Deep learning coupling constraint-based gravity-magnetic joint inversion method, system and terminal

Abstract

The application belongs to the technical field of gravity and magnetic exploration, and discloses a gravity and magnetic joint inversion method, system and terminal based on deep learning coupling constraint, which generates sample data sets through forward modeling and Gaussian blur simulation based on a gravity and magnetic anomaly body model, and constructs a Trans_FCN inversion network; the sample data sets are used to train the Trans_FCN gravity and magnetic model conversion network to realize the'soft' mapping between gravity and magnetic data; finally, the trained deep neural network provides a 'correct' reference model for joint inversion. The application can accurately invert the position and physical property of a gravity and magnetic anomaly body, has strong learning ability and certain generalization ability, and can effectively solve the gravity and magnetic anomaly data inversion problem. Different physical properties and size models are designed, the model position is moved, different physical property combinations are set, and a large amount of data is obtained through Gaussian blur, so that the physical property coupling and structure coupling relationship between gravity and magnetic models can be learned at the same time, and the joint inversion effect can be effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of gravity and magnetic exploration technology, and discloses a gravity and magnetic joint inversion method, system, terminal and medium based on deep learning coupling constraints. Background Technology

[0002] Currently, in recent years, single-property inversion methods have matured. Adding depth weighting, smoothing constraints, and property constraints to traditional methods can make the inversion results more reliable, mitigating the ambiguity problem in geophysical inversion to some extent. However, because different property inversion methods are subject to varying degrees of external conditions during data acquisition, single-property inversion has limitations. Since inversion results from different properties in the same region are complementary, multi-property joint inversion can be used to jointly interpret the acquired geophysical data, reducing the impact of observational errors on the reliability of the inversion results and further reducing the ambiguity of geophysical inversion.

[0003] In the process of geophysical data acquisition, gravity and magnetic data are often acquired simultaneously, providing the necessary conditions for joint gravity and magnetic inversion. Gravity data can invert the morphology of deep basements and the distribution of deep faults, while magnetic data can solve the problem of the distribution of different rock types and also invert shallow structural information. Moreover, joint gravity and magnetic inversion has a stronger resolution in depth than inversion using single gravity and magnetic data. Due to the complementarity between gravity and magnetic data in geophysical data inversion, it has become a hot topic of research for many scholars. For example, Zhou Junjie's doctoral dissertation studied the joint inversion methods of structural coupling and physical property coupling to solve the problem of three-dimensional joint gravity and magnetic inversion, and performed cross-gradient joint inversion of gravity and magnetic data under the conditions of considering remanence and normalized magnetic source intensity. Fregoso E et al. and Gross L used the cross-gradient method for joint gravity and magnetic inversion, Bosch M et al. used lithological constraints for joint gravity and magnetic inversion, and Guo Lianghui et al. extended the three-dimensional imaging method of gravity and gravity gradient data to the field of magnetic exploration, laying a solid foundation for joint gravity and magnetic inversion.

[0004] Thanks to advancements in big data storage, massive parallelization, and computational optimization, deep learning methods are widely used in the field of single-property inversion. Examples include using neural networks to invert gravity and magnetic data separately, using deep learning to invert seismic data, using deep learning for joint inversion of gravity anomalies and gravity gradient anomalies, and using deep learning for joint inversion of magnetic anomalies and magnetic gradient anomalies. However, deep learning is less commonly used in multi-property joint inversion.

[0005] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:

[0006] 1) Existing gravity and magnetic inversion methods rely on reference models, have long computation times, and have limited data, which cannot meet the demand for large amounts of gravity and magnetic data required when using deep learning to solve geophysical inversion problems.

[0007] (2) Existing gravity and magnetic joint inversion methods may encounter “skin effect” and multiple solutions problem when reconstructing the magnetic susceptibility model or density model of geological bodies.

[0008] (3) Existing physical property coupling or structural coupling joint inversion methods have their own defects and cannot accurately invert the specific location and shape of the underground model. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides a gravity and magnetic field joint inversion method, system, and terminal based on deep learning coupling constraints.

[0010] This invention is implemented as follows: a joint gravity and magnetic inversion method based on deep learning coupling constraints, comprising:

[0011] A sample dataset was generated using a gravity and magnetic anomaly model, and a gravity and magnetic model transformation network (Trans_FCN) was constructed.

[0012] The gravity and magnetism model conversion network Trans_FCN is trained based on the sample dataset;

[0013] The Trans_FCN model transformation network is used to provide a "correct" reference model for the joint gravity and magnetic inversion, ultimately resulting in more reliable inversion results.

[0014] Furthermore, the gravity and magnetic field joint inversion method based on deep learning coupling constraints includes the following steps:

[0015] Step 1: Design a large number of gravity and magnetic anomaly models using Matlab language, and generate a sample dataset by performing Gaussian blurring.

[0016] Step 2: Construct a gravity and magnetism model conversion network, and use the generated sample dataset to train the constructed Trans_FCN network to optimize the network parameters of the Trans_FCN network;

[0017] Step 3: Input the individual gravity and magnetic inversion results into the trained Trans_FCN network to provide a "correct" reference model for the next gravity and magnetic inversion, realize the interaction of gravity and magnetic property models, improve the reliability of joint gravity and magnetic inversion, and repeat step 3 to finally obtain the gravity and magnetic inversion results.

[0018] Furthermore, the specific process in step one is as follows:

[0019] (1.1) Given the shape of the anomaly: First, determine a subsurface model area of ​​a specified size, then mesh the area, and finally design some different anomaly blocks and traverse every position of the subsurface mesh model. When traversing the subsurface mesh model, anomaly models near the edge will be discarded.

