Geomagnetic field downward continuation method, system, device, and medium

Abstract

The application relates to a geomagnetic field downward continuation method, system, device and medium, the method comprising: acquiring vector geomagnetic field data at a reference height and a downward continuation distance value; inputting the vector geomagnetic field data at the reference height and the downward continuation distance value into a geomagnetic field downward continuation neural network model to obtain vector geomagnetic field data at a target height; wherein the height value of the target height is the difference between the height value of the reference height and the downward continuation distance value, the geomagnetic field downward continuation neural network model is obtained by training a training set, and each training sample in the training set comprises vector geomagnetic field data at a training reference height, a training downward continuation distance value and vector geomagnetic field data at a training target height as a supervision label. The geomagnetic field downward continuation method in the application can effectively improve the available distance of downward continuation, and supports flexible setting of the downward continuation distance at non-fixed point intervals within a preset continuation distance range.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geomagnetic field continuation, and in particular to a geomagnetic field downward continuation method, system, device and medium. BACKGROUND

[0002] As a passive navigation method that does not rely on satellite signals, geomagnetic navigation has received widespread attention in recent years in complex electromagnetic environments and high-security scenarios. In order to achieve high-precision geomagnetic matching navigation at different altitudes and in different regions, it is usually necessary to pre-construct a geomagnetic anomaly field database with high spatial resolution, wherein "downward continuation" of observation data obtained at a higher altitude or survey line position to a target altitude or near-space is one of the important means for constructing a three-dimensional geomagnetic anomaly field model and representing the distribution of the geomagnetic anomaly field at the target altitude.

[0003] In engineering applications, downward continuation methods based on mathematical models or data-driven models are usually used to process observation data to provide a basis for geomagnetic navigation, resource exploration, etc. However, in related technologies, the geomagnetic anomaly field downward continuation technology in the vector data scenario is still difficult to meet the application requirements of long-distance, non-fixed point distance downward continuation in engineering as a whole. SUMMARY

[0004] The present application provides a geomagnetic field downward continuation method, system, device and medium to solve some or all of the deficiencies in related technologies.

[0005] According to a first aspect of an embodiment of the present application, a geomagnetic field downward continuation method is provided, comprising:

[0006] Obtaining vector geomagnetic field data at a reference altitude and a downward continuation distance value; wherein the downward continuation distance value is any value within a preset continuation distance range, the lower limit of the continuation distance range is not less than 0 and the upper limit is not greater than the altitude value of the reference altitude;

[0007] Inputting the vector geomagnetic field data at the reference altitude and the downward continuation distance value into a geomagnetic field downward continuation neural network model to obtain vector geomagnetic field data at a target altitude; wherein the altitude value of the target altitude is the difference between the altitude value of the reference altitude and the downward continuation distance value, the geomagnetic field downward continuation neural network model is obtained by training a training set, and each training sample in the training set includes vector geomagnetic field data at a training reference altitude, a training downward continuation distance value and vector geomagnetic field data at a training target altitude as a supervision label.

[0008] Optionally, the vector geomagnetic field data at the training reference height of each training sample in the training set is the vector geomagnetic field data at a layered height value corresponding to the training reference height in the target vector geomagnetic field data set in the vector geomagnetic field true value library.

[0009] The vector geomagnetic field true value library comprises at least one vector geomagnetic field data set, each vector geomagnetic field data set corresponds to a vector geomagnetic field, and each vector geomagnetic field data set comprises vector geomagnetic field data at a plurality of layered height values, the plurality of layered height values cover a preset height range in the height direction, and the difference between any two adjacent layered height values is a preset height interval.

[0010] For each layered height value, the vector geomagnetic field data set comprises vector geomagnetic field values at a plurality of sampling positions in the target area, and the interval of the plurality of sampling positions in the horizontal direction is a preset horizontal point distance.

[0011] Optionally, the training downward extrapolation distance value is greater than 0 and not greater than the height value of the training reference height, and the height value of the training target height is the difference between the height value of the training reference height and the training downward extrapolation distance value.

[0012] The vector geomagnetic field data at the training target height is determined according to the vector geomagnetic field data at the first layered height and the vector geomagnetic field data at the second layered height.

[0013] The vector geomagnetic field data at the first layered height and the vector geomagnetic field data at the second layered height both belong to the target vector geomagnetic field data set.

[0014] The height value of the first layered height is less than the height value of the training target height and adjacent to the training target height in the height direction.

[0015] The height value of the second layered height is greater than the height value of the training target height and adjacent to the training target height in the height direction.

[0016] Optionally, the vector geomagnetic field data at the training target height is a linear combination of the vector geomagnetic field data at the first layered height and the vector geomagnetic field data at the second layered height, and the weight coefficient of the linear combination is determined according to the height difference between the training target height and the first layered height and the second layered height.

[0017] Optionally, the vector geomagnetic field data at each layered height value comprises an X component, a Y component and a Z component along three-dimensional space coordinate axes.

[0018] The vector geomagnetic field component at the training reference height of each training sample in the training set is one of the X component, the Y component and the Z component;

[0019] The downward continuation of the geomagnetic field neural network model is configured to downward continue the X component, the Y component and the Z component respectively.

[0020] Optionally, the vector geomagnetic field corresponding to at least one vector geomagnetic field data set in the vector geomagnetic field true value library is a vector geomagnetic field obtained through numerical simulation.

[0021] In the numerical simulation, the established vector geomagnetic field satisfies the following parameter conditions:

[0022] The number of magnetic sources arranged within the preset horizontal range is not less than a first number and not greater than a second number; wherein the first number is an integer not less than 2.

[0023] The magnetic source includes a basement type magnetic source and / or a ore body type magnetic source; wherein the burial depth of the basement type magnetic source is within a first depth range and the magnetization intensity is within a first intensity range, and the burial depth of the ore body type magnetic source is within a second depth range and the magnetization intensity is within a second intensity range.

[0024] The geomagnetic field parameters used in the numerical simulation include a magnetic inclination angle and / or a magnetic declination angle; wherein the value of the magnetic inclination angle is within a first angle range, and the value of the magnetic declination angle is within a second angle range.

[0025] Optionally, the downward continuation of the geomagnetic field neural network model is a deep neural network model, and the deep neural network model comprises:

[0026] An input layer, configured to receive the vector geomagnetic field data at the reference height and the downward continuation distance value;

[0027] A multi-stage encoder, an input end of which is connected to an output end of the input layer, configured to perform step-by-step encoding on the vector geomagnetic field data at the reference height and the downward continuation distance value;

[0028] A multi-stage decoder, an input end of which is connected to an output end of the multi-stage encoder, configured to perform step-by-step decoding on the encoded vector geomagnetic field data at the reference height and the encoded downward continuation distance value;

[0029] A jump connection structure, arranged between each encoding stage in the multi-stage encoder and a corresponding decoding stage in the multi-stage decoder, configured to connect the output of each encoding stage with the input of the corresponding decoding stage;

[0030] A receptive field enhancement module is arranged in at least one stage of the multi-stage encoder and / or the multi-stage decoder, and is configured to expand an effective receptive field of the deep neural network model for the vector geomagnetic field data.

[0031] According to a second aspect of the embodiments of the present application, a geomagnetic field downward continuation system is provided, comprising:

[0032] An input data acquisition module is configured to acquire vector geomagnetic field data at a reference height and a downward continuation distance value; wherein the downward continuation distance value is an arbitrary value in a preset continuation distance range, and a lower limit of the continuation distance range is not less than 0 and an upper limit of the continuation distance range is not greater than a height value of the reference height.

[0033] A vector geomagnetic field data acquisition module is configured to input the vector geomagnetic field data at the reference height and the downward continuation distance value into a geomagnetic field downward continuation neural network model to obtain vector geomagnetic field data at a target height; wherein a height value of the target height is a difference between the height value of the reference height and the downward continuation distance value, and the geomagnetic field downward continuation neural network model is obtained by training a training set, and each training sample in the training set comprises vector geomagnetic field data at a training reference height, a training downward continuation distance value, and vector geomagnetic field data at a training target height as a supervision label.

