Seismic fault identification method based on Bayesian optimization and interpretable tree model

Abstract

The invention discloses a seismic fault identification method based on Bayesian optimization and an interpretable tree model, and belongs to the technical field of seismic exploration, and the method comprises the following steps: S1, synthesizing seismic data through a convolution model; s2, extracting seismic attributes related to a fault by using the synthesized seismic data obtained in S1; s3, performing feature selection on the seismic attributes by using the maximum information coefficient to obtain a feature subset; s4, constructing an integrated tree learning model of Bayesian optimization, and automatically searching an optimal hyper-parameter combination through a Bayesian optimization algorithm; s5, explaining the optimal integrated tree model; and S6, preprocessing the actual seismic data set, inputting the actual seismic data set into the final model, and obtaining a fault identification result. According to the seismic fault identification method based on Bayesian optimization and the interpretable tree model, the fault identification precision and reliability are improved, the fault identification accuracy is improved, and the specific influence of different types of seismic attributes on the prediction result is explained.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration technology, and in particular to a seismic fault identification method based on Bayesian optimization and an interpretable tree model. Background Art

[0002] Faults are an important geological factor affecting coal mining safety. Fault structures developed in coal seams are often associated with a series of accidents, including coal explosions, gas outbursts, water inrush, and roof falls. In addition, faults destroy the continuity of coal seams and significantly affect the effectiveness of intelligent coal mining. Therefore, it is crucial to determine the distribution of faults in the structural interpretation of coalfield seismic exploration.

[0003] The accuracy of fault identification in seismic interpretation is affected by many factors, including seismic resolution, time-to-depth conversion, and expert interpretation bias. Traditional methods rely too much on manual work, resulting in discontinuous identification results, time-consuming and labor-intensive, and easily affected by the experience of interpreters. In recent decades, scholars have proposed many computer-assisted fault identification methods. These methods jointly detect faults by combining one or more seismic attributes onto seismic profiles. Commonly used seismic attributes include coherence, curvature, variance, and ant bodies. Researchers have been seeking to discover new seismic attributes that can be used for fault identification, such as geometric attributes, structural tensors, and texture attributes. However, how to select a suitable set of optimal seismic attributes is a challenging task, because different combinations of seismic attributes may lead to multiple interpretations of fault identification.

[0004] Compared with traditional methods, machine learning-based methods have significant advantages in efficiency, intelligence, and automation. These innovative technologies reduce the dependence on professional seismic interpreters, enhance adaptability to large-scale data growth, and can further reveal hidden patterns in the data. Many machine learning algorithms, including BP neural networks, support vector machines (SVMs), and convolutional neural networks (CNNs), have been proven to be effective in the application of fault identification tasks. Although machine learning-based methods have shown good performance in earthquake fault identification, there are still some technical challenges that need to be addressed. These include building high-quality training datasets, optimizing hyperparameters, and model interpretability. Summary of the invention

[0005] The purpose of the present invention is to provide an earthquake fault identification method based on Bayesian optimization and interpretable tree model, to improve the precision and reliability of fault identification, to improve the accuracy of fault identification, and to explain the specific impact of different types of earthquake attributes on the prediction results.

[0006] To achieve the above object, the present invention provides a method for identifying earthquake faults based on Bayesian optimization and an interpretable tree model, comprising the following steps: S1. Construct a 3D geological model based on the geological profile and logging data of the target area, and then synthesize seismic data through the convolution model; S2, extracting seismic attributes related to the fault using the synthetic seismic data obtained in step S1; S3. Use the maximum information coefficient to perform feature selection on seismic attributes to remove redundant and irrelevant attributes to obtain feature subsets. The obtained feature subsets are split in a 1:1 ratio, 50% of which is used as a training data set for building and training the model; the remaining 50% is used as a test data set; S4, further dividing the data set obtained in S3 into a training set and a verification set, constructing a Bayesian optimized ensemble tree learning model, training the model using the training set, and optimizing the model's hyperparameters using the verification set, and automatically searching for the optimal hyperparameter combination through the Bayesian optimization algorithm; S5. Explain the optimal ensemble tree model using the SHAP method; S6. Preprocess the actual earthquake data set, extract earthquake attributes, and input them into the final model to obtain the fault identification results.

