In-well earthquake prediction method and system based on deep learning

Abstract

The invention discloses an in-well earthquake prediction method and system based on deep learning, and the method comprises the steps: obtaining the historical earthquake data of a target region, and the historical earthquake data comprise earth sound data and geomagnetic data; performing corresponding feature extraction and selection on the earth sound data and the geomagnetic data to form earth sound feature data and geomagnetic feature data; training a pre-constructed earthquake prediction model based on the earth sound feature data and the geomagnetic feature data; the earthquake prediction model is constructed based on deep learning through a Stacking integration method, the earthquake prediction model comprises two layers, the first layer is a heterogeneous primary learner group, the second layer is a meta learner, and the output of the first layer is used as the input of the second layer; and periodically acquiring to-be-measured earth sound data and terrestrial magnetism data of the target area, inputting the to-be-measured earth sound data and the terrestrial magnetism data into the trained earthquake prediction model, and outputting a prediction result. According to the invention, the prediction accuracy is improved, and more reliable early warning is provided.

Description

Technical Field

[0001] This invention relates to the field of earthquake prediction technology, specifically to a method and system for well-drilled earthquake prediction based on deep learning. Background Technology

[0002] Earthquake prediction technology is an extremely challenging task, and existing technologies have certain limitations, with the accuracy of predictions needing to be improved.

[0003] First, traditional methods often rely on a single data source, such as analyzing only ground acoustic data or only geomagnetic data. Ground acoustic data reflects the sound wave signals generated by the fracturing and friction of underground rocks, exhibiting significant time-series characteristics; geomagnetic data is closely related to changes in underground rock stress, possessing spatial field characteristics. However, earthquakes are complex physical processes involving the accumulation and transmission of underground stress fields until the rock mass fractures. A single data source can only reflect one aspect of this process, resulting in incomplete information and leading to low prediction accuracy and a high false alarm rate.

[0004] Second, even though some studies attempt to integrate multi-source data, powerful Temporally correlated geosonic data and spatially structured geomagnetic data are typical heterogeneous data, exhibiting significant differences in data format, physical meaning, and feature distribution—a phenomenon known as the "modal gap." Current technologies lack a model framework capable of effectively mining and fusing deep, nonlinear precursor information from these two types of heterogeneous data. Simple feature stitching or using a single type of model, such as RNNs or CNNs alone, fails to simultaneously capture the long-term dependencies of time series and the local correlations of spatial features, let alone effectively establish the complex mapping relationship between these two factors and key earthquake parameters (occurrence time, location, magnitude). Furthermore, in the specific application scenario of "in-well" earthquakes, the limited space for sensor deployment and strong environmental noise interference place higher demands on data quality and model robustness, while general-purpose earthquake prediction models are often difficult to apply directly.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] The technical problem this invention aims to solve is the modal gap caused by the heterogeneous data sources of ground acoustic data and geomagnetic data in existing well-drilled seismic prediction, resulting in low prediction accuracy and high false alarm rates. The purpose of this invention is to provide a deep learning-based well-drilled seismic prediction method and system. By designing a dedicated heterogeneous learner fusion framework, it deeply mines and effectively fuses the complementary precursory information contained within these data sources. Through the fused seismic prediction model, the joint prediction accuracy and precision of the four key parameters of earthquake occurrence—time, location, magnitude, and depth—are significantly improved, providing more reliable early warnings. Furthermore, a complete technical solution from data preprocessing and feature engineering to model training and deployment is provided. This solution is specifically optimized for the noise characteristics and spatial constraints of well-drilled data, possessing stronger engineering practicality and scenario adaptability.

[0007] This invention is achieved through the following technical solution:

[0008] In a first aspect, the present invention provides a deep learning-based method for predicting seismic activity in wells, the method comprising:

[0009] Acquire historical earthquake data for the target area, including ground acoustic and geomagnetic data;

[0010] The ground acoustic data and geomagnetic data are respectively subjected to corresponding feature extraction and selection to form ground acoustic feature data and geomagnetic feature data;

[0011] Based on ground acoustic feature data and geomagnetic feature data, a pre-constructed earthquake prediction model is trained to obtain a trained earthquake prediction model. The earthquake prediction model is constructed using the Stacking ensemble method to build a deep learning-based earthquake prediction model. The earthquake prediction model consists of two layers: the first layer is a heterogeneous primary learner group, and the second layer is a meta-learner. The output of the first layer is used as the input of the second layer.

[0012] The system periodically acquires ground acoustic and geomagnetic data of the target area, inputs these data into a trained earthquake prediction model, and outputs prediction results. The prediction results include prediction data such as earthquake occurrence time, location latitude and longitude, magnitude, and depth.

[0013] Furthermore, the ground acoustic data includes at least one or more of the following: ground acoustic wave amplitude, frequency, and duration;

[0014] Geomagnetic data includes at least one or more of the following: intensity, direction, and rate of change of the geomagnetic field.

[0015] Furthermore, the method also includes:

[0016] Preprocessing is performed on the ground acoustic data and geomagnetic data to obtain preprocessed ground acoustic data and geomagnetic data; the preprocessing includes filtering of high-frequency noise in the well, denoising based on wavelet transform, and normalization.

[0017] Furthermore, the heterogeneous primary learner group includes a first primary learner and a second primary learner configured in parallel.

[0018] The first primary learner consists of a parallel LSTM model and an Xgboost model, used to process ground acoustic feature data with strong time-dependent characteristics and output the first prediction vector after concatenation.

[0019] The second primary learner comprises a parallel CNN model and a random forest model, used to process geomagnetic feature data with spatial structure characteristics, and outputs a concatenated second prediction vector.

[0020] Furthermore, the first primary learner includes:

[0021] LSTM model: Its input is ground acoustic feature data represented in time series. It captures the long-term temporal dependence of ground acoustic feature data through gating units, including forget gate, input gate and output gate; and outputs the first preliminary prediction result of seismic parameters.

[0022] The XGBoost model takes ground acoustic feature data as input, organized in the form of a feature table, and uses a gradient boosting decision tree to evaluate and predict the importance of features. The output is a second preliminary prediction result of the seismic parameters. The second preliminary prediction result is consistent with the first preliminary prediction result in terms of dimension.

[0023] First splicing unit: splices the first preliminary prediction result with the second preliminary prediction result and outputs the spliced ​​first prediction vector.

[0024] Furthermore, the second primary learner includes:

[0025] CNN model: Its input is geomagnetic feature data in a two-dimensional matrix organized in a spatial grid. It extracts local spatial anomaly patterns through convolutional layers using 3×3 convolutional kernels, compresses feature dimensions through pooling layers, and maps spatial features into high-level semantic representations through fully connected layers; the output is a third preliminary prediction result containing local spatial structure information.

[0026] Random Forest Model: Its input is geomagnetic feature data organized in the form of a feature table. It is trained in parallel with multiple decision trees, calculates the feature importance of each feature, and outputs a fourth preliminary prediction result based on the global feature importance assessment. The feature importance results of the random forest calculation model are fed back to optimize the input feature selection of the CNN model, forming a synergistic enhancement effect.

[0027] The second splicing unit splices the third preliminary prediction result with the fourth preliminary prediction result and outputs the spliced ​​second prediction vector.

[0028] Furthermore, the meta-learner employs a Logistic regression learner to assign weights to each primary learner by learning the complex nonlinear relationship between the primary prediction results of each primary learner in the first layer and the real earthquake labels; and based on the assigned weights, performs weighted fusion and final decision-making to output the final decision result.

[0029] Furthermore, the method also includes:

[0030] Based on the 5-fold cross-validation method, the geosonic and geomagnetic feature data in historical earthquake data are divided into training and validation sets.

[0031] The performance of the earthquake prediction model under different parameters was evaluated, and the optimal parameters of the first and second primary learners were adjusted using a grid search strategy.

[0032] Secondly, this invention provides a deep learning-based well seismic prediction system, which includes:

[0033] The acquisition unit is used to acquire historical earthquake data of the target area, including ground acoustic data and geomagnetic data.

[0034] The feature extraction and selection unit is used to extract and select corresponding features from the ground acoustic data and the geomagnetic data respectively, forming ground acoustic feature data and geomagnetic feature data;

[0035] The prediction model unit is used to train a pre-built earthquake prediction model based on geosonic and geomagnetic feature data to obtain a trained earthquake prediction model. The earthquake prediction model is a deep learning-based earthquake prediction model constructed using the Stacking ensemble method. The earthquake prediction model consists of two layers: the first layer is a heterogeneous primary learner group, and the second layer is a meta-learner. The output of the first layer is used as the input of the second layer.

[0036] The earthquake prediction and early warning unit is used to periodically acquire ground acoustic data and geomagnetic data of the target area, input the ground acoustic data and geomagnetic data into the trained earthquake prediction model, and output prediction results. The prediction results include prediction data such as earthquake occurrence time, location latitude and longitude, magnitude and depth; and early warning is issued based on the prediction results.

[0037] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned deep learning-based well seismic prediction method.

[0038] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0039] 1. This invention innovatively proposes a modeling approach of "heterogeneous learner synergy": targeting the time series characteristics of ground acoustic data and the spatial structure characteristics of geomagnetic data, it matches the two most advantageous learner combinations, LSTM model + XGBoost model and CNN model + random forest model, respectively, for feature extraction and preliminary prediction. This overcomes the shortcomings of single models in processing multimodal data and achieves accurate mining of deep information from heterogeneous data.

[0040] 2. This invention achieves decision-level fusion of two primary prediction results through the Stacking integration framework: the meta-learner (LR) can learn the complementary relationship between the primary learner outputs of local acoustic and geomagnetic data, effectively utilizing the synergistic effect between different data modes and precursor information, thereby achieving an accuracy superior to any single model or simple model combination in the joint prediction of earthquake spatiotemporal strong parameters.

[0041] 3. The entire technical solution of this invention is closely integrated with the "underground" application scenario: from the acquisition of sensor data in the well and the preprocessing of noise, to the feature selection and model optimization for downhole spatial constraints, a high-precision and robust earthquake prediction solution specifically for underground safety monitoring is formed, which has important practical value. Attached Figure Description

[0042] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0043] Figure 1 This is a flowchart of the deep learning-based well seismic prediction method of the present invention;

[0044] Figure 2 This is a structural block diagram of the deep learning-based well seismic prediction system of the present invention;

[0045] Figure 3 This is a detailed structural block diagram of the deep learning-based well seismic prediction system of the present invention; Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0047] This invention constructs a deep learning-based earthquake prediction model using a stacking ensemble method. The core of this earthquake prediction model lies in its design of a divide-and-conquer and fusion framework, addressing the heterogeneous characteristics of ground acoustic and geomagnetic data. Specifically:

[0048] For time-series ground acoustic data: This invention creatively combines the LSTM model and the XGBoost model as the first primary learner. The LSTM model excels at capturing long-term temporal dependencies in acoustic signals, while the XGBoost model efficiently filters out the key features most relevant to earthquakes from high-dimensional time-series features. Both models process the same ground acoustic feature data in parallel. The LSTM model receives the time-series input of the ground acoustic feature data, extracts temporal dependency features through its gating units (forget gate, input gate, and output gate), and outputs a preliminary prediction vector for earthquake parameters. XGBoost model: Receives the same ground acoustic feature data (in feature table form), uses a gradient boosting decision tree to evaluate and predict feature importance, and outputs a preliminary prediction vector for seismic parameters. Parallel processing mechanism: The LSTM model and the XGBoost model are processed independently in parallel, and their outputs... and They are identical in dimension and then concatenated into a vector. Input to the second-layer meta-learner (Logistic Regression); Meta-learner fusion: Logistic Regression model learning and The complementary relationship between these elements is used to output the final earthquake parameter prediction results through weighted fusion. Information can be extracted from two different dimensions: "temporal evolution pattern" and "feature importance ranking," forming a complementary relationship.

[0049] For geomagnetic data in the spatial domain: This invention creatively combines a CNN model and a random forest model as a second primary learner. The CNN model receives spatial grid input of geomagnetic feature data (e.g., organizing the geomagnetic field strength, direction, and rate of change of multiple monitoring points into a two-dimensional matrix). Local spatial anomaly patterns are extracted using 3×3 convolutional kernels in convolutional layers, and feature dimensions are compressed using pooling layers. The final output is a prediction vector P containing spatial structure information. CNN Random Forest Model: Receives the same geomagnetic feature data, but organizes it in the form of a feature table (each column corresponds to a monitoring point or feature dimension). It trains multiple decision trees in parallel, calculates the importance index (MDI) of each feature, and outputs a prediction vector P based on the feature importance assessment. RFParallel Processing and Feature Synergy: CNN and Random Forest models process geomagnetic feature data in parallel, mining precursor information from two dimensions: "local spatial pattern" and "global feature importance," respectively. The feature importance results calculated by the Random Forest model can be fed back to optimize the input feature selection of the CNN model, forming a synergistic enhancement effect. Output Concatenation and Meta-Learning: [The text abruptly ends here, likely due to an incomplete translation or source material.] and Concatenate into a vector The output of the first primary learner, along with the output of the Logistic Regression meta-learner, is fed into the final decision fusion. The CNN model effectively extracts local anomaly patterns in the spatial distribution of the geomagnetic field through convolutional kernels, while the random forest model evaluates the importance of each spatial feature point through metrics such as MDI, playing a role in feature selection and stability improvement. The combination of these two approaches achieves a collaborative analysis from "local spatial patterns" to "global feature importance."

