Automatic identification method of gravity faults based on deep learning

Abstract

The present invention provides a deep learning-based method for automatic gravity fault identification. The method comprises the following steps: Step 1: constructing a three-dimensional fault geological model; Step 2: performing density filling on the three-dimensional geological model; Step 3: constructing three-dimensional fault gravity sample data; Step 4: constructing gravity fault sample labels; Step 5: constructing a deep learning-based gravity fault identification network; Step 6: using an optimization algorithm to train network parameters to obtain a gravity fault automatic identification network model; and Step 7: outputting and applying the gravity fault automatic identification network model. This deep learning-based method for automatic gravity fault identification utilizes a deep learning algorithm to fully exploit the fault structure information contained in gravity anomalies, improving the efficiency and accuracy of fault identification and providing a new technical tool for gravity fault identification.

Description

Technical Field

[0001] The present invention relates to the field of exploration geophysical technology, and in particular to a method for automatic identification of gravity faults based on deep learning. Background Art

[0002] In the field of oil and gas exploration, faults, as common underground geological structures, play a very important role in the conduction, migration, and sealing of oil and gas. It is of great significance to use various geophysical data to identify and characterize faults. Using gravity data to identify faults is an important technical means to characterize the distribution characteristics of underground faults. It is mainly based on potential field edge detection techniques such as derivative calculations in different directions and gradient calculations, and has good results in identifying the occurrence of large-scale faults. However, due to the superposition effect of gravity anomaly sources and underground density bodies, the anomalies generated by faults are relatively weak and easily interfered with by other geological bodies, observation noise, and processing noise. As a result, conventional fault identification technology has certain limitations. In addition, there is a lack of quantitative characterization when continuously optimizing fault extraction parameters, resulting in low efficiency and strong subjectivity, which affects the accuracy of fault identification and characterization.

[0003] Chinese patent application CN201911015974.6 discloses a method for identifying gravity fault images based on tectonic context, including: Step 1: Preprocessing Bouguer gravity anomalies to form basic data for processing and interpretation; Step 2: Separating the source field of the gravity difference trend surface to quantitatively characterize the gravity anomaly response of the target stratum; Step 3: Enhancement and processing of the gravity anomaly to enhance weak linear structural information and highlight the linear images of small-scale faults; Step 4: Combining multiple processing and conversion results to establish a nonlinear objective function for narrowing gradient bands; Step 5: Comprehensive analysis of the fault linear images to determine the occurrence and combination patterns of faults at different levels. This method for identifying gravity fault images based on tectonic context improves the effectiveness and practicality of fine-grained characterization of fault targets using gravity data, enhances lateral resolution, highlights fault boundary features, and enables clearer identification of the fault structural framework.

[0004] Chinese patent application CN201910664831.1 discloses a method for earthquake fault identification based on variable-neighborhood sliding window machine learning. The method converts raw earthquake data into a three-dimensional matrix data volume; creates sliding windows of different scales for each data point in the three-dimensional matrix data volume, and combines sliding windows of the same scale into a corresponding scale matrix; applies a corresponding discriminant model to the scale matrix to perform fault identification, generating separate identification result tensors, and then integrates the three identification result tensors into a matrix; and applies a decision tree model to the matrix to identify faults. Compared with existing technologies, the method boasts high parallelism and greater efficiency than other global optimal algorithms, making better use of multiple processor cores on large computers. The method also eliminates the difficulty in determining physical dimensions in traditional earthquake data research, as it utilizes a variable neighborhood rather than a fixed neighborhood, allowing for the consideration of multiple scale effects through multiple fields of view.

[0005] Chinese patent application CN201911159916.0 discloses a method for identifying faults based on the total horizontal gravity gradient of the target layer, which belongs to the field of petroleum geophysical exploration. The method includes the following steps: gridding Bouguer gravity anomalies; upward analytical extension of the target layer to obtain the target layer regional background gravity anomaly and the target layer residual gravity anomaly; processing the total horizontal gravity gradient of the target layer to initially extract fault information; enhancing the target layer's inclination angle to ultimately extract fault information in the target layer; restoring the structural background; and tracing and interpreting the target layer's faults.