[0034] According to a third aspect of the embodiments of the present application, a computer program product is provided, which, when executed by a processor, performs the aforementioned geomagnetic field downward continuation method.

[0035] According to a fourth aspect of the embodiments of the present application, an electronic device is provided, comprising a processor and a memory, wherein the memory stores instructions, and when the processor executes the instructions, the electronic device performs the aforementioned geomagnetic field downward continuation method.

[0036] According to a fifth aspect of the embodiments of the present application, a non-transitory computer readable storage medium is provided, and the storage medium stores a computer program, and when the computer program is executed, the computer device performs the aforementioned geomagnetic field downward continuation method.

[0037] The technical solutions provided by the embodiments of the present application can include the following beneficial effects:

[0038] The geomagnetic field downward continuation method in the application inputs the vector geomagnetic field data at the reference height and the downward continuation distance value which can be freely set within the preset continuation distance range into the geomagnetic field downward continuation neural network model, so that the downward continuation distance value is input as an explicit condition of the model, supporting flexible adjustment of the downward continuation distance within the distance range from the input side; at the same time, the training set containing multiple groups of training vector geomagnetic field data at the reference height, different training downward continuation distance values and corresponding training vector geomagnetic field data at the target height is used to train the geomagnetic field downward continuation neural network model, so that the geomagnetic field downward continuation neural network model learns the corresponding relationship between "vector geomagnetic field data at the reference height" and "downward continuation distance value" and "vector geomagnetic field data at the target height" within the entire preset continuation distance range, and supports long-distance and non-fixed-point-distance downward continuation from the model capability.

[0039] Based on the above setting, the geomagnetic field downward continuation method in the application can output the vector geomagnetic field data at the target height for any downward continuation distance value within the preset continuation distance range, compared with the geomagnetic anomaly field downward continuation technology in the related art which mainly adopts fixed-point-distance or limited continuation distance, the application can effectively improve the available distance of downward continuation in the vector geomagnetic field scenario, and support flexible setting of the downward continuation distance in non-fixed-point-distance within the preset continuation distance range, so as to better meet the application requirement of long-distance and non-fixed-point-distance downward continuation in engineering. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0041] Figure 1 is a flowchart of a geomagnetic field downward continuation method according to an embodiment of the application;

[0042] Figure 2 is a structure diagram of a geomagnetic field downward continuation neural network model according to an embodiment of the application;

[0043] Figure 3 is Figure 2 is a structure diagram of a receptive field enhancement module in the geomagnetic field downward continuation neural network model shown in

[0044] Figures 4A-4F is a spatial distribution diagram of typical vector geomagnetic field data under different height and plane source number conditions obtained by numerical simulation according to an embodiment of the application;

[0045] Figure 5 is a schematic diagram of the relationship between different downward continuation distances and mean absolute error (MAE) according to an embodiment of the present application;

[0046] Figures 6A-6L is a comparison schematic diagram of the distribution of the Z component of the vector geomagnetic anomaly field data at different starting altitudes and before and after downward continuation by 1000 m according to an embodiment of the present application;

[0047] Figure 7 is a structural schematic block diagram of a geomagnetic field downward continuation system according to an embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated.

[0049] If the present application has involved directional indications or positional relationships in the embodiments, such as up, down, left, right, front, back, top, bottom, center, vertical, horizontal, longitudinal, transverse, etc., such terms are only used to explain the relative positional relationship, movement condition, etc. between the components in a certain posture; if the posture changes, the directional indications or positional relationships also change accordingly. In addition, the terms "first", "second", etc. in the embodiments of the present application are only used for convenience of description, and cannot be understood as indicating or implying relative importance.

[0050] As a new passive navigation method, geomagnetic navigation has important significance. Building a spatial geomagnetic database is the basis for realizing geomagnetic navigation, and potential field continuation is an effective method for solving the construction of the geomagnetic database. The spatial conversion method of potential field data includes upward continuation and downward continuation. The upward continuation method belongs to the Dirichlet problem in mathematics and is a well-posed problem, which has a stable solution. In actual calculation, as long as enough points are involved in the calculation, a relatively accurate result can be obtained. The downward continuation is an ill-posed problem and cannot be solved directly, and is sensitive to noise and observation error. Therefore, many scholars at home and abroad have proposed various solutions for downward continuation.

[0051] As early as in 1940s, Evjen used Taylor series method to analytically continue the potential field data, and then Peters derived the downward continuation formula using this method. In 1958, Dean first introduced the downward continuation problem in spatial domain, which cannot be solved directly, into frequency domain using Fourier transform, and derived the continuation operator in frequency domain, but this kind of method can only be used for downward continuation of about 2 times the point distance. In 2002, Fedi proposed the ISVD (integrated second vertical derivative) method for downward continuation, which uses different solving methods for odd and even order vertical derivatives when Taylor series expansion is performed: the odd order is first solved by vertical integration, and then solved by Laplace equation, while the even order is directly solved by Laplace equation. This method can stably realize downward continuation of about 5 times the point distance.

[0052] In engineering applications, Xu Shizhe proposed the integral-iterative potential field continuation method. This method vertically projects the measured potential field values on the relief surface to the lower part of a horizontal surface as the initial value of the potential field on the horizontal surface. Then, according to the initial value on the horizontal surface, the integral method is used to calculate the potential field value on the relief surface, and the difference between the measured value and the calculated value on the relief surface is used to correct the potential field value on the horizontal surface. This iteration is repeated until the difference between the measured value and the calculated value on the relief surface is small enough to be ignored. With the potential field value on the horizontal surface, the integral method or other methods can be used to calculate the potential field value on any curved surface or horizontal surface above the horizontal surface. This method converts the unstable downward continuation problem into a relatively stable upward continuation problem, improves the maximum depth of continuation, and effectively extends the depth of continuation to 20 times the data point distance. It is a kind of practical method commonly used in large-depth downward continuation of potential field data. The principle of this method is relatively simple, it does not need to solve linear algebraic equations, and it has high calculation speed and can effectively continue 20 times the point distance. However, in large-scale applications, multiple iterations will affect the calculation speed, too few iterations will affect the continuation accuracy, and the iteration process will amplify the noise signal to some extent, thus limiting the use of this method to some extent.

[0053] To solve the problems of unstable solution, difficult parameter selection, and rapid amplification of high-frequency signals to produce high-frequency noise in the downward continuation of geomagnetic anomaly data, Ge et al. proposed a geomagnetic anomaly data stable downward continuation method based on the SRGAN (Super-Resolution Generative Adversarial Network, single image super-resolution generative adversarial network) deep learning network. This method constructs a geomagnetic anomaly data downward continuation network structure based on SRGAN, combines CBAM (Convolutional Block Attention Module, convolutional block attention module) spatial attention mechanism and adaptive Bayesian threshold multi-wavelet transform to comprehensively improve the continuation accuracy of the network, and proposes to use Kriging super-resolution algorithm and sliding sub-region extraction technology to make training data set, which can effectively improve the learning effect of the network on the detail features of the magnetic anomaly. This kind of geomagnetic anomaly data downward continuation method based on deep learning network has certain advantages in suppressing noise amplification and improving continuation results, but the published method only supports integer multiple point distance continuation, cannot support arbitrary distance continuation, has poor flexibility, and has low application value; at the same time, the continuation point distance in the experiment is about 9 times, and the effective continuation distance is limited.

[0054] In summary, the continuation distance of the above related methods is not more than 20 times the point distance, and there is still a lack of stable continuation scheme in long-distance downward continuation, and some related methods based on deep learning network only support integer multiple point distance, cannot support arbitrary distance downward continuation, have insufficient flexibility, and the application range is limited.

[0055] The present application provides a geomagnetic field downward continuation method, system, device and medium. The geomagnetic field downward continuation method of the present application will be described in detail below in conjunction with the drawings. In the case of no conflict, the embodiments and features in the embodiments described below can be combined with each other.

[0056] Figure 1 The present application illustrates a geomagnetic field downward continuation method according to an exemplary embodiment. As shown in Figure 1 The geomagnetic field downward continuation method includes the following steps:

[0057] S101, obtaining vector geomagnetic field data at a reference height and a downward continuation distance value.

[0058] The downward extrapolation distance value is any value in a preset extrapolation distance range, the lower limit of the extrapolation distance range is not less than 0, and the upper limit is not greater than the height value of the reference height. In this application, the reference height is used to represent the height position at which the vector geomagnetic field data is obtained, and can be a certain fixed height layer above the ground, or an actual height layer of a flight platform or an observation platform, and is the starting height when downward extrapolation is performed. The downward extrapolation distance value is used to represent the distance downward from the reference height in the vertical direction, and is a scalar parameter. By subtracting the distance value from the height value of the reference height, the height value of the target height can be determined, and thus the height position to which the vector geomagnetic field data needs to be extrapolated can be determined. The preset extrapolation distance range is used to limit the value range of the downward extrapolation distance value, and the lower limit is not less than 0 and the upper limit is not greater than the height value of the reference height, so that the downward extrapolation distance value can be arbitrarily selected within the range, thereby ensuring that the height value of the target height is not higher than the reference height and falls within the actual required spatial range. On the one hand, upward extrapolation or exceeding the effective height range is avoided, and on the other hand, the downward extrapolation distance can be flexibly set within the distance range during the training and application stages, so as to support the downward extrapolation requirements between different heights.

[0059] S102, input the vector geomagnetic field data at the reference height and the downward extrapolation distance value into the geomagnetic field downward extrapolation neural network model to obtain the vector geomagnetic field data at the target height.

[0060] The height value of the target height is the difference between the height value of the reference height and the downward extrapolation distance value, and the geomagnetic field downward extrapolation neural network model is obtained by training a training set. Each training sample in the training set includes vector geomagnetic field data at a training reference height, a training downward extrapolation distance value, and vector geomagnetic field data at a training target height as a supervision label.

[0061] In the present application, the vector geomagnetic field data at the training reference height, the training downward extrapolation distance value, and the vector geomagnetic field data at the training target height correspond to the values of the vector geomagnetic field data at the reference height, the downward extrapolation distance value, and the vector geomagnetic field data at the target height in the training stage, respectively. Specifically, the vector geomagnetic field data at the training reference height is used to represent the vector geomagnetic field sample at a certain reference height selected when the training set is constructed, and is one of the input data of the geomagnetic field downward extrapolation neural network model in the training stage. The training downward extrapolation distance value is used to represent the distance downwardly extrapolated along the vertical direction with respect to the training reference height, and its value is located within the preset training extrapolation distance range. By subtracting the training downward extrapolation distance value from the height value of the training reference height, the height value of the training target height can be determined. The preset training extrapolation distance range is used to limit the value interval of the training downward extrapolation distance value in the training stage, and the lower limit thereof is not less than 0, and the upper limit thereof is not greater than the height value of the corresponding training reference height. The preset extrapolation distance range can be set as all or part of the aforementioned preset extrapolation distance range, so as to cover the expected downward extrapolation distance distribution in the training process. The vector geomagnetic field data at the training target height is used to represent the vector geomagnetic field distribution at the aforementioned training target height, and is the supervision label matched with the vector geomagnetic field data at the corresponding training reference height and the training downward extrapolation distance value. By combining the vector geomagnetic field data at different training reference heights, the training downward extrapolation distance values, and the corresponding vector geomagnetic field data at the training target height in the training set construction process, the geomagnetic field downward extrapolation neural network model can learn the relationship from the reference height and the downward extrapolation distance value to the vector geomagnetic field data at the target height in the training stage, so as to output the corresponding vector geomagnetic field data at the target height according to the actual vector geomagnetic field data at the reference height and the downward extrapolation distance value in the application stage. The geomagnetic field downward extrapolation method of the present application inputs the vector geomagnetic field data at the reference height and the downward extrapolation distance value freely settable within the preset extrapolation distance range into the geomagnetic field downward extrapolation neural network model, so that the downward extrapolation distance value is input as an explicit condition of the model, which supports flexible adjustment of the downward extrapolation distance within the distance range from the input side. At the same time, the training set containing multiple groups of vector geomagnetic field data at the training reference height, different training downward extrapolation distance values, and the corresponding vector geomagnetic field data at the training target height is used to train the geomagnetic field downward extrapolation neural network model, so that the geomagnetic field downward extrapolation neural network model learns the corresponding relationship between the "vector geomagnetic field data at the reference height" and the "downward extrapolation distance value" and the "vector geomagnetic field data at the target height" within the entire preset extrapolation distance range, and supports long-distance and non-fixed-point downward extrapolation from the model capability.

[0062] Based on the above settings, the geomagnetic field downward extension method in this application can output vector geomagnetic field data at the target height for any downward extension distance value within a preset extension distance range. Compared with related technologies that mostly use geomagnetic anomaly field downward extension techniques with fixed point spacing or limited extension distance, this method can effectively improve the available downward extension distance in vector geomagnetic field scenarios. It also supports flexibly setting the downward extension distance within a preset extension distance range according to non-fixed point spacing, so as to better meet the application needs of long-distance, non-fixed point spacing downward extension in engineering.

[0063] In one optional embodiment, the geomagnetic field downward extension neural network model is a deep neural network model, including: an input layer, a multi-level encoder, a multi-level decoder, a jump connection structure, and a receptive field enhancement module, used to perform feature encoding and decoding on the input reference height magnetic anomaly field and downward extension distance, and output the magnetic anomaly field at the target height.

[0064] The input layer receives vector geomagnetic field data at a reference altitude and downward extension distance values. In the specific implementation, it combines... Figure 2 As shown, the input layer can be implemented by the initial feature extraction layer: the magnetic anomaly field at the reference height and the downward extension distance are input into the network, fused by the initial feature extraction layer and the initial feature extraction to obtain a size of... The feature map is defined as follows, where H and W represent the number of grid rows and columns in the horizontal and vertical directions, respectively, and 32 represents the number of feature channels at each grid position.

[0065] The input of the multi-stage encoder is connected to the output of the input layer, and it is used to encode the vector geomagnetic field data at the reference height and the downward extension distance values ​​step by step. Combined with... Figure 2 As shown, the multi-level encoder consists of four encoding stages and a deepest intermediate stage. The four encoding stages are, in order, the first, second, third, and fourth encoding stages. To progressively reduce spatial resolution and expand the number of feature channels at each encoding stage, a downsampling module can be set between each encoding stage to perform spatial downsampling and channel number transformation on the feature map. Simultaneously, to expand the effective receptive field and enhance feature extraction capabilities, a receptive field enhancement module can be arranged in each encoding stage of the multi-level encoder. This module is inserted as an independent functional module into the corresponding encoding stage to extract feature information related to the magnetic anomaly field and downward extension distance at the current scale. In this embodiment, as... Figure 2 As shown, one receptive field enhancement module is inserted in each of the first, second, and third encoding stages, and eight receptive field enhancement modules are inserted in series in the fourth encoding stage. After the first encoding stage, spatial downsampling and channel expansion are performed through a downsampling module, reducing the spatial size of the feature map to its original size in both the row and column directions. , the number of channels is expanded to 2 times of the original, and a feature map with a size of is obtained. After feature extraction by the second encoding stage, the feature map is again passed through the downsampling module to obtain a feature map with a size of . Similarly, after feature extraction by the third encoding stage, the feature map is passed through the downsampling module to obtain a feature map with a size of . After deep feature extraction by the multiple receptive field enhancement modules in the fourth encoding stage, the spatial size is further reduced and the number of channels is further expanded by the downsampling module to obtain a feature map with a size of . In the above multi-level encoder, , , , and , , , represent the step-by-step downsampling in the row and column directions, and 32, 64, 128, 256, 512 represent the number of feature channels at different encoding layers. An intermediate stage is set at the deepest layer of the multi-level encoder, which also includes a receptive field enhancement module for further optimizing and integrating high-level features on a feature map with a size of . It can be regarded as a centralized enhancement of the output features of the multi-level encoder, and the output of the intermediate stage is used as the input of the multi-level decoder.

[0066] The input end of the multi-level decoder is connected to the output end of the multi-level encoder, and is used to decode the vector geomagnetic field data at the reference height and the downward continuation distance value step by step. As shown in Figure 2 , the multi-level decoder is composed of four decoding stages, which are the first decoding stage, the second decoding stage, the third decoding stage and the fourth decoding stage. In order to recover the spatial resolution and compress the number of feature channels step by step in the decoding stage, an upsampling module can be set between each decoding stage to perform spatial upsampling and channel number transformation on the feature map. At the same time, in order to enhance the feature fusion and reconstruction capability in the decoding process, a receptive field enhancement module can be inserted into part or all of the decoding stages of the multi-level decoder, and the receptive field enhancement module is embedded as an independent functional unit into the corresponding decoding stage. First, the feature with a size of output by the intermediate stage is 2 times upsampled in the spatial dimension by the upsampling module, and the number of channels is compressed to in the channel dimension, to obtain a feature map with a size of , and input into the first decoding stage for feature fusion and optimization. Then, the feature map with a size of is obtained by the upsampling module, and input into the second decoding stage. The feature map with a size of is obtained by further upsampling, and input into the third decoding stage. The feature map with a size of The feature map is input to the fourth decoding stage. Each decoding stage also includes a receptive field enhancement module to fuse and enhance features while progressively restoring spatial resolution. To facilitate the conversion of the feature map recovered by the multi-level decoder into a magnetic anomaly field at the target height, an output mapping layer is provided at the output of the multi-level decoder in this embodiment. This output mapping layer can be implemented using a fully connected layer structure and is used to determine the size of the output from the fourth decoding stage. The feature map is mapped point by point to predict the vector geomagnetic field data at the target altitude.

[0067] A jump-connection structure is placed between each encoding stage in a multi-level encoder and the corresponding decoding stage in a multi-level decoder, used to connect the output of each encoding stage to the input of the corresponding decoding stage. Figure 2 As shown, the skip connection structure is specifically manifested in the dashed connections between the first encoding stage and the fourth decoding stage, the second encoding stage and the third decoding stage, the third encoding stage and the second decoding stage, and the fourth encoding stage and the first decoding stage. The skip connection structure connects and fuses the output features of the corresponding encoding stage with the upsampled features of the decoding stage, enabling the multi-level decoder to simultaneously utilize high-level abstract information and shallow-level detail information when recovering the magnetic anomaly field distribution of the target height.

[0068] The receptive field enhancement module, as an independent functional module distinct from multi-level encoders, multi-level decoders, and skip-connection structures, can be structurally embedded into each encoding and / or decoding stage. In this embodiment, one receptive field enhancement module is embedded in each of the first, second, and third encoding stages, eight receptive field enhancement modules are embedded in series in the fourth encoding stage, and one receptive field enhancement module is also embedded in each of the first to fourth decoding stages, to enhance feature extraction capabilities and expand the effective receptive field at different scales. Figure 3 It shows Figure 2 The internal structure of the receptive field enhancement module in the code is as follows: This module starts with the input features and contains two main branches and one direct jump connection. It expands the effective receptive field and enhances important features through multiple convolutions and gating operations. On the left main branch, the input features are first normalized, and then sequentially processed through convolutions and gating operations. depth Separable convolutions are used for channel transformation and local spatial feature extraction. Then, non-linear gating is performed through a gating unit structure. Next, the spatial-channel attention mechanism is introduced via the SCA (Spatial-Channel Attention) module. Finally, [the process is described in the original text]. Convolution integrates the features; on the right branch, the input features are also normalized by layers and then passed sequentially through two layers. The cavity convolution and the intermediate gating unit complementarily extract features in different directions and scales. The receptive field enhancement module internally forms a short circuit path through the left and right branches and the weighted summation node at the bottom. The input features and the features processed by each branch are fused by weighted summation at the "⊕" node, and the features of different scales and different processing paths are integrated. Through the combination of the above multi-branch convolution, gating, attention, and weighted fusion structure, the receptive field enhancement module expands the effective receptive field of the deep neural network for magnetic anomaly features without significantly increasing the computational amount, and improves the response capability to key details, thereby supporting the aforementioned downward continuation of the geomagnetic field neural network model in long-distance downward continuation.

[0069] By using the above-mentioned geomagnetic field downward continuation neural network model including an input layer, a multi-level encoder, a multi-level decoder, a skip connection structure, and a receptive field enhancement module, the vector geomagnetic field data at the reference height and the downward continuation distance value are jointly used as inputs, different scale spatial features are extracted step by step in the encoding stage, and the spatial resolution is recovered step by step in the decoding stage. At the same time, the shallow detail information is retained by the skip connection structure, and the effective receptive field of the reference height vector geomagnetic field data is expanded by the receptive field enhancement module, and the response to key structures is strengthened, so that the network can learn the complex nonlinear mapping relationship between the reference height and the target height in a larger spatial range. In the vector geomagnetic field downward continuation application scenario of the present application, this structure is beneficial to stably output the vector geomagnetic field data at the target height under the condition of a given reference height and downward continuation distance value, improves the result stability and detail retention capability in long-distance downward continuation, weakens the amplification effect of observation noise in the continuation process, and supports any downward continuation distance within a preset continuation distance range in the same network. Therefore, the comprehensive requirements of continuation distance, precision and flexibility for constructing high-precision vector geomagnetic field three-dimensional database and serving geomagnetic navigation and other engineering applications are better met.

[0070] In an optional embodiment, the vector geomagnetic field data at the training reference height of each training sample in the training set is the vector geomagnetic field data at the layered height value corresponding to the training reference height in the target vector geomagnetic field data set in the vector geomagnetic field true value library.

[0071] The vector geomagnetic field true value library includes at least one vector geomagnetic field data set, each vector geomagnetic field data set corresponds to a vector geomagnetic field; each vector geomagnetic field data set includes vector geomagnetic field data at a plurality of layered height values, the plurality of layered height values cover a preset height range in the height direction, and the difference between any two adjacent layered height values is a preset height interval. For each layered height value, the vector geomagnetic field data set includes vector geomagnetic field values at a plurality of sampling positions in the target area, and the spacing of the plurality of sampling positions in the horizontal direction is a preset horizontal point distance.

[0072] In the present application, the layered height value is used to represent the height corresponding to each height layer when the vector geomagnetic field is sampled in the height direction, and the set thereof gives the value of the vector geomagnetic field at multiple discrete height layers. The preset height range is used to represent the overall height interval covered by the above-mentioned multiple layered height values in the height direction, for example, it can extend from the near-earth height to a certain height in the upper air, for limiting the effective distribution range of the vector geomagnetic field data in the height direction in the true value library. The preset height interval is used to represent the height difference between any two adjacent layered height values, which is usually set to a fixed value in a vector geomagnetic field data set, so as to form multiple height layers with equal intervals in the preset height range, for example, a height interval of 100m is uniformly used. In different vector geomagnetic field data sets or different application scenarios, the preset height interval can also be set to different values according to the needs, so as to balance the data amount and the calculation overhead while ensuring the height resolution. The target region is used to represent the spatial region in the horizontal direction for sampling the vector geomagnetic field, which can be a given geographical region or a navigation working area, and a plurality of sampling positions are arranged in the region according to a preset horizontal point distance, so as to obtain the vector geomagnetic field values distributed at multiple points in the target region at each layered height value.

[0073] In the present embodiment, in order to discretely represent the vector geomagnetic field in the target region, the vector geomagnetic field data set can be grid set in the height direction and the horizontal direction. Specifically, each vector geomagnetic field data set includes vector geomagnetic field data at multiple layered height values, the multiple layered height values cover a preset height range in the height direction, and the difference between any two adjacent layered height values is a preset height interval, thereby forming a series of discrete layered height values in the height direction, for describing the vector geomagnetic field distribution at different height layers.

[0074] For each layered height value, the vector geomagnetic field data set includes vector geomagnetic field values at multiple sampling positions in the target region, and the spacing of the multiple sampling positions in the horizontal direction is a preset horizontal point distance. Specifically, grid lines can be arranged in the east-west direction and the north-south direction of the target region according to the preset horizontal point distance, and the intersection points of the grid lines constitute the multiple sampling positions, each sampling position corresponds to a spatial position in the target region, for recording the vector geomagnetic field values at different layered height values at the position. Through the combination of multiple layered height values, the preset height interval, and multiple sampling positions and the preset horizontal point distance, the vector geomagnetic field true value distribution in the target region can be discretely represented in three-dimensional space.

[0075] In specific applications, the preset height range, the preset height interval, the horizontal range of the target region, and the preset horizontal point distance can be set according to the size of the survey area and the required spatial resolution. For example, a preset height range of several kilometers and a preset height interval of several hundred meters can be selected, and a target region horizontal range of several kilometers and a preset horizontal point distance of several meters can be selected. The present application is not limited to a specific numerical combination.

[0076] The vector geomagnetic field true value library is constructed in the above manner, and the vector geomagnetic field data at the training reference height is extracted therefrom, so that the training samples are all derived from the vector geomagnetic field data set obtained based on numerical simulation or other means, sampled at layered height values in the height direction, and sampled at a preset horizontal point distance in the horizontal direction. On the one hand, the plurality of layered height values cover the preset height range in the height direction and adopt the preset height interval, which is conducive to the geomagnetic field downward continuation neural network model learning the rules of the variation of the vector geomagnetic field with height and the correlation between different height layers at the same time in the training stage; on the other hand, a plurality of sampling positions are arranged in the target region according to the preset horizontal point distance, which can reflect the detailed changes of the vector geomagnetic field in the horizontal plane while ensuring the spatial resolution. Based on the three-dimensional vector geomagnetic field true value data regularly sampled in the height direction and the horizontal direction, the training set is constructed, which is helpful to improve the generalization ability and stability of the model at different heights and different spatial positions, and more accurately reconstruct the vector geomagnetic field distribution at the target height when long-distance downward continuation, thereby providing a reliable data basis for subsequent construction of a high-precision vector geomagnetic field database.

[0077] In order to enable the geomagnetic field downward continuation neural network model to adapt to different combinations of reference height and downward continuation distance, a random height and a random downward continuation distance can be used to obtain diversified data distribution when constructing the training set. Specifically, a plurality of training reference height values are randomly selected within the preset training reference height range, and a training downward continuation distance value is randomly sampled within the available continuation distance range corresponding to each training reference height. The vector geomagnetic field data at the training reference height randomly selected is combined with the training downward continuation distance value, and the vector geomagnetic field data at the corresponding training target height is used as a supervision label to form a plurality of training samples. Through the above method of randomly combining the reference height and the downward continuation distance to construct the training samples, a more rich height combination and continuation distance distribution can be covered within the entire preset continuation distance range, so that the geomagnetic field downward continuation neural network model fully learns the mapping relationship between different heights and different downward continuation distances in the training process, thereby better supporting the setting of any downward continuation distance and improving the generalization ability of the model in actual application.

[0078] Further, in order to improve the robustness and universality of the geomagnetic field downward continuation neural network model to observation noise and data disturbance, random noise can be superimposed on the input magnetic field data during training for data enhancement. Specifically, Gaussian noise with a magnitude within a predetermined range can be randomly added to the input magnetic field component data during the training phase, for example, the noise amplitude can be selected as a relative disturbance of 0% to 2%, so that the model can still learn the downward continuation mapping relationship stably under certain noise interference.

[0079] In terms of training parameter settings, the AdamW (Adam with Decoupled Weight Decay) optimizer can be used to update the network parameters; the initial learning rate can be set to 0.001, and a polynomial decay strategy can be used to gradually reduce the learning rate; the training iteration number can be set to 600,000. The above noise amplitude, optimizer type, learning rate value and its decay strategy, and training iteration number are all adjustable according to specific application requirements, and are not limited to the above values.

[0080] In an optional embodiment, the training downward continuation distance value is greater than 0 and not greater than the training reference height value, and the training target height value is the difference between the training reference height value and the training downward continuation distance value. The vector geomagnetic field data at the training target height is determined according to the vector geomagnetic field data at the first layer height and the vector geomagnetic field data at the second layer height. The vector geomagnetic field data at the first layer height and the vector geomagnetic field data at the second layer height both belong to the target vector geomagnetic field data set. The first layer height value is less than the training target height value and adjacent to the training target height in the height direction. The second layer height value is greater than the training target height value and adjacent to the training target height in the height direction.

[0081] When the training target height is located between two adjacent layer heights, a dynamic interpolation method can be used to obtain the vector geomagnetic field data at the training target height by linear combination of the vector geomagnetic field data at the first layer height and the vector geomagnetic field data at the second layer height. Specifically, the vector geomagnetic field data at the training target height can be determined by the following formula:

[0082]

[0083] wherein, is the vector geomagnetic field data at the first layer height, is the vector geomagnetic field data at the second layer height, is the vector geomagnetic field data at the training target height, is the first weight coefficient, is a second weight coefficient, satisfying .

[0084] By limiting the training downward extrapolation distance value to be greater than 0 and not greater than the height value of the training reference height, upward extrapolation or zero-distance degenerate samples can be avoided, the training target height is always located in the effective height range below the training reference height, and the height value of the training target height is uniquely determined by the difference between the training reference height and the training downward extrapolation distance value. On this basis, the training target height is limited between two adjacent layered heights in the height direction in the target vector geomagnetic field data set, and the vector geomagnetic field data at the training target height is determined using the vector geomagnetic field data at the first layered height and the second layered height. This is conducive to constructing supervised label samples at intermediate heights based on existing layered height values, so that the training samples are expanded from discrete layers to dense samples covering continuous height intervals in the height direction. On the one hand, it can make full use of the data at adjacent layered heights in the vector geomagnetic field true value library to improve the use efficiency of training data, and on the other hand, it is helpful for the geomagnetic field downward extrapolation neural network model to learn the smooth change rule and interpolation relationship between different height layers, so as to more accurately estimate the vector geomagnetic field data at the target height between the preset layered heights when facing any set downward extrapolation distance value in the application stage, and improve the accuracy and stability of the downward extrapolation at any distance.

[0085] Further, the weight coefficients of the linear combination and are determined according to the height difference between the training target height and the first layered height and the second layered height. The weight coefficients and are used to reflect the relative closeness of the training target height to the first layered height and the second layered height, and the corresponding weight size can be determined according to the height difference between the training target height and the two layered heights, and the weight can be normalized as needed, so that the sum of and is a constant. By setting the weight based on the height difference, the layered height closer to the training target height occupies a larger weight in the linear combination, and the layered height farther away occupies a smaller weight, so that a smooth interpolation conforming to the height change rule is formed between the vector geomagnetic field data at the first layered height and the second layered height. The vector geomagnetic field data at the training target height determined in this way can better reflect the continuous change characteristics of the vector geomagnetic field in the height direction, which is conducive to improving the learning effect of the geomagnetic field downward extrapolation neural network model on the transition characteristics between different heights, and reducing the interpolation error of the extrapolation result at any target height. Specifically, the weight coefficients of the linear combination can be determined by the following formula:

[0086]

[0087]

[0088] in, As the first weighting coefficient, This is the second weighting coefficient. The distance from the second layer height to the reference height. To train the distance from the target height to the reference height, This is the distance from the first layer height to the reference height. The difference between any two adjacent layer height values ​​is the preset height interval.

[0089] Under the above weighting settings, the closer the training target height is to the first layer height, ( The larger the number of ) The larger the relative height; the closer the training target height is to the second layer height, ( The larger the number of ) The larger the relative size.

[0090] When the difference between any two adjacent layer height values ​​is a preset height interval of 100 meters, that is, the distance between the first layer height and the second layer height is 100 meters, the vector geomagnetic field data at the training target height can be determined by the following formula:

[0091]

[0092] in, This is the vector geomagnetic field data at the first layer height. This is the vector geomagnetic field data at the second layer height. To train vector geomagnetic field data at the target altitude, The distance from the second layer height to the reference height. To train the distance from the target height to the reference height, This is the distance from the first layer height to the reference height.

[0093] It should be noted that the above inverse distance weighting is only one example of how to determine the linear combination weight coefficients based on the height difference. In other implementations, other functional relationships related to the height difference can also be used to determine the weights. and As long as the height difference between the training target height and a certain layer height is smaller, the weight coefficient of the vector geomagnetic field data at that layer height in the linear combination will be larger.

[0094] In an optional embodiment, the vector geomagnetic field data at each layer height value includes an X component, a Y component and a Z component along three-dimensional spatial coordinate axes. The vector geomagnetic field component at the training reference height of each training sample in the training set is one of the X component, the Y component and the Z component. The geomagnetic field downward continuation neural network model is configured to respectively perform downward continuation on the X component, the Y component and the Z component.

[0095] In order to enable the geomagnetic field downward continuation neural network model to respectively perform downward continuation on the X, when constructing the training set, the three-component corresponding magnetic field distribution is split into three scalar magnetic field data sets, and one of the X component, the Y component and the Z component is randomly selected each time a training sample is generated, the magnetic field distribution of the component at the training reference height and the training downward continuation distance value are taken as the model input, and the magnetic field distribution of the same component at the training target height is taken as the supervision label. By mixing training on the above randomly selected single-component magnetic field data, the geomagnetic field downward continuation neural network model can learn the mapping relationship of respectively performing downward continuation on different component magnetic field data under the premise of sharing a set of network parameters, so that the same model can output the downward continuation result of the corresponding component at the target height when receiving the X component, the Y component or the Z component as input. In order to obtain stable continuation results, when constructing the training set, the X component, the Y component and the Z component at the same spatial position are respectively sampled, so that the three components take turns as input components to participate in training in different training samples, which is equivalent to a kind of data enhancement processing on the component dimension. On the one hand, modeling respectively according to the components can avoid the interference caused by mixing multiple components in a single training sample, so that the network can more accurately learn the continuation law of the change of each component with height; on the other hand, training the X component, the Y component and the Z component under the same network structure and sharing parameters is beneficial to improving the generalization ability of the model to different components and different spatial positions, and enhancing the robustness to observation noise and data difference. In the application stage, the network respectively performs downward continuation on the X component, the Y component and the Z component, and combines the three components to form complete vector geomagnetic field data at the target height, so as to obtain more stable and reliable continuation results in the vector geomagnetic field downward continuation scene.

[0096] In an optional embodiment, the vector geomagnetic field corresponding to at least one vector geomagnetic field data set in the vector geomagnetic field true value library is a vector geomagnetic field obtained through numerical simulation. Numerical simulation is used to divide the space into grids according to the height value and horizontal point distance under the condition of given magnetic body distribution and geomagnetic field parameters, and to calculate the numerical value of the vector geomagnetic field at each grid sampling position, thereby obtaining the vector geomagnetic field data at multiple height layers and multiple sampling positions. In specific implementation, numerical simulation can be implemented based on existing three-dimensional magnetic field forward calculation programs, for example, using MATLAB software, Python combined with numerical calculation library, or COMSOL general electromagnetic simulation software, under the preset magnetic source parameters and geomagnetic field parameters, the space grid divided according to the height value and horizontal point distance is subjected to magnetic field forward calculation, and the vector geomagnetic field data set directly used to fill the vector geomagnetic field true value library is generated. The software platform and implementation mode of the above numerical simulation are optional examples, and those skilled in the art can implement them according to needs using other software or self-developed programs with similar numerical calculation capabilities. The vector geomagnetic field data is used to reflect the fluctuation and uneven distribution of the geomagnetic field at different spatial positions. This part of the local feature changing with the spatial position is usually called geomagnetic anomaly, so the corresponding data can also be called vector geomagnetic anomaly field data. Hereinafter, the vector geomagnetic field and the vector geomagnetic anomaly field can be understood as equivalent without causing ambiguity. The vector geomagnetic field established in numerical simulation satisfies the following parameter conditions:

[0097] (1) The number of magnetic sources arranged in the preset horizontal range is not less than a first number and not greater than a second number. The first number is an integer not less than 2. The preset horizontal range is used to represent the area in the horizontal direction for magnetic source arrangement and vector geomagnetic field calculation, for example, it can be set as a rectangular area of several kilometers square to limit the plane range of numerical simulation and sampling point distribution.

[0098] In specific implementation, the preset horizontal range can be set as 30km×30km, a plurality of magnetic sources are arranged in the horizontal range, and the number of magnetic sources is set as 2-20, so as to form rich vector geomagnetic anomaly field combinations in the same spatial range through the number change. At the same time, in order to match the commonly used flight height of navigation, 61 degrees layers can be arranged in the height direction in the range of 0-6km with a height interval of 100m, the vector geomagnetic field at multiple height layers in the preset horizontal range is sampled, the vector geomagnetic field data obtained at each height layer is combined to form a vector geomagnetic field data set, which is used to represent the vector geomagnetic field distribution in the height range of 0-6km in the preset horizontal range.

[0099] (2) The magnetic source includes a base-type magnetic source and / or a ore-body-type magnetic source. The base-type magnetic source has a burial depth within a first depth range and a magnetization intensity within a first intensity range, while the ore-body-type magnetic source has a burial depth within a second depth range and a magnetization intensity within a second intensity range. The first and second depth ranges are used to define the burial depth ranges for different types of magnetic sources, and the first and second intensity ranges are used to define the magnetization intensity ranges for different types of magnetic sources. These ranges can be set according to specific simulation requirements, and can differ to reflect differences such as weak magnetic fields at depth and strong magnetic fields at shallow depth, or they can overlap or have the same value in some ranges. This application does not impose any limitations on this.

[0100] In practical implementation, a base-type magnetic source can simulate deep, large-scale, low-magnetization structures. Its burial depth and magnetization intensity fall within a first depth range and a first intensity range, respectively. For example, a base-type magnetic source with a burial depth of 5–10 km corresponds to a first depth range of 5–10 km, and a magnetization intensity of 1–5 A / m corresponds to a first intensity range of 1–5 A / m. Its size can be on the order of kilometers. A orebody-type magnetic source can simulate shallow, small-scale, highly magnetized orebodies. Its burial depth and magnetization intensity fall within a second depth range and a second intensity range, respectively. For example, a orebody-type magnetic source with a burial depth of 100–1000 m corresponds to a second depth range of 100–1000 m, and a magnetization intensity of 10–50 A / m corresponds to a second intensity range of 10–50 A / m. Its size can be on the order of hundreds of meters. By combining different parameter values ​​of base-type and orebody-type magnetic sources in the same numerical simulation, a vector geomagnetic anomaly field distribution that more closely resembles actual geological conditions can be obtained.

[0101] (3) The geomagnetic field parameters used in the numerical simulation include magnetic inclination and / or magnetic declination. The magnetic inclination value is within a first angular range, and the magnetic declination value is within a second angular range. The first and second angular ranges are used to define the value intervals of the magnetic inclination and magnetic declination in the numerical simulation. These ranges can be set according to the geomagnetic field distribution of the target region, for example, selecting a typical magnetic inclination range covering mid-latitude regions and a magnetic declination range covering different tectonic orientations. The two angular ranges can be different or overlap in some intervals; this application does not impose any limitations on this.

[0102] In practice, the magnetic inclination is taken within a first angular range, for example, 40° to 65°, to simulate the geomagnetic field characteristics of mid-latitude regions. The magnetic declination is taken within a second angular range, for example, covering 0° to 360°, thereby introducing geomagnetic field backgrounds from different directions into the numerical simulation to reflect various vector geomagnetic field morphologies under complex geological structures.

[0103] By the above method, the number of magnetic sources in the numerical simulation is limited to be between the first number and the second number, and the base type magnetic source and the ore body type magnetic source are arranged comprehensively within the preset horizontal range, while different burial depths and magnetization strengths are set in the first depth range, the second depth range, and the first intensity range, the second intensity range, respectively, and then the magnetic inclination and the magnetic declination in the first angle range and the second angle range are changed, so that the vector geomagnetic field data under the conditions of superposition of various magnetic source types, combination of different burial depths and strengths, and different geomagnetic field parameters can be constructed in the same numerical simulation framework. Compared with the simulation field with single magnetic source type and idealized parameter setting commonly used in existing research, the vector geomagnetic field obtained by the above multi-parameter joint design is closer to the actual geomagnetic field characteristics at the navigation flight height, and can form a vector geomagnetic anomaly field data with diverse distribution and rich shape in the vector geomagnetic field true value library, which provides training samples covering various spatial configurations and parameter combinations for the geomagnetic field downward continuation neural network model, and is beneficial to improve the generalization ability and stability of the model under complex scenes and different parameter conditions, so as to obtain more reliable downward continuation results in actual navigation application.

[0104] In an exemplary embodiment, in order to intuitively show the distribution of the vector geomagnetic anomaly field obtained by numerical simulation under the above parameter conditions, the X component, the Y component and the Z component in the vector geomagnetic anomaly field data can be selected for illustration. Figures 4A-4F The simulation results of the X component, the Y component and the Z component in the vector geomagnetic anomaly field data at the height of 0m and 3000m under the condition that the preset horizontal range is 30kmx30km and the number of magnetic sources is 5 and 19 are shown. Figure 4A The distribution of the X component in the vector geomagnetic anomaly field data at the height of 0m and 3000m when the number of magnetic sources is 5 is shown. Figure 4B The distribution of the Y component in the vector geomagnetic anomaly field data at the height of 0m and 3000m when the number of magnetic sources is 5 is shown. Figure 4C The distribution of the Z component in the vector geomagnetic anomaly field data at the height of 0m and 3000m when the number of magnetic sources is 5 is shown. Figure 4D The distribution of the X component in the vector geomagnetic anomaly field data at the height of 0m and 3000m when the number of magnetic sources is 19 is shown. Figure 4E The distribution of the Y component in the vector geomagnetic anomaly field data at the height of 0m and 3000m when the number of magnetic sources is 19 is shown. Figure 4FThe distribution of the Z component in the vector geomagnetic anomaly field data at the height of 0 m and 3000 m when the number of magnetic sources is 19 is shown. The horizontal axis in the figure is the east coordinate "East", with a unit of m; the vertical axis is the north coordinate "North", with a unit of m, and the horizontal and vertical coordinate ranges are both -15000 m to 15000 m, corresponding to a square region with a horizontal size of 30 km x 30 km, and the coordinate origin is located at the center of the horizontal region. The right color bar represents the numerical value of the corresponding component at each grid point, with a unit of nT, and the color changes from blue to red to represent the magnetic field strength from low to high, which is used to visually display the distribution of the three components of the vector geomagnetic anomaly field in the entire 30 km x 30 km region at each height plane. By comparing the distribution of different magnetic source numbers, different height planes, and different components in the same coordinate range, the influence of the number of magnetic sources and the change of height on the three components in the vector geomagnetic anomaly field data can be intuitively reflected. Figures 4A-4F

[0105] It should be understood that Figures 4A-4F only two height layers of 0 m and 3000 m and different magnetic source numbers are selected as examples to illustrate the distribution of the numerical simulation results on the typical height plane. When constructing the vector geomagnetic field true value library, the vector geomagnetic anomaly field data at each height layer can be sampled according to the aforementioned 61 height layers in the range of 0 m to 6000 m with an interval of 100 m, and the X component, Y component and Z component data in the vector geomagnetic anomaly field data at each height layer are combined to form a vector geomagnetic field data set, thereby representing the three-dimensional distribution of the vector geomagnetic anomaly field in the height range of 0 m to 6000 m in the preset horizontal range.

[0106] In an embodiment, to verify the applicability and stability of the downward continuation method of the geomagnetic field proposed in the present application in the long-distance and high-point-distance scenario, a comparative experiment is performed on the vector geomagnetic anomaly field data generated by the aforementioned numerical simulation. In the experiment, a vector geomagnetic anomaly field data set with a spatial range of 30 km x 30 km and a point distance of 25 m is constructed, and on this basis, different downward continuation distances, noise conditions and preset height intervals are set to analyze the downward continuation error and stability of the method.

[0107] As Figure 5 shown, Figure 5 the curve of MAE (Mean Absolute Error) changing with distance under different downward continuation distances is given. Figure 5 ​The horizontal axis represents the downward extension distance in meters (m), and the vertical axis represents MAE in nT. The blue curve represents the model configuration with a preset height interval of 100m and 2% Gaussian noise added to the input magnetic field data; the red curve represents the model configuration with a preset height interval of 100m and no noise added; and the green dots represent the model configuration with a preset height interval of 500m and no noise added. It can be seen that in the noise condition comparison experiment, the error of the blue curve increases more slowly than that of the red curve, indicating that after adding noise to the input magnetic field data and training, the model performs more stably during the downward extension process.

[0108] For example, when the downward extension distance does not exceed 1000m, corresponding to approximately 40 times the point spacing, the downward extension error of the model trained with added noise can be kept within 5nT. In the preset height interval comparison experiment, taking a downward extension distance of 500m or less as an example, corresponding to approximately 20 times the point spacing, the error can be controlled within 1nT. When comparing the preset height intervals of 100m and 500m, it can be found that the model with a preset height interval of 500m shows relatively better stability at longer distances, such as 1500m, at approximately 60 times the point spacing, while there is no significant advantage between the two preset height interval configurations at distances below 1000m. As can be seen from the above results, this method can still maintain high accuracy and good stability in long-distance and high-multiple point spacing downward extension tasks. The relevant distances, point spacings, and error values ​​are a set of test results in this embodiment, which can be adjusted according to actual application needs and do not constitute a limitation on the scope of protection of this application.

[0109] like Figures 6A-6L As shown, to further analyze the sensitivity of this method to changes in downward extension distance, tests were conducted on the downward extension effect at different starting heights. Figures 6A-6L The horizontal axis represents longitude, measured in degrees East (°E); the vertical axis represents latitude, measured in degrees North (°N). In this embodiment, the selected longitude and latitude range corresponds to a horizontal region of approximately 30km × 30km, to maintain consistency with the vector geomagnetic anomaly field dataset constructed through the aforementioned numerical simulation. The color bars on the right side of each figure represent the values ​​of the Z component Bz at each spatial location within the corresponding height plane, measured in nT. The colors, from blue to red, indicate increasing magnetic field strength, visually demonstrating the changes in the magnitude of the Z component of the vector geomagnetic anomaly field at different heights and planar locations.

[0110] Specifically, Figures 6A-6F The vector geomagnetic anomaly field distribution of the Z component in the test data at heights of 6000m, 5000m, 4000m, 3000m, 2000m, and 1000m is shown. Figures 6G-6LThe vector magnetic anomaly field distribution obtained by downward continuation 1000m under the corresponding height by the method is shown. Taking 1000m as the unified downward continuation distance, the Z component in the test data is downward continued 1000m at the above six heights, respectively, the corresponding MAE is calculated, and the obtained MAE is about 2.41nT, 2.82nT, 3.45nT, 4.27nT, 5.44nT and 6.88nT, respectively. Figures 6A-6L From the visual contrast and the above numerical values, it can be seen that at the same downward continuation distance, the lower the initial height, the relatively larger the downward continuation error, but the overall error changes relatively smoothly with the height, and the performance of the downward continuation method at different heights has good stability. The error values in the embodiment are used to illustrate a typical performance of the method, and different error results can be obtained in actual application according to specific data and application environment.

[0111] Figure 7 A geomagnetic field downward continuation system according to an example embodiment of the present application is shown. As shown in Figure 7 the geomagnetic field downward continuation system 1 includes an input data acquisition module 2 and a vector geomagnetic field data acquisition module 3. The input data acquisition module is used to acquire vector geomagnetic field data at a reference height and a downward continuation distance value. The downward continuation distance value is any value in a preset continuation distance range, the lower limit of the continuation distance range is not less than 0 and the upper limit is not greater than the height value of the reference height. The vector geomagnetic field data acquisition module is used to input the vector geomagnetic field data at the reference height and the downward continuation distance value into a geomagnetic field downward continuation neural network model to obtain the vector geomagnetic field data at the target height. The height value of the target height is the difference between the height value of the reference height and the downward continuation distance value, the geomagnetic field downward continuation neural network model is obtained by training the training set, and each training sample in the training set includes the vector geomagnetic field data at the training reference height, the training downward continuation distance value and the vector geomagnetic field data at the training target height as a supervision label.

[0112] The present application provides a computer program product, including computer programs / instructions, which are executed by a processor to implement the method described in any one of the above.

[0113] The present application also provides a computer program stored in a computer readable storage medium, such as a storage medium, and when the processor executes the computer program, the processor executes the method described above.

[0114] The application can take the form of a computer program product implemented on one or more computer-readable storage media (including, but not limited to, disk memory, CD-ROM, optical memory, etc.) containing program code. The computer-readable storage media include permanent and non-permanent, removable and non-removable media, and can be implemented in any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer-readable storage media include, but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic disk storage or other magnetic storage device, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0115] The electronic device described above can execute the method provided by the embodiments herein. The electronic device described above can include the geomagnetic field downward continuation system described above, such as one or more of a server device and a PC (Personal Computer) device. The server device and the PC device can each include, but are not limited to, a server, a desktop computer, a tablet computer, or a notebook computer.

[0116] It should be noted that the technical solutions or technical features described in the above embodiments can be combined or supplemented with each other without conflict. The scope of protection of the present application is not limited to the precise structure described in the above embodiments and shown in the accompanying drawings; any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A downward continuation method of the geomagnetic field, characterized by, The method comprises: obtaining vector geomagnetic field data at a reference height and a downward continuation distance value; wherein the downward continuation distance value is any value within a preset continuation distance range, the lower limit of the continuation distance range is not less than 0, and the upper limit of the continuation distance range is not greater than the height value of the reference height; inputting the vector geomagnetic field data at the reference height and the downward continuation distance value into a geomagnetic field downward continuation neural network model to obtain vector geomagnetic field data at a target height; wherein the height value of the target height is the difference between the height value of the reference height and the downward continuation distance value, the geomagnetic field downward continuation neural network model is obtained by training a training set, and each training sample in the training set comprises vector geomagnetic field data at a training reference height, a training downward continuation distance value, and vector geomagnetic field data at a training target height as a supervision label.

2. The geomagnetic field downward continuation method of claim 1, wherein the vector geomagnetic field data at the training reference height of each training sample in the training set is the vector geomagnetic field data at the corresponding layered height value of the target vector geomagnetic field data set in the vector geomagnetic field true value library; wherein the vector geomagnetic field true value library comprises at least one vector geomagnetic field data set, each vector geomagnetic field data set corresponds to a vector geomagnetic field; each vector geomagnetic field data set comprises vector geomagnetic field data at a plurality of layered height values, the plurality of layered height values cover a preset height range in the height direction, and the difference between any two adjacent layered height values is a preset height interval; for each layered height value, the vector geomagnetic field data set comprises vector geomagnetic field values at a plurality of sampling positions in a target area, and the interval of the plurality of sampling positions in the horizontal direction is a preset horizontal point distance.

3. The geomagnetic field downward continuation method of claim 2, wherein, the training downward continuation distance value is greater than 0 and not greater than the height value of the training reference height, and the height value of the training target height is the difference between the height value of the training reference height and the training downward continuation distance value; the vector geomagnetic field data at the training target height is determined according to the vector geomagnetic field data at a first layered height and the vector geomagnetic field data at a second layered height; wherein the vector geomagnetic field data at the first layered height and the vector geomagnetic field data at the second layered height both belong to the target vector geomagnetic field data set; the height value of the first layered height is less than the height value of the training target height and adjacent to the training target height in the height direction; the height value of the second layered height is greater than the height value of the training target height and adjacent to the training target height in the height direction.

4. The geomagnetic field downward continuation method of claim 3, wherein the vector geomagnetic field data at the training target height is a linear combination of the vector geomagnetic field data at the first layered height and the vector geomagnetic field data at the second layered height; wherein the weight coefficient of the linear combination is determined according to the height difference between the training target height and the first layered height and the second layered height.

5. The downward continuation method of the geomagnetic field according to claim 2, wherein the vector geomagnetic field data at each layer height value comprises an X component, a Y component and a Z component along three-dimensional spatial coordinate axes; the vector geomagnetic field component at the training reference height of each training sample in the training set is one of the X component, the Y component and the Z component; the downward continuation neural network model of the geomagnetic field is configured to perform downward continuation on the X component, the Y component and the Z component respectively.

6. The downward continuation method of the geomagnetic field according to claim 2, wherein the vector geomagnetic field corresponding to at least one vector geomagnetic field data set in the vector geomagnetic field true value library is a vector geomagnetic field obtained by numerical simulation; wherein the vector geomagnetic field established in the numerical simulation satisfies the following parameter conditions: the number of magnetic sources arranged within a preset horizontal range is not less than a first number and not more than a second number; wherein the first number is an integer not less than 2; the magnetic sources include base-type magnetic sources and / or ore body-type magnetic sources; wherein the base-type magnetic sources have a burial depth within a first depth range and a magnetization intensity within a first intensity range, and the ore body-type magnetic sources have a burial depth within a second depth range and a magnetization intensity within a second intensity range; the geomagnetic field parameters used in the numerical simulation include a magnetic inclination and / or a magnetic declination; wherein the magnetic inclination has a value within a first angle range, and the magnetic declination has a value within a second angle range.

7. The downward continuation method of the geomagnetic field according to claim 2, wherein the downward continuation neural network model of the geomagnetic field is a deep neural network model, and the deep neural network model comprises: an input layer configured to receive the vector geomagnetic field data at the reference height and the downward continuation distance value; a multi-stage encoder having an input end connected to an output end of the input layer and configured to perform stage-by-stage encoding on the vector geomagnetic field data at the reference height and the downward continuation distance value; a multi-stage decoder having an input end connected to an output end of the multi-stage encoder and configured to perform stage-by-stage decoding on the encoded vector geomagnetic field data at the reference height and the encoded downward continuation distance value; a skip connection structure arranged between each encoding stage in the multi-stage encoder and a corresponding decoding stage in the multi-stage decoder, and configured to connect an output of each encoding stage with an input of the corresponding decoding stage; a receptive field enhancement module arranged in at least one stage of the multi-stage encoder and / or the multi-stage decoder, and configured to expand an effective receptive field of the deep neural network model for the vector geomagnetic field data. comprises: an input data acquisition module configured to acquire the vector geomagnetic field data at the reference height and the downward continuation distance value; wherein the downward continuation distance value is an arbitrary value within a preset continuation distance range, and a lower limit of the continuation distance range is not less than 0 and an upper limit of the continuation distance range is not greater than a height value of the reference height; ​ 8. A geomagnetic field downward continuation system characterized by, ​ ​ The vector geomagnetic field data acquisition module is configured to input the vector geomagnetic field data at the reference height and the downward extrapolation distance value into a geomagnetic field downward extrapolation neural network model to obtain vector geomagnetic field data at a target height; wherein the height value of the target height is the difference between the height value of the reference height and the downward extrapolation distance value, the geomagnetic field downward extrapolation neural network model is obtained by training a training set, and each training sample in the training set includes vector geomagnetic field data at a training reference height, a training downward extrapolation distance value, and vector geomagnetic field data at a training target height as a supervision label.

9. A computer program product, when executed by a processor, performs the geomagnetic field downward extrapolation method of any one of claims 1-7.

10. An electronic device, comprising: The electronic device includes a processor and a memory, and the memory stores instructions, which, when executed by the processor, cause the electronic device to perform the geomagnetic field downward extrapolation method of any one of claims 1-7.

11. A non-transitory computer-readable storage medium, comprising: The storage medium stores a computer program, which, when executed, causes the computer device to perform the geomagnetic field downward extrapolation method of any one of claims 1-7.

Images