[0020] (1.2) Given the magnetic parameters of the anomalous body: Assign values ​​to the anomalous body model in step (1.1) according to the known relationship between the magnetic susceptibility and density of the underground medium in a certain place, and obtain the corresponding anomalous body density model and magnetic susceptibility model;

[0021] (1.3) Gaussian blurring: Since the smooth constraint is added in the Gauss-Newton method inversion process, the inversion result has no sharp boundary. Therefore, Gaussian filtering is applied to the anomaly density model and magnetic susceptibility model generated in step (1.2) to blur the inversion transformation result of the gravity and magnetic subsurface model. This step also increases the generalization ability of Trans_FCN.

[0022] Furthermore, since the density and magnetic susceptibility of the anomalous body are selected based on prior knowledge, the deep neural network can learn the physical property coupling relationship between gravity and magnetism. In step (1.3), Gaussian blurring is applied to both the input and output physical property models of the neural network before training, so that the neural network can learn the structural coupling relationship between gravity and magnetism data from the training set.

[0023] The two inversion constraint networks proposed in this invention can share the same training set, which greatly reduces the computational cost required to build the training set data.

[0024] Furthermore, the gravity and magnetic property conversion network Trans_FCN includes:

[0025] The input data size of Trans_FCN is M×N×K, and the output data size is also M×N×K, where M, N, and K are the number of grids in the x, y, and z directions, respectively.

[0026] The Trans_FCN network consists of three parts: an encoding part, a decoding part, and skip connections. The encoding part includes multiple 3×3 convolutional layers, batch normalization layers, ReLU activation functions, and 2×2 max pooling layers. Meanwhile, the decoding part uses 2×2 deconvolutional layers, multiple 3×3 convolutional layers, batch normalization layers, and ReLU activation functions. The corresponding decoding recovery part represents the output.

[0027] Skip connections are used to connect encoding and decoding paths.

[0028] Furthermore, step two includes:

[0029] First, the sample dataset is preprocessed by randomly dividing it into a training set and a test set.

[0030] Then, set the parameters, train the Trans_FCN network using the training set, and use the test set to check whether the network can achieve mutual conversion between gravity and magnetic underground anomaly physical property models without distortion. If not, modify the parameters and train again until the set maximum number of iterations is reached.

[0031] Finally, if the trained network achieves the expected results, the density model is obtained by inputting the magnetic susceptibility model, or the magnetic susceptibility model is obtained by inputting the density model.

[0032] Furthermore, the parameter modification in step two includes:

[0033] Given the objective function, and using mean squared error to compare the difference between the predicted and actual results, the loss function is:

[0034]

[0035] Where f i For the prediction results of the neural network, l i For a realistic physical property model, N t The number of iterations is 5000, representing the number of meshes.

[0036] Furthermore, the gravity and magnetic field inversion process in step three includes:

[0037] (3.1) Calculate the objective function using gravity and magnetic geophysical data, which includes observation data, initial model, reference model, depth weighting, smoothing constraints, and physical property constraints;

[0038] (3.2) Differentiate the objective function and calculate whether the derivative value reaches the threshold;

[0039] (3.3) If the threshold or the maximum number of iterations is reached, a separate inversion result is obtained. If the threshold or the maximum number of iterations is not reached, the objective function is solved using the Gauss-Newton method and the initial model is updated. Step (3.1) is repeated until the threshold is reached.

[0040] The specific deep learning-coupled gravity and magnetic field joint inversion process is as follows: Figure 2 As shown, DFCNG2M and DFCMNM2G are the neural networks (hereinafter referred to as Trans_FCN) used in this invention for converting underground physical property models.

[0041] Furthermore, the gravity and magnetic field joint inversion process based on deep learning coupling constraints also includes:

[0042] First, the gravity anomaly data inversion process is solved using the Gauss-Newton method through a gravity geophysical model, and the density model is updated accordingly. Then, the density model is converted into a magnetic susceptibility model using a trained DFCNG2M deep neural network to update the magnetic anomaly inversion reference model and participate in the solution of the objective function for magnetic anomaly inversion.

[0043] Then, the magnetic susceptibility model is updated by using the Gauss-Newton method to solve the magnetic anomaly data inversion process.

[0044] The magnetic susceptibility model is converted into a density model using a trained DFCMN2G deep neural network to update the gravity anomaly inversion reference model and participate in the solution of the objective function for gravity anomaly inversion.

[0045] Finally, repeat the above process N times until the maximum number of iterations is reached, to obtain the final density model and magnetic susceptibility model, and the inversion ends.

[0046] The process of performing a separate inversion using the Gauss-Newton method is as follows: Figure 7 As shown, the objective function for gravity and magnetic inversion that needs to be solved using the Gauss-Newton method in this invention is:

[0047] Φ(m)=Φ d (m)+Φ m (m)+Φ l (m)

[0048] Where Φ d (m) represents the data fitting term, Φ m (m) represents the model constraint term, Φ l (m) is the smoothing constraint term. Expanding the above formula further, we get:

[0049]

[0050]

[0051]

[0052] Among them || || 2 For L2 norm, C d C m G represents the covariance matrix of the data constraint term and the model constraint term, respectively. z W is a depth-weighted kernel matrix, where d0 represents the observed data. z The total depth weighting matrix is ​​given by m0, which is the reference model; D x D y D z These are the difference operator matrices for the three axial directions, C lx C ly Clz Let this be its corresponding covariance matrix. Smoothness constraints and physical property constraints are used here to mitigate the non-uniqueness and instability issues in the inversion process. The inversion result can then be obtained by iteratively solving the objective function using the Gauss-Newton method.

[0053] Another object of the present invention is to provide a gravity and magnetic joint inversion system based on deep learning coupling, which implements the aforementioned gravity and magnetic joint inversion method based on deep learning coupling constraints, comprising:

[0054] The sample dataset generation module is used to design a large number of gravity and magnetic anomaly models using the Matlab language and generate sample datasets by performing Gaussian blurring.

[0055] The network training module is used to construct the Trans_FCN gravity-magnetism model conversion network, i.e., the Trans_FCN network, and to train the constructed Trans_FCN network using the generated sample dataset to optimize the network parameters of the Trans_FCN network.

[0056] The inversion result acquisition module is used to obtain the inversion results. First, the underground density model is obtained through gravity inversion. The Trans_FCN gravity and magnetic model conversion network converts the density model into a magnetic susceptibility model, which is used as the reference model for the next magnetic inversion. The magnetic anomaly is then inverted separately to obtain the magnetic susceptibility model. The Trans_FCN gravity and magnetic model conversion network converts the magnetic susceptibility model into a density model, which is used as the reference model for the next gravity inversion, until the maximum number of iterations is reached.

[0057] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the gravity and magnetic joint inversion method based on deep learning coupling constraints.

[0058] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the gravity and magnetic joint inversion method based on deep learning coupling constraints.

[0059] Another objective of this invention is to provide an information data processing terminal for implementing the gravity and magnetic joint inversion system based on deep learning coupling constraints.

[0060] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0061] This invention utilizes forward modeling of gravity and magnetic anomalies and Gaussian fuzzing simulation to generate sample datasets, constructing a Trans_FCN inversion network. The Trans_FCN gravity and magnetic model transformation network is trained using the sample datasets to achieve property correlations and structural similarities between gravity and magnetic models. Finally, a trained deep neural network provides a reliable reference model for joint inversion, providing location and property values. This invention can accurately invert the location and properties of gravity and magnetic anomalies, exhibiting strong learning and generalization capabilities, effectively solving the problem of gravity and magnetic anomaly data inversion. By designing models with different properties and sizes, and by moving the model positions, setting different property combinations, and performing Gaussian fuzzing to obtain a large amount of data, this invention can simultaneously learn the property coupling and structural coupling relationships between gravity and magnetic models, effectively improving the joint inversion effect.

[0062] The observation data obtained in this invention were not found in the training set. The observation data were input into the gravity and magnetic joint inversion system based on deep learning coupling constraints to obtain the predicted underground model. The results showed that the predicted underground model had a shape and location that were roughly the same as the real underground model, and the physical properties were numerically close to the real model. It can simultaneously take into account the structural coupling and physical property coupling relationship between the gravity and magnetic physical property models. The inversion results are more accurate and focused in predicting the location of underground anomalies compared with the cross-gradient gravity and magnetic joint inversion method and the gravity and magnetic inversion method alone. This provides a new idea for the application of deep learning in joint inversion.

[0063] This invention constructs paired magnetic susceptibility and density property models by assuming a relationship between magnetic susceptibility and density properties in a certain location, and performs Gaussian blurring to make the training model as similar as possible to the results of gravity and magnetic inversion alone, thereby improving the generalization ability of Trans_FCN.

[0064] This invention obtains the Trans_FCN gravity and magnetic model network by adjusting parameters and conducting iterative training. This network provides a "correct" reference model for gravity and magnetic inversion, enabling interactive gravity and magnetic inversion. The trial inversion results show that the underground model is well recovered, accurately reversing the specific location and shape of the model. There is also no "adhesion" phenomenon between the two models in the horizontal direction.

[0065] Model calculations show that Trans_FCN can simultaneously take into account the physical property coupling and structural coupling between gravity and magnetic models, and enable gravity and magnetic models to interact, thereby improving the reliability of the final inversion results.

[0066] This invention can accurately invert the location and physical properties of gravity and magnetic anomalies, has strong learning ability and a certain generalization ability, and can effectively solve the problem of joint inversion of gravity and magnetic anomaly data.

[0067] The technical solution of this invention fills a technical gap in the industry both at home and abroad: This invention is the first to apply deep learning coupling to the field of gravity and magnetic joint inversion, and can simultaneously combine the physical property coupling and structural coupling relationship between different physical property models in the process of joint inversion of multiple physical properties, providing a new idea for the application of deep learning in joint inversion. Attached Figure Description

[0068] Figure 1 This is a flowchart of the gravity and magnetic field joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention;

[0069] Figure 2 This is a schematic diagram illustrating the forward and inverse relationships of geophysical data properties provided in this embodiment of the invention;

[0070] Figure 3 This is a graph showing the relationship between the training set density and magnetic susceptibility of the gravity and magnetic joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention.

[0071] Figure 4 This is a schematic diagram of a set of training data for the Trans_FCN method based on deep learning coupling constraints for joint gravity and magnetic inversion provided in an embodiment of the present invention.

[0072] Figure 5 This is a schematic diagram of the Trans_FCN network structure of the gravity and magnetic field joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention;

[0073] Figure 6 This is a schematic diagram of underground grid cell partitioning using a gravity and magnetic joint inversion method based on deep learning coupling constraints, provided in an embodiment of the present invention:

[0074] Figure 7 This is a flowchart of a separate inversion using the Gauss-Newton method provided in an embodiment of the present invention;

[0075] Figure 8 This is an inversion flowchart of the gravity and magnetic field joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention;

[0076] Figure 9 The following are the actual positions and observation data of the trial model 1 of the gravity and magnetic field joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention: (a) gravity data; (b) magnetic measurement data; (c) horizontal slice of density model; (d) horizontal slice of magnetic susceptibility model; (e) vertical slice of density model; (f) vertical slice of magnetic susceptibility model.

[0077] Figure 10The following are the actual positions and observation data of the trial model 2 of the gravity and magnetic joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention: (a) gravity data; (b) magnetic measurement data; (c) horizontal slice of density model; (d) horizontal slice of magnetic susceptibility model; (e) vertical slice of density model; (f) vertical slice of magnetic susceptibility model.