[0050] Final decision fusion: Two heterogeneous primary learner groups provide preliminary predictive evidence from temporal and spatial perspectives, respectively. The second-layer Logistic Regression meta-learner then learns the complex nonlinear relationship between these primary predictions and real earthquake labels, automatically assigning appropriate weights to evidence from different sources and with varying degrees of reliability, achieving optimal ensemble decision-making, and thus accurately predicting earthquake parameters.

[0051] Example 1

[0052] like Figure 1 As shown, the present invention provides a deep learning-based well seismic prediction method, which includes:

[0053] Step 1: Obtain historical earthquake data for the target area. The historical earthquake data includes the earthquake occurrence time, location (latitude and longitude), magnitude, depth, and corresponding ground acoustic and geomagnetic data collected by sensors in the well.

[0054] In this embodiment, the ground acoustic data is acquired through a ground sensor array and includes at least the ground acoustic wave amplitude, frequency, and duration.

[0055] Geomagnetic data is acquired through a geomagnetic sensor array and includes at least the intensity, direction, and rate of change of the geomagnetic field.

[0056] Step 2: Preprocess the ground acoustic data and geomagnetic data to obtain preprocessed ground acoustic data and geomagnetic data;

[0057] In this embodiment, the preprocessing includes filtering of high-frequency noise in the well, denoising based on wavelet transform, and normalization.

[0058] Step 3: Perform corresponding feature extraction and selection on the preprocessed ground acoustic data and geomagnetic data to form ground acoustic feature data and geomagnetic feature data respectively;

[0059] Step 4: Based on the ground acoustic feature data and geomagnetic feature data, train the pre-constructed earthquake prediction model to obtain the trained earthquake prediction model. The earthquake prediction model is a deep learning-based earthquake prediction model constructed using the Stacking ensemble method. The earthquake prediction model consists of two layers: the first layer is a heterogeneous primary learner group, and the second layer is a meta-learner. The output of the first layer is used as the input of the second layer.

[0060] In this embodiment, the heterogeneous primary learner group includes a first primary learner and a second primary learner configured in parallel.

[0061] The first primary learner consists of a parallel LSTM model and an Xgboost model, used to process ground acoustic feature data with strong time-dependent characteristics and output the first prediction vector after concatenation.

[0062] The second primary learner comprises a parallel CNN model and a random forest model, used to process geomagnetic feature data with spatial structure characteristics, and outputs a concatenated second prediction vector.

[0063] The first and second primary learners described above are at the same level and processed in parallel within the Stacking ensemble learning framework, together forming the first primary learner group, which is specifically designed to process ground acoustic feature data. Both the first and second primary learners receive the same ground acoustic feature data as input, process it independently, and output their respective prediction results. Finally, the two prediction results are fed into a meta-learner (e.g., Logistic Regression) for fusion decision-making.

[0064] In this embodiment, the first primary learner includes:

[0065] The LSTM model takes ground acoustic feature data as input, represented as a time series, with each time step containing multiple features (such as amplitude, frequency, and duration). It captures the long-term temporal dependencies of the ground acoustic feature data through gating units, including a forget gate, input gate, and output gate. The output is a preliminary prediction of the seismic parameters. This preliminary prediction is a hidden state vector that integrates the temporal evolution information of the entire sequence, representing a preliminary prediction or intermediate feature representation of key seismic parameters (time, location, magnitude, and depth).

[0066] The XGBoost model takes ground acoustic feature data as input, organized in the form of a feature table (e.g., each sample is a row, and each column represents the statistical features within a certain time window: mean, variance, maximum, etc.). As a gradient boosting decision tree model, XGBoost builds a tree model through multiple iterations, automatically filtering important features (e.g., which ground acoustic features are most relevant to earthquakes). During processing, XGBoost uses its objective function (including a loss function and a regularization term) to optimize predictions, emphasizing feature importance ranking and preventing overfitting; that is, the XGBoost model evaluates and predicts feature importance through a gradient boosting decision tree. The output is a second preliminary prediction of earthquake parameters (similar to the output format of the LSTM model); and the second preliminary prediction is dimensionally consistent with the first preliminary prediction.

[0067] First splicing unit: splices the first preliminary prediction result with the second preliminary prediction result and outputs the spliced ​​first prediction vector.

[0068] The above technical solutions employ two approaches: the LSTM model focuses on capturing dynamic changes over time (e.g., the evolution trend of acoustic signals over time), while the XGBoost model focuses on extracting key indicators from the perspective of feature importance (e.g., which ground acoustic features contribute most to earthquake prediction). Combining the LSTM and XGBoost models achieves complementary analysis from both "temporal evolution" and "feature importance" dimensions. The outputs of the LSTM and XGBoost models (such as their respective magnitude predictions) are concatenated into a feature vector, which is then input into the second-layer meta-learner (e.g., Logistic regression).

[0069] During the training phase, the LSTM and XGBoost models are trained independently (without sharing parameters) but use the same geosonic feature data labels (real earthquake parameters). After training, they work in parallel during the inference phase, processing new geosonic feature data in real time. In practical applications, the feature importance evaluation results of the XGBoost model can be used to guide feature selection of geosonic data, optimize the feature set input to the LSTM model, and improve overall efficiency.

[0070] Specifically, the LSTM model is a special type of recurrent neural network (RNN) that addresses the vanishing and exploding gradient problems that traditional RNNs encounter when processing long-sequence data, enabling the network to learn long-term dependencies within the sequence. Geosound data is time-series data containing acoustic signals that vary over time, and these signals may be closely related to seismic activity. The LSTM model, through its gating mechanism (including input gate, forget gate, and output gate), can capture the long-term dependencies in this time-series data, which is particularly important for predicting seismic activity. Furthermore, this invention has found that the XGBoost model performs exceptionally well in handling high-dimensional features and large-scale datasets; that is, the XGBoost model excels at processing geosound feature data. Therefore, combining the XGBoost model with the LSTM model to process geosound feature data further improves the accuracy of earthquake prediction.

[0071] Preferably, the LSTM model consists of a forget gate, an input gate, and an output gate, and the processing and update process of these three components is as follows:

[0072] ;

[0073] ;

[0074] ;

[0075] ;

[0076] in It is a time series representation of geographic feature data. It is the Forgotten Gate that decides which information to discard. The input gate determines the retention of the current input information. express The layer creates the current cell state based on the current input and the previous output. The output gate determines the output value of the current cell. This indicates the output of the current cell, and The sigmoid function represents the output of the unit at the previous time step. It is an activation function. , , , It is a weight vector. , , , It is the bias vector.

[0077] Preferably, the Xgboost model uses the following objective function during construction:

[0078] ;

[0079] in ;

[0080] In the formula, Representing the The true values ​​of each sample, where T is the node tree of the tree model. Representing the A tree model, The representative is the leaf node weight vector. For custom constant coefficients, here Represents the predicted value and the true value loss function, penalty term Used to limit the complexity of the Xgboost model.

[0081] In this embodiment, the second primary learner includes:

[0082] CNN Model: Its input is a two-dimensional matrix of geomagnetic feature data organized in a spatial grid, such as measurements of geomagnetic field strength, direction, and rate of change at multiple locations, which can be organized into a two-dimensional matrix similar to an image. Processing: Convolutional layers use 3×3 convolutional kernels to extract local spatial anomaly patterns (such as abrupt changes in the magnetic field or gradient changes in a specific region), pooling layers compress the feature dimension, downsample the features, and retain key spatial information; fully connected layers map the spatial features into a high-level semantic representation. The output is a feature vector representing a preliminary prediction of key earthquake parameters, emphasizing precursor information extracted from the perspective of spatial structure. The output feature vector is a third preliminary prediction result containing local spatial structure information.

[0083] Random Forest Model: Its input is geomagnetic feature data organized in the form of a feature table (measurements at each location point are in one column, and each time / spatial sample is in one row); Processing: Random Forest constructs multiple decision trees to evaluate the importance of geomagnetic features (using indicators such as MDI); It selects the spatial feature points most relevant to earthquakes from a global perspective (e.g., which locations have the most indicative geomagnetic changes); Through parallel training of multiple decision trees, it calculates the feature importance results of each feature. The output is a fourth preliminary prediction result based on the global feature importance evaluation, emphasizing the judgment made from the perspective of feature importance.

[0084] The feature importance results of the random forest computation model are fed back to optimize the input feature selection of the CNN model, forming a synergistic enhancement effect;

[0085] The second splicing unit splices the third preliminary prediction result with the fourth preliminary prediction result and outputs the spliced ​​second prediction vector.

[0086] The above technical solutions utilize CNN models, which excel at capturing local spatial correlations and patterns (e.g., specific forms of magnetic field changes within a region). Random forest models excel at global feature importance assessment and stability prediction (e.g., identifying which monitoring points contribute most to earthquake prediction across the entire region). Combining CNN and random forest models achieves collaborative analysis from "local spatial pattern recognition" to "global feature importance ranking." Actual collaborative mode: CNN and random forest models process raw geomagnetic feature data independently, each outputting its own prediction results. The random forest model directly uses the original feature table, without relying on CNN's feature extraction. Collaborative feature selection: The feature importance (MDI index) calculated by the random forest model can be used to optimize the input features of the CNN model, for example, selecting highly important monitoring point data as the input channel for the CNN model, improving computational efficiency and model performance. Output fusion method: The outputs of the CNN and random forest models (such as their respective prediction probabilities of epicenter location) are concatenated into a feature vector, which is then input into the second-layer meta-learner. The meta-learner learns to assign weights to these two outputs, comprehensively considering the spatial pattern recognition results and feature importance assessment results. Optimization for geomagnetic feature data: The convolution kernel design of the CNN model can be optimized to account for the spatial continuity of the geomagnetic field. The tree depth and number of trees in the random forest model can be adjusted according to the dimensions of the geomagnetic features to prevent overfitting.

[0087] Specifically, CNN models excel at processing data with a grid-like structure, such as images, because CNNs can capture local spatial features through convolutional layers. In geomagnetic feature data, spatial features are crucial information because changes in the geomagnetic field may be related to stress variations in underground rocks. CNNs can extract these local features through their convolutional layers, thus effectively identifying and analyzing spatial variations in the geomagnetic field. Additionally, random forest models also have unique advantages in processing geomagnetic data. This invention creatively combines these two models to process geomagnetic feature data, achieving a further improvement in prediction accuracy.

[0088] Preferably, the CNN model consists of convolutional layers, pooling layers, fully connected layers, and a softmax output layer, and its processing procedure is as follows: Assume that the CNN model has a total of Words, vector dimension is The input matrix of the convolutional layer is then represented as Convolutional layers use multiple 3x3 convolutional kernels to perform convolution operations to obtain various data features. Here, multiple 3x3 convolution kernels (smaller convolution kernels) are used for convolution operations to increase the depth and complexity of the network while reducing the number of parameters; , for function, For the weight vector, It is a bias vector. For the first Data feature information; feature information after convolution The input pooling layer further compresses the information, and then the fully connected layer fuses the feature information. Finally, the softmax output layer is used to output the predicted classification information.

[0089] Preferably, the random forest model uses the MDI (Mean Importance Index) as an indicator to evaluate feature importance. It selects the top-ranked features from the original geomagnetic feature data whose total importance reaches a preset threshold as input features, thus completing feature selection. The MDI index is defined as follows:

[0090] ;

[0091] in, Represents all trees in a random forest. Represents a node in a tree. Indicates whether the node uses the first... Variables As a dividing feature For nodes The sample proportion at which the Gini coefficient is used as an indicator of impurity. This indicates a decrease in Gini impurity before and after the division. This represents the optimal split node.

[0092] In this embodiment, the meta-learner adopts a Logistic regression learner, which is used to learn the complex nonlinear relationship between the primary prediction results of each primary learner in the first layer and the real earthquake labels, assign weights to each primary learner, and perform weighted fusion and final decision based on the assigned weights to output the final decision result.

[0093] In the above technical solution, the LSTM model and XGBoost model of the first primary learner process the ground acoustic feature data in parallel, and each outputs a prediction vector (for example, each vector contains 4 values, corresponding to the preliminary prediction of earthquake time, latitude and longitude, magnitude, and depth, respectively). These two prediction vectors are concatenated or averaged within the group to form a unified output vector O for the group. Acoustic (For example, the dimension is 8, assuming each sub-model outputs 4 dimensions). After the CNN model and the random forest model of the second primary learner process the geomagnetic feature data in parallel, they each output a prediction vector; these two prediction vectors are also concatenated or averaged within the group to form the unified output vector O of the group. Magnetic(The dimensions are the same as the ground sound group, e.g., 8-dimensional). The meta-learner (e.g., Logistic Regression) does not receive four independent outputs, but rather two output vectors that have undergone preliminary in-group integration: Meta-learner input = [O Acoustic O Magnetic The concatenated vector has a total dimension of 16 (assuming each output has 8 dimensions). The Logistic Regression model learns the complex relationship between the two sets of inputs through maximum likelihood estimation, assigning weight coefficients to the dimensions of the input vector.

