CN120910810B - Gravity Anomaly Inversion Method Based on PMU-Net Deep Learning Network - Google Patents
Gravity Anomaly Inversion Method Based on PMU-Net Deep Learning NetworkInfo
- Publication number
- CN120910810B CN120910810B CN202511439014.8A CN202511439014A CN120910810B CN 120910810 B CN120910810 B CN 120910810B CN 202511439014 A CN202511439014 A CN 202511439014A CN 120910810 B CN120910810 B CN 120910810B
- Authority
- CN
- China
- Prior art keywords
- gravity
- inversion
- dimensional
- gravity anomaly
- density
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V7/00—Measuring gravitational fields or waves; Gravimetric prospecting or detecting
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Geophysics (AREA)
- Geophysics And Detection Of Objects (AREA)
Abstract
本发明公开了一种基于PMU‑Net深度学习网络的重力异常体反演方法，属于重力反演领域，包括：通过联合反演两种不同重力异常数据，增加观测方程组数量，全面呈现地下地质结构。借助两个独立通道处理数据，一通道使用Mamba模块捕捉观测点间依赖，另一通道利用空间预训练模型捕捉观测点与地下密度的位置关联。处理后的特征融合输入一维卷积构建的U型神经网络，减少计算量并提升效率。还包含模拟数据生成模块，通过嵌入先验知识解决深度学习在重力反演中的可解释性弱问题。本发明有效提高了复杂地质条件下反演的准确性和可靠性，为地球物理勘探提供了一种高效、准确的新方法。
This invention discloses a gravity anomaly inversion method based on the PMU-Net deep learning network, belonging to the field of gravity inversion. It includes: jointly inverting two different gravity anomaly datasets to increase the number of observation equations and comprehensively present the subsurface geological structure. Data is processed using two independent channels: one channel uses a Mamba module to capture dependencies between observation points, and the other channel utilizes a spatially pre-trained model to capture the positional correlation between observation points and subsurface density. The processed features are fused and input into a U-shaped neural network constructed using one-dimensional convolutions, reducing computational load and improving efficiency. A simulated data generation module is also included, which addresses the weak interpretability problem of deep learning in gravity inversion by embedding prior knowledge. This invention effectively improves the accuracy and reliability of inversion under complex geological conditions, providing a new, efficient, and accurate method for geophysical exploration.
Description
Technical Field
The invention belongs to the technical field of gravity inversion, and particularly relates to a gravity anomaly inversion method based on a PMU-Net deep learning network.
Background
Gravity inversion is a key technology in geophysical exploration, and has the remarkable advantages of low cost and high efficiency by measuring the surface gravity field and deducing the underground density distribution by means of an algorithm model. In recent years, with the development of artificial intelligence technology, people begin to predict gravity data by deep learning to rapidly image a subsurface geological structure. The early method mainly uses single gravity field data as input to directly predict the underground density body, but the mode can only reflect the gravity abnormality of the underground density body in a specific aspect, the information is extremely one-sided, and the uncertainty of inversion solution is higher.
In order to improve the situation, a GRAINVNET model based on U-Net is designed, and gravity anomaly data and gravity gradient anomaly data are adopted for joint inversion, so that a good prediction result is obtained. However, the method is simple in processing different gravity anomaly data, and complex features and correlations between the data are difficult to capture. Later, different characteristics are respectively predicted by different decoders, so that the information capturing capacity is further improved, and the problem of insufficient association capturing between observation points and the underground density model exists. Fusion of the unet++ network architecture with LeakyReLU activation functions using an n×n convolution kernel has also been proposed to improve the accuracy of the inversion result, but the high computational complexity of two-dimensional convolution limits the efficiency of the model.
In summary, the existing deep learning model has shortcomings in fully describing the underground geological structure, mining the relevance between different data, learning the relationship between observation points and density grids, calculating efficiency and the like, and needs to be improved to meet the requirements of high-precision and high-efficiency gravity inversion under complex geological conditions.
Disclosure of Invention
The invention provides a gravity anomaly inversion method based on a PMU-Net deep learning network, which aims to solve the problems in the prior art.
In order to achieve the above purpose, the invention provides a gravity anomaly inversion method based on a PMU-Net deep learning network, which comprises the following steps:
Acquiring two different types of gravity anomaly data as input;
Respectively processing two kinds of gravity anomaly data through two independent channels, mining the correlation and complementarity between the data, wherein one channel captures the dependence between observation points by using Mamba modules, and the other channel captures the position correlation between the observation points and the underground density through a space pre-training model;
And 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.
Optionally, the two different types of gravity anomaly data include gravity anomaly data and gravity gradient anomaly data.
Optionally, the Mamba module includes:
Carrying out relative coordinate coding on each measuring point of the gravity abnormal data to construct a space association matrix;
extracting the space mode characteristics of the gravity anomaly data through a multi-scale convolution kernel;
Based on a dynamic state space model, constructing a measuring point evolution sequence by using a plane coordinate grid index, and evolving the physical state of each measuring point according to a mass conservation law;
Mapping the gravity anomaly signals through a response matrix of the underground physical structure, and tensor coupling the mapping result and the initial geological state through a state transition matrix;
Projecting the updated geological state to a detectable physical quantity space through an observation operator, and introducing a lithology-density conversion matrix to reserve direct influence items of an original input signal;
and cooperative constraint of the structural evolution dynamic process and the density physical property static distribution is carried out through a double-path fusion mechanism.
Optionally, the spatial pre-training model includes:
Constructing a physical association frame of gravitational field characteristics and inversion space, and performing cross-domain characteristic coupling of observation signals and underground structures;
Extracting multi-scale local features in the gravity field diagram through F-Encoder, capturing the spatial correlation of observation points, and generating a low-dimensional vector representing the gravity field features;
Performing geometric transformation alignment on the inversion space position coordinates through P-Encoder to generate a high-dimensional vector containing the geometric structure information of the target body;
synchronously processing gradient change caused by gravity abnormal density difference and geometric constraint of inversion space by cooperating the low-dimensional vector and the high-dimensional vector to form a field-space characteristic coupling mechanism;
Regularization constraints are provided for the inversion process by the field-space feature coupling mechanism.
Optionally, the U-shaped neural network adopts a one-dimensional convolution layer to replace a traditional two-dimensional convolution layer, and is used for processing one-dimensional sequence characteristics of gravity anomaly data.
Optionally, the U-shaped neural network further comprises a full-scale jump connection mechanism for fusing features of different scales in the encoder and decoder.
Optionally, the full-scale jump connection mechanism adjusts the features of different scales to the same scale through linear mapping, then performs fusion, and further integrates through one-dimensional convolution.
Optionally, generating simulated training data is further included:
Initializing inversion space parameters and geologic body characteristic parameters;
generating a measuring network, setting grid parameters and constructing a structured sampling grid;
Constructing a density distribution model in the inversion space according to the geologic body characteristic parameters;
Calculating gravity anomaly data corresponding to the density distribution model according to a forward formula;
and storing the gravity anomaly data and the corresponding density distribution model to form a training data set.
The invention also provides a computer device comprising a memory, a processor and a computer program stored on the memory, the processor executing the computer program to carry out the steps of the method.
The invention also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor realizes the steps of the method.
Compared with the prior art, the invention has the following advantages and technical effects:
The gravity anomaly inversion method based on the PMU-Net deep learning network of the space pre-training is obviously superior to the prior art in multiple aspects, namely, firstly, in the aspect of inversion precision, two kinds of gravity anomaly data Gz and Gxz are subjected to joint inversion, and the data are respectively processed by means of two independent channels, so that the system can capture the characteristics of the underground geological structure more comprehensively, and the key indexes such as average gravity anomaly offset error, average gravity anomaly accuracy and the like are obviously improved. Secondly, PMU-Net is excellent in edge information restoring capability, and the problems of edge blurring and position deviation common in the prior art are effectively avoided. This mainly benefits from the introduction of pre-training models and attention mechanisms, and the design of full-scale jump connections, enabling the system to better capture the dependency between density anomaly locations and measurement point locations while reducing feature loss during downsampling. In addition, in the aspect of calculation efficiency, the U-shaped neural network constructed by adopting one-dimensional convolution remarkably reduces calculation complexity, improves the calculation efficiency of a model, and provides a light-weight solution for rapid inversion of a gravity field. Meanwhile, through full-scale jump connection, the system can more comprehensively integrate gravity characteristics of different scales, semantic information from fine granularity to coarse granularity is effectively captured, and the accuracy of inversion results is further improved. Finally, by embedding priori knowledge in training data, the system can better combine geophysical constraint and gravity data driving, further improve the prediction capability of deep learning, and effectively solve the problem of weak interpretability of the deep learning in gravity inversion. In conclusion, the PMU-Net deep learning network provided by the invention has remarkable superiority in gravity anomaly inversion under complex geological conditions, not only improves the inversion precision and reliability, but also greatly improves the calculation efficiency, and provides a new efficient and accurate method for geophysical exploration.