[0006] In recent years, deep learning technology has demonstrated great application potential in fields such as image recognition and speech recognition, and has received widespread attention in the field of geophysics. For example, it has achieved good application results in fault identification, river channel detection, and reservoir inversion interpretation of seismic data. However, no relevant research has been carried out in the field of gravity fault identification.

[0007] The above existing technologies are significantly different from the present invention and fail to solve the technical problem we want to solve. To this end, we have invented a new automatic identification method of gravity faults based on deep learning. Summary of the Invention

[0008] The purpose of this invention is to provide a deep learning-based gravity fault automatic identification method that provides technical support for mineral resource exploration and deployment by finding more efficient and reliable technical means and processes for automatic identification of gravity faults.

[0009] The purpose of the present invention can be achieved by the following technical measures: a method for automatic identification of gravity faults based on deep learning, the method for automatic identification of gravity faults based on deep learning comprising:

[0010] Step 1, constructing a three-dimensional fault geological model;

[0011] Step 2: Fill the 3D geological model with density;

[0012] Step 3, constructing three-dimensional fault gravity sample data;

[0013] Step 4: construct gravity fault sample labels;

[0014] Step 5: construct a gravity fault identification network based on deep learning;

[0015] Step 6: Using the optimization algorithm to train the network parameters and obtain the gravity fault automatic identification network model;

[0016] Step 7: Output and apply the gravity fault automatic identification network model.

[0017] The purpose of the present invention can also be achieved by the following technical measures:

[0018] In step 1, key parameters such as undulating structure, fault dip, fault azimuth, and fault burial depth are set to generate a large number of three-dimensional fault geological models.

[0019] In step 1, simulate the undulating structure and randomly set a three-dimensional geological geometric structure model H containing multiple strata. Assume that the dimensions of the model to be constructed in the x, y, and z directions are n. x 、n y 、n z In order to eliminate the boundary effect caused by the undulating structure, n xyout grids, expanding n downwards along the z direction zout grids, that is, the initial size of the model H becomes n x +2n xyout 、n y +2n xyout 、n z +n zout ;

[0020] The structural undulation effect is applied to the model H. The undulating structure is composed of the surfaces generated by the Gaussian source. The specific formula is:

[0021]

[0022] Among them, b k 、c k d k , σ k are the amplitude, X-direction center position, Y-direction center position and variance of the k-th Gaussian source respectively; is the depth coordinate after structural undulation.

[0023] In step 1, simulate a 3D cross section; randomly set the cross section center point (x0, y0, z0), cross section strike azimuth, and cross section inclination parameters, and establish a new coordinate system with the center point (x0, y0, z0) as the reference point, as shown below:

[0024]

[0025] Among them, R is a three-dimensional rotation transformation matrix, which is in the form of:

[0026]

[0027] Among them, φ is the section strike azimuth, θ is the section dip; to simulate the curved section, multiple sine functions are combined and constructed in the new coordinate system.

[0028]

[0029] Among them, f(x,y) is the curved section, A and B are the amplitude of the sine function, w, are the frequency and initial phase of the sine function respectively, K and L are the number of sine functions;

[0030] After the cross section is constructed, the displacement d of the strata on both sides of the cross section is calculated in the new coordinate system.

[0031] In step 1, the calculated results of the new coordinate system after displacement are transformed back to the original coordinate system; the stratum model is interpolated three-dimensionally to obtain a three-dimensional fault model in the original coordinate system, and the corresponding points near the cross section f(x, y) are used as the three-dimensional fault sample label positions.

[0032] In step 2, according to the stratum distribution in the established three-dimensional geological geometric structure model H, random residual density is assigned to different strata according to the density range of the actual situation. In this process, a certain low-frequency trend of increasing density with depth is considered.

[0033] In step 3, the gravity anomaly generated by the three-dimensional geological density body is calculated using the three-dimensional forward gravity formula, and noise interference is added to it to simulate the measured gravity anomaly.

[0034] In step 3, gravity forward modeling is performed according to the right hexahedron formula; gravity anomalies are calculated for each right hexahedron unit of the underground three-dimensional model and superimposed and summed to obtain the forward simulation results of the gravity anomaly; in order to eliminate the boundary effect during the forward modeling process, the horizontal expansion grid size of the formation model H is increased by a certain proportion, and the simulated density model is infinitely extended horizontally outward to eliminate the boundary effect and highlight the gravity anomaly caused by the tectonic changes in the gravity and magnetic observation area; finally, a certain amount of noise interference is added to the gravity anomaly to ensure that the forward simulation data is closer to the real data.

[0035] In step 4, the plane position of the fault sample label corresponding to the gravity anomaly forward modeling area is determined based on the fault distribution in the 3D fault geological model. Since the section is a 3D volume and the gravity anomaly is a 2D plane data, the 3D fault label data volume is summed along the depth direction to obtain the 2D label projection result, which is matched with the gravity observation point. Considering that the gravity anomaly and the section projection point are not strictly corresponding, the points near the center line of the fault plane are used as the final fault plane sample labels.

[0036] In step 5, the fault sample gravity data is used as input and the fault plane label is used as output to construct a deep learning network structure; the identification of the fault plane position from gravity plane data is regarded as an image semantic segmentation problem; the fault plane sample label position is calibrated to 1, and other non-fault positions are calibrated to 0; and the U-shaped convolutional neural network structure is used to automatically identify gravity faults.

[0037] In step 5, the input data is the gravity plane data with random noise added, and the output sample label is the two-dimensional fault plane sample label with the same size as the gravity plane data; a three-layer network structure is used to complete encoding and decoding; in the encoding process, each layer uses a pooling layer for downsampling, and then uses a convolution layer to extract features, using ReLU as the activation function; in the decoding process, the transposed convolution operation and ReLU activation function are used to reconstruct the features; the first, second, and third layers of the decoding process are combined and spliced with the third, second, and first layers of the encoding process in the channel direction, respectively, to introduce feature information at the corresponding scale, providing multi-scale and multi-level information for subsequent fault identification; after the decoding layer, a 1×1 convolution operation is applied to match the size of the hidden layer output to the input data size; finally, the SoftMax layer is used to convert it into a probability map, and the Generalized Dice loss is used as the loss function, and the network parameters are optimized through Adam.

[0038] In step 6, a large amount of gravity forward modeling sample data and fault sample labels are used to train the convolutional neural network, and a certain validation data set is retained to test the reliability of the deep learning network model. At the same time, considering that the gravity anomaly values in different regions may vary greatly in actual situations, all gravity data for training, validation, and subsequent actual testing are normalized.

[0039]

[0040] Where Δg represents the original gravity data, μ Δg represents the mean amplitude of gravity data, σ Δg represents the gravity anomaly variance value, and Δg′ represents the normalization processing result of the gravity data volume;

[0041] Initialize the network weights, set training parameters such as iteration rounds and descent rate, and use the Adam method for optimization.

[0042] In step 7, the actual input data is gridded and preprocessed according to needs, including curve flattening, anomaly separation, and denoising. The trained automatic recognition model is then used to characterize the fault. During the prediction process, a sliding window prediction is performed on the actual data according to the size of the training model input. Finally, the recognition results of each sliding window are spliced to obtain the fault automatic recognition result.