[0094] Specifically, the Logistic Regression learner models the log odds of events.

[0095] Preferably, the Logistic regression learner includes:

[0096] ;

[0097] in, For weight vectors, For the input vector, This is the bias value. Indicates in Under known conditions, The probability of occurrence; when the Logistic Regression learner learns, the parameter estimates of the weight vector and bias values ​​are solved by maximum likelihood estimation.

[0098] Step 5: Periodically acquire the ground acoustic data and geomagnetic data of the target area, input the ground acoustic data and geomagnetic data into the trained earthquake prediction model, and output the prediction results; the prediction results include the earthquake occurrence time, location latitude and longitude, magnitude and depth.

[0099] As a further implementation, the method also includes:

[0100] Based on the 5-fold cross-validation method, the geosonic and geomagnetic feature data in historical earthquake data are divided into training and validation sets.

[0101] The performance of the earthquake prediction model under different parameters was evaluated, and the optimal parameters of the first and second primary learners were adjusted using a grid search strategy.

[0102] This invention addresses the challenge of heterogeneous fusion of geosonic and geomagnetic data in well environments by designing a heterogeneous learner collaborative fusion framework. A first primary learner, composed of an LSTM model and an Xgboost model, focuses on the temporal characteristics of the geosonic data; a second primary learner, composed of a CNN model and a random forest model, focuses on the spatial characteristics of the geomagnetic data; and a Logistic Regression meta-learner is then used to weight and fuse the primary prediction results. This invention effectively mines complementary information from multi-source precursor data, significantly improving the accuracy and reliability of well-based earthquake prediction.

[0103] Example 2

[0104] like Figure 2 and Figure 3 As shown, the difference between this embodiment and Embodiment 1 is that this embodiment provides a deep learning-based well seismic prediction system, which corresponds one-to-one with the deep learning-based well seismic prediction method in Embodiment 1; the system includes:

[0105] The acquisition unit is used to acquire historical earthquake data of the target area, including ground acoustic data and geomagnetic data.

[0106] The feature extraction and selection unit is used to extract and select corresponding features from the ground acoustic data and the geomagnetic data respectively, forming ground acoustic feature data and geomagnetic feature data;

[0107] The prediction model unit is used to train a pre-built earthquake prediction model based on geosonic and geomagnetic feature data to obtain a trained earthquake prediction model. The earthquake prediction model is a deep learning-based earthquake prediction model constructed using the Stacking ensemble method. The earthquake prediction model consists of two layers: the first layer is a heterogeneous primary learner group, and the second layer is a meta-learner. The output of the first layer is used as the input of the second layer.

[0108] The earthquake prediction and early warning unit is used to periodically acquire ground acoustic data and geomagnetic data of the target area, input the ground acoustic data and geomagnetic data into the trained earthquake prediction model, and output prediction results. The prediction results include prediction data such as earthquake occurrence time, location latitude and longitude, magnitude and depth; and early warning is issued based on the prediction results.

[0109] The execution process of each unit can be carried out according to the steps of the deep learning-based well seismic prediction method in Example 1, and will not be described in detail in this example.

[0110] Meanwhile, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned deep learning-based well seismic prediction method.

[0111] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0114] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0115] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based well seismic prediction method, characterized in that, The method includes: Acquire historical earthquake data for the target area, including ground acoustic data and geomagnetic data; The ground acoustic data and geomagnetic data are respectively subjected to corresponding feature extraction and selection to form ground acoustic feature data and geomagnetic feature data; Based on the aforementioned geosonic and geomagnetic feature data, a pre-constructed earthquake prediction model is trained to obtain a trained earthquake prediction model. The earthquake prediction model is constructed using the Stacking ensemble method to build a deep learning-based earthquake prediction model. The earthquake prediction model includes two layers: the first layer is a heterogeneous primary learner group, and the second layer is a meta-learner. The output of the first layer is used as the input of the second layer. The system periodically acquires ground acoustic data and geomagnetic data of the target area, inputs the ground acoustic data and geomagnetic data into a trained earthquake prediction model, and outputs prediction results; the prediction results include prediction data including earthquake occurrence time, location latitude and longitude, magnitude and depth.

