Gravity anomalous body inversion method based on PMU-Net deep learning network

Abstract

The invention discloses a gravity anomalous body inversion method based on a PMU-Net deep learning network, and belongs to the field of gravity inversion, and the method comprises the steps: increasing the number of observation equation sets through the joint inversion of two different types of gravity anomalous data, and comprehensively presenting an underground geologic structure. Data are processed by means of two independent channels, one channel uses a Mama module to capture dependency between observation points, and the other channel uses a spatial pre-training model to capture position association between the observation points and underground density. And the processed features are fused and input into the U-shaped neural network constructed by one-dimensional convolution, so that the calculation amount is reduced and the efficiency is improved. The system further comprises a simulation data generation module, and the problem of weak interpretability of deep learning in gravity inversion is solved by embedding priori knowledge. According to the invention, the accuracy and reliability of inversion under complex geological conditions are effectively improved, and an efficient and accurate new method is provided for geophysical exploration.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of gravity inversion, and particularly relates to a gravity anomaly body inversion method based on a PMU-Net deep learning network. BACKGROUND

[0002] Gravity inversion is a key technology in geophysical exploration, which measures the gravity field on the ground and deduces the underground density distribution by means of algorithm model, and has the significant advantages of low cost and high efficiency. In recent years, with the development of artificial intelligence technology, people began to use deep learning to predict gravity data to quickly image the underground geological structure. The early method mainly uses single gravity field data as input to directly predict the underground density body, but this way can only reflect the gravity anomaly of the underground geological body in a particular aspect, the information is extremely one-sided, and the uncertainty of the inversion solution is high.

[0003] To improve this situation, some people designed a GraInvNet model based on U-Net, which uses gravity anomaly data and gravity gradient anomaly data for joint inversion, and obtains good prediction results. However, this method is simple in processing different gravity anomaly data, and it is difficult to capture the complex features and mutual relationships between data. Later, some people predicted different features through different decoders to further improve the information capture ability, but there are still problems of insufficient capture of the dependence between observation points and the position correlation between observation points and underground density model. Some people also proposed the fusion of UNet++ network architecture with N×N convolution kernel and LeakyReLU activation function to improve the accuracy of the inversion result, but the high computational complexity of two-dimensional convolution limits the efficiency of the model.

[0004] In summary, the existing deep learning model still has deficiencies in comprehensively describing the underground geological structure, mining the correlation between different data, learning the relationship between observation points and density grid, and computational efficiency, and needs to be improved to meet the needs of high-precision and high-efficiency gravity inversion under complex geological conditions. SUMMARY

[0005] The application provides a gravity anomaly body inversion method based on a PMU-Net deep learning network to solve the problems in the prior art.

[0006] To achieve the above purpose, the application provides a gravity anomaly body inversion method based on a PMU-Net deep learning network, which comprises the following steps: Obtain two different types of gravity anomaly data as input; Two kinds of gravity anomaly data are processed through two independent channels to mine the correlation and complementarity between the data, one channel uses the Mamba module to capture the dependence between observation points, and the other channel captures the position correlation between observation points and underground density through a spatial pre-training model; The processed features are fused, and the fused features are input into a U-shaped neural network constructed by one-dimensional convolution to obtain the underground density distribution result.

[0007] Optionally, the two different types of gravity anomaly data include gravity anomaly data and gravity gradient anomaly data.

[0008] Optionally, the Mamba module includes: The relative coordinate coding is performed on each measurement point of the gravity anomaly data to construct a spatial correlation matrix; The spatial pattern features of the gravity anomaly data are extracted through a multi-scale convolution kernel; Based on the dynamic state space model, the measurement point evolution sequence is constructed based on the plane coordinate grid index, and the physical state of each measurement point is evolved according to the law of conservation of mass; The gravity anomaly signal is mapped through the underground physical property structure response matrix, and the mapping result is coupled with the initial geological state through the state transition matrix; The updated geological state is projected to the detectable physical quantity space through the observation operator, and the lithology-density conversion matrix is introduced to retain the direct influence term of the original input signal; Through the dual-path fusion mechanism, the tectonic evolution dynamics process and the density physical static distribution are cooperatively constrained.