[0078] Figure 11 The following are the actual positions and observation data of the trial model 3 of the gravity and magnetic joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention: (a) gravity data; (b) magnetic measurement data; (c) horizontal slice of density model; (d) horizontal slice of magnetic susceptibility model; (e) vertical slice of density model; (f) vertical slice of magnetic susceptibility model.

[0079] Figure 12 The following are gravity inversion results provided by the single anomaly model in this embodiment of the invention: (a) horizontal profile of gravity anomaly inversion density alone; (b) horizontal profile of cross-gradient gravity and magnetic joint inversion density; (c) horizontal profile of deep learning coupled gravity and magnetic joint inversion density; (d) horizontal profile of magnetic susceptibility model; (e) horizontal profile of cross-gradient gravity and magnetic joint inversion density; (f) horizontal profile of deep learning coupled gravity and magnetic joint inversion density.

[0080] Figure 13 The following are the magnetic inversion results of the single anomaly model provided in this embodiment of the invention: (a) vertical profile of density inverted by gravity anomaly alone; (b) vertical profile of density inverted by cross-gradient gravity and magnetic joint; (c) vertical profile of density inverted by deep learning coupled gravity and magnetic joint; (d) vertical profile of magnetic susceptibility model; (e) vertical profile of density inverted by cross-gradient gravity and magnetic joint; (f) vertical profile of density inverted by deep learning coupled gravity and magnetic joint.

[0081] Figure 14 The following are gravity inversion results provided by the dual-anomaly single-property model in this embodiment of the invention: (a) horizontal profile of gravity anomaly inversion density alone; (b) horizontal profile of density inversion obtained by cross-gradient gravity and magnetic joint inversion; (c) horizontal profile of density inversion obtained by deep learning coupled gravity and magnetic joint inversion; (d) horizontal profile of magnetic susceptibility model; (e) horizontal profile of density inversion obtained by cross-gradient gravity and magnetic joint inversion; (f) horizontal profile of density inversion obtained by deep learning coupled gravity and magnetic joint inversion.

[0082] Figure 15 The following are the magnetic inversion results of the dual-anomaly single-property model provided in this embodiment of the invention: (a) vertical profile of density inverted by gravity anomaly alone; (b) vertical profile of density inverted by cross-gradient gravity and magnetic joint; (c) vertical profile of density inverted by deep learning coupled gravity and magnetic joint; (d) vertical profile of magnetic susceptibility model; (e) vertical profile of density inverted by cross-gradient gravity and magnetic joint; (f) vertical profile of density inverted by deep learning coupled gravity and magnetic joint.

[0083] Figure 16 The following are gravity inversion results provided by the dual-anomaly dual-property model in this embodiment of the invention: (a) horizontal profile of gravity anomaly inversion density alone; (b) horizontal profile of density inversion obtained by cross-gradient gravity and magnetic joint inversion; (c) horizontal profile of density inversion obtained by deep learning coupled gravity and magnetic joint inversion; (d) horizontal profile of magnetic susceptibility model; (e) horizontal profile of density inversion obtained by cross-gradient gravity and magnetic joint inversion; (f) horizontal profile of density inversion obtained by deep learning coupled gravity and magnetic joint inversion.

[0084] Figure 17 The magnetic inversion results of the dual-anomaly dual-property model provided in this embodiment of the invention are as follows: (a) Vertical profile of density inverted by gravity anomaly alone; (b) Vertical profile of density inverted by cross-gradient gravity and magnetic joint; (c) Vertical profile of density inverted by deep learning coupled gravity and magnetic joint; (d) Vertical profile of magnetic susceptibility model; (e) Vertical profile of density inverted by cross-gradient gravity and magnetic joint; (f) Vertical profile of density inverted by deep learning coupled gravity and magnetic joint.

[0085] Figure 18 The following are observation data of the dual-anomaly dual-property model with added noise and the results of gravity and magnetic joint inversion coupled by deep learning provided in this embodiment of the invention: (a) is the gravity anomaly data with added noise; (b) is the magnetic anomaly data with added noise; (c) is the horizontal profile of the density model of the density model coupled by deep learning with gravity and magnetic joint inversion; (e) is the vertical profile of the density model of the density model coupled by deep learning with gravity and magnetic joint inversion; (d) is the horizontal profile of the magnetic susceptibility model coupled by deep learning with gravity and magnetic joint inversion; and (f) is the horizontal and vertical profiles of the magnetic susceptibility model coupled by deep learning with gravity and magnetic joint inversion. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0087] To enable those skilled in the art to fully understand how the present invention is specifically implemented, this section provides an explanatory description of the embodiments that expand upon the technical solutions of the claims.

[0088] like Figure 5 As shown, the forward and inverse model diagram of the present invention specifically includes:

[0089] Forward modeling involves using surface observation data to derive information from a known subsurface geological model, while inverse modeling involves using surface observation data to predict the subsurface geological model. Forward and inverse modeling can be represented as a mapping relationship between model space and data space. The gravity and magnetic joint inversion method based on deep learning coupling constraints provided in this invention includes:

[0090] A sample dataset is generated based on the gravity and magnetic anomaly model. A Trans_FCN gravity and magnetic model conversion network is constructed based on the generated sample dataset. The Trans_FCN gravity and magnetic model conversion network is trained using the sample dataset. The Trans_FCN gravity and magnetic model conversion network is used to provide a "correct" reference model for the joint gravity and magnetic inversion, and finally a more reliable inversion result is obtained.

[0091] like Figure 1 As shown, the gravity and magnetic field joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention includes the following steps:

[0092] (1.1) Calculate the objective function using gravity and magnetic geophysical data (such as observation data, initial model, reference model, depth-weighted, smoothing constraints, physical property constraints, etc.).

[0093] (1.2) Differentiate the objective function and calculate whether the derivative value reaches the threshold;

[0094] (1.3) If the threshold or the maximum number of iterations is reached, a separate inversion result is obtained. If the threshold or the maximum number of iterations is not reached, the objective function is solved using the Gauss-Newton method and the physical property model (i.e. the original initial model) is updated. Step (1.3) is repeated until the threshold is reached.

[0095] like Figure 6 As shown, this embodiment of the invention provides a schematic diagram of underground grid cell partitioning using a gravity and magnetic joint inversion method based on deep learning coupling constraints:

[0096] Forward modeling of gravity anomalies, as follows Figure 6 As shown, the underground model is first uniformly divided into multiple regular cuboid elements. Then, the anomaly of each cuboid element to each observation point on the observation plane is calculated. Finally, the anomaly of the entire underground model to the observation plane is obtained by summing the anomaly effects of all cuboid elements.

[0097] The formula for calculating the gravity anomaly at the observation point P(x, y, z) corresponding to a single cuboid element is as follows:

[0098]

[0099] Where x i =x-ξ i y i =y-η j , z k =z-ζ k , G0 is the gravitational constant, ρ is the density of the cuboid model element, Δg is the gravitational anomaly, and Q() in the figure is the center of the subdivided element volume.

[0100] Forward modeling of magnetic anomalies is similar to that of gravity anomalies. The formula for calculating the magnetic anomaly at the observation point P(x, y, z) corresponding to a single cuboid element is as follows:

[0101]

[0102] If remanence is neglected: where μ0 is the free permeability, K1 = 2MN, K2 = 2NL, K3 = 2ML, K4 = L 2 K5 = M 2 K6 = -N 2 , L=cos I cos D, M=cos I sin D, N=sin I, r=[(ξ-x) 2 +(η-y) 2 +(ζ-z) 2 ] 1 / 2 In the formula: I is the magnetic inclination angle, D is the magnetic declination angle; in the figure, Q() is the center of the subdivided unit volume element, () is the corner coordinate position of the underground cuboid unit, and r is the distance from the cuboid grid unit to the observation point.

[0103] In this embodiment of the invention, the selected underground space has a length and width of 1000m and a depth of 500m, and is uniformly divided into 20*20*10 grid cells, each with a size of 50m*50m*50m. The experiments in this paper will use the controlled variable method, employing individual inversion, cross-gradient joint inversion, and deep learning coupled joint inversion methods in four embodiments (see Table 1) to verify the effectiveness of the deep learning coupled joint inversion method. Depth weighting, physical property constraints, and smoothness constraints are added during the inversion process to make the inversion results more reliable. It is worth noting that the model shape and physical property combination in embodiments three and four are not present in the training set of the deep neural network, in order to test the generalization ability of this method (i.e., the inversion ability of unlearned data).

[0104] Table 1. Example Model

[0105]

[0106] The large number of gravity and magnetic anomaly models required for design using Matlab language and the Gaussian blur generation sample dataset provided in this embodiment of the invention include:

[0107] (1) Given the shape of the anomaly: First, determine a subsurface model area of ​​a specified size, then mesh this area, design some different anomaly blocks and traverse every position of the subsurface mesh model. In this paper, when traversing the subsurface mesh model, the anomaly models near the edge will be discarded.

[0108] (2) Given the magnetic parameters of the anomalous body: According to the relationship between density and magnetic susceptibility obtained from prior knowledge, assign values ​​to the anomalous body model in step (1) to obtain the corresponding anomalous body density model and magnetic susceptibility model.

[0109] (3) Gaussian blurring: Since the smoothing constraint is added during the Gaussian-Newton inversion process, the inversion result has no sharp boundaries. Therefore, we will apply Gaussian filtering to the anomaly density model and magnetic susceptibility model generated in step (2) to blur the inversion transformation result of the gravity and magnetic subsurface model. This step also increases the generalization ability of Trans_FCN.

[0110] Since the density and magnetic susceptibility of the anomalous body in step (2) are selected based on prior knowledge, the deep neural network can learn the physical property coupling relationship between gravity and magnetism. In step (3), Gaussian blurring is applied to both the input and output physical property models of the neural network before training, enabling the neural network to learn the structural coupling relationship between gravity and magnetism data from the training set. The two inversion constraint networks proposed in this invention can share the same training set, which greatly reduces the computational cost required to establish the training set data.

[0111] like Figure 3 (Generated based on the relationship between rock density and magnetic susceptibility in a certain region, as shown in Table 2) The training dataset of this embodiment consists of several sets of training data. Each set of training data is a pair of gravity and magnetic models, corresponding to the density model and magnetic susceptibility model of a certain anomalous body. Figure 4 As shown, first give the size and position of the anomaly (e.g., a centered 4*4*4), then from... Figure 3 By extracting a density-magnetic susceptibility relationship and applying Gaussian blur, a set of training data can be generated. Figure 4 The density selected for a set of 4*4*4 anomaly models is 2426 kg / m³. 3 The combination of physical properties with a magnetic susceptibility of 436e-5 (SI), and the cross-sectional schematic diagram obtained after Gaussian blurring, which is a set of training data.

[0112] Table 1 Relationship between rock density and magnetic susceptibility in a certain region

[0113]

[0114] Furthermore, the settings for other hyperparameters are shown in Table 2.

[0115] Table 2 Trans_FCN Network Parameter Settings

[0116]

[0117] like Figure 2As shown, the Trans_FCN network structure of a gravity and magnetic field joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention specifically includes:

[0118] A novel gravity and magnetism model transformation network (Trans_FCN) was designed, drawing inspiration from the classic U-net network. In this invention, the Trans_FCN has an input size of M×N×K and an output size of M×N×K, where M, N, and K represent the number of grid cells in the x, y, and z directions, respectively. Its network structure comprises three parts: encoding, decoding, and skip connections. The decoding recovery part represents the output, and the skip connections connect the encoding and decoding paths. The encoding part repeatedly uses 3×3 convolutions, batch normalization, ReLU activation functions, and 2×2 max-pooling layers. Simultaneously, the decoding part employs 2×2 deconvolution layers and repeatedly uses 3×3 convolutions, batch normalization layers, and ReLU activation functions.

[0119] Furthermore, the step of training the constructed Trans_FCN network using the generated sample dataset and optimizing the network parameters of the Trans_FCN network includes:

[0120] The sample dataset is preprocessed by randomly dividing it into a training set and a test set. Parameters are set, and the Trans_FCN network is trained using the training set. The test set is used to check whether the network achieves the expected results. If not, the parameters are modified and the network is trained again. If the trained network achieves the expected results, the magnetic susceptibility model is input to obtain the density model, or the density model is input to obtain the magnetic susceptibility model.

[0121] Furthermore, the modified parameters include:

[0122] 1) Determine the objective function:

[0123] We use mean squared error to compare the difference between the predicted results and the actual results, and the loss function is:

[0124]

[0125] Where f i For the prediction results of the neural network, l i For a realistic physical property model, N t This represents the number of meshes.

[0126] 2) Set the number of iterations to 5000.

[0127] like Figure 8 As shown, the training flowchart of the gravity and magnetic field joint inversion method based on deep learning coupling constraints provided in this embodiment of the invention specifically includes:

[0128] Step ② is the process of solving the gravity anomaly data inversion using the Gauss-Newton method, updating the density model in step ③.

[0129] Step 4 uses the trained DFCNG2M deep neural network to convert the density model from step 3 into a magnetic susceptibility model to update the magnetic anomaly inversion reference model in step 5, and participates in the solution of the objective function for magnetic anomaly inversion in step 6.

[0130] Step 6 is the process of solving the magnetic anomaly data inversion using the Gauss-Newton method, updating the magnetic susceptibility model in step 7.

[0131] Step 9 uses the trained DFCMN2G deep neural network to convert the magnetic susceptibility model from step 7 into a density model to update the gravity anomaly inversion reference model in step 10, and participates in the solution of the objective function for gravity anomaly inversion in step 2.

[0132] Repeat steps ②-⑩ N times to obtain the final density model. and magnetic susceptibility model The inversion is complete.

[0133] This invention provides a gravity and magnetic field joint inversion system based on deep learning coupling constraints, comprising:

[0134] The sample dataset generation module is used to design a large number of gravity and magnetic anomaly models using the Matlab language and generate sample datasets by performing Gaussian blurring.

[0135] The network training module is used to construct the Trans_FCN gravity-magnetism model conversion network, i.e., the Trans_FCN network, and to train the constructed Trans_FCN network using the generated sample dataset to optimize the network parameters of the Trans_FCN network.

[0136] The inversion result acquisition module first obtains the underground density model through gravity inversion. The Trans_FCN gravity and magnetic model conversion network converts the density model into a magnetic susceptibility model, which is then used as the reference model for the next magnetic inversion. The magnetic anomaly is then separately inverted to obtain the magnetic susceptibility model. This magnetic susceptibility model is then converted into a density model through the Trans_FCN gravity and magnetic model conversion network, which is then used as the reference model for the next gravity inversion. The above steps are repeated until the maximum number of iterations is reached, and the inversion result is obtained.

[0137] To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides specific product or related technology application examples of the technical solution claimed.

[0138] The embodiments used in this invention will compare the inversion results of the traditional gravity and magnetic field inversion method and the cross-gradient gravity and magnetic field joint inversion method with the method of this invention using the controlled variable method.

[0139] Example 1: The single anomaly model selected in this example is located in the center of the grid, at a depth of 100m, with a north, east, and vertical extension of 200×200×150m. The residual density of the anomaly is 2800kg / m³. 3 The corresponding magnetic susceptibility is 1680e-5 (SI), and the geomagnetic field dip and deflection are 50° and -5°, respectively. The magnetization direction of the magnetic body is the same as that of the geomagnetic field (there is no remanent magnetization). An observation data area of ​​1950m x 1950m was selected, with 40 x 40 observation points and a distance of 50m between adjacent observation points. The underground model we are discussing is located at the exact center of the observation data area. Forward modeling of this model yielded the observation data. Figure 9 As shown in (a) and (b) in the figure. Since the model is symmetrical along the north-south direction, a vertical section at Northing = 500m can be selected, as shown in the figure. Figure 9 As shown in (c) and (d), a comparison is made with the cross-section at depth = 150m, as shown in the figure. Figure 9 As shown in (e) and (f).

[0140] Example 2:

[0141] The example in this section uses a dual-anomaly single-property model located in the center of the grid. The distance between the two anomalies is 200m, the burial depth is 100m, and the north, east, and vertical extensions are 200×150×150m. The residual density of each anomaly is 2800kg / m³. 3 The corresponding magnetic susceptibility is 1680e⁻⁵ (SI), and the geomagnetic field dip and deflection are 50° and -5°, respectively. The magnetization direction of the magnetic body is the same as that of the geomagnetic field (there is no remanent magnetization). An observation data area of ​​1950m*1950m was selected, with 40*40 observation points and a distance of 50m between adjacent observation points. The underground model we are discussing is located at the exact center of the observation data area. Forward modeling of this model yielded the observation data. Figure 10 As shown in (a) and (b) in the figure. Since the model is symmetrical along the north-south direction, a vertical section at northing = 500m can be selected, as shown in the figure. Figure 10 As shown in (c) and (d), a comparison is made with the cross-section at depth = 150m, as shown in the figure. Figure 10 As shown in (e) and (f).

[0142] Example 3:

[0143] The dual-anomaly dual-property model selected in this example is located in the center of the grid, with a distance of 200m between the two anomalies, a burial depth of 100m, and extensions of 200×150×150m in the north, east, and vertical directions. The density of the anomalies is 2800kg / m³. 3 and 1400kg / m 3The corresponding magnetic susceptibility is 1680e-5 (SI) and 840e-5 (SI), with geomagnetic field dip and deflection angles of 50° and -5°, respectively. The magnetization direction of the magnetic body is the same as that of the geomagnetic field (there is no remanent magnetization). An observation data area of ​​1950m x 1950m was selected, with 40 x 40 observation points and a spacing of 50m between adjacent observation points. The underground model we are discussing is located at the exact center of the observation data area. Forward modeling of this model yielded the observation data. Figure 11 As shown in (a) and (b) in the figure. Since the model is symmetrical along the north-south direction, a vertical section at Northing = 500m can be selected, as shown in the figure. Figure 11 As shown in (c) and (d), a comparison is made with the cross-section at depth = 150m, as shown in the figure. Figure 11 As shown in (e) and (f).

[0144] Example 4: The example selected in this section has the same position and shape as the example model in Example 3. The difference is that 20% random noise was added to the observation data to verify the noise resistance of the method in this paper.

[0145] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented using hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or using software executed by various types of processors, or using a combination of the above-described hardware circuitry and software, such as firmware.

[0146] The embodiments of the present invention have achieved some positive results during the research and development or use process, and have indeed great advantages compared with the prior art. The following content describes the experimental process with data, charts and other information.

[0147] like Figures 9 to 11 As shown, the actual location and observation data of the trial model in the embodiment of the present invention are shown. The trial model 3 is the same as the trial model 4 in terms of physical properties and location settings, and 20% random noise is added.

[0148] like Figures 12 to 18As shown, the trial model inversion results of the gravity and magnetic field joint inversion method using deep learning coupling constraints provided in this embodiment of the invention specifically include:

[0149] Figure 12 The figures show the gravity anomaly inversion results obtained by gravity anomaly inversion alone, cross-gradient joint inversion, and deep learning coupling constraint joint inversion. As can be seen from the figures, compared with gravity anomaly inversion alone and cross-gradient joint inversion, the inversion results obtained by our proposed deep learning coupling constraint method are more focused, more accurate in location, have a higher degree of model convergence, and the physical parameters are closer to the true values.

[0150] Figure 13 This is the result of gravity anomaly inversion, which is achieved through magnetic anomaly inversion alone, cross-gradient joint inversion, and deep learning coupled constraint joint inversion. Figure 13 Similar results can be observed. Figure 12 , Figure 13 The results show that the method of the present invention is generally superior to gravity / magnetic inversion alone and gravity / magnetic joint inversion with cross gradients.

[0151] Figure 14 The results are gravity anomaly inversion results obtained by gravity anomaly inversion alone, cross-gradient joint inversion, and deep learning coupling constraint joint inversion. As can be seen from the figure, under the premise of the same observation data, the gravity and magnetic joint inversion results obtained by the method of this invention are closer to the real physical properties in terms of physical properties, and can separate two similar anomalies horizontally. The model has a higher degree of convergence, the inversion results are more focused, and the location is more accurate. It also has the advantage of cross-gradient joint inversion in keeping the gravity and magnetic model consistent as a whole, and makes up for the deficiency of low resolution of cross-gradient joint inversion with a small amount of observation data.

[0152] Figure 15 This is the result of gravity anomaly inversion, which is achieved through magnetic anomaly inversion alone, cross-gradient joint inversion, and deep learning coupled constraint joint inversion. Figure 15 Similar results can be observed. Figure 14 , Figure 15 The results show that the method of the present invention is generally superior to gravity / magnetic inversion alone and gravity / magnetic joint inversion with cross gradients.

[0153] Figure 16 This is the result of gravity anomaly inversion using individual inversion, cross-gradient joint inversion, and deep learning coupled joint inversion, with a magnetic susceptibility of 840e-5 (SI) and a density of 1400 kg / m³. 3 The abnormal combination of physical properties did not appear Figure 4In the combination of gravity and magnetic property relationships, the joint inversion results of deep learning coupling constraints can make the inversion results more focused compared with the cross-gradient joint inversion and gravity and magnetic inversion methods. It can also basically separate the high-density high magnetic susceptibility anomaly models with similar locations from the low-density low magnetic susceptibility anomaly models, further verifying the stability and generalization ability of the Trans_FCN network proposed in this invention.

[0154] Figure 17 This is the result of gravity anomaly inversion using magnetic anomaly inversion alone, cross-gradient joint inversion, and deep learning coupled joint inversion. Figure 17 Similar results can be observed. Figure 16 , Figure 17 The results show that the method of the present invention is generally superior to gravity / magnetic inversion alone and gravity / magnetic joint inversion with cross gradients.

[0155] Figure 18 The figure shows the observation data and inversion results after adding noise to the model in Example 3. As can be seen from the figure, even with noise, the method proposed in this paper can still separate two anomalies with different physical properties and their positions are accurate. This further verifies that the gravity and magnetic joint inversion method based on deep learning coupling proposed in this invention has a certain noise resistance capability.

[0156] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A joint gravity and magnetic inversion method based on deep learning coupling constraints, characterized in that, include: A sample dataset was generated using a gravity and magnetic anomaly model, and a gravity and magnetic model transformation network (Trans_FCN) was constructed. The gravity and magnetism model conversion network Trans_FCN is trained based on the sample dataset; The Trans_FCN gravity and magnetic model transformation network is used to provide a "correct" reference model for the joint gravity and magnetic inversion, ultimately resulting in more reliable inversion results. The gravity and magnetic field joint inversion method based on deep learning coupling constraints includes the following steps: Step 1: Design a large number of gravity and magnetic anomaly models using Matlab language, and generate a sample dataset by performing Gaussian blurring. Step 2: Construct a gravity and magnetism model conversion network, and use the generated sample dataset to train the constructed Trans_FCN network to optimize the network parameters of the Trans_FCN network; Step 3: Input the individual gravity and magnetic inversion results into the trained Trans_FCN network to provide a "correct" reference model for the next gravity and magnetic inversion, realize the interaction of gravity and magnetic property models, improve the reliability of joint gravity and magnetic inversion, and repeat step 3 to finally obtain the gravity and magnetic inversion results. The gravity and magnetism model transformation network Trans_FCN includes: The input data size of Trans_FCN is M×N×K, and the output data size is also M×N×K, where M, N, and K are the number of grids in the x, y, and z directions, respectively. The Trans_FCN network consists of three parts: an encoding part, a decoding part, and skip connections. The encoding part includes multiple 3×3 convolutional layers, batch normalization layers, ReLU activation functions, and 2×2 max pooling layers. Meanwhile, the decoding part uses 2×2 deconvolutional layers, multiple 3×3 convolutional layers, batch normalization layers, and ReLU activation functions. The corresponding decoding recovery part represents the output. The process of separate gravity and magnetic inversion in step three includes: (3.1) Calculate the objective function using gravity and magnetic geophysical data, which includes observation data, initial model, reference model, depth weighting, smoothing constraints, and physical property constraints; (3.2) Differentiate the objective function and calculate whether the derivative value reaches the threshold; (3.3) If the threshold or the maximum number of iterations is reached, the inversion result is obtained separately. If the threshold or the maximum number of iterations is not reached, the objective function is solved by the Gauss-Newton method and the initial model is updated. Then step (3.1) is repeated until the threshold is reached.

2. The gravity and magnetic field joint inversion method based on deep learning coupling constraints as described in claim 1, characterized in that, The specific process in step one is as follows: (1.1) Given the shape of the anomaly: First, determine a subsurface model area of ​​a specified size, then mesh the area, and finally design some different anomaly blocks and traverse every position of the subsurface mesh model. When traversing the subsurface mesh model, anomaly models near the edge will be discarded. (1.2) Given the values ​​of the magnetic parameters of the anomalous body: Assign values ​​to the anomalous body model in step (1.1) according to the known relationship between the magnetic susceptibility and density of the underground medium in a certain place, and obtain the corresponding anomalous body density model and magnetic susceptibility model; (1.3) Gaussian blurring: Since the smooth constraint is added in the Gauss-Newton method inversion process, the inversion result has no sharp boundary. Therefore, Gaussian filtering is applied to the anomaly density model and magnetic susceptibility model generated in step (1.2) to blur the inversion transformation result of the gravity and magnetic subsurface model.

3. The gravity and magnetic field joint inversion method based on deep learning coupling constraints as described in claim 2, characterized in that, The density and magnetic susceptibility of the anomalous body are selected based on prior knowledge, so the deep neural network can learn the physical property coupling relationship between gravity and magnetism. In step (1.3), Gaussian blurring is applied to the physical property models of the neural network input and output before training, so that the neural network can learn the structural coupling relationship between gravity and magnetism data from the training set.

4. The gravity and magnetic field joint inversion method based on deep learning coupling constraints as described in claim 1, characterized in that, Step two includes: First, the sample dataset is preprocessed by randomly dividing it into a training set and a test set. Then, set the parameters, train the Trans_FCN network using the training set, and use the test set to check whether the network can achieve mutual conversion between gravity and magnetic underground anomaly physical property models without distortion. If not, modify the parameters and train again until the set maximum number of iterations is reached. Finally, if the trained network achieves the expected results, the density model is obtained by inputting the magnetic susceptibility model, or the magnetic susceptibility model is obtained by inputting the density model.

5. The gravity and magnetic field joint inversion method based on deep learning coupling constraints as described in claim 1, characterized in that, The parameter modifications in step two include: Given the objective function, and using mean squared error to compare the difference between the predicted and actual results, the loss function is: ； in The prediction results of the neural network, For realistic physical property models, The number of meshes is 5000.

6. The gravity and magnetic field joint inversion method based on deep learning coupling constraints as described in claim 5, characterized in that, The gravity and magnetic joint inversion process based on deep learning coupling constraints also includes: First, the process of gravity anomaly data inversion is solved using the Gauss-Newton method through a gravity geophysical model to update the density model. Then, the density model is converted into a magnetic susceptibility model through a trained DFCNG2M deep neural network to update the magnetic anomaly inversion reference model and participate in the solution of the objective function of magnetic anomaly inversion. Then, the magnetic anomaly data inversion process is solved using the Gauss-Newton method to update the magnetic susceptibility model; the magnetic susceptibility model is converted into a density model through a trained DFCMN2G deep neural network to update the gravity anomaly inversion reference model and participate in the objective function solution of gravity anomaly inversion. Finally, repeat the above steps N times until the maximum number of iterations is reached to obtain the final density model and magnetic susceptibility model, and the inversion ends. The Gauss-Newton method was used for separate inversion.

7. A gravity and magnetic joint inversion system based on deep learning coupling, implementing the gravity and magnetic joint inversion method based on deep learning coupling constraints as described in any one of claims 1 to 6, characterized in that, include: The sample dataset generation module is used to design a large number of gravity and magnetic anomaly models using the Matlab language and generate sample datasets by performing Gaussian blurring. The network training module is used to construct the Trans_FCN gravity-magnetism model conversion network, i.e., the Trans_FCN network, and to train the constructed Trans_FCN network using the generated sample dataset to optimize the network parameters of the Trans_FCN network. The inversion result acquisition module is used to obtain the inversion results. First, the underground density model is obtained through gravity inversion. The Trans_FCN gravity and magnetic model conversion network converts the density model into a magnetic susceptibility model, which is used as the reference model for the next magnetic inversion. The magnetic anomaly is then inverted separately to obtain the magnetic susceptibility model. The Trans_FCN gravity and magnetic model conversion network converts the magnetic susceptibility model into a density model, which is used as the reference model for the next gravity inversion, until the maximum number of iterations is reached.

Images