A boundary identification method based on intelligent gravity and magnetic field data denoising

Abstract

The present application belongs to the technical field of gravity and magnetic prospecting data processing, and relates to a boundary identification method based on intelligent gravity and magnetic field data denoising, which comprises the following steps: step 1: generating an original geophysical gravity and magnetic field data set by using a random combination forward method; step 2: adding a random percentage of Gaussian noise to all data in the original geophysical gravity and magnetic field data set to obtain a noisy data set; step 3: constructing a denoising prediction network model; step 4: training the denoising prediction network model using the noisy data; step 5: evaluating the effectiveness of the trained denoising prediction network model until a qualified denoising prediction network model is obtained; step 6: inputting the noisy data into the qualified denoising prediction network model to obtain a denoising result; and step 7: using the denoising result to identify the boundary and obtain a correlation coefficient matrix, and then determining the positions of ore bodies and structures. The method has the beneficial effects of high denoising precision, clear boundary and great significance for improving processing and interpretation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gravity and magnetic prospecting data processing, and particularly relates to a boundary identification method based on intelligent gravity and magnetic field data denoising. BACKGROUND

[0002] The geophysical boundary identification method is an important content in gravity and magnetic field data processing and interpretation, and can determine the horizontal distribution of the anomaly body.

[0003] The commonly used methods at present include total horizontal derivative method, vertical derivative method and analytical signal method. However, these numerical calculation methods are prone to noise interference. In the past research, the mainstream denoising method is the frequency domain method, including low-pass filtering, Bayesian filtering, Kalman filtering and the like. The characteristic is to calculate in the frequency domain to remove most of the noise. However, the problem of this kind of method is that the denoising result depends on parameter adjustment and the processing effect on medium and high amplitude data is poor.

[0004] In order to obtain ideal denoising effect, artificial intelligence technology has been applied to the denoising processing of geophysical data by virtue of its application experience in the image field, such as using DnCNN (Denoising Convolutional Neural Network) to remove Gaussian noise of gravity and magnetic data, using a hole convolution generative adversarial network to remove seismic data noise, using SVM (Support Vector Machine) dictionary learning method to remove electrical data noise and the like. Different from other geophysical data, gravity and magnetic field data can realize regular distribution of observation points through gridding and be expressed by a matrix. Therefore, gravity and magnetic field data can be processed as image data; compared with the traditional method, the intelligent method can restore the characteristics of geophysical data more completely. However, in the field of gravity and magnetic data denoising, due to the limitation of network structure and data set construction, some methods have the problems of poor robustness and inability to remove multi-level noise. In the aspect of boundary identification, the traditional boundary identification method has the problem of insufficient identification accuracy. SUMMARY

[0005] Technical problems to be solved

[0006] In view of the above-mentioned defects and deficiencies of the prior art, the present application provides a boundary identification method based on intelligent gravity and magnetic field data denoising, which solves the technical problems of poor robustness and inability to remove multi-level noise of some methods in the field of gravity and magnetic data denoising due to the limitation of network structure and data set construction.

[0007] Technical scheme

[0008] In order to achieve the above object, the main technical scheme adopted by the present application comprises:

[0009] The present application provides a boundary identification method based on intelligent gravity and magnetic field data denoising, comprising the following steps:

[0010] Step 1: using a random combination forward method to generate an original geophysical gravity and magnetic field data set;

[0011] Step 2: adding a random percentage of Gaussian noise to all data in the original geophysical gravity and magnetic field data set to obtain a noisy data set;

[0012] Step 3: constructing a denoising prediction network model;

[0013] Step 4: training the denoising prediction network model using noisy data;

[0014] Step 5: evaluating the effectiveness of the trained denoising prediction network model until a qualified denoising prediction network model is obtained;

[0015] Step 6: inputting noisy data into the qualified denoising prediction network model to obtain a denoising result;

[0016] Step 7: using the denoising result for boundary identification to obtain a correlation coefficient matrix and then determining the location of the ore body and structure.

[0017] As a further improvement of the present application, in step 1, a random combination forward method is used to randomly generate random models of different sizes, shapes, positions, densities and magnetic susceptibilities within the determined survey space, and the generated random models include cuboids, spheres and random walk bodies, the gravity and magnetic field anomaly data of which are calculated to constitute the original geophysical gravity and magnetic field data set.

[0018] As a further improvement of the present application, the random model parameter setting formula is expressed as follows:

[0019] First, the center position of the random model M is determined, and M is set as a random value with a value range conforming to the following formula:

[0020] ;

[0021] In the formula, respectively represent the coordinates of the center position of M in three directions, respectively represent the range of the observation space in the x, y and z axis directions;

[0022] If it is a random cuboid, the parameters of the random model are:

[0023] ;

[0024] In the formula, Parameters of the cuboid model, l, w and h represent the length, width and height of the cuboid, respectively;

[0025] If it is a random sphere, the parameters of the random model are:

[0026] ;

[0027] In the formula, Parameters of the cuboid model and the sphere model, d represents the radius of the sphere model;

[0028] If it is a random walker, the parameters of the random model are:

[0029] ;

[0030] In the formula, Parameters of the random walker, wherein And t represent the size of the edge of a single cubic element and the number of cubes, respectively.

[0031] As a further improvement of the present application, the noisy data set obtained in step 2 is specifically: after adding random percentage Gaussian noise to all data in the original geophysical gravity and magnetic field data set respectively to generate noisy data corresponding to the original data, the noisy data set is obtained, and the training set and the test set are divided according to the preset proportion.

[0032] As a further improvement of the present application, the noisy data set is obtained by adding random percentage Gaussian noise with a maximum value between 1% and 7% to all data in the original geophysical gravity and magnetic field data set, and the noisy data set is obtained according to the following formula:

[0033] ;

[0034] In the formula, And Respectively represent the obtained noisy data and the data to be added noise, Is a random integer between 1 and 7.

[0035] As a further improvement of the present application, the constructed denoising prediction network model includes a full convolution network layer and an encoder-decoder Net; wherein the convolution layer contained in the full convolution network layer is at least 4 layers, and the structure of the Net at least includes two down-sampling modules and two up-sampling modules; the up-sampling module at least includes a convolution layer and an up-sampling layer, and the down-sampling module at least includes a convolution layer and a down-sampling layer.

[0036] As a further improvement of the present application, step 4 of training the denoising prediction network model using noisy data specifically includes the following steps:

[0037] Step 4.1: preprocessing the training set data to obtain a preprocessed data set;

[0038] Step 4.2: inputting the preprocessed data set into the denoising prediction network model for training and testing;

[0039] In the training and testing process, the loss function is calculated as follows:

[0040] ;

[0041] In the formula, n is the number of elements of the matrix x , is the output denoising result, is the normalized noisy data.

[0042] As a further improvement of the present application, the trained denoising prediction network model in step 5 is evaluated for effectiveness until a qualified denoising prediction network model is obtained, which is: first, the trained denoising prediction network model is preliminarily evaluated for effectiveness by the loss function, and then it is verified again using multiple sets of theoretical model data until a qualified denoising prediction network model is obtained.

[0043] As a further improvement of the present application, the noisy data is input into the qualified denoising prediction network model in step 6 to obtain the denoising result, which is: after the noisy data is input into the fully convolutional network, the noise weighting matrix is obtained, and the noisy data is input into Net to obtain the denoising matrix , the final denoising result is obtained by calculating the sum of the denoising matrix and , and the specific calculation formula is as follows:

[0044]

[0045] ;

[0046] ;

[0047] In the formula, represents an activation function, and L represents the number of fully convolutional layers used.

[0048] As a further improvement of the present application, the denoising result is used for boundary identification in step 7 to obtain a correlation coefficient matrix, and then the positions of the ore body and the structure are determined, which specifically includes the following steps:

[0049] ​Step 7.1: Obtain the vertical derivative using the vertical derivative method and the Theta map method on the obtained denoising result, and obtain the vertical derivative And the Theta map method parameters The calculation formula is respectively:

[0050] ;

[0051] ;

[0052] Step 7.2: Obtain the correlation coefficient matrix of the obtained And Calculate the correlation coefficient matrix The calculation formula is:

[0053] ;

[0054] In the formula, A is the number of matrix elements.

[0055] Step 7.3: Determine the location of the ore body and the structure through the correlation coefficient matrix.

[0056] The beneficial effects of the present application are:

[0057] Introduce various modeling methods to construct high-complexity data sets, which are more suitable for complex geological conditions, increase the applicability and robustness of the trained model, and realize multi-level noise prediction and removal through random noise mode and the established denoising prediction network. It has been verified that the MSE value of the geophysical gravity and magnetic field data denoised by the method of the present application generated by the theoretical model is the lowest, which is 0.0309. It can solve the denoising problem of geophysical gravity and magnetic data, and compared with other methods, it has the characteristics of high precision, high robustness and not easy to be disturbed by human factors. In the boundary identification result, combined with the characteristics of the vertical derivative method and the Theta map method, the delineated ore body boundary is clearer and more accurate. The method of the present application is of great significance to improve the processing precision of geophysical gravity and magnetic field data and high-precision interpretation. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 A flowchart of a boundary identification method based on intelligent gravity and magnetic field data denoising provided by the embodiment of the present application;

[0059] Figure 2 The theoretical model position and gravity gradient Original data image, wherein (a) is the theoretical model position, (b) is the gravity gradient Original data image;

[0060] Figure 3The figure is a denoising comparison of different levels of noise data in the embodiment of the present application, wherein (a) is a noisy data image obtained by adding 1% Gaussian noise of the maximum value of the data to the original data; (b) is a data image denoised by using the present application; (c) is a data image denoised by using the median filter; and (d) is a data image denoised by using the low-pass filter.

[0061] Figure 4 The figure is a denoising comparison of different levels of noise data in the embodiment of the present application, wherein (a) is a noisy data image obtained by adding 3% Gaussian noise of the maximum value of the data to the original data; (b) is a data image denoised by using the present application; (c) is a data image denoised by using the median filter; and (d) is a data image denoised by using the low-pass filter.

[0062] Figure 5 The figure is a denoising comparison of different levels of noise data in the embodiment of the present application, wherein (a) is a noisy data image obtained by adding 5% Gaussian noise of the maximum value of the data to the original data; (b) is a data image denoised by using the present application; (c) is a data image denoised by using the median filter; and (d) is a data image denoised by using the low-pass filter.

[0063] Figure 6 The figure is a measured magnetic anomaly data of Macheng iron mine area used in the embodiment of the present application, wherein (a) is a data image to be denoised, and (b) is a data image after denoising.

[0064] Figure 7 The figure is a result of boundary identification of the ore body of Macheng iron mine area by using the correlation coefficient boundary identification method in the embodiment of the present application, wherein (a) is an input image after denoising, (b), (c) and (d) are respectively the results of the vertical derivative method, the Theta graph method and the correlation coefficient method. DETAILED DESCRIPTION

[0065] In order to better explain the present application, the following will be described in detail by specific embodiments in combination with the drawings.

[0066] In order to better understand the above technical solutions, the following will describe the exemplary embodiments of the present application in more detail with reference to the drawings. Although the exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer, more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0067] In the present embodiment, a boundary identification method based on intelligent gravity and magnetic field data denoising, as shown in Figure 1 includes the following steps:

[0068] Step 1: generate an original geophysical gravity and magnetic field data set using a random combination forward method;

[0069] Specifically, in order to simulate as much as possible the data situation that the boundary identification may face, a random combination forward method is used to randomly generate random models of different sizes, different shapes, different positions, different densities and different magnetic susceptibilities in the determined survey area.

[0070] The generated random models include cuboids, spheres and random walk bodies, and the gravity and magnetic field anomaly data thereof is calculated to constitute the original geophysical gravity and magnetic field data set;

[0071] The gravity and magnetic field anomaly data mentioned here includes gravity gradient data, gravity anomaly data, magnetic gradient data and magnetic anomaly data;

[0072] The original geophysical gravity and magnetic field data set is composed of various gravity and magnetic field anomaly data, wherein each set of gravity and magnetic field anomaly data represents an anomaly generated by a random model;

[0073] When generating the model body, the generated random model will correspond to a random integer from 1 to 3, and the random integer is 1, 2 and 3 respectively corresponding to the cuboid, the sphere and the random walk body, wherein the cuboid model is used to simulate the anomalies of layered and quasi-layered ore bodies such as BIF iron ore; the sphere is used to simulate the anomalies of salt dome and magma capsule; the random walk body is used to simulate the anomalies of vein-shaped ore bodies.

[0074] The above random model parameter setting formula is expressed as follows:

[0075] First, the center position of the random model M is determined, and M is set as a random value, the value range of which conforms to the following formula:

[0076] (1)

[0077] In the formula, respectively represent the coordinates of the center position of M in three directions, respectively represent the range of the observation space in the x, y and z axis directions;

[0078] If it is a random cuboid, the parameters of the random model are:

[0079] (2)

[0080] In the formula, represent the parameters of the cuboid model, and l, w and h respectively represent the long side, the short side and the height of the cuboid;

[0081] If it is a random sphere, the parameters of the random model are:

[0082] (3)

[0083] wherein, denote the parameters of the cuboid model and the sphere model, wherein l, w and h denote the length, the width and the height of the cuboid respectively, and d denotes the radius of the sphere model;

[0084] wherein, denote the parameters of the cuboid model and the sphere model, and d denotes the radius of the sphere model;

[0085] If it is a random walker, the parameters of the random model are:

[0086] (4)

[0087] wherein, denote the parameters of the random walker, wherein and t denote the size of the edge of a single cubic voxel and the number of cubes respectively;

[0088] wherein, all the variable subscripts max and min denote the maximum value and the minimum value of the corresponding variable, which are set before the random model is generated;

[0089] In this embodiment, the forward used is calculated in the Cartesian coordinate system, and the forward method used is the random cuboid forward, the random sphere forward and the random walker forward, and there are multiple models distributed in a survey area. Among them, when the random cuboid is generated, the length, the width and the height of the cuboid are randomly assigned; the sphere model is randomly assigned to the sphere radius; the random walker mainly randomly assigns the initial seed position, the seed growth direction and the growth number. The distribution range of the generated data on the x, y and z axes is 0m to 6400m, 0m to 6400m and 0m to 1000m respectively. Among them, the distances between the survey points and the survey lines are both 50m. The gravity anomaly data , the gravity gradient data , , , , , and the total magnetic field anomaly values are calculated for the established random combination body theoretical model, and the original geophysical gravity and magnetic field data set used contains groups of data.

[0090] Step 2: Add a random percentage of Gaussian noise to all the data in the original geophysical gravity and magnetic field data set to obtain a noisy data set;

[0091] Specifically, after adding random percentage Gaussian noise to all data in the original geophysical gravity and magnetic field data set respectively to generate noisy data corresponding to the original geophysical gravity and magnetic field data, a noisy original geophysical gravity and magnetic field data set, i.e., a noisy data set, is obtained, and then a certain proportion is extracted to generate a test set, and a noisy training set and a test set are obtained.

[0092] Random percentage Gaussian noise with a maximum value between 1% and 7% is added to all data in the original geophysical gravity and magnetic field data set, and the noise is added according to the following formula:

[0093] (5)

[0094] In the formula, and respectively represent the obtained noisy data and the data to be added with noise, is a random integer between 1 and 7, which is used to simulate measurement noise and noise generated by deep geological bodies, and obtain the noisy original geophysical gravity and magnetic field data set. At this time, the data set contains multiple groups of data, and each group of data includes original data and its corresponding noisy data.

[0095] The proportion of the extracted test set should be at least 40% of the noisy gravity and magnetic field data set, and the same type of data should be ensured; the number of different abnormal data in the training set and the test set should be the same, and both contain noiseless data.

[0096] In this embodiment, random percentage Gaussian noise is added to all data in the original geophysical gravity and magnetic field data set, and the number of training set data used is 60,000 groups, and the number of test set data is 40,000 groups.

[0097] Step 3: Construct a denoising prediction network model;

[0098] Specifically, the constructed denoising prediction network model includes a fully convolutional network layer and an encoder-decoder Net.

[0099] In order to improve the accuracy of calculation, the convolutional layer included in the fully convolutional network layer should not be less than 4 layers, and the structure of the Net should at least include two down-sampling modules and two up-sampling modules. Whether it is an up-sampling module or a down-sampling module, it should at least include a convolutional layer and a down-sampling layer (up-sampling layