and the vertical principal stress, are constructed simultaneously. Feature learning is carried out uniformly through a shared feature extraction layer, and each task branch outputs the corresponding stress parameter prediction value. Further, a comprehensive loss function composed of regression errors of each sub-task and physical consistency constraints of geostress is constructed. The network parameters are uniformly optimized through joint backpropagation, thereby achieving the coordinated inversion and accuracy improvement of multiple geostress parameters. The trained deep learning model is used to perform geostress inversion.

2. The geostress inversion method based on the LSTM-Transformer hybrid deep learning model according to claim 1, characterized in that, Preprocessing of the measured geostress data includes: Missing values ​​in the vertical principal stress are marked with a preset special symbol and a binary classification label is generated to indicate whether there are missing values ​​in the current sample, thus separating the original vertical principal stress data into a numerical part and a missing value label part. Several derived features are constructed based on measured geostress data, including pressure ratio, pressure difference, and burial depth pressure ratio. All numerical features are standardized, and the original data samples are expanded.

3. The geostress inversion method based on the LSTM-Transformer hybrid deep learning model according to claim 2, characterized in that, Expanding the original data sample includes: Gaussian noise enhancement, interpolation enhancement, SMOTE enhancement, and rotation enhancement are applied to the original data samples.

4. The geostress inversion method based on the LSTM-Transformer hybrid deep learning model according to claim 1, characterized in that, The deep learning model includes: Long Short-Term Memory (LSTM) networks are used to extract temporal dependency information from input features; Transformer encoders are used to capture global feature associations through a multi-head self-attention mechanism; Fully connected layers are used to output prediction results; Regression model: Stress direction, maximum secondary principal stress, and minimum secondary principal stress are all output using regression models; for vertical principal stress, a classification model is first used to determine missing values, and then a regression model is used to predict its value.

5. The geostress inversion method based on the LSTM-Transformer hybrid deep learning model according to claim 4, characterized in that, Residual connections and layer normalization are provided between the long short-term memory network and the Transformer encoder.

6. The geostress inversion method based on the LSTM-Transformer hybrid deep learning model according to claim 2, characterized in that, Using a trained deep learning model, geostress inversion includes: The standardized input features are fed into the trained deep learning model to obtain the predicted values ​​of stress direction, maximum secondary principal stress, minimum secondary principal stress and vertical principal stress, respectively. The prediction of the vertical principal stress employs a two-step method: The classification sub-model is used to determine whether it is the preset special symbol; If the value is numerical, a regression sub-model is used for prediction, and the results are combined with the classification results to obtain the final predicted value of the vertical principal stress.

7. The geostress inversion method based on the LSTM-Transformer hybrid deep learning model according to claim 1, characterized in that, Training the deep learning model further includes: An early stopping mechanism is used during model training; The model uses mean square error and coefficient of determination as evaluation indicators when predicting stress direction, maximum secondary principal stress and minimum secondary principal stress; and uses accuracy, confusion matrix and classification report as evaluation indicators when predicting vertical principal stress classification.

Images