[0007] Preferably, in S1, folds are constructed by two-dimensional Gaussian transformation, and fault structures are generated by translation transformation. ; In the formula, , , , Represents different fold structural parameters. By randomly setting the fold structural parameters, a convolution model with different folds is obtained. is the horizontal coordinate of the transformation plane, is the ordinate of the transformation plane, is the number of Gaussian transformations, is the index, is a natural constant.

[0008] Preferably, the seismic attributes in S2 need to be preprocessed by standardization, and the standardization formula is as follows: ; In the formula, represents the data after normalization. represents the data before standardization. , represent the mean and standard deviation of the original data respectively.

[0009] Preferably, the specific operation steps in S4 are as follows: S41. Select a suitable ensemble learning method, construct multiple decision trees and integrate their prediction results to form an ensemble tree model. The basic form is as follows: ; In the formula, Represented as training set A sample of Indicates the prediction result of the sample; is the number of decision trees in the model; represents a decision tree, Expressed as The parameters of each decision tree Divide its feature space into Regions , and each region corresponds to a fixed output value ; represents the indicator function; S42, using the training set to perform preliminary training on the selected ensemble tree learning model, and setting default hyperparameters of the ensemble tree learning model, the hyperparameters including the number of trees, the maximum depth, and the minimum number of sample splits; S43. Set the hyperparameter search space and use the Bayesian optimization algorithm to automatically find the optimal hyperparameter combination in the search space. The objective function of the Bayesian optimization is as follows: ; In the formula, represents the objective function of the ensemble tree learning model, represents a set of hyperparameters, represents the search space of the model's hyperparameters, represents the optimal combination of hyperparameters; S44. Retrain the ensemble tree learning model under the optimal hyperparameter combination and evaluate its performance using the validation set.

[0010] Preferably, the following loss function is used in the training phase of the ensemble tree learning model in S42: As the optimization target, the AdamW algorithm is used as the optimizer to iteratively update the model parameters; ; In the formula, is the number of training samples, For the The probability of predicting an instance as a fault, For the The true category of the instance, with a value of 1 indicating a fault and 0 indicating a non-fault.

[0011] Therefore, the present invention adopts the above-mentioned earthquake fault identification method based on Bayesian optimization and interpretable tree model, which has the following beneficial effects: (1) An interpretable ensemble tree model based on Bayesian optimization is used to identify fault structures in earthquakes, thereby improving the accuracy and reliability of fault identification; (2) The Bayesian optimization algorithm is used to optimize the ensemble tree learning model, which effectively improves the overall performance of the model and thus improves the accuracy of fault identification; (3) With the help of SHAP technology, we have a more detailed understanding of the decision-making process inherent in the proposed model, making fault prediction more transparent, extracting earthquake attributes from multiple dimensions, and explaining in detail the specific impact of different categories of earthquake attributes on the prediction results, providing richer interpretability.

[0012] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of an earthquake fault identification method based on Bayesian optimization and interpretable tree model of the present invention; Figure 2 It is a three-dimensional synthetic seismic schematic diagram of an earthquake fault identification method based on Bayesian optimization and an interpretable tree model of the present invention; wherein, Figure 2 (a) is a 3D seismic data view. Figure 2 (b) is the seismic cross-section view. Figure 2 (c) in the figure is the fault label corresponding to the seismic section; Figure 3 This is an example diagram of earthquake attribute statistical results of an earthquake fault identification method based on Bayesian optimization and an interpretable tree model of the present invention; Figure 4 This is an example diagram of a MIC feature selection method of an earthquake fault identification method based on Bayesian optimization and an interpretable tree model of the present invention; Figure 5 It is an example diagram of the Bayesian optimization process of an earthquake fault identification method based on Bayesian optimization and an interpretable tree model of the present invention; Figure 5 (a) in the above formula is GBDT. Figure 5 (b) in the figure is XGBoost. Figure 5 (c) in the figure is LightGBM. Figure 5 (d) in the figure is CatBoost. Figure 5 (e) in the equation is MLP. Figure 5 (f) in the figure is SVM; Figure 6 It is a comparison diagram of experimental results of various integrated tree models of an earthquake fault identification method based on Bayesian optimization and interpretable tree model of the present invention; Figure 6 (a) is the ROC curve of the hyperparameter optimization experiment. Figure 6(b) in the figure is the AUC value of different sampling methods. Figure 6 (c) in the figure shows the performance of different models on the test dataset. Figure 6 (d) in the figure is the validation loss and AUC curve of BO-CatBoost; Figure 7 It is a fault identification result diagram of the fault identification model of the earthquake fault identification method based on Bayesian optimization and interpretable tree model on the verification set of the present invention; wherein, Figure 7 (a) in the figure is the original seismic profile in the validation set. Figure 7 (b) in the figure is the fault label in the validation set. Figure 7 (c) in the figure is the fault identification result; Figure 8 It is a characteristic analysis diagram of an earthquake fault identification method based on Bayesian optimization and interpretable tree model based on SHAP in the present invention; Figure 8 (a) is the SHAP summary graph for each feature. Figure 8 (b) in the figure is the feature importance ranking graph; Fig. 9 It is a local interpretation diagram of the model performance of an earthquake fault identification method based on Bayesian optimization and an interpretable tree model of the present invention; Fig. 9 (a) in the example is a sample predicted to be a fault. Fig. 9 (b) is an example predicted as a non-fault sample. DETAILED DESCRIPTION

[0014] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0015] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0016] Embodiment 1 like Figure 1 As shown, the present invention provides a method for identifying earthquake faults based on Bayesian optimization and an interpretable tree model, comprising the following steps: S1. Construct a 3D geological model based on the geological profile and logging data of the target area, and then synthesize seismic data through the convolution model; The three-dimensional geological model should include the structure of the strata, the distribution of lithology and possible fault information. The synthetic seismic data through the convolution model can be replaced by the synthetic seismic data through the wave equation. When synthesizing seismic data through the wave equation, it is necessary to further obtain the seismic profile through the migration method.

[0017] Folds are constructed by two-dimensional Gaussian transformation, and fault structures are generated by translation transformation. ; In the formula, , , , Represents different fold structural parameters. By randomly setting the fold structural parameters, a convolution model with different folds is obtained. is the horizontal coordinate of the transformation plane, is the ordinate of the transformation plane, is the number of Gaussian transformations, is the index, is a natural constant.

[0018] The three-dimensional synthetic seismic response is calculated by convolution model using Ricker wavelet and adding appropriate amount of noise to make it closer to the actual earthquake. The introduction of randomness is achieved by randomly setting the model parameters, including fault location, dip, inclination and strike. This method aims to simulate the diversity and complexity observed in seismic data. As shown in Table 1, the embodiment of the present invention constructs a data set including 5 pairs of size The synthetic seismic data and their corresponding fault labels are obtained by adding random noise. The signal-to-noise ratio (SNR) is calculated as follows: ; In the formula, represents the amplitude of seismic signal; Represents the noise signal amplitude.

[0019] Table 1 Description of synthetic earthquake dataset ;

[0020] like Figure 2 As shown, Figure 2 The 3D seismic volumes with different types of faults and their seismic sections in three directions are shown. In addition, a corresponding binary mask is generated for each seismic image, in which fault data is marked with 1 and non-fault data is marked with 0. Figure 2 (a) is a 3D view of the seismic data volume. Figure 2 (b) shows a seismic section of the synthetic 3D seismic volume, and the corresponding fault labels are shown in Figure 2 As shown in (c), the fault outline is clearly outlined.

[0021] S2, extracting seismic attributes related to the fault using the synthetic seismic data obtained in step S1; Seismic attributes include amplitude, frequency, phase, coherence and texture attributes, which can reflect the structural changes of the strata and the existence of faults. Each seismic attribute is properly preprocessed, such as standardization or normalization, to eliminate the impact of dimensional differences on subsequent analysis.

[0022] Ten main lines were randomly selected from the 3D seismic data volume, and the constructed data set included 163,840 samples, including 4,912 fault samples and 158,928 non-fault samples. A total of 40% of the data were selected as data for the training phase. The number of non-fault samples is much larger than the number of fault samples. Then, 13 seismic attributes of 4 categories listed in Table 2 were obtained on each sample.