[0009] Optionally, the spatial pre-training model includes: The physical correlation framework of the gravity field features and the inversion space is constructed to perform cross-domain feature coupling of the observation signal and the underground structure; The multi-scale local features in the gravity field map are extracted through the F-Encoder to capture the spatial correlation of the observation points, and a low-dimensional vector representing the gravity field features is generated; The position coordinates of the inversion space are geometrically transformed and aligned through the P-Encoder to generate a high-dimensional vector containing the geometric structure information of the target body; The low-dimensional vector and the high-dimensional vector are cooperated to synchronously process the gradient changes caused by the gravity anomaly density difference and the geometric constraints of the inversion space, forming a field-space feature coupling mechanism; The field-space feature coupling mechanism provides regularization constraints for the inversion process.

[0010] Optionally, the U-shaped neural network uses a one-dimensional convolution layer to replace the traditional two-dimensional convolution layer to process the one-dimensional sequence characteristics of the gravity anomaly data.

[0011] Optionally, the U-shaped neural network further comprises a full-scale skip connection mechanism for fusing features of different scales in the encoder and the decoder.

[0012] Optionally, the full-scale skip connection mechanism adjusts features of different scales to the same scale through linear mapping and then fuses them, and further integrates them through one-dimensional convolution.

[0013] Optionally, the method further comprises generating simulated training data: initializing inversion space parameters and geological body feature parameters; generating a survey network, setting grid parameters and constructing a structured sampling grid; constructing a density distribution model in the inversion space according to the geological body feature parameters; calculating gravity anomaly data corresponding to the density distribution model according to the forward formula; storing the gravity anomaly data and the corresponding density distribution model to form a training data set.

[0014] The application further provides a computer device comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to realize the steps of the method.

[0015] The application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to realize the steps of the method.

[0016] Compared with the prior art, the application has the following advantages and technical effects: The gravity anomaly body inversion method based on the space pre-training PMU-Net deep learning network of the present application is significantly superior to the prior art in many aspects: first, in terms of inversion accuracy, by jointly inverting two kinds of gravity anomaly data Gz and Gxz and processing these data through two independent channels respectively, the system can more comprehensively capture the characteristics of the underground geological structure, thereby significantly improving key indicators such as average gravity anomaly offset error and average gravity anomaly accuracy. Secondly, in terms of edge information restoration ability, PMU-Net performs well, effectively avoiding the common edge blur and position offset problems in the prior art. This is mainly due to the introduction of the pre-training model and the attention mechanism, as well as the design of the full-scale skip connection, which enables the system to better capture the dependency between the position of the density anomaly body and the position of the measuring point, while reducing the feature loss in the downsampling process. In addition, in terms of computational efficiency, the U-shaped neural network constructed by one-dimensional convolution significantly reduces the computational complexity and improves the computational efficiency of the model, providing a lightweight solution for fast gravity field inversion. At the same time, through full-scale skip connection, the system can more comprehensively integrate gravity features of different scales, effectively capturing semantic information from fine-grained to coarse-grained, further improving the accuracy of the inversion result. Finally, by embedding prior knowledge in the training data, the system can better combine geophysical constraints and gravity data-driven, further improving the prediction ability of deep learning, effectively solving the problem of weak explainability of deep learning in gravity inversion. In summary, the PMU-Net deep learning network of the present application has shown significant superiority in gravity anomaly body inversion under complex geological conditions, not only improving the accuracy and reliability of the inversion, but also significantly improving the computational efficiency, providing an efficient and accurate new method for geophysical exploration. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and their Figure 1 is a feature fusion schematic diagram of an embodiment of the present application; Figure 2 is a Mamba module schematic diagram of an embodiment of the present application; Figure 3 is a space pre-training model schematic diagram of an embodiment of the present application; Figure 4 is a full-scale skip connection schematic diagram of an embodiment of the present application; Figure 5Fig. 1 is a comparison chart of inversion results of embodiments of the present application, wherein (a1) is a three-dimensional visual image of a real model, (a2) is an x-z section view of the real model along y=1800 meters, (a3) is a y-z section view of the real model along x=700 meters, (b1) is a three-dimensional visual image of the U-Net inversion result, (b2) is an x-z section view of the U-Net inversion result along y=1800, (b3) is a y-z section view of the U-Net inversion result along x=700, (c1) is a three-dimensional visual image of the TransUnet inversion result, (c2) is an x-z section view of the TransUnet inversion result along y=1800, (c3) is a y-z section view of the TransUnet inversion result along x=700, (d1) is a three-dimensional visual image of the PMU-Net inversion result, (d2) is an x-z section view of the PMU-Net inversion result along y=1800, and (d3) is a y-z section view of the PMU-Net inversion result along x=700; Figure 6 Fig. 2 is a data set generation flowchart of embodiments of the present application; Figure 7 Fig. 3 is a gravity forward modeling corresponding schematic diagram of embodiments of the present application. DETAILED DESCRIPTION