Drawings
The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:
FIG. 1 is a schematic diagram of feature fusion according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of Mamba modules according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a spatial pre-training model according to an embodiment of the present invention;
FIG. 4 is a full-scale jump connection schematic of an embodiment of the present invention;
FIG. 5 is a comparison plot of inversion effects for an embodiment of the invention, wherein (a 1) is a three-dimensional visual image of a real model, (a 2) is an x-z plot of a real model along y=1800 meters, (a 3) is a y-z plot of a real model along x=700 meters, (b 1) is a three-dimensional visual image of a U-Net inversion result along y=1800, (b 2) is an x-z plot of a U-Net inversion result along x=700, (c 1) is a three-dimensional visual image of a TransUnet inversion result along y=700, (c 2) is an x-z plot of a TransUnet inversion result along y=1800, (c 3) is a TransUnet inversion result along y-z plot of x=700, (d 1) is a three-dimensional visual image of a PMU-Net inversion result along y=1800, (d 3) is a y-z plot of a PMU-Net inversion result along x=700;
FIG. 6 is a flow chart of data set generation according to an embodiment of the present invention;
fig. 7 is a schematic diagram of a gravity forward modeling according to an embodiment of the present invention.
Detailed Description
It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other. The application will be described in detail below with reference to the drawings in connection with embodiments.
It should be noted that the steps illustrated in the flowcharts of the figures may be performed in a computer system such as a set of computer executable instructions, and that although a logical order is illustrated in the flowcharts, in some cases the steps illustrated or described may be performed in an order other than that illustrated herein.
Example 1
The embodiment provides a gravity anomaly inversion method based on a PMU-Net deep learning network, which comprises the following steps:
Acquiring two different types of gravity anomaly data as input;
Respectively processing two kinds of gravity anomaly data through two independent channels, mining the correlation and complementarity between the data, wherein one channel captures the dependence between observation points by using Mamba modules, and the other channel captures the position correlation between the observation points and the underground density through a space pre-training model;
And 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.
Specifically, the model uses the gravity component Gz and the gravity gradient component Gxz as inputs to the model, and the two different types of gravity anomaly data are processed separately through two separate channels. The gravity anomaly graph Gz solves the problem of long-distance dependence through relative position coding, and captures the dependence relationship between each observation point through a Mamba module, and the gravity anomaly data Gxz captures the position association between each observation point and each density grid of the underground density model through a space pre-training model. Then, the two types of processed features are fused, then a U-shaped neural network constructed by one-dimensional convolution is input, and features with different layers and different scales can be fused by using full-scale jump connection. As shown in fig. 1.
Further, mamba modules include:
Carrying out relative coordinate coding on each measuring point of the gravity abnormal data to construct a space association matrix;
extracting the space mode characteristics of the gravity anomaly data through a multi-scale convolution kernel;
Based on a dynamic state space model, constructing a measuring point evolution sequence by using a plane coordinate grid index, and evolving the physical state of each measuring point according to a mass conservation law;
Mapping the gravity anomaly signals through a response matrix of the underground physical structure, and tensor coupling the mapping result and the initial geological state through a state transition matrix;
Projecting the updated geological state to a detectable physical quantity space through an observation operator, and introducing a lithology-density conversion matrix to reserve direct influence items of an original input signal;
and cooperative constraint of the structural evolution dynamic process and the density physical property static distribution is carried out through a double-path fusion mechanism.
Specifically, mamba modules are shown in FIG. 2.
In fig. 2, the mathematical expression is:
Wherein h represents a geological state, A represents a state transition matrix, t represents a measuring point, B represents a response matrix, X represents a gravity anomaly signal, Y represents an observation operator, Z represents a lithology-density transition matrix, and Y represents an observation output.
The embodiment constructs a gravity data characteristic extraction method based on dynamic state space modeling, constructs a space association matrix through relative coordinate coding of gravity data Gz measuring points, and accurately extracts space mode characteristics of regional gravity anomaly by adopting a multi-scale convolution kernel corresponding to a typical construction scale of 5-50 km. After processing, a characteristic tensor (batch dimension 16×measurement point sequence 1681×characteristic dimension 512) containing 512-dimensional density physical parameters is formed, and the tensor completely characterizes key geological information such as salt structure development degree of a research area.
In the dynamic modeling stage, a plane coordinate grid index is used for constructing a measuring point evolution sequence (t=0-1680), and the physical state evolution of each measuring point strictly follows the law of conservation of mass. Taking the measurement point t=1 as an example, the gravity anomaly signal x (1) is mapped through the response matrix B (dimension 1024×512) of the underground physical structure, and the matrix element B ij quantitatively characterizes the sensitivity degree of the j-th type density physical parameter to the i-th type geological structure deformation. The mapping result is tensor coupled with the initial geological state h (0) (containing the regional structural stress field background information) through a state transition matrix a (dimensions 1024 x 1024). Wherein, matrix A kl describes the mutual coupling relation between k-type geologic structure units and l-type lithofacies density units caused by geological processes such as fault activity, magma invasion and the like. The updated geological state h (1) is projected to a detectable physical quantity space through an observation operator Y (512 multiplied by 1024), and the row vector of the operator corresponds to the gravity response kernel function of the interfaces with different densities. At the same time, a lithology-density transformation matrix Z (512×512) is introduced, preserving the direct influence terms of the original input signal x (1). Through a double-path fusion mechanism, the cooperative constraint of the structural evolution dynamic process and the density physical property static distribution is realized.