[0023] Table 2 Earthquake attribute information table ;

[0024] like Figure 3 As shown, a box plot is used to display the statistical results of the initial data set. It should be noted that the blue asterisks represent outliers, and half of the seismic attributes have some outliers. For most seismic attributes, the median is not located in the center of the box, indicating that the distribution of these data is asymmetric. In order to prevent the size of the eigenvalues ​​and outliers from affecting the results, data normalization is used. The normalization formula is as follows: ; In the formula, represents the data after normalization. represents the data before standardization. , represent the mean and standard deviation of the original data respectively.

[0025] S3. Use the maximum information coefficient (MIC) to select features of seismic attributes to remove redundant and irrelevant attributes to obtain feature subsets. The obtained feature subsets are split in a 1:1 ratio, with 50% used as a training data set for building and training models, and the remaining 50% as a test data set. The MIC coefficient between any two columns of data in the seismic attribute data set obtained by S2 is presented in the form of a heat map as follows Figure 4 As shown. The range of MIC is [0,1]. The larger the MIC value, the stronger the correlation between the two columns of data. A series of different MIC thresholds are selected, and a set of seismic attributes are identified as feature vectors for model training. In this example, a threshold of 0.6 is set, and attributes below this threshold are selected for model training. Subsequently, the importance of these features will be analyzed through the SHAP interpretation framework.

[0026] S4, further dividing the data set obtained in S3 into a training set and a verification set, constructing a Bayesian optimized ensemble tree learning model, training the model using the training set, and optimizing the model's hyperparameters using the verification set, and automatically searching for the optimal hyperparameter combination through the Bayesian optimization algorithm; The specific steps are as follows: S41, selecting a suitable ensemble learning method, constructing multiple decision trees and integrating their prediction results to form an ensemble tree model; Ensemble learning methods such as gradient boosted tree (GBDT), XGBoost, LightGBM or CatBoost, which improve the accuracy and robustness of the model by building multiple decision trees (such as Classification And Regression Tree, CART) and integrating their prediction results. The earthquake attribute set was divided into a training set and a validation set in a ratio of 8:2. Four GBDT algorithms (GBDT, XGBoost, LightGBM and CatBoost) were trained and tested, and Bayesian optimization was used for hyperparameter optimization. Finally, four optimal models were obtained, namely BO-GBDT, BO-XGBoost, BO-LightGBM and BO-CatBoost. At the same time, SVM and MLP were used as experimental comparison models.

[0027] Ensemble Tree Model The basic form is as follows: ; In the formula, Represented as training set A sample of Indicates the prediction result of the sample; is the number of decision trees in the model; represents a decision tree, Expressed as The parameters of each decision tree Divide its feature space into Regions , and each region corresponds to a fixed output value ; represents the indicator function; S42, using the training set to perform preliminary training on the selected ensemble tree learning model, and setting default hyperparameters of the ensemble tree learning model, the hyperparameters including the number of trees, the maximum depth, and the minimum number of sample splits; The following loss function is used in the training phase of the ensemble tree learning model As the optimization target, the AdamW algorithm is used as the optimizer to iteratively update the model parameters; ; In the formula, is the number of training samples, For the The probability of predicting an instance as a fault, For the The true category of the instance, with a value of 1 indicating a fault and 0 indicating a non-fault.

[0028] S43, set the hyperparameter search space, and use the Bayesian optimization algorithm to automatically find the optimal hyperparameter combination in the search space. The process of using Bayesian optimization for hyperparameter configuration consists of two main aspects: a probabilistic surrogate model and an acquisition function, where the probabilistic surrogate model is responsible for fitting a complex objective function based on previous evaluations, and the acquisition function is responsible for generating new parameter combinations. The Tree-structured Parzen Estimator (TPE) is used as a probabilistic surrogate model and the Expected Improvement (EI) is used as an acquisition function. The TPE method uses two density functions to define , as shown below: ; In the formula, is the threshold value, represents the observation results corresponding to the selected parameters, represents the density formed using the searched parameter observations, is the density formed by the remaining parameters.