[0043] The deep learning-based automatic gravity fault identification method proposed in this paper automatically generates a large number of realistic fault geometry and density models, obtains corresponding gravity data through three-dimensional forward modeling, and then uses deep learning algorithms to automatically analyze the nonlinear mapping relationship between gravity anomaly characteristics and fault spatial positions, fully exploring the fault structural information associated with gravity anomalies. In comparison, conventional gravity fault interpretation involves the selection of different algorithms and requires extensive parameter optimization, resulting in a complex and subjective process. The proposed method, however, directly obtains fault positions in all directions from end to end, boasting the advantages of high efficiency and reliability, and can provide new technical solutions for regional gravity structural interpretation. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flowchart of a specific embodiment of the method for automatic identification of gravity faults based on deep learning of the present invention;

[0045] Figure 2 A diagram of a three-dimensional fault model with complex structures according to a specific embodiment of the present invention;

[0046] Figure 3 A density model and a gravity anomaly map corresponding to the three-dimensional fault model according to a specific embodiment of the present invention;

[0047] Figure 4 A superimposed diagram of gravity forward simulation data and fault plane sample labels according to a specific embodiment of the present invention;

[0048] Figure 5 This is a structural diagram of a deep convolutional neural network for automatic identification of gravity faults according to a specific embodiment of the present invention;

[0049] Figure 6 A residual gravity anomaly map used in the test described in a specific embodiment of the present invention;

[0050] Figure 7 This is a diagram showing the automatic fault identification results of residual gravity anomalies used in a test of the present invention. DETAILED DESCRIPTION

[0051] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0052] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations and / or combinations thereof.

[0053] like Figure 1 As shown, Figure 1 The flowchart of the method for automatic identification of gravity faults based on deep learning of the present invention is as follows:

[0054] 1) Construction of 3D fault geological models. Set key parameters such as undulating structure, fault dip, fault azimuth, and fault burial depth to generate a large number of 3D fault geological models;

[0055] (2) Density filling of 3D geological model. Based on the construction of 3D fault geological model, fill in the density value of the stratum according to the actual situation;

[0056] (3) Construction of 3D fault gravity sample data. Using the 3D forward gravity formula, the gravity anomaly generated by the 3D geological density body is calculated, and noise interference is added to it to simulate the measured gravity anomaly;

[0057] (4) Construction of gravity fault sample labels. According to the fault distribution in the three-dimensional fault geological model, the plane position of the fault sample labels corresponding to the gravity anomaly forward modeling area is determined;

[0058] (5) Construction of gravity fault identification network based on deep learning. Take the fault sample gravity data as input and the fault plane label as output to build a deep learning network structure;

[0059] (6) Using the optimization algorithm to train the network parameters and obtain the network model for automatic identification of gravity faults;

[0060] (7) Output and application of the network model for automatic identification of gravity faults. The measured data are processed and the fault plane position corresponding to the processed gravity anomaly is obtained using the trained network model.

[0061] The following are several specific embodiments of the present invention.

[0062] Example 1:

[0063] In a specific embodiment 1 of the present invention, the deep learning-based automatic gravity fault identification method of the present invention includes the following steps:

[0064] Step 1: Set key parameters such as undulating structure, fault dip, fault azimuth, and fault burial depth to generate a large number of three-dimensional fault geological models;

[0065] First, simulate the undulating structure. Randomly set up a three-dimensional geological geometric structure model H containing multiple strata, assuming that the dimensions of the model to be constructed in the x, y, and z directions are n x 、n y 、n z In order to eliminate the boundary effect caused by the undulating structure, n xyout grids, expanding n downwards along the z direction zout grids, that is, the initial size of the model H becomes n x +2n xyout 、n y +2n xyout 、n z +n zout .

[0066] The structural undulation effect is applied to the model H. The undulating structure is composed of the surfaces generated by the Gaussian source. The specific formula is:

[0067]

[0068] Among them, b k 、c k d k , σ k are the amplitude, X-direction center position, Y-direction center position and variance of the k-th Gaussian source respectively; is the depth coordinate after structural undulation.

[0069] After that, simulate the 3D cross section. Randomly set the cross section center point (x0, y0, z0), cross section strike azimuth, cross section inclination and other parameters. With the center point (x0, y0, z0) as the reference point, establish a new coordinate system as shown below:

[0070]

[0071] Among them, R is a three-dimensional rotation transformation matrix, which is in the form of:

[0072]

[0073] Among them, φ is the cross-section strike azimuth, and θ is the cross-section inclination. To simulate the curved cross-section, multiple sine functions are combined and constructed in the new coordinate system.

[0074]

[0075] Among them, f(x,y) is the curved section, A and B are the amplitude of the sine function, w, are the frequency and initial phase of the sine function respectively, and K and L are the number of sine functions.

[0076] After the cross section is constructed, the displacement d of the strata on both sides of the cross section is calculated in the new coordinate system.

[0077] Again, the calculated results of the new coordinate system after displacement are transformed back to the original coordinate system.

[0078] Finally, the stratum model is interpolated three-dimensionally to obtain the three-dimensional fault model in the original coordinate system, and the corresponding points near the cross section f(x, y) are used as the three-dimensional fault sample label positions.

[0079] Step 2: Based on the construction of the three-dimensional fault geological model, fill in the density value of the stratum according to the actual situation.

[0080] According to the stratum distribution in the established three-dimensional geological geometric structure model H, random residual density is assigned to different strata according to the actual density range. In this process, a certain low-frequency trend of increasing density with depth can be considered.

[0081] Step 3: Use the three-dimensional gravity forward modeling formula to calculate the gravity anomaly generated by the three-dimensional geological density body, and add noise interference to it to simulate the measured gravity anomaly.

[0082] Gravity forward modeling is performed using the hexahedron formula. Gravity anomalies are calculated for each hexahedron element in the three-dimensional subsurface model and summed up to obtain the forward modeling results. To eliminate boundary effects during the forward modeling process, the horizontal grid size of the stratigraphic model H can be increased proportionally. This simulates the density model's horizontal extension, eliminating boundary effects and highlighting gravity anomalies caused by tectonic changes in the gravity and magnetic observation area. Finally, a certain amount of noise interference is added to the gravity anomaly to ensure that the forward modeled data is closer to the actual data.

[0083] Step 4: According to the fault distribution in the three-dimensional fault geological model, the plane position of the fault sample label corresponding to the gravity anomaly forward modeling area is determined.

[0084] Since the cross section is a three-dimensional volume, while the gravity anomaly is a two-dimensional plane data, the three-dimensional fault label data volume is summed along the depth direction to obtain the two-dimensional label projection result, which is then matched with the gravity observation points. Considering that the gravity anomaly and the cross section projection points do not strictly correspond, the points near the centerline of the fault plane are used as the final fault plane sample labels.

[0085] Step 5: Take the fault sample gravity data as input and the fault plane label as output to build a deep learning network structure.

[0086] Identifying fault plane locations from gravity plane data is treated as an image semantic segmentation problem. Sample locations on the fault plane are labeled as 1, while other non-fault locations are labeled as 0. A U-shaped convolutional neural network architecture is used for automatic gravity fault identification.

[0087] The input data consists of gravity plane data with random noise added, and the output sample labels are two-dimensional fault plane sample labels of the same size as the gravity plane data. A three-layer network structure is used for encoding and decoding. During the encoding process, each layer uses a pooling layer for downsampling, followed by a convolutional layer to extract features, using ReLU as the activation function. During the decoding process, features are reconstructed using a transposed convolution operation and a ReLU activation function. The first, second, and third layers of the decoding process are combined and spliced with the third, second, and first layers of the encoding process, respectively, along the channel direction, incorporating feature information at corresponding scales and providing multi-scale and multi-level information for subsequent fault identification. After the decoding layer, a 1×1 convolution operation is applied to match the size of the hidden layer output to the input data size. Finally, a SoftMax layer is used to convert the hidden layer output into a probability map, and the generalized dice loss (GDL) is used as the loss function. The network parameters are optimized using Adam.

[0088] Step 6: Use the optimization algorithm to train the network parameters and obtain the gravity fault automatic identification network model.

[0089] A large amount of gravity forward modeling data and fault sample labels were used to train the convolutional neural network, and a certain validation dataset was retained to test the reliability of the deep learning network model. At the same time, considering that gravity anomalies in different regions may vary significantly in actual situations, all gravity data for training, validation, and subsequent actual testing were normalized.

[0090]

[0091] Where Δg represents the original gravity data, μ Δg represents the mean amplitude of gravity data, σ Δg represents the gravity anomaly variance, and Δg′ represents the normalized processing result of the gravity data volume.

[0092] Initialize the network weights, set training parameters such as iteration rounds and descent rate, and use the Adam method for optimization.

[0093] Step 7: Output and application of the gravity fault automatic identification network model. Process the measured data and use the trained network model to obtain the fault plane position corresponding to the processed gravity anomaly.

[0094] The input data is gridded and preprocessed as needed, including curve flattening, anomaly separation, and denoising. The trained automatic recognition model is then used to characterize faults. During the prediction process, a sliding window prediction is performed on the actual data based on the size of the training model input. Finally, the recognition results of each sliding window are spliced together to obtain the automatic fault recognition result.

[0095] Example 2:

[0096] In the second specific embodiment of the present invention, the specific steps of the deep learning-based gravity fault automatic identification method of the present invention are as follows:

[0097] (1) Construction of three-dimensional fault geological models. Set key parameters such as undulating structure, fault dip, fault azimuth, and fault burial depth to generate a large number of three-dimensional fault geological models;

[0098] (2) Density filling of 3D geological model. Based on the construction of 3D fault geological model, fill in the density value of the stratum according to the actual situation;

[0099] (3) Construction of 3D fault gravity sample data. Using the 3D forward gravity formula, the gravity anomaly generated by the 3D geological density body is calculated, and noise interference is added to it to simulate the measured gravity anomaly;

[0100] (4) Construction of gravity fault sample labels. According to the fault distribution in the three-dimensional fault geological model, the plane position of the fault sample labels corresponding to the gravity anomaly forward modeling area is determined;

[0101] (5) Construction of gravity fault identification network based on deep learning. Take the fault sample gravity data as input and the fault plane label as output to build a deep learning network structure;

[0102] (6) Using the optimization algorithm to train the network parameters and obtain the network model for automatic identification of gravity faults;

[0103] (7) Output and application of the network model for automatic identification of gravity faults. The measured data are processed and the fault plane position corresponding to the processed gravity anomaly is obtained using the trained network model.

[0104] Example 3

[0105] The present invention will be further described below with reference to specific embodiments.

[0106] Figure 2This is a 3D fault model with complex structures. As shown in the figure, the subsurface half-space segmentation ranges from -8 to 8 kilometers in the X and Y directions, with intervals of 250 meters. The subsurface half-space segmentation ranges from 0 to 10 kilometers in the Z direction, with intervals of 500 meters. The fault model includes three randomly generated sections.

[0107] Figure 3 The density model and gravity anomaly map corresponding to the 3D fault model are shown below. Density is randomly distributed along the vertical direction. The two sides of the fault in the same stratum have vertical displacement, which causes the density distribution to change with burial depth, thus causing the gravity anomaly to change.

[0108] Figure 4 This is an overlay of gravity forward modeling data and fault plane sample labels. The 3D fault volume data is overlaid along the depth direction and projected onto a 2D plane. There is a certain correlation between the marked fault locations and the changes in gravity anomalies.

[0109] Figure 5 Diagram of the deep convolutional neural network architecture for automatic gravity fault identification. The input data consists of gravity plane data with random noise added, and the output sample labels are two-dimensional fault plane sample labels of the same size as the gravity plane data. A three-layer network architecture is used for encoding and decoding. During the encoding process, each layer includes a 2×2 pooling layer (MaxPooL) for downsampling, followed by a 3×3 convolution (Conv2D) for feature extraction, using ReLU as the activation function. The number of channels per layer is 64, 128, and 256, respectively. During the decoding process, features are reconstructed using a 3×3 transposed convolution (DeConv2D) operation and ReLU activation function, with the number of channels per layer being 256, 128, and 64, respectively. The first, second, and third layers of the decoding process are combined and concatenated with the third, second, and first layers of the encoding process, respectively, along the channel direction, incorporating feature information at corresponding scales and providing multi-scale and multi-layer information for subsequent fault identification. After the decoding layer, a 1×1 convolution is applied to scale the hidden layer output to the input data size. Finally, the SoftMax layer is used to convert it into a probability map, and the generalized dice loss (GDL) is used as the loss function to optimize the network parameters through Adam.