Further, the spatial pre-training model comprises:
Constructing a physical association frame of gravitational field characteristics and inversion space, and performing cross-domain characteristic coupling of observation signals and underground structures;
Extracting multi-scale local features in the gravity field diagram through F-Encoder, capturing the spatial correlation of observation points, and generating a low-dimensional vector representing the gravity field features;
Performing geometric transformation alignment on the inversion space position coordinates through P-Encoder to generate a high-dimensional vector containing the geometric structure information of the target body;
synchronously processing gradient change caused by gravity abnormal density difference and geometric constraint of inversion space by cooperating the low-dimensional vector and the high-dimensional vector to form a field-space characteristic coupling mechanism;
Regularization constraints are provided for the inversion process by the field-space feature coupling mechanism.
Specifically, the influence of inversion space and network measurement space position information on gravity inversion accuracy cannot be realized in the inversion process of traditional CNN, U-Net and other network models. Therefore, the invention aims to enable the model to more accurately capture the space position characteristics of each gravity anomaly data in the observation network and each density unit in the underground density inversion space and what dependency relationship exists between the characteristics, thereby improving the inversion precision and reliability.
As shown in fig. 3, the spatial pre-training network model proposed in the present embodiment implements cross-domain feature coupling between an observation signal and a subsurface structure by constructing a physical correlation frame of gravitational field features and inversion space. F-Encoder extracts multi-scale local features in a gravity field diagram and captures the spatial correlation of observation points based on convolution and self-attention modules, the model is essentially 'physical signal coding' of geologic body field responses such as density interface fluctuation, construction boundaries and the like, a low-dimensional vector F implying gravitational field features is generated, however P-Encoder performs geometric transformation alignment on inversion space position coordinates through introducing a T-Net module of PointNet, a high-dimensional vector P containing target body geometric structure information such as salt dome and the like is generated, and 'space priori coding' of morphological features of underground structures is realized. The physical significance of the model is that a separation processing mode of 'gravitational field characteristics' and 'density anomaly body space positions' in traditional inversion is broken through, namely, gradient changes caused by density differences of gravity anomalies and geometric constraints of inversion space are synchronously considered through cooperation of F-Encoder and P-Encoder to form a 'field-space characteristic coupling' mechanism, structural representation of position coordinates by a T-Net module is equivalent to introduction of geologic body morphology priori, spatial covariance of the gravitational field is implicitly learned by modeling of observation point association by a self-attention layer, and a cross-domain mapping relation established by pre-training provides regularization constraint for an inversion process, so that non-uniqueness of solutions is effectively restrained, and an inversion result is more close to a real geological structure.
The design of the P-Encoder is integrated with a PointNet architecture T-Net module, and the core function of the model is to convert the position coordinate information of the inversion space in the simulation data into a high-dimensional characteristic vector P after a series of transformation and alignment operations, so that the geometric invariance of the model is enhanced. The module comprises an input transformation network and a feature transformation network, wherein the input transformation network and the feature transformation network are respectively used for extracting features of input spatial position features through a multi-layer perceptron (MLP), aggregating the features into global feature vectors through maximum pooling, outputting a transformation matrix with corresponding dimensions through a full connection layer, and finally utilizing the transformation matrix to spatially align input data to realize transformation of feature space.
Further, the multi-force anomaly data feature fusion is as follows:
In the field of geophysical gravity anomaly inversion, gravity anomaly data Gz and gravity gradient components Gxz are two common types of observation data that can reflect different characteristics of subsurface anomalies. The gravity anomaly data Gz generally contains more low-frequency information, and the low-frequency part corresponds to a geologic body structure with a larger scale. The analysis and the processing of the low-frequency information can effectively extract the approximate form and the distribution range of the underground deep large-scale geologic body, and provide macroscopic constraint conditions for inversion. Gxz contain more high frequency information corresponding to edges and detail portions of the image or signal. In geophysical inversion, the high-frequency information is helpful for identifying fine structures such as boundaries, faults and the like of underground geologic bodies, and the fine structures of the geologic bodies can be characterized by extracting and analyzing the high-frequency information. The method for jointly inverting the two gravity anomaly data can comprehensively depict the underground geological structure.
Therefore, after the Mamba and two independent channels of space and training are adopted to extract the characteristics, the model uses the splicing (concatenation) operation to fuse the multi-scale information extracted by the two channels, so that the geophysical constraint and gravity data driven collaborative optimization is realized. In addition, the Mamba module extracts low-frequency components of gravity anomaly data Gz to reflect macroscopic morphology and distribution range of deep large-scale geologic bodies and provide inverted regional background constraint, and the space pre-training module extracts high-frequency components of gravity gradient components Gxz to be sensitive to fine structures such as shallow fault boundaries and the like, so that boundary resolution capability is enhanced. Through feature vector splicing, two field map information forms a multi-scale fusion mechanism for constructing low-frequency control background and high-frequency fixed details, provides complete physical constraints for inversion and covering different geological depths and construction scales, and also provides richer input characterization for a subsequent network layer.
Further, the U-shaped neural network adopts a one-dimensional convolution layer to replace a traditional two-dimensional convolution layer and is used for processing the one-dimensional sequence characteristic of the gravity anomaly data.
Specifically, the U-shaped network structure is originally derived from a classical document U-Net Convolutional Networks for Biomedical Image Segmentation in the biomedical image segmentation field, and the U-shaped network structure realizes the efficient fusion of image multi-scale characteristics through a U-shaped topological structure and a cross-layer jump connection mechanism, and shows excellent performance in medical image segmentation tasks. When migrating to the field of inversion of geophysical models, a two-dimensional convolution architecture adopted by a traditional U-shaped network faces significant challenges, namely, when inverting gravity anomaly data, whether the convolution is in a1 XN row direction or in an N XN two-dimensional area, the model inversion accuracy is improved only in a limited way and the calculation complexity is high due to the fact that redundant characteristic coupling calculation is introduced.
In a geophysical gravity inversion scene, gravity anomaly data essentially form a one-dimensional ordered sequence, and the gravity anomaly value of each measuring point corresponds to one element in the sequence, and the physical meaning of the gravity anomaly value is mainly reflected in the spatial correlation of the measuring points. In view of the above, the embodiment provides an improved scheme for adapting the geophysical data characteristics, wherein one-dimensional convolution is used for replacing a two-dimensional convolution layer in a traditional U-shaped network, and the computing efficiency is remarkably improved while the multi-scale feature fusion capability is maintained.
From parameter optimization, the input characteristic dimension is set as I (such as multi-source data channels including gravity anomaly, terrain correction value and the like), the output characteristic dimension is set as O, the one-dimensional convolution kernel length is set as K, and the learnable parameter number of the one-dimensional convolution layer is as follows:
In contrast, the parameter amounts of the two-dimensional convolution layer (convolution kernel size k_h×k_w) are:
When processing one-dimensional sequence data of equal scale, the parameter quantity of the two-dimensional convolution increases in square order with the kernel size, while the one-dimensional convolution only increases linearly. The characteristics enable the one-dimensional convolution to have remarkable advantages in geophysical long-sequence data processing, not only can effectively capture the space correlation of gravity anomaly data through convolution kernel sliding operation, but also can avoid redundant calculation introduced by two-dimensional convolution, so that model parameter and calculation complexity are greatly reduced on the premise of guaranteeing inversion accuracy, and a lightweight solution is provided for rapid inversion of a gravity field.
Further, the U-shaped neural network also comprises a full-scale jump connection mechanism for fusing the characteristics of different scales in the encoder and the decoder. The full-scale jump connection adjusts the features of different scales to the same scale through linear mapping, then carries out fusion, and further integrates through one-dimensional convolution.
In particular, in the gravity density anomaly inversion process of geophysics using conventional U-Net, features of different scales may provide different information. The low-level feature receptive field is small, and can capture boundary texture information in the field map, and the high-level feature receptive field is large, and can capture semantic information in the field map. The traditional U-Net can cause the loss of partial edge information when carrying out layer-by-layer downsampling, the lost characteristics can not be retrieved by upsampling, and although the U-Net uses jump connection to carry out multi-scale information fusion, thereby realizing the retrieval of partial edge information, the decoder still has difficulty in fully utilizing the full-scale characteristic information extracted by the encoder in downsampling in the upsampling process. To solve this problem, the present embodiment employs full-scale jump connection.
As shown in fig. 4, the gravitational field data will pass through a U-shaped architecture consisting of an encoder and a decoder. The encoder of the first half is dedicated to feature extraction, while the decoder of the second half is responsible for feature fusion and mapping. The embodiment adopts a full-scale jump connection strategy and aims at fusing gravity characteristics from different scales. The above sampling layer D2 is described in detail as an example. As illustrated in section t of the figure, D2 differs from the traditional jump connection in that it does not rely solely on D3 to be feature fused directly with E2 by deconvolution, but instead consists of four parts. And D2, adjusting the features of all scales extracted by each layer of the neural network to the same scale through linear mapping, and then fusing. Finally, the fused features are further integrated by a one-dimensional convolution operation.
The mathematical expression formula is as follows:
Wherein function C represents convolution operation, function H represents feature aggregation mechanism (one convolution layer+one BN+one ReLU), function D and function u represent up-sampling and down-sampling operation, respectively, "[. Cndot ]" represents channel dimension concatenation fusion, E is feature of each stage of the encoder, i marks "encoder stage of current interest", k is temporary variable traversing "earlier encoder stage", for aggregating multi-scale features.
In this way, each decoder module fuses small-scale and co-scale feature maps from the encoder, as well as large-scale feature maps from the previous decoder, which capture fine-grained and coarse-grained semantics at full scale, with fewer parameters. In the earth physics, the adoption of the full-scale jump connection mode can effectively reduce the situation that the edge positioning accuracy of an abnormal body is reduced because the boundary information of different geology cannot be captured due to the loss of fine granularity space information.
In order to illustrate the effectiveness of the network model, the embodiment designs a relatively simple simulation model (fig. 7) for test analysis, and adopts other depth learning inversion methods for comparison test, the results are shown in fig. 5, the (a 1) in fig. 5 is a three-dimensional visual image of the real model, the (a 2) in fig. 5 is an x-z section view of the real model along y=1800 m, the (a 3) in fig. 5 is a three-dimensional visual image of the real model along y=700 m, the (b 1) in fig. 5 is a three-dimensional visual image of the U-Net inversion result, the (b 2) in fig. 5 is an x-z section view of the U-Net inversion result along y=1800 m, the (b 3) in fig. 5 is a three-dimensional visual image of the U-Net inversion result along x=700, the (c 1) in fig. 5 is a three-dimensional visual image of the TransUnet inversion result, the (c 2) in fig. 5 is an x-z section view of the y=1800 m, the (c 3) in fig. 5 is a three-dimensional visual image of the inversion result along y=700 m, the (b 2) in fig. 5 is a three-dimensional visual image of the U-z section view of the U-Net inversion result along y=700, and the three-dimensional visual image of the three-dimensional visual image along the three-z section of the three-dimensional inversion result in fig. 5 (d=700, and the three-dimensional visual image along the three-dimensional visual image of the three-dimensional visual image in the three-dimensional inversion result can be observed along the three-dimensional visual image along the three-z=x=3 plane. Although the three-dimensional visualization model of the U-Net prediction is generally consistent with the simulation model, the accuracy of the two-dimensional field map is significantly affected due to the blurring of edges. This phenomenon may be due to the lack of U-Net in extracting global information.
On the other hand, transUNet has lower accuracy in the three-dimensional inversion space, but the accuracy of the two-dimensional field map exceeds UNet. This is because TransUNet introduced a self-attention (self-attention) mechanism that enabled it to capture global information, more precisely portraying edge details. Thus, the field pattern density contours generated by TransUNet are smoother. However, due to the fact that the model is insufficient in position information learning, the position of the predicted density anomaly is shifted, and the accuracy of the three-dimensional inversion space is further affected. It is noted that the above-mentioned U-Net also has limitation in learning position information, but deviation in abnormal body position prediction is not common, and TransUNet cases in the drawings are special cases, and are very rare.
The PMU-Net network provided by the embodiment shows excellent restoring capability for inversion space edge information on an inversion result. The method effectively solves the problems of information loss, position offset and the like frequently encountered by the traditional U-Net and TransUNet in the process of recovering the edge information. This significant advantage is mainly due to the pretrained model and the attention mechanism employed by the PMU-Net network, which enables it to better capture the internal dependencies between the density anomaly locations and the site locations, and the interdependencies between the two, that are ignored by U-Net and TransUNet. Meanwhile, the full-scale jump connection adopted by the PMU-Net network can more comprehensively capture coarse-granularity semantic information and fine-granularity semantic information, and feature loss in the downsampling process is effectively reduced. Therefore, the density structure predicted by the inversion method of the network has higher consistency with the simulation model.
Further, the method also includes generating simulated training data:
Initializing inversion space parameters and geologic body characteristic parameters;
generating a measuring network, setting grid parameters and constructing a structured sampling grid;
Constructing a density distribution model in the inversion space according to the geologic body characteristic parameters;
Calculating gravity anomaly data corresponding to the density distribution model according to a forward formula;
and storing the gravity anomaly data and the corresponding density distribution model to form a training data set.
Specifically, the integration of priori knowledge into training data and input data of a neural network is an effective method for applying constraints to data-driven deep learning, so that the defect of poor interpretability of the deep learning in gravity inversion can be effectively overcome, and the prediction capability of the deep learning is further improved. Generating simulation data is an important way to embed a priori knowledge in training data, and deep learning has been a challenge in the gravity inversion field in the absence of data. Thus, it is very important to generate good analog data.
In theory, the gravity anomaly observed by the survey net is caused by a density anomaly in the inversion space, and is independent of the background surrounding rock, and the gravity anomaly is positive anomaly. And the actual geologic body is usually irregular, but the influence of the simple geologic body on gravity anomaly can be simulated by generating some regular density anomaly bodies, and finally the complex geologic body can be simulated by superposition of some regular anomaly bodies.
Fig. 6 is a data set generation flow chart. The flow chart illustrates a data set generation process for geophysical inversion studies using a deep learning method. The process starts with an "initialization configuration", which defines inversion space parameters (including x-axis, y-axis, z-axis ranges, where mp_inter is the dividing interval of the inversion space, num_block is the number of dividing blocks of the inversion space) and geologic features (x 1、x2、y1、y2、z1、z2 is the start and stop coordinates of the x, y, z-axis, respectively, q represents the filling density of the geologic volume), and lays a foundation for subsequent modeling. Next is "grid generation", which sets grid parameters (the grid is partitioned with ob_inter as the separation distance), and constructs a structured sampling grid by traversing X/Y direction to assign values for obx, oby, obz. Subsequently, the "inversion space generation" resets the inversion space parameters (the geologic volume is partitioned with mp_inter=200 as the separation distance), traversing all the volumes by setting the density of the volume to q when its spatial position satisfies x 1≤i≤x2、y1≤j≤y2、z1≤k≤z2, thereby constructing a density distribution model. The step of forward calculation calculates gravity anomaly data by using a forward formula, and simulates an observation result based on a defined density model. Finally, creating a storage catalog, generating a file name and writing in gravity anomaly data by the data storage and file management, and ensuring systematic storage of the synthesized data. The data are critical to training a deep learning model, can be used for inversion and estimation of actual density distribution from observation anomalies, and connect forward modeling and inversion tasks in the geophysical field.
Wherein gravity anomalies are caused by changes in the gravitational field. As shown in FIG. 7, when the inversion space has uniform density, the point A is only acted by the gravity force, and the stress direction is vertical and downward, however, when an abnormal density body X with different density from the surrounding environment exists in the inversion space, the X generates a new gravity force, so that the point A is influenced by the gravity force of the abnormal density body X while being influenced by the gravity force of the gravity force, and the gravity field of the point A is changed due to superposition of the two gravity forces, so that gravity force abnormality is generated. Respectively solving the gravity anomaly values through a forward formula:
where V represents the gravity level, x, y, z represents the spatial coordinates of the observation point, for determining the position of the observation point in space, and x ', z' represents the source point (anomaly) coordinate variable, for determining the position of the mass source (e.g., subsurface density anomaly, etc.) generating the gravity field in space. r represents a position vector of a field point, represents a spatial position of an observation point, can be composed of coordinates (x, y, z), r ' represents a position vector of a source point, represents a spatial vector of a position of a quality source, and |r ' -r| represents a distance between the field point and the source point, namely, a position vector, and a modular length of a difference between r ' and r reflects a spatial distance between the two. ρ (r') represents the mass density at the source point, describing the mass distribution per unit volume of the mass source. Gamma denotes a proportionality constant related to gravity, and relates to the physical law of the gravitational field, a unit system, etc. (for example, a gravitational constant G is associated). Representing the triple volume integral, integrating the volume element dv' where the mass source is located, in order to accumulate the contributions of all the mass source elements to the second partial derivative of the gravitational potential at the observation point (gravity gradient component), σ representing the mass areal density (here the integral is the volume integral, more precisely understood as mass density, representing the mass per volume, describing the density characteristics of the mass distribution that produces the gravitational field). ζ, η, ζ represents the spatial coordinates of the source point, i.e. the spatial coordinates of the location of the mass element that produces the gravitational field. dζ, dη, dζ represent volume elements, which are tiny volume portions in the integration region for differential accumulation of the mass distribution.
In this example, a measurement grid with a size of 4000m x 4000m was used, and the grid pitch was set to 200m, so that the measurement grid was precisely divided into 400 observation points. For the convenience of further research and discussion, it is assumed that the measuring net is a plane, that is, the z coordinates of all the observation points are all equal to 0. Gz and Gxz obtained by forward formulas are used as input of the deep neural network, so that geological information in the model can be extracted more accurately. By the method, the cause of gravity anomaly and the relation between the gravity anomaly value and inversion space density can be further understood, and a solid theoretical basis is provided for subsequent inversion work.
The present embodiment also provides a computer apparatus comprising a memory, a processor and a computer program stored on the memory, the processor executing the computer program to carry out the steps of the method.
The present embodiment also provides a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of the method.
The present application is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present application are intended to be included in the scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.