[0029] The objective function of Bayesian optimization is as follows: ; In the formula, represents the objective function of the ensemble tree learning model, represents a set of hyperparameters, represents the search space of the model's hyperparameters, represents the optimal combination of hyperparameters; like Figure 5 As shown, Figure 5 The following figure shows an example of the Bayesian optimization process of an embodiment of the present invention. In order to prevent the model from overfitting during training, the parameters of the early stopping mechanism are set to 50. Table 3 shows the final results of the parameters common to the four GBDT algorithms optimized by the Bayesian optimization method, where is the number of boosted trees, is the learning rate, is the maximum number of leaves in a tree, is the maximum depth of the tree, is the sample fraction used to fit a single base learner. The changes between hyperparameters and objective function values ​​during automatic optimization are as follows: Figure 5 (a) Figure 5 (b) Figure 5 (c) and Figure 5As shown in (d) in the figure. The horizontal axis represents the parameter name, and the first and last values ​​of the corresponding column represent its value range. The color of the line represents the evaluation indicator value of different parameter combinations during the validation set parameter optimization process, corresponding to the value in the first column.

[0030] Table 3 Hyperparameter optimization results of ensemble tree model ;

[0031] S44. Retrain the ensemble tree learning model under the optimal hyperparameter combination and use the validation set to evaluate its performance. like Figure 6 As shown, Figure 6 The figure shows the comparison of experimental results of various ensemble tree models according to the embodiments of the present invention. Figure 6 (a) in Figure 2 shows the ROC curve of the hyperparameter optimization experiment. Figure 6 As shown in (b) in Figure 1, a comparative experiment was designed to study the impact of different sampling methods on the prediction results. The first group was not treated as a control group. The second group was subjected to random undersampling (RU) processing, whereby non-fault points were randomly sampled to achieve an equal ratio of non-fault points to fault points. The third group used the SMOTE method to interpolate the fault point data and increase the number of fault points until the samples were balanced. The fourth group of data was processed using the HS method. Compared with other GBDT methods, BO-CatBoost has the lowest sensitivity to training set size and class imbalance, and its detection results are highly robust. Figure 6 (c) in the figure shows the comparison results of various ensemble tree models under five different indicators. Figure 6 (d) shows the loss curve and AUC curve of BO-CatBoost in the validation phase. After 200 training cycles, the early stopping mechanism triggers the termination of model training before the model overfits the training data.

[0032] like Figure 7 As shown, Figure 7 The differences between the indicators of different ensemble tree models are shown in a more intuitive way. Figure 7 (a) in the figure is the original seismic profile and fault label in the validation set. Figure 7 (b) is the fault identification result of the GBDT model. Figure 7 (c) in the figure is the fault identification result of the XGBoost model. Figure 7 (d) in the figure is the fault identification result of the LightGBM model. Figure 7(e) in the figure is the fault identification result of the CatBoost model. The experimental results show that the four ensemble tree models all show good performance in fault identification. Among them, the BO-CatBoost model performs best, with an AUC of 83.53%, which is the best among the four ensemble tree models. This result further confirms the superior performance of the proposed model in dealing with complex fault data classification problems, and provides a valuable reference for subsequent research in related fields.

[0033] S5. Use SHAP (Shapley Additive Explanations) method to explain the optimal ensemble tree model; For the black-box model BO-CatBoost obtained by Bayesian optimization, the prediction results of the model are explained globally and locally based on the SHAP explanation framework. In addition, the rationality of the explanation results is comprehensively analyzed by comparing the experimental results. First, the Shapley values ​​of all features included in each sample on the test set are calculated. Subsequently, the global and local explanations are analyzed and applied, and the value of the influencing factors of each individual feature is analyzed.

[0034] like Figure 8 As shown, Figure 8 It is a SHAP summary graph and feature importance ranking graph in the embodiment of the present invention, which illustrates the global impact of various input features on model prediction. Each row along the vertical axis represents a feature, arranged in descending order of feature importance from top to bottom. The horizontal axis represents the SHAP value, Figure 8 Each point in (a) represents a feature of a sample. The redder the color, the greater the feature value of the point. The positive and negative values ​​of SHAP indicate the direction of influence of various factors on the predicted fault results. A positive value means that an increase in the factor leads to an increase in the probability of predicting a fault, while a negative value indicates its inhibitory effect. When points overlap, they are stacked vertically. Obviously, GEN, LBP, and GLCM_CON2 (the subcomponent of GON in the 90° direction) have higher values.

[0035] Figure 8 (b) shows that the feature with the greatest impact on model prediction is GEN, which is consistent with the Figure 1 In contrast, GLCM_ENE1 (the subcomponent of GEN at 45°) and GCO are the features with the least influence.

[0036] Fig. 9 The SHAP interpretation method is used to interpret the local performance of the model in the embodiment of the present invention. Fig. 9 (a) in the figure shows a sample predicted to be a fault. Fig. 9 (b) Example prediction for a non-fault sample.

[0037] S6. Preprocess the actual earthquake data set, extract earthquake attributes, and input them into the final model to obtain the fault identification results.

[0038] Therefore, the present invention adopts the above-mentioned earthquake fault identification method based on Bayesian optimization and interpretable tree model to improve the precision and reliability of fault identification, improve the accuracy of fault identification, and explain the specific influence of different types of earthquake attributes on the prediction results.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for earthquake fault identification based on Bayesian optimization and interpretable tree model, characterized in that: The following steps are involved: S1. Construct a 3D geological model based on the geological profile and logging data of the target area, and then synthesize seismic data through the convolution model; S2, extracting seismic attributes related to the fault using the synthetic seismic data obtained in step S1; S3. Use the maximum information coefficient to perform feature selection on seismic attributes to remove redundant and irrelevant attributes to obtain feature subsets. The obtained feature subsets are split in a 1:1 ratio, 50% of which is used as a training data set for building and training the model; the remaining 50% is used as a test data set; S4, further dividing the data set obtained in S3 into a training set and a verification set, constructing a Bayesian optimized ensemble tree learning model, training the model using the training set, and optimizing the model's hyperparameters using the verification set, and automatically searching for the optimal hyperparameter combination through the Bayesian optimization algorithm; S5. Explain the optimal ensemble tree model using the SHAP method; S6. Preprocess the actual earthquake data set, extract earthquake attributes, and input them into the final model to obtain fault identification results.

2. The earthquake fault identification method based on Bayesian optimization and interpretable tree model according to claim 1, characterized in that: In S1, folds are constructed by two-dimensional Gaussian transformation, and fault structures are generated by translation transformation. ； In the formula, , , , Represents different fold structural parameters. By randomly setting the fold structural parameters, a convolution model with different folds is obtained. is the horizontal coordinate of the transformation plane, is the ordinate of the transformation plane, is the number of Gaussian transformations, is the index, is a natural constant.

3. The earthquake fault identification method based on Bayesian optimization and interpretable tree model according to claim 1, characterized in that: The seismic attributes in S2 need to be standardized and preprocessed. The standardization formula is as follows: ； In the formula, represents the data after normalization. represents the data before standardization. , represent the mean and standard deviation of the original data respectively.

4. The earthquake fault identification method based on Bayesian optimization and interpretable tree model according to claim 1, characterized in that: The specific operation steps in S4 are as follows: S41. Select a suitable ensemble learning method, construct multiple decision trees and integrate their prediction results to form an ensemble tree model. The basic form is as follows: ； In the formula, Represented as training set A sample of Indicates the prediction result of the sample; is the number of decision trees in the model; represents a decision tree, Expressed as The parameters of each decision tree Divide its feature space into Regions , and each region corresponds to a fixed output value ; represents the indicator function; S42, using the training set to perform preliminary training on the selected ensemble tree learning model, and setting default hyperparameters of the ensemble tree learning model, the hyperparameters including the number of trees, the maximum depth, and the minimum number of sample splits; S43. Set the hyperparameter search space and use the Bayesian optimization algorithm to automatically find the optimal hyperparameter combination in the search space. The objective function of the Bayesian optimization is as follows: ； In the formula, represents the objective function of the ensemble tree learning model, represents a set of hyperparameters, represents the search space of the model's hyperparameters, represents the optimal combination of hyperparameters; S44. Retrain the ensemble tree learning model under the optimal hyperparameter combination and evaluate its performance using the validation set.

5. The earthquake fault identification method based on Bayesian optimization and interpretable tree model according to claim 4, characterized in that: The following loss function is used in the training phase of the ensemble tree learning model in S42 As the optimization target, the AdamW algorithm is used as the optimizer to iteratively update the model parameters; ； In the formula, is the number of training samples, For the The probability of predicting an instance as a fault, For the The true category of the instance, with a value of 1 indicating a fault and 0 indicating a non-fault.

Images