[0018] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0019] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0020] Embodiment one The embodiment provides a gravity anomaly body inversion method based on a PMU-Net deep learning network, comprising the following steps: Obtaining two different types of gravity anomaly data as input; Processing the two types of gravity anomaly data through two independent channels to mine the correlation and complementarity between the data, wherein one channel uses a Mamba module to capture the dependence between observation points, and the other channel captures the positional correlation between observation points and underground density through a spatial pre-training model; Fusing the processed features, and inputting the fused features into a U-shaped neural network constructed by one-dimensional convolution for inversion to obtain an underground density distribution result.

[0021] Specifically, the model takes the gravity component Gz and the gravity gradient component Gxz as the input of the model, and processes two different types of gravity anomaly data through two independent channels. Among them, the gravity anomaly map Gz is first processed by relative position coding to solve the long-distance dependence problem, and then the dependence between observation points is captured by the Mamba module; the gravity anomaly data Gxz is processed by a spatial pre-training model to capture the position relationship between the observation points and the underground density model. Then, the two types of features processed by the above are fused, and then input into the U-shaped neural network constructed by one-dimensional convolution, and the full-scale jump connection can fuse the features of different levels and different scales. As shown in Figure 1 .

[0022] Further, the Mamba module includes: performing relative coordinate coding on each measurement point of the gravity anomaly data to construct a spatial correlation matrix; extracting spatial pattern features of the gravity anomaly data through a multi-scale convolution kernel; based on a dynamic state space model, constructing a measurement point evolution sequence based on a plane coordinate grid index, and evolving the physical state of each measurement point according to the law of conservation of mass; mapping the gravity anomaly signal through an underground physical structure response matrix, and coupling the mapping result with the initial geological state through a state transition matrix; projecting the updated geological state to the detectable physical quantity space through an observation operator, and introducing a lithology-density conversion matrix to retain the direct influence term of the original input signal; through a double-path fusion mechanism, the tectonic evolution dynamics process and the static distribution of density physical properties are cooperatively constrained.

[0023] Specifically, the Mamba module is as shown in Figure 2 .

[0024] In Figure 2 , the mathematical expression is: In the formula, h represents the geological state, A represents the state transition matrix, t represents the measurement point, B represents the response matrix, X represents the gravity anomaly signal, Y represents the observation operator, Z represents the lithology-density conversion matrix, and y represents the observation output.

[0025] The embodiment constructs a gravity data feature extraction method based on dynamic state space modeling. A spatial correlation matrix is constructed by encoding the relative coordinates of the gravity data Gz measurement points. A multi-scale convolution kernel corresponding to a typical structure scale of 5-50 km is used to accurately extract the spatial pattern features of regional gravity anomalies. After processing, a feature tensor containing 512-dimensional density physical parameters is formed (batch dimension 16 x measurement point sequence 1681 x feature dimension 512). The tensor completely represents key geological information such as the development degree of salt structures in the study area.

[0026] In the dynamic modeling stage, the measurement point evolution sequence (t=0-1680) is constructed using a planar coordinate grid index. The physical state evolution of each measurement point strictly follows the law of conservation of mass. Taking the measurement point t=1 as an example, its gravity anomaly signal x(1) is mapped through the underground physical structure response matrix B (dimension 1024x512). The matrix elements B ij The mapping result is coupled with the initial geological state h(0) (containing regional tectonic stress field background information) through the state transition matrix A (dimension 1024x1024). Among them, the matrix A kl The coupling relationship between the kth geological structure unit and the lth lithofacies density unit caused by geological processes such as fault activity and magma intrusion is described. The updated geological state h(1) is projected to the detectable physical quantity space through the observation operator Y (512x1024). The row vectors of the operator correspond to the gravity response kernel functions of different density interfaces. At the same time, the lithology-density conversion matrix Z (512x512) is introduced to retain the direct influence term of the original input signal x(1). Through the dual-path fusion mechanism, the tectonic evolution dynamics and the static distribution of density physical properties are cooperatively constrained.

[0027] Further, the spatial pre-training model comprises: A physical association framework is constructed between the gravity field features and the inversion space to couple the cross-domain features of the observation signal and the underground structure. The F-Encoder extracts multi-scale local features in the gravity field map and captures the spatial correlation of the observation points to generate a low-dimensional vector representing the gravity field features. The P-Encoder performs geometric transformation alignment on the position coordinates of the inversion space to generate a high-dimensional vector containing the geometric structure information of the target body. The low-dimensional vector and the high-dimensional vector are cooperated to synchronously process the gradient changes caused by gravity anomaly density differences and the geometric constraints of the inversion space, forming a field-space feature coupling mechanism. The field-space feature coupling mechanism provides regularization constraints for the inversion process.

[0028] Specifically, the traditional CNN, U-Net and other network models often fail to recognize the influence of the inversion space and the position information of the survey network space on the gravity inversion accuracy in the inversion process. Therefore, the research goal of the present application is to enable the model to more accurately capture the spatial position characteristics of each gravity anomaly data in the observation network and each density unit in the inversion space, and the dependence between these characteristics, thereby improving the accuracy and reliability of the inversion.

[0029] As Figure 3 shown, the spatial pre-training network model proposed in the present embodiment realizes the cross-domain feature coupling of the observation signal and the underground structure by constructing a physical association framework of the gravity field features and the inversion space. The model designs a double-encoder mechanism: the F-Encoder extracts the multi-scale local features in the gravity field map and captures the spatial correlation of the observation points based on the convolution and self-attention module, which is essentially the "physical signal coding" of the field response of the density interface fluctuation, structure boundary and other geological bodies, generating a low-dimensional vector F containing the gravity field features; however, the P-Encoder generates a high-dimensional vector P containing the geometric structure information of the target body such as salt dome by introducing the T-Net module of PointNet to perform geometric transformation alignment on the position coordinates of the inversion space, realizing the "spatial prior coding" of the morphological features of the underground structure. The physical meaning of the model lies in breaking the separation processing mode of "gravity field features" and "density anomaly body spatial position" in the traditional inversion: that is, through the cooperation of F-Encoder and P-Encoder, the gradient changes caused by the density differences of gravity anomalies and the geometric constraints of the inversion space are considered simultaneously, forming a "field-space feature coupling" mechanism; the structured representation of the position coordinates by the T-Net module is equivalent to introducing the morphological prior of the geological body, and the self-attention layer models the correlation of the observation points, which implicitly learns the spatial covariance of the gravity field; the cross-domain mapping relationship established by pre-training provides regularization constraints for the inversion process, effectively suppressing the non-uniqueness of the solution, so that the inversion result is closer to the real geological structure.

[0030] The T-Net module of the PointNet architecture is integrated into the design of the P-Encoder, and its core function is to convert the position coordinate information of the inversion space in the simulation data into a high-dimensional feature vector P after a series of transformation and alignment operations, thereby enhancing the geometric invariance of the model. The module includes an input transformation network and a feature transformation network: first, the spatial position features of the input are extracted by a multi-layer perceptron (MLP), and the global feature vector is aggregated by maximum pooling, and then a transformation matrix of the corresponding dimension is output by a fully connected layer; finally, the input data is spatially aligned by using the transformation matrix, realizing the transformation of the feature space.

[0031] Further, the multiple gravity anomaly data features are fused as follows: In the field of geophysical gravity anomaly body inversion, gravity anomaly data Gz and gravity gradient component Gxz are two common types of observation data that can reflect different characteristics of underground anomaly bodies. Gravity anomaly data Gz usually contains more low-frequency information, and the low-frequency part corresponds to large-scale geological body structures. Analysis and processing of these low-frequency information can effectively extract the general shape and distribution range of large-scale deep geological bodies, providing macro constraints for inversion. Gxz contains more high-frequency information, and the high-frequency information corresponds to the edges and details of the image or signal. In geophysical inversion, these high-frequency information helps to identify the boundaries and faults of underground geological bodies. Through extraction and analysis of high-frequency information, the fine structure of geological bodies can be described. The joint inversion method using these two types of gravity anomaly data can comprehensively depict the underground geological structure.

[0032] Therefore, after using Mamba and space and training two independent channels to extract features, the model uses concatenation operation to fuse the multi-scale information extracted by the two channels to realize the collaborative optimization of geophysical constraints and gravity data driving. In addition, the Mamba module extracts the low-frequency components of gravity anomaly data Gz to reflect the macro shape and distribution range of deep large-scale geological bodies, providing regional background constraints for inversion; the spatial pre-training module extracts the high-frequency components of the gravity gradient component Gxz, which is sensitive to fine structures such as shallow faults and boundaries, and enhances the boundary resolution capability. Through feature vector concatenation, the two kinds of field information form a multi-scale fusion mechanism of “low-frequency control background and high-frequency detail”, providing complete physical constraints covering different geological depths and structural scales for inversion, and providing more abundant input representations for subsequent network layers.

[0033] Further, the U-shaped neural network uses a one-dimensional convolutional layer to replace the traditional two-dimensional convolutional layer for processing one-dimensional sequence characteristics of gravity anomaly data.

[0034] Specifically, the U-shaped network structure is originally derived from the classic literature “U-Net: Convolutional Networks for Biomedical Image Segmentation” in the field of biomedical image segmentation. Through the U-shaped topology structure and cross-layer skip connection mechanism, it realizes efficient fusion of multi-scale features of images, and shows excellent performance in medical image segmentation tasks. When migrated to the field of geophysical model inversion, the two-dimensional convolutional architecture used by the traditional U-shaped network faces significant challenges: for gravity anomaly data inversion, whether it is a 1×N row direction convolution or an N×N two-dimensional regional convolution, both introduce redundant feature coupling calculations, resulting in limited model inversion accuracy and high computational complexity.

[0035] In the geophysical gravity inversion scene, the gravity anomaly data essentially constitutes a one-dimensional ordered sequence, and the gravity anomaly value of each measuring point corresponds to an element in the sequence, and the physical meaning is mainly reflected in the spatial correlation of the measuring point. In view of this, the embodiment proposes an improved scheme suitable for the characteristics of geophysical data: replace the two-dimensional convolution layer in the traditional U-shaped network with a one-dimensional convolution, while maintaining the multi-scale feature fusion capability, significantly improving the computational efficiency.

[0036] From the perspective of parameter optimization, assuming that the input feature dimension is I (such as including gravity anomaly, terrain correction value and other multi-source data channels), the output feature dimension is O, and the one-dimensional convolution kernel length is K, the learnable parameter quantity of the one-dimensional convolution layer is: In comparison, the parameter quantity of the two-dimensional convolution layer (convolution kernel size K_h*K_w) is: When processing one-dimensional sequence data of the same scale, the parameter quantity of the two-dimensional convolution increases in a square level with the kernel size, while the one-dimensional convolution only increases linearly. This feature makes the one-dimensional convolution have a significant advantage in geophysical long sequence data processing, which can effectively capture the spatial correlation of gravity anomaly data through convolution kernel sliding operation, and avoid the redundant calculation introduced by two-dimensional convolution, thereby greatly reducing the model parameter quantity and computational complexity on the premise of ensuring the inversion accuracy, providing a lightweight solution for fast gravity field inversion.

[0037] Further, the U-shaped neural network also includes a full-scale skip connection mechanism for fusing features of different scales in the encoder and the decoder. The full-scale skip connection adjusts the features of different scales to the same scale through linear mapping and then fuses them, and further integrates them through one-dimensional convolution.

[0038] Specifically, in the process of using the traditional U-Net for geophysical gravity density anomaly body inversion, different scale features can provide different information. Low-level features have small receptive fields and can capture boundary texture information in the field map; high-level features have large receptive fields and can capture semantic information in the field map. When the traditional U-Net performs layer-by-layer downsampling, part of the edge information is lost. This lost feature cannot be recovered by upsampling. Although the U-Net uses skip connection to fuse multi-scale information, it is difficult for the decoder to fully utilize the full-scale feature information extracted by the encoder during the upsampling process. To solve this problem, the embodiment uses full-scale skip connection.