[0110] Figure 6 Residual gravity anomaly map used for testing.

[0111] Figure 7 The fault automatic identification result diagram of the residual gravity anomaly used in the test. Figure 6After the residual gravity data used in the test were normalized, the corresponding fault plane identification results were obtained using the trained recognition model. They have a good correspondence with the actual fault distribution, and the breakpoint results are clear, providing a strong reference for subsequent structural interpretation.

[0112] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art may modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features therein. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0113] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. A method for automatic identification of gravity faults based on deep learning, characterized in that: The deep learning-based automatic gravity fault identification method includes: Step 1, constructing a three-dimensional fault geological model; Step 2, performing density filling of the three-dimensional fault geological model; Step 3, constructing three-dimensional fault gravity sample data; Step 4: construct gravity fault sample labels; Step 5: construct a gravity fault identification network based on deep learning; Step 6: Using the optimization algorithm to train the network parameters and obtain the gravity fault automatic identification network model; Step 7: Output and apply the gravity fault automatic identification network model; In step 1, key parameters such as undulating structure, fault dip, fault azimuth, and fault burial depth are set to generate a large number of three-dimensional fault geological models; In step 1, simulate the undulating structure and randomly set a three-dimensional geological geometric structure model H containing multiple strata. Assume that the dimensions of the model to be constructed in the x, y, and z directions are n. x 、n y 、n z In order to eliminate the boundary effect caused by the undulating structure, n xyout grids, expanding n downwards along the z direction zout grids, that is, the initial size of the model H becomes n x +2n xyout 、n y +2n xyout 、n z +n zout ; The structural undulation effect is applied to the three-dimensional geological geometric structure model H. The undulating structure is composed of the surfaces generated by the Gaussian source. The specific formula is: Among them, b k 、c k d k , σ k are the amplitude, X-direction center position, Y-direction center position and variance of the k-th Gaussian source respectively; is the depth coordinate after structural undulation.

2. The method for automatic identification of gravity faults based on deep learning according to claim 1, characterized in that: In step 1, simulate a 3D cross section; randomly set the cross section center point (x0, y0, z0), cross section strike azimuth, and cross section inclination parameters, and establish a new coordinate system with the center point (x0, y0, z0) as the reference point, as shown below: Among them, R is a three-dimensional rotation transformation matrix, which is in the form of: Among them, φ is the section strike azimuth, θ is the section dip; to simulate the curved section, multiple sine functions are combined and constructed in the new coordinate system. Among them, f(x,y) is the curved section, A and B are the amplitude of the sine function, w, are the frequency and initial phase of the sine function respectively, K and L are the number of sine functions; After the cross section is constructed, the displacement d of the strata on both sides of the cross section is calculated in the new coordinate system.

3. The method for automatic identification of gravity faults based on deep learning according to claim 2, characterized in that: In step 1, the calculated results of the new coordinate system after displacement are inversely transformed back to the original coordinate system; the stratum model is interpolated three-dimensionally to obtain a three-dimensional fault geological model in the original coordinate system, and the corresponding points near the cross section f(x, y) are used as the three-dimensional fault sample label positions.

4. The method for automatic identification of gravity faults based on deep learning according to claim 1, characterized in that: In step 2, according to the stratum distribution in the established three-dimensional geological geometric structure model H, random residual density is assigned to different strata according to the density range of the actual situation. In this process, a certain low-frequency trend of increasing density with depth is considered.

5. The method for automatic identification of gravity faults based on deep learning according to claim 1, characterized in that: In step 3, the gravity anomaly generated by the three-dimensional geological density body is calculated using the three-dimensional forward gravity formula, and noise interference is added to it to simulate the measured gravity anomaly.

6. The method for automatic identification of gravity faults based on deep learning according to claim 5, characterized in that: In step 3, gravity forward modeling is performed according to the right hexahedron formula; gravity anomalies are calculated for each right hexahedron unit of the underground three-dimensional geological geometric structure model H and are superimposed and summed to obtain the forward modeling result of gravity anomaly; In order to eliminate the boundary effect during the forward modeling process, the horizontal expansion grid size of the three-dimensional geological geometric structure model H of the stratum is increased according to a certain proportion, and the simulated density model is infinitely extended horizontally outward to eliminate the boundary effect and highlight the gravity anomaly caused by the tectonic changes in the gravity and magnetic observation area. Finally, a certain amount of noise interference is added to the gravity anomaly to ensure that the forward simulation data is closer to the real data.

7. The method for automatic identification of gravity faults based on deep learning according to claim 6, characterized in that: In step 4, according to the fault distribution in the three-dimensional fault geological model, the plane position of the gravity fault sample label corresponding to the gravity anomaly forward modeling area is determined; since the cross section is a three-dimensional body, and the gravity anomaly is a two-dimensional plane data; Therefore, the sample label data volume of the three-dimensional fault gravity fault is summed along the depth direction to obtain the two-dimensional label projection result, which is matched with the gravity observation point; Considering that the gravity anomaly and the cross-section projection points are not strictly corresponding, the points near the center line of the fault plane are used as the final two-dimensional fault plane sample labels.

8. The method for automatic identification of gravity faults based on deep learning according to claim 7, characterized in that: In step 5, the gravity fault sample data is used as input and the fault plane sample label is used as output to construct a deep learning network structure; the identification of the fault plane position from the gravity plane data is regarded as an image semantic segmentation problem; the fault plane sample label position is calibrated to 1, and other non-fault positions are calibrated to 0; and the U-shaped convolutional neural network structure is used to automatically identify gravity faults.

9. The method for automatic identification of gravity faults based on deep learning according to claim 8, characterized in that: In step 5, the input data is the gravity plane data with random noise added, and the output sample label is the two-dimensional fault plane sample label with the same size as the gravity plane data; a three-layer network structure is used to complete the encoding and decoding; During the encoding process, each layer uses a pooling layer for downsampling, and then a convolution layer is used to extract features, using ReLU as the activation function. During the decoding process, the features are reconstructed using a transposed convolution operation and a ReLU activation function. The first, second, and third layers of the decoding process are combined and spliced with the third, second, and first layers of the encoding process, respectively, in the channel direction, introducing feature information at the corresponding scale, providing multi-scale and multi-level information for subsequent fault identification. After the decoding layer, a 1×1 convolution operation is applied to match the size of the hidden layer output to the input data size; finally, the SoftMax layer is used to convert it into a probability map, and Generalized Dicelos s is used as the loss function, and Adam is used to optimize the network parameters.

10. The method for automatic identification of gravity faults based on deep learning according to claim 1, characterized in that: In step 6, a large amount of gravity forward modeling sample data and gravity fault sample labels are used to train the convolutional neural network, and a certain validation data set is retained to test the reliability of the deep learning network model. At the same time, considering that the gravity anomaly values in different regions may vary greatly in actual situations, all gravity data for training, validation, and subsequent actual testing are normalized. Where Δg represents the original gravity data, μ Δg represents the mean amplitude of gravity data, σ Δg represents the gravity anomaly variance value, and Δg′ represents the normalization processing result of the gravity data volume; Initialize the network weights, set the training parameters such as iteration rounds and descent rate, and use the Adam method for optimization.

11. The method for automatic identification of gravity faults based on deep learning according to claim 1, characterized in that: In step 7, the input data is gridded and preprocessed according to the requirements, including curve flattening, anomaly separation, and denoising. The trained automatic recognition model is then used to characterize the fault. During the prediction process, a sliding window prediction is performed on the actual data according to the size of the training model input, and finally the recognition results of each sliding window are spliced to obtain the automatic fault recognition result.

Images