2. The deep learning-based well seismic prediction method according to claim 1, characterized in that, The ground acoustic data includes at least one or more of the following: ground acoustic wave amplitude, frequency, and duration; The geomagnetic data includes at least one or more of the following: intensity, direction, and rate of change of the geomagnetic field.

3. The well-based seismic prediction method based on deep learning according to claim 1, characterized in that, The method also includes: The ground acoustic data and geomagnetic data are preprocessed to obtain preprocessed ground acoustic data and geomagnetic data; the preprocessing includes filtering of high-frequency noise in the well, denoising based on wavelet transform, and normalization processing.

4. The deep learning-based well seismic prediction method according to claim 1, characterized in that, The heterogeneous primary learner group includes a first primary learner and a second primary learner configured in parallel. The first primary learner includes a parallel LSTM model and an Xgboost model, used to process ground acoustic feature data with strong time-dependent characteristics, and outputs a first prediction vector after concatenation. The second primary learner includes a parallel CNN model and a random forest model, used to process geomagnetic feature data with spatial structure characteristics, and outputs a concatenated second prediction vector.

5. The deep learning-based well seismic prediction method according to claim 4, characterized in that, The first primary learner includes: LSTM model: Its input is ground acoustic feature data represented in time series. It captures the long-term temporal dependencies of the ground acoustic feature data through gating units, including forget gate, input gate and output gate; and outputs the first preliminary prediction result of seismic parameters. The XGBoost model takes ground acoustic feature data as input, organized in the form of a feature table, and uses a gradient boosting decision tree to evaluate and predict the importance of features. The output is a second preliminary prediction result of the seismic parameters. The second preliminary prediction result is consistent with the first preliminary prediction result in terms of dimension. First splicing unit: splices the first preliminary prediction result with the second preliminary prediction result and outputs the spliced ​​first prediction vector.

6. The well-based seismic prediction method based on deep learning according to claim 4, characterized in that, The second primary learner includes: CNN model: Its input is geomagnetic feature data in a two-dimensional matrix organized in a spatial grid. It extracts local spatial anomaly patterns through convolutional layers using 3×3 convolutional kernels, compresses feature dimensions through pooling layers, and maps spatial features into high-level semantic representations through fully connected layers; the output is a third preliminary prediction result containing local spatial structure information. Random Forest Model: Its input is geomagnetic feature data organized in the form of a feature table. It is trained in parallel with multiple decision trees, calculates the feature importance of each feature, and outputs a fourth preliminary prediction result based on global feature importance evaluation. The feature importance results of the random forest calculation model are fed back to optimize the input feature selection of the CNN model. The second splicing unit splices the third preliminary prediction result with the fourth preliminary prediction result and outputs the spliced ​​second prediction vector.

7. The deep learning-based well seismic prediction method according to claim 1, characterized in that, The meta-learner employs a Logistic regression learner, which learns the complex nonlinear relationship between the primary prediction results of each primary learner in the first layer and the real earthquake labels, assigns weights to each primary learner, and performs weighted fusion and final decision based on the assigned weights, outputting the final decision result.

8. The deep learning-based well seismic prediction method according to claim 4, characterized in that, The method also includes: The historical earthquake data is divided into training and validation sets based on the 5-fold cross-validation method, using ground acoustic and geomagnetic feature data. The performance of the earthquake prediction model under different parameters is evaluated, and the optimal parameters of the first and second primary learners are adjusted using a grid search strategy.

9. A deep learning-based well seismic prediction system, characterized in that, The system includes: An acquisition unit is used to acquire historical earthquake data of the target area, the historical earthquake data including ground acoustic data and geomagnetic data; The feature extraction and selection unit is used to extract and select corresponding features from the ground acoustic data and the geomagnetic data respectively, to form ground acoustic feature data and geomagnetic feature data; The prediction model unit is used to train a pre-constructed earthquake prediction model based on the geosonic feature data and geomagnetic feature data to obtain a trained earthquake prediction model. The earthquake prediction model is a deep learning-based earthquake prediction model constructed using the Stacking ensemble method. The earthquake prediction model includes two layers: the first layer is a heterogeneous primary learner group, and the second layer is a meta-learner. The output of the first layer is used as the input of the second layer. An earthquake prediction and early warning unit is used to periodically acquire ground acoustic data and geomagnetic data of a target area, input the ground acoustic data and geomagnetic data into a trained earthquake prediction model, and output prediction results; the prediction results include prediction data including earthquake occurrence time, location latitude and longitude, magnitude and depth; and to issue early warnings based on the prediction results.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based well seismic prediction method as described in any one of claims 1 to 8.

Images