[0039] As Figure 4As shown, gravity field data is processed through a U-shaped architecture consisting of an encoder and a decoder. The encoder in the first half focuses on feature extraction, while the decoder in the second half is responsible for feature fusion and mapping. This embodiment employs a full-scale skip connection strategy to fuse gravity features from different scales. The abovesampling layer D2 is used as an example for detailed explanation. As shown in part t in the figure, D2 differs from traditional skip connections; it doesn't simply rely on D3 directly fusing features with E2 after deconvolution, but consists of four parts. D2 first adjusts the features extracted from all scales of the neural network to the same scale through linear mapping before fusing them. Finally, a one-dimensional convolution operation is used to further integrate these fused features.

[0040] The mathematical expression is as follows: In this context, function C represents the convolution operation, function H represents the feature aggregation mechanism (one convolutional layer + one BN + one ReLU), function D and function u represent the upsampling and downsampling operations respectively, "[·]" represents channel dimension splicing and fusion, E is the feature of each stage of the encoder; i marks the "current encoder stage of interest"; k is a temporary variable for traversing "earlier encoder stages" and is used to aggregate multi-scale features.

[0041] In this way, each decoder module fuses small-scale and same-scale feature maps from the encoder, as well as large-scale feature maps from previous decoders. These feature maps capture both fine-grained and coarse-grained semantics across the entire scale, and require fewer parameters. In geophysics, employing full-scale skip connections can effectively reduce the loss of fine-grained spatial information, which can prevent the capture of boundary information across different geological zones and lead to decreased accuracy in anomaly edge localization.

[0042] To illustrate the effectiveness of the network model, this embodiment designs a relatively simple simulation model ( Figure 7 Tests and analyses were conducted, and comparative experiments were performed using other deep learning inversion methods. The results are as follows: Figure 5 As shown, Figure 5 (a1) in the image represents a 3D visual image of the real model. Figure 5 (a2) in the diagram is the xz cross-section of the actual model along y=1800 meters. Figure 5 (a3) is a cross-sectional view of the actual model along the yz plane at x=700 meters. Figure 5 (b1) in the image is a three-dimensional visual image of the U-Net inversion result. Figure 5 (b2) in the image is the xz cross-section of the U-Net inversion result along y=180°. Figure 5 (b3) in the figure is the yz cross-section of the U-Net inversion result along x=700.Figure 5 (c1) in the image represents a 3D visual image of the TransUnet inversion result. Figure 5 (c2) in the figure is the xz cross-section of the TransUnet inversion result along y=180°. Figure 5 (c3) in the figure is the yz cross-section of the TransUnet inversion result along x=700. Figure 5 (d1) in the image represents a 3D visual image of the PMU-Net inversion result. Figure 5 (d2) in the figure is the xz cross-section of the PMU-Net inversion result along y=180°. Figure 6 (d3) in the image is the yz section plot of the PMU-Net inversion result along x=700. Figure 7 It is clearly observed that U-Net did not achieve ideal accuracy in both the 3D inversion space and the 2D field map. Although the 3D visualization model predicted by U-Net largely matches the simulation model, the accuracy of the 2D field map was significantly affected due to blurred edges. This phenomenon may be attributed to U-Net's shortcomings in extracting global information.

[0043] On the other hand, TransUNet has lower accuracy in 3D inversion space, but its accuracy in 2D field maps surpasses that of UNet. This is because TransUNet introduces a self-attention mechanism, enabling it to capture global information and more accurately depict edge details. Therefore, the density contour lines in the field maps generated by TransUNet are smoother. However, due to the model's limitations in learning location information, the predicted density anomaly locations are shifted, thus affecting the accuracy in 3D inversion space. It is worth noting that while UNet also has limitations in learning location information, its deviation in anomaly location prediction is uncommon; the TransUNet case shown in the illustration is a special case and extremely rare.

[0044] The PMU-Net network proposed in this embodiment exhibits excellent edge information restoration capability on the inversion result. It effectively makes up for the information loss and position offset problems often encountered by traditional U-Net and TransUNet in restoring edge information. This significant advantage is mainly due to the pre-trained model and attention mechanism adopted by the PMU-Net network, which enables it to better capture the internal dependency between the density anomaly body position and the station position that U-Net and TransUNet ignore, as well as the mutual dependency between the two. At the same time, the full-scale skip connection adopted by the PMU-Net network can more comprehensively capture coarse-grained semantic information and fine-grained semantic information, effectively reducing the feature loss in the downsampling process. Therefore, the inversion method of the network predicts a density structure that has a higher degree of agreement with the simulation model.

[0045] Further, the method further comprises generating simulation training data: initializing inversion space parameters and geological body feature parameters; generating a survey network, setting grid parameters and constructing a structured sampling grid; constructing a density distribution model in the inversion space according to the geological body feature parameters; calculating the gravity anomaly data corresponding to the density distribution model according to the forward formula; storing the gravity anomaly data and the corresponding density distribution model to form a training data set.

[0046] Specifically, integrating prior knowledge into the training data and input data of the neural network is an effective method to impose constraints on data-driven deep learning, which can effectively solve the weakness of weak interpretability of deep learning in gravity inversion and further improve the prediction ability of deep learning. Generating simulation data is an important way to embed prior knowledge in training data, and deep learning has always faced the challenge of lack of data in the field of gravity inversion. Therefore, it is very important to generate good simulation data.

[0047] Theoretically, the gravity anomaly observed by the survey network is caused by the density anomaly body in the inversion space, and is irrelevant to the background country rock. The gravity anomaly is a positive anomaly. And the real geological body is usually irregular, but the influence of simple geological body on gravity anomaly can be simulated by generating some regular density anomaly bodies, and finally the superposition of some regular anomaly bodies is used to simulate complex geological bodies.

[0048] ​A flowchart is generated for the dataset. The flowchart shows the dataset generation process for geophysical inversion research using deep learning methods. The flow starts with "initial configuration", which defines the inversion space parameters (including x-axis, y-axis, z-axis range, where mp_inter is the division interval of the inversion space, and num_block is the number of inversion space division blocks) and geological body characteristics (x1, x2, y1, y2, z1, z2 are the starting and ending coordinates of the x, y, z axes, and q represents the filling density of the geological body), laying the foundation for subsequent modeling. Next is "measurement network generation", which sets the grid parameters (the measurement network is divided by an interval distance of ob_inter), and assigns values to obx, obz by traversing the X / Y direction, constructing a structured sampling grid. Then, "inversion space generation" resets the inversion space parameters (the geological body is divided by an interval distance of mp_inter=200), and traverses all blocks: when the block space position satisfies x1≤i≤x2, y1≤j≤y2, z1≤k≤z2, its density is set to q, thereby constructing a density distribution model. The "forward calculation" step uses the forward formula to calculate the gravity anomaly data, simulating the observation results based on the defined density model. Finally, "data storage and file management" creates a save directory, generates a file name and writes the gravity anomaly data, ensuring the systematic saving of synthetic data. These data are crucial for training deep learning models and can be used for subsequent inversion and estimation of actual density distribution from observed anomalies, bridging the forward modeling and inversion tasks in the field of geophysics.

[0049] where the gravity anomaly is caused by the change of the gravitational field. As shown in ​ When the density of the inversion space is uniform, point A is only subject to the force of the Earth's gravity, and its force direction is vertically downward; however, when there is a density anomaly body X in the inversion space that is different from the surrounding environment, X will produce a new gravity, causing point A to be affected by the gravity of the Earth and the gravity of the density anomaly body X. The superposition of these two forces causes the gravitational field of point A to change, resulting in a gravity anomaly. The gravity anomaly value is calculated by the forward formula: In the formula, V represents the gravity potential, x, y, z represent the spatial coordinates of the observation point, which are used to determine the position of the observation point in space, x', z' represent the source point (anomaly body) coordinate variables, which are used to determine the position of the mass source (such as underground density anomaly body, etc.) generating the gravity field in space. r represents the position vector of the field point, which represents the spatial position of the observation point and can be composed of coordinates (x, y, z), r' represents the position vector of the source point, which represents the spatial vector of the position of the mass source, |r'-r| represents the distance between the field point and the source point, that is, the modulus of the difference between the position vectors r' and r, reflecting the spatial distance between the two. p(r') represents the mass density at the source point, which describes the mass distribution in the unit volume of the mass source. γ represents a proportional constant related to gravity, which is related to the physical law of the gravity field, the unit system, etc. (for example, the gravitational constant G is related). represents a triple volume integral, which is integrated with respect to the volume element dv' of the mass source, and the purpose is to accumulate the contribution of all mass source elements to the second-order partial derivative of the gravity potential of the observation point (gravity gradient component). σ represents the mass surface density (here the integral is a volume integral, which can be more accurately understood as a mass density, which represents the mass in a unit volume and describes the density characteristics of the mass distribution generating the gravity field). ξ, η, ζ represent the spatial coordinates of the source point, that is, the spatial coordinates of the position of the mass element generating the gravity field. dξ, dη, dζ represent the volume element, which is a small volume part in the integral region, used to accumulate the mass distribution.

[0050] In this embodiment, a survey network with a size of 4000m*4000m is used, and the grid spacing is set to 200m, so that it can be accurately divided into 400 observation points. For the convenience of further research and discussion, it is assumed that the survey network is a plane, that is, the z coordinates of all observation points are equal to 0. Gz and Gxz obtained by the forward formula are used as the input of the deep neural network, so that the geological information in the model can be more accurately extracted. Through this method, the causes of gravity anomalies and the relationship between gravity anomaly values and inverted spatial density can be better understood, and a solid theoretical basis can be provided for subsequent inversion work.

[0051] The embodiment also provides a computer device, which comprises a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method.

[0052] The embodiment also provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to realize the steps of the method.

[0053] The above merely provides the preferred embodiments of the present application, and the protection scope of the present application is not limited thereto, and any changes or substitutions within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A gravity anomaly body inversion method based on a PMU-Net deep learning network, characterized in that, The method comprises the following steps: obtaining two different types of gravity anomaly data as input; processing the two types of gravity anomaly data through two independent channels to mine the correlation and complementarity between the data, wherein one channel uses a Mamba module to capture the dependence between observation points, and the other channel captures the positional correlation between observation points and underground density through a spatial pre-training model; fusing the processed features, and inputting the fused features into a U-shaped neural network constructed by one-dimensional convolution to obtain the underground density distribution result.

2. The method of claim 1, wherein, The two different types of gravity anomaly data include gravity anomaly data and gravity gradient anomaly data.

3. The method of claim 1, wherein, The Mamba module comprises: performing relative coordinate coding on each measurement point of the gravity anomaly data to construct a spatial correlation matrix; extracting spatial pattern features of the gravity anomaly data through a multi-scale convolution kernel; based on a dynamic state space model, constructing an observation point evolution sequence with a plane coordinate grid index, and evolving the physical state of each observation point according to the law of conservation of mass; mapping the gravity anomaly signal through an underground physical property structure response matrix, and tensor coupling the mapping result and the initial geological state through a state transition matrix; projecting the updated geological state to the detectable physical quantity space through an observation operator, and introducing a lithology-density conversion matrix to retain the direct influence term of the original input signal; performing a double-path fusion mechanism to cooperatively constrain the tectonic evolution dynamics process and the static distribution of density physical properties.

4. The method of claim 1, wherein, The spatial pre-training model comprises: constructing a physical correlation framework of gravity field features and inversion space to perform cross-domain feature coupling of observation signals and underground structures; extracting multi-scale local features in the gravity field map through an F-Encoder and capturing the spatial correlation of observation points to generate a low-dimensional vector representing the gravity field features; performing geometric transformation alignment on the position coordinates of the inversion space through a P-Encoder to generate a high-dimensional vector containing the geometric structure information of the target body; synchronously processing the gradient changes caused by gravity anomaly density differences and the geometric constraints of the inversion space by cooperating the low-dimensional vector and the high-dimensional vector to form a field-space feature coupling mechanism; providing regularization constraints for the inversion process through the field-space feature coupling mechanism.

5. The method of claim 1, wherein, The U-shaped neural network replaces the traditional two-dimensional convolution layer with a one-dimensional convolution layer to process the one-dimensional sequence characteristics of the gravity anomaly data.

6. The method of claim 1, wherein, The U-shaped neural network further comprises a full-scale skip connection mechanism for fusing features of different scales in the encoder and the decoder.

7. The method of claim 6, wherein, The full-scale skip connection mechanism adjusts the features of different scales to the same scale through linear mapping and then fuses them, and further integrates them through one-dimensional convolution.

8. The method of claim 1, wherein, It also includes generating simulated training data: initializing inversion space parameters and geological body feature parameters; generating a measurement network, setting grid parameters and constructing a structured sampling grid; constructing a density distribution model in the inversion space according to the geological body feature parameters; calculating the gravity anomaly data corresponding to the density distribution model according to the forward formula; storing the gravity anomaly data and the corresponding density distribution model to form a training data set.

9. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 8. The processor executes the computer program to implement the steps of the method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1-8.

Images