Claims (8)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202511439014.8A CN120910810B (en) | 2025-10-10 | 2025-10-10 | Gravity Anomaly Inversion Method Based on PMU-Net Deep Learning Network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202511439014.8A CN120910810B (en) | 2025-10-10 | 2025-10-10 | Gravity Anomaly Inversion Method Based on PMU-Net Deep Learning Network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN120910810A CN120910810A (en) | 2025-11-07 |
| CN120910810B true CN120910810B (en) | 2025-12-26 |
Family
ID=97548159
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202511439014.8A Active CN120910810B (en) | 2025-10-10 | 2025-10-10 | Gravity Anomaly Inversion Method Based on PMU-Net Deep Learning Network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN120910810B (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119247489A (en) * | 2024-10-23 | 2025-01-03 | 中国地质大学（北京） | A deep learning network structure and method for magnetic anomaly inversion |
| CN120746833A (en) * | 2025-06-26 | 2025-10-03 | 中国人民解放军国防科技大学 | Submarine topography super-resolution reconstruction method based on window displacement multi-source fusion |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CA3122684C (en) * | 2018-12-11 | 2023-12-19 | Exxonmobil Upstream Research Company | Data augmentation for seismic interpretation systems and methods |
| WO2020123099A2 (en) * | 2018-12-11 | 2020-06-18 | Exxonmobil Upstream Research Company | Automated seismic interpretation-guided inversion |
| CN120448732B (en) * | 2025-04-25 | 2025-10-31 | 安徽省国防科技情报研究所 | Geological disaster automatic identification system and method based on multi-source remote sensing data |
| CN120409258A (en) * | 2025-04-28 | 2025-08-01 | 成都理工大学 | A gravity inversion method and system based on dual network drive |
-
2025
- 2025-10-10 CN CN202511439014.8A patent/CN120910810B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN119247489A (en) * | 2024-10-23 | 2025-01-03 | 中国地质大学（北京） | A deep learning network structure and method for magnetic anomaly inversion |
| CN120746833A (en) * | 2025-06-26 | 2025-10-03 | 中国人民解放军国防科技大学 | Submarine topography super-resolution reconstruction method based on window displacement multi-source fusion |
Also Published As
| Publication number | Publication date |
|---|---|
| CN120910810A (en) | 2025-11-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bi et al. | Deep relative geologic time: A deep learning method for simultaneously interpreting 3‐D seismic horizons and faults | |
| EP4010740B1 (en) | Petrophysical inversion with machine learning-based geologic priors | |
| CA3122488C (en) | Subsurface models with uncertainty quantification | |
| Bisdom et al. | An integrated workflow for stress and flow modelling using outcrop-derived discrete fracture networks | |
| EP3371629B1 (en) | Representing structural uncertainty in a mesh representing a geological environment | |
| CN114117906A (en) | A Multiscale Unsupervised Seismic Velocity Inversion Method Based on Self-encoding of Observation Data | |
| CN112731522B (en) | Intelligent recognition method, device and equipment for seismic stratum and storage medium | |
| EP4030198B1 (en) | Building accurate training images for automatic seismic interpretation | |
| Zhang et al. | SaltISCG: Interactive salt segmentation method based on CNN and graph cut | |
| CN118378319B (en) | Three-dimensional unit body attribute modeling method and system based on multi-source heterogeneous information fusion | |
| Wang et al. | Seismic stratum segmentation using an encoder–decoder convolutional neural network | |
| CN118112663B (en) | A method and system for rapid aviation transient electromagnetic imaging | |
| CN119986799A (en) | AI large model reservoir parameter prediction method based on 3D seismic subvolume data drive | |
| CN109154674A (en) | The displacement between seismic image is determined using light stream | |
| WO2024192162A1 (en) | Systems and methods for efficiently performing stochastic inversion methods in seismic exploration applications | |
| Lang et al. | Semi-supervised seismic impedance inversion with convolutional neural network and lightweight transformer | |
| CN117192643A (en) | Electrical and seismic joint inversion method and storage medium based on deep learning constraints | |
| Alzahrani et al. | MicroGraphNets: Automated characterization of the micro-scale wettability of porous media using graph neural networks | |
| CN119832179B (en) | Method and system for improving the accuracy of transparent geological model based on tunneling navigation data elements | |
| CN120910810B (en) | Gravity Anomaly Inversion Method Based on PMU-Net Deep Learning Network | |
| CN116413783A (en) | Seismic data interpretation method and device | |
| Lopez et al. | 2.5 D Hexahedral Meshing for Reservoir Simulations | |
| CN118151236A (en) | Seismic velocity deep learning inversion method and system based on data-space domain fusion | |
| CN117132735A (en) | A multi-source information geological surface reconstruction method and system | |
| CN112199859B (en) | Gravity gradient data joint inversion method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |