Method and device for processing magnetic anomaly data

Abstract

The invention provides a magnetic anomaly data processing method and device, and the method comprises the steps: forming a data set according to a sphere magnetic anomaly forward modeling formula; a U-net neural network is constructed; training a U-net neural network according to the data set to obtain a magnetic anomaly data processing model; acquiring magnetic field data of the observation plane and a height difference between the observation plane and the observation position; and according to the height difference between the observation plane and the observation position, selecting a magnetic anomaly data processing model or an improved derivative iteration method to carry out continuation on the magnetic field data of the observation plane.

Description

Technical Field

[0001] The present invention belongs to the field of magnetic anomaly data processing, and particularly relates to a method and device for processing magnetic anomaly data. Background Technique

[0002] Aeromagnetic observations are usually carried out at fluctuating flight altitudes. And in order to save the detection time cost, traditional aerial surveys can only measure the magnetic anomaly data of one survey line or a certain observation point. However, when performing quantitative interpretation, the data needs to be reduced to a plane at the same altitude. Therefore, it is of great significance to extend the magnetic field data at different altitudes to the same altitude for subsequent magnetic anomaly analysis. According to the requirements of spatial data conversion, the extension can be divided into upward extension and downward extension. The upward extension has a stable and unique solution. However, since the downward extension operator has a high-pass filtering property, it will quickly amplify high-frequency noise and cause the result to diverge. So, the downward extension is a typical ill-posed problem [3].

[0003] The key to studying the downward extension problem lies in finding its approximate linear operator, while ensuring the extension accuracy and improving the extension height and the stability of the extension result as much as possible. Many scholars at home and abroad have proposed various solutions for this. Xu Shizhe et al. and Liu Dongjia et al. transformed the downward extension process into an iterative process of stable upward extension, but the noise will be amplified with the increase of the number of iterations; Wang Yanguo et al., Bateman et al. and Evjen et al. applied the Taylor series expansion to the improvement of the extension operator, but the result is prone to divergence affected by noise; The equivalent source method is used to perform forward magnetic field calculation by simulating field sources, but because it is necessary to solve a large system of equations, the calculation efficiency is often low; Fedi et al. obtained a stable vertical derivative based on the integrated second vertical derivative (ISVD) method, avoiding the problem of Taylor series divergence. In addition, adding a regularization term can also improve the stability of the downward extension. For example, the Tikhonov iterative regularization method and the Landweber iterative regularization method

[11] . The most crucial step is to select the optimal regularization parameter.

[0004] In summary, the current methods for potential field downward extension have low accuracy, and how to improve the accuracy is an urgent problem to be solved.

[0005] Iterative algorithms do not require matrix inversion, but as the number of iterations increases, noise is rapidly amplified; the continuation results of the regularization method are sensitive to the choice of regularization parameters and are prone to loss of high-frequency signals. Although many studies have been done on the analysis and comparison of continuation methods, there are still few comparative studies on the continuation method based on deep learning and traditional methods. Therefore, based on the U-Net neural network method, this paper uses the magnetic anomaly data generated by the combined model at different heights for network training to construct a network model suitable for magnetic field continuation. By comparing the performance of three traditional continuation methods, namely the integral iteration method, the improved derivative iteration method, and the Tikhonov regularization iteration method, at different continuation depths and different noise amplitudes, the accuracy and stability of different methods are explored. Summary of the Invention

[0006] The object of the present invention is to provide a method and device for processing magnetic anomaly data to improve the continuation accuracy of magnetic anomaly data.

[0007] In the first aspect, a method for processing magnetic anomaly data is provided, including: Forming a data set according to the magnetic anomaly forward formula of the sphere model; Constructing a U-net neural network; Training the U-net neural network according to the data set to obtain a magnetic anomaly data processing model; Obtaining the magnetic field data of the observation plane and the height difference between the observation plane and the observation position; According to the height difference between the observation plane and the observation position, select the magnetic anomaly data processing model or the improved derivative iteration method to continue the magnetic field data of the observation plane.

[0008] Optionally, the step of selecting the magnetic anomaly data processing model or the improved derivative iteration method to continue the magnetic field data of the observation plane according to the height difference between the observation plane and the observation position includes: According to the height difference between the observation plane and the observation position and the grid spacing in the data set to select the magnetic anomaly data processing model or the improved derivative iteration method to process the magnetic field data of the observation plane; wherein, if the ratio of the height difference between the observation plane and the observation position and the grid spacing in the data set is less than the first threshold, the improved derivative iteration method is selected to process the magnetic field data on the ground. If the ratio of the height difference between the observation plane and the observation position and the grid spacing in the data set is greater than or equal to the first threshold, the magnetic anomaly data processing model is selected to process the magnetic field data of the observation plane.

[0009] Optionally, the value range of the first threshold is 10 to 20.

[0010] Optionally, according to the forward formula of magnetic anomaly of the sphere model, the steps of forming the data set include: According to the forward formula of magnetic anomaly of the sphere, the combined model of multiple spheres is obtained by superimposing magnetic anomalies; By adjusting the radius, burial depth, and magnetic susceptibility of the combined model of the spheres, grid data at different heights from the ground are obtained, and the grid spacing is ; Among them, the magnetic anomaly data and the data at the ground at the same height are used as the input and output of a network structure. After normalizing the input data and output data, a data set is obtained.

[0011] Optionally, the U-net neural network includes: Input layer; Downsampling layer, including the first convolutional layer, the first ReLU layer, the second convolutional layer and the second ReLU layer, the third convolutional layer and the third ReLU layer; the convolutional kernels of the first convolutional layer, the second convolutional layer, and the third convolutional layer are 3×3, and the convolutional stride is 2; Upsampling layer, including the first deconvolutional layer, the fourth ReLU layer, the second deconvolutional layer, the fifth ReLU layer, the third deconvolutional layer and the sixth ReLU layer; among them, the convolutional kernels of the first deconvolutional layer, the second deconvolutional layer, and the third deconvolutional layer are 3×3; a skip link is provided between the first ReLU layer and the third deconvolutional layer, and a skip link is provided between the second ReLU layer and the second deconvolutional layer; Output layer.

[0012] Optionally, the steps of training the U-net according to the data set to obtain the magnetic anomaly data processing model include: Taking the magnetic anomaly data at the ground as the training label, extracting the data features of each height plane of the magnetic field through the convolutional layer in the U-net, and through the result comparison of the downsampling encoding and upsampling decoding stages, the network parameters are adjusted and optimized in the reverse direction; Through iterative optimization, the training error of the U-net converges to a stable value to form a magnetic anomaly data processing model.

[0013] Optionally, the steps of processing the magnetic field data of the observation plane by using the improved derivative iteration method include: The potential field data of the observation plane is transformed to the wavenumber domain through Fourier transform to obtain its wave spectrum ; Taking the potential field wave spectrum of the observation plane as the initial value of the potential field of the plane to be continued downward ; Taking Upward continuation height to the observation plane to obtain ; Extend upward by height and height respectively to obtain the upward continuation values and ; Through the vertical first derivative relationship, the error correction term is obtained using the formula: ; After processing the error correction term through the upward continuation operator, update the plane wave spectrum to be downward continued ; Repeat the steps of upward continuation height and height , the calculation steps of the error correction value, and the calculation steps of the downward plane wave spectrum, and continuously iterate to correct the continuation value until the error is reached or the maximum number of iterations is reached, then stop the iteration; The plane wave spectrum to be continued obtained is transformed to the spatial domain through inverse Fourier transform to obtain the potential field data of the downward continued plane .

[0014] In a second aspect, a device for processing magnetic anomaly data is provided, including: A data set formation module for forming a data set according to the forward magnetic anomaly formula of the sphere model; A model construction module for constructing a U-net neural network; A training module for training the U-net neural network according to the data set to obtain a magnetic anomaly data processing model; An acquisition module for acquiring the magnetic field data of the observation plane and the height difference between the observation plane and the observation position; An extension module for extending the magnetic field data of the observation plane according to the height difference between the observation plane and the observation position by selecting a magnetic anomaly data processing model or an improved derivative iteration method.

[0015] In a third aspect, an electronic device is provided, including the device for processing magnetic anomaly data as described above.

[0016] In a fourth aspect, a computer-readable storage medium is provided, in which at least one program code is stored, and the program code is executed by a processor to implement the magnetic anomaly data processing method as described in any one of the above.

[0017] The unexpected technical effect of the technical solution provided by the present invention is: The present invention has studied various extension methods (including integral iteration method, improved derivative iteration method, Tikhonov iteration regularization method, U-Net neural network method, etc.). In the prior art, the extension accuracy of the method decreases with the increase of the multiple of the extension point distance. When the multiple of the extension point distance reaches 30 times the point distance, the extension result of the existing method will show serious distortion. Based on this, the problem of how to improve the extension accuracy at a high multiple of the point distance is specifically studied. It is found that the downward extension method based on the U-Net neural network can adjust the extension height by adjusting the height of the sample set data. Therefore, the problem of serious distortion in the extension of more than 20 times the point distance by the traditional extension method can be forcibly solved. The results show that the U-Net neural network method has the smallest error and stable results when the extension height is 30 times the point distance. In addition, the improved derivative iteration method applies the observed data of two upward extension planes, which is equivalent to connecting a low-pass filter in series. Therefore, it has a good effect on suppressing noise, but also causes the loss of some high-frequency signals, and its extension accuracy at a low multiple of the point distance is relatively good. By combining the improved derivative iteration method and U-net for magnetic field data extension according to the range of the multiple of the extension point distance (i.e., the extension height), the accuracy of the extended magnetic field data can be improved to a great extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 It is a flow chart of a method for processing magnetic anomaly data provided by the present invention; Figure 2 It is a schematic diagram of a sphere model provided by the present invention; Figure 3 It is a schematic diagram of the structure of a U-net provided by the present invention; Figure 4 It is a schematic diagram of the parameters of a combined sphere model provided by the present invention; Figure 5 It is a magnetic anomaly isogram provided by the present invention; Figure 6 It is a comparison table of the extension accuracy of a magnetic field data extension method provided by the present invention; Figure 7 It is a schematic diagram of the extension result of the integral iteration method provided by the present invention; Figure 8 It is a schematic diagram of the extension result of the improved derivative iteration method provided by the present invention; Figure 9 Schematic diagram of the extension result of a Tikhonov iterative regularization method provided by the present invention; Figure 10 Schematic diagram of the extension result of a U-Net neural network method provided by the present invention; Figure 11 Comparison table of the extension accuracy of another magnetic field data extension method provided by the present invention.

[0020] Figure 12 Structural block diagram of a magnetic anomaly data processing device provided by the present invention; Figure 13 Structural block diagram of an electronic device provided by the present invention.

[0021] The reference signs are as follows: 11: Dataset formation module; 12: Model construction module; 13: Training module; 14: Acquisition module; 15: Extension module; 21 Processor; 22: Memory. Detailed implementation manners

[0022] In order to make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] Figure 1 Flowchart of a magnetic anomaly data processing method provided by the present invention. Refer to Figure 1 , the method steps include: S101. Form a dataset according to the magnetic anomaly forward modeling formula of the sphere model.

[0024] In one example, step S101 includes: The first step. According to the magnetic anomaly forward modeling formula of the sphere, superimpose magnetic anomalies to obtain a combined model of multiple spheres.

[0025] Refer to Figure 2 Schematic diagram of a sphere model provided by the present invention.

[0026] Let the buried depth of the sphere center be , the magnetization intensity be , the volume be , the magnetic moment ; the center coordinates of the sphere are , the distance from the center of the sphere to any point in space , and its magnetic field expression is:

[0027]

[0028]

[0029]

[0030]

[0031]

[0032] The relationship between the three components of the magnetic anomaly and the magnetic anomaly is:

[0033] Substituting the three components of the magnetic anomaly ( , , ) into the above magnetic anomaly formula, the forward formula of the magnetic anomaly of the sphere can be obtained.

[0034] Second, by adjusting the radius, burial depth, and magnetic susceptibility of the combined model of the sphere, grid data at different heights from the ground are obtained, and the grid spacing is ; among them, the magnetic anomaly data at the same height and the data at the ground are used as the input and output of a network structure.

[0035] Third, after normalizing the input data and the output data, a data set is obtained.

[0036] In this embodiment, the deep learning network can be regarded as a kind of data fitting, aiming to establish a mapping relationship between the input and output data. The traditional neural network has good fitting performance, but due to factors such as computing power and training efficiency, large-scale neural networks are usually not adopted. Due to the differences in the dimensions and magnitudes of the input data, in order to reduce the training difficulty and prevent the activation function from saturating, it is necessary to normalize both the input and output data to the interval.

[0037] The normalization formula is:

[0038] Among them, is the normalized data, is the original data, and are the minimum and maximum values of the data set interval respectively. According to the normalization parameters obtained after normalization, inverse normalization processing needs to be performed according to these normalization parameters when using the network output data.

[0039] If the range of magnetic anomaly data to be processed is large, the idea of image processing can be considered, and the grid data to be continued can be segmented and stored in several small regions.

[0040] The data set used for training in the present invention is the magnetic anomaly data of each height plane of the forward modeling of the mixed sphere model. According to the forward modeling formula of the sphere magnetic anomaly, the combined model of multiple spheres is obtained by superimposing the magnetic anomalies. By adjusting the parameters such as the radius, burial depth, and magnetic susceptibility of the sphere model, the grid data within the range of 200m×200m at the ground (h = 0m), h = 10m, h = 20m, and h = 30m are obtained, and the grid spacing is 1m.

[0041] Considering that the non - mapping relationships corresponding to different continuation heights are different, the magnetic anomaly data at the same height and the data at the ground need to be used as the input and output of a network structure. There are three continuation heights in the model experiment of the present invention, so three suitable U - Net network models need to be constructed.

[0042] S102. Construct a U - net neural network.

[0043] See Figure 3 , in an example, the U - net includes: Input layer; Downsampling layer, including the first convolutional layer, the first ReLU layer, the second convolutional layer, the second ReLU layer, the third convolutional layer, and the third ReLU layer; the convolutional kernels of the first convolutional layer, the second convolutional layer, and the third convolutional layer are 3×3, and the convolutional stride is 2; Upsampling layer, including the first deconvolutional layer, the fourth ReLU layer, the second deconvolutional layer, the fifth ReLU layer, the third deconvolutional layer, and the sixth ReLU layer; among them, the convolutional kernels of the first deconvolutional layer, the second deconvolutional layer, and the third deconvolutional layer are 3×3; there is a skip link between the first ReLU layer and the third deconvolutional layer, and there is a skip link between the second ReLU layer and the second deconvolutional layer; Output layer.

[0044] The U-net network constructed in the present invention is developed for magnetic field downward continuation and includes two parts: downsampling and upsampling. For the downsampling part, a three-layer stacked convolutional structure is adopted, with a total of 3 convolutional layers and 3 rectified linear unit (ReLU) layers. The convolution kernel is set to 3×3, and the convolution stride is set to 2, thereby realizing the function of downsampling in the pooling layer and expanding the receptive field of the convolutional layer. The upsampling part is composed of 3 transposed convolutional structures, with a total of 3 transposed convolution layers and 3 ReLU layers. The convolution kernel of each transposed convolution layer is 3×3. In each downsampling and upsampling convolution process, the resolution of the feature map can maintain structural symmetry, and two skip links are established between downsampling and upsampling to prevent the phenomenon of gradient vanishing during network training and fuse the data of downsampling and upsampling.

[0045] S103. Train the U-net neural network according to the data set to obtain a magnetic anomaly data processing model.

[0046] In one example, step S103 includes: First step: Use the magnetic anomaly data at the ground as training labels, extract the data features of the magnetic field at each height plane through the convolutional layer in the U-net, and through the result comparison of the downsampling encoding and upsampling decoding stages, reversely optimize and adjust the network parameters.

[0047] Second step: Through iterative optimization, make the training error of the U-net converge to a stable value to form a magnetic anomaly data processing model.

[0048] S104. Obtain the magnetic field data of the observation plane and the height difference between the observation plane and the observation position.

[0049] Among them, the observation position is the position where the observation device is located, and the observation plane is the area detected by the observation device.

[0050] S105. According to the height difference between the observation plane and the observation position, select the magnetic anomaly data processing model or the improved derivative iteration method to extend the magnetic field data of the observation plane.

[0051] In one example, step S105 includes: According to the height difference between the observation plane and the observation position and the grid spacing in the data set ratio, select the magnetic anomaly data processing model or the improved derivative iteration method to process the magnetic field data of the observation plane.

[0052] Among them, if the ratio of the height difference between the observation plane and the observation position and the grid spacing in the data set is less than the first threshold, then select the improved derivative iteration method to process the magnetic field data on the ground. If the height difference between the observation plane and the observation position and the grid spacing in the dataset is greater than or equal to the first threshold, then select the magnetic anomaly data processing model to process the magnetic field data of the observation plane.

[0053] In one example, the value range of the first threshold is 10 to 20. For example, the first threshold is 15.

[0054] In this embodiment, the extension steps of the improved derivative iteration method are as follows: First step, transform the potential field data of the observation plane to the wavenumber domain through Fourier transform to obtain its wave spectrum .

[0055] Second step, use the potential field wave spectrum of the observation plane as the initial value of the potential field of the plane to be extended downward .

[0056] Third step, extend the upward extension height to the observation plane to obtain .

[0057] Fourth step, extend upward by the height and the height respectively to obtain the upward extension values and .

[0058] Fifth step, through the vertical first derivative relationship, use the formula to obtain the error correction term: ; Sixth step, after processing the error correction term through the upward extension operator, update the wave spectrum of the plane to be extended downward.

[0059] Seventh step, repeat the steps of the upward extension height and the height , the error correction value calculation step, and the calculation step of the downward plane wave spectrum (that is, repeat the fourth to sixth steps), continuously iterate and correct the extension value until the error is met or the maximum number of iterations is reached, then stop the iteration.

[0060] Eighth step, transform the obtained wave spectrum of the plane to be extended through inverse Fourier transform to the spatial domain to obtain the potential field data of the plane extended downward.

[0061] See Figure 4 , Figure 4Schematic diagram of the parameters of a combined sphere model provided by the present invention.

[0062] According to the forward model of sphere magnetic anomaly, a superposition model of four combined spheres with different parameters was constructed. The model parameters of the spheres are as Figure 4 shown (assuming the downward direction of the z-axis is positive). The background field intensity is 50000 nT, the point spacing and line spacing are both 1 m, and the grid size is 201×201. The X-direction range is (-100 m, 100 m), and the Y-direction is (-100 m, 100 m). Considering that the continuation height of the traditional continuation method is limited by the point spacing and often cannot exceed 20 times the point spacing, to compare the variation of the continuation accuracy of the U-Net neural network method with the point spacing, the magnetic anomaly data of the combined model at the ground level (h = 0 m), a height of 10 m (10 times the point spacing), a height of 20 m (20 times the point spacing), and a height of 30 m (30 times the point spacing) were calculated.

[0063] See Figure 5 , Figure 5 Magnetic anomaly contour map provided by the present invention. Among them, Figure 5 in (a) represents the magnetic anomaly contour map at h = 0 m, Figure 5 in (b) represents the magnetic anomaly contour map obtained after adding noise at h = 10 m, Figure 5 in (c) represents the magnetic anomaly contour map obtained after adding noise at h = 20 m, Figure 5 in (d) represents the magnetic anomaly contour map obtained after adding noise at h = 30 m.

[0064] In the experiment, the magnetic anomaly data at the ground level (h = 0 m) were used as the observed data. To simulate the actual data processing, Gaussian white noise with a zero mean and a standard deviation of 5% of the absolute mean of the theoretical magnetic anomaly at that height was added to the magnetic anomaly data at the three height planes of h = 10 m, h = 20 m, and h = 30 m to test the anti-noise ability of each method. The magnetic anomaly data contour map of the theoretical model is as shown in Figure 5. To analyze the accuracy of different continuation methods at different continuation depths, four methods including the integral iteration method, the improved derivative iteration method, the Tikhonov iteration regularization method, and the U-Net neural network method were used to Figure 5 the magnetic anomaly data with added noise at three different heights in (b), Figure 5 the magnetic anomaly data with added noise at three different heights in (c), Figure 5 the magnetic anomaly data with added noise at three different heights in (d) were continued to the ground level. The mean square error and correlation error were calculated using the obtained continuation values and the magnetic anomaly data at the ground level

[0065] as accuracy evaluation indicators to analyze the continuation errors under different noise levels and different continuation heights.

[0066] The calculation formula for the correlation error is as follows:

[0067] Figure 6 This is a comparison table of the extension accuracy of the first magnetic field data extension method provided by the present invention. Among them, the maximum number of iterations of the integral iteration method is selected as 200, and the step size of each iteration is 0.5; the maximum number of iterations of the improved derivative iteration method is selected as 200; the number of iterations of the Tikhonov iteration regularization method is 50, and the regularization parameter is selected by the L-curve method, which is less than 0.5. Before these three traditional extension methods are extended, considering the boundary effect, edge expansion processing is required. When training the U-Net model, in terms of dataset construction, the sample dataset is the magnetic anomaly data obtained by forward modeling of spherical anomaly models with different sizes and different occurrence states. Since the sample data is calculated from the theoretical model and the data has high accuracy, the number of samples used to train a single network model in this paper is 2000, the ratio of the training set to the validation set is 9:1, the number of training rounds is 8 rounds, and the number of iterations per round is 100 times.

[0068] Figure 7 (a1)-(a3) in are the extension results of downwardly extending (b)-(d) in 5 by 10 times the point distance, 20 times the point distance, and 30 times the point distance to the ground using the integral iteration method. Figure 8 (b1)-(b3) in Figure 9 (c1)-(c3) in Figure 10 (d1)-(d3) in are the extension results of the improved derivative iteration method, the Tikhonov iteration regularization method, and the U-Net neural network method, respectively.

[0069] It can be seen that as the extension depth increases, compared with the downward extension method of the U-Net neural network, the extension errors of the integral iteration method, the improved derivative iteration method, and the Tikhonov iteration regularization method increase significantly. When the U-Net neural network method extends by 10 times the point distance, the error is 1.2394 nT, and when it extends by 30 times the point distance, the error still remains at 1.4329 nT.

[0070] When extended to 30 times the point spacing, serious distortions occurred in the extension results of the integral iteration method, the improved derivative iteration method, and the Tikhonov iteration regularization method. The noise in the extension result of the integral iteration method was severely amplified, drowning out the effective signal, and the error reached 30.5600 nT. Since the improved derivative iteration method applied the observed data of two upward extension planes, which was equivalent to cascading a low-pass filter, it had a good effect on suppressing noise, but also caused the loss of some high-frequency signals, and the error value was 33.5520 nT. When the optimal regularization parameter was selected according to the L-curve method in the Tikhonov iteration regularization method, its extension result was significantly better than that of the integral iteration method, and the number of iterations was less. However, this method could not distinguish high-frequency noise from high-frequency signals, and it would also lead to the loss of high-frequency signals, with an error value of 26.8577 nT. Since the downward extension method based on the U-Net neural network could adjust the extension height by adjusting the height of the sample set data, it forcibly solved the problem of serious distortion when the traditional extension method was extended by more than 20 times the point spacing. The results showed that when the extension height was 30 times the point spacing, the U-Net neural network method had the smallest error and stable results, and the errors were reduced by 5.5386 nT, 7.4110 nT, and 9.1570 nT compared with the other three methods respectively.

[0071] To further test the robustness of the U-Net neural network method, numerical experiment comparative analysis was carried out using test data with a 10% amplitude random error added, and the results are as Figure 11 shown.

[0072] The results showed that although sample data with a 5% random error was introduced during the training of the U-Net neural network, for data with a large noise amplitude, the network model designed in this paper still had good performance. It indicated that the U-Net deep neural network method had a certain degree of robustness. When the extension point spacing was 30 times the point spacing, the errors of the U-Net neural network method were reduced by 7.1763 nT, 8.2115 nT, and 8.9619 nT compared with the other three methods respectively. However, the extension accuracy of the four extension methods decreased as the noise amplitude of the data increased.

[0073] Figure 12 This is the structural block diagram of a magnetic anomaly data processing device provided by the present invention. Refer to Figure 12 , including: A data set forming module 11, configured to form a data set according to a mixed sphere model and a sphere magnetic anomaly forward formula; A model construction module 12, configured to construct a U-net; A training module 13, configured to train the U-net according to the data set to obtain a magnetic anomaly data processing model; An acquisition module 14, configured to acquire magnetic field data of an observation plane and the height difference between the observation plane and the observation position; An extension module 15, configured to select a magnetic anomaly data processing model or an improved derivative iteration method to extend the magnetic field data of the observation plane according to the height difference between the observation plane and the observation position.

[0074] Figure 13 The block diagram of a structure of an electronic device provided by the present invention. Refer to Figure 13 , the electronic device may include Figure 12 The processing device for magnetic anomaly data as described above. Generally, the electronic device includes: a processor 21 and a memory 22.

[0075] The processor 21 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 21 may be implemented in at least one of the following hardware forms: DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 21 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the wake state, also known as the CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. The memory 22 may include one or more computer-readable storage media, and the computer-readable storage media may be non-transitory. The memory 22 may further include a high-speed random access memory and a non-volatile memory, such as one or more disk storage devices and flash storage devices. In some embodiments, the non-transitory computer-readable storage media in the memory 22 is used to store at least one instruction, and the at least one instruction is used to be executed by the processor 21 to implement the method for processing magnetic anomaly data executed by the electronic device provided in the method embodiments of the present application.

[0076] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for processing magnetic anomaly data, characterized in that, Including: Form a data set according to the forward magnetic anomaly formula of the sphere model; Construct a U-net neural network; Train the U-net neural network according to the data set to obtain a magnetic anomaly data processing model; Obtain the magnetic field data of the observation plane and the height difference between the observation plane and the observation position; According to the height difference between the observation plane and the observation position, select the magnetic anomaly data processing model or the improved derivative iteration method to extend the magnetic field data of the observation plane.

2. The method for processing magnetic anomaly data according to claim 1, wherein, The steps of selecting the magnetic anomaly data processing model or the improved derivative iteration method to extend the magnetic field data of the observation plane according to the height difference between the observation plane and the observation position include: According to the ratio of the height difference between the observation plane and the observation position , the grid spacing in the dataset , select a magnetic anomaly data processing model or an improved derivative iteration method to process the magnetic field data of the observation plane; Among them, if the height difference between the observation plane and the observation position and the grid spacing in the dataset has a ratio less than the first threshold, then the improved derivative iteration method is selected to process the magnetic field data on the ground. If the height difference between the observation plane and the observation position and the grid spacing in the dataset has a ratio greater than or equal to the first threshold, then the magnetic anomaly data processing model is selected to process the magnetic field data on the observation plane.

3. The method for processing magnetic anomaly data according to claim 1, wherein The value range of the first threshold is 10 to 20.

4. The method for processing magnetic anomaly data according to any one of claims 1 to 3, wherein The steps of forming a data set according to the forward magnetic anomaly formula of the sphere model include: According to the forward magnetic anomaly formula of the sphere, superimpose the magnetic anomalies to obtain a combined model of multiple spheres; By adjusting the radius, burial depth, and magnetic susceptibility of the combined model of the sphere, grid data of the model at different heights from the ground are obtained, and the grid spacing is ; among them, the magnetic anomaly data and the data at the ground at the same height are used as the input and output of a network structure. After normalizing the input data and output data, obtain a data set.

5. The method for processing magnetic anomaly data according to any one of claims 1 to 3, characterized in that The U-net neural network includes: Input layer; Downsampling layer, including a first convolutional layer, a first ReLU layer, a second convolutional layer and a second ReLU layer, a third convolutional layer and a third ReLU layer; the convolutional kernels of the first convolutional layer, the second convolutional layer, and the third convolutional layer are 3×3, and the convolutional stride is 2; Upsampling layer, including a first transposed convolutional layer, a fourth ReLU layer, a second transposed convolutional layer, a fifth ReLU layer, a third transposed convolutional layer and a sixth ReLU layer; wherein, the convolutional kernels of the first transposed convolutional layer, the second transposed convolutional layer, and the third transposed convolutional layer are 3×3; there is a skip link between the first ReLU layer and the third transposed convolutional layer, and there is a skip link between the second ReLU layer and the second transposed convolutional layer; Output layer.

6. The method for processing magnetic anomaly data according to any one of claims 1 to 3, characterized in that, The steps of training the U-net neural network according to the data set to obtain a magnetic anomaly data processing model include: Use the magnetic anomaly data at the ground as the training label, extract the data features of each height plane of the magnetic field through the convolutional layer in the U-net, and through the result comparison of the downsampling encoding and upsampling decoding stages, reversely optimize and adjust the network parameters; Through iterative optimization, make the training error of the U-net converge to a stable value to form a magnetic anomaly data processing model.

7. The method for processing magnetic anomaly data according to any one of claims 1 to 3, characterized in that, The steps of processing the magnetic field data of the observation plane by using the improved derivative iteration method include: The observed plane potential field data is Fourier-transformed into the wavenumber domain to obtain its wave spectrum ; Using the potential field wave spectrum of the observation plane as the initial value of the potential field for the plane to be continued downward ; Extend the upward continuation height to the observation plane to obtain ; Extend upward by heights and respectively to obtain upward continuation values and ; Through the vertical first derivative relationship, use the formula to obtain an error correction term: ； After processing the error correction term with the upward continuation operator, update the plane wave spectrum to be downward continued ; Repeat Upward continuation height And height Steps, error correction value calculation steps, downward plane wave spectrum calculation steps, continuously iterate to correct the continuation value until the error When it reaches or reaches the maximum number of iterations, stop the iteration; The obtained plane wave spectrum to be continued downward is inverse Fourier transformed to the spatial domain to obtain the potential field data of the downward continued plane .

8. A processing device for magnetic anomaly data, characterized in that, Including: Data set formation module, used to form a data set according to the forward magnetic anomaly formula of the sphere model; Model construction module, used to construct a U-net neural network; Training module, used to train the U-net neural network according to the data set to obtain a magnetic anomaly data processing model; Acquisition module, used to obtain the magnetic field data of the observation plane and the height difference between the observation plane and the observation position; Extension module, used to select the magnetic anomaly data processing model or the improved derivative iteration method to extend the magnetic field data of the observation plane according to the height difference between the observation plane and the observation position.

9. An electronic device, characterized in that, Including the magnetic anomaly data processing device according to claim 8.

10. A computer-readable storage medium, characterized in that, At least one program code is stored in the computer-readable storage medium, and the program code is executed by a processor to implement the method for processing magnetic anomaly data according to any one of claims 1 to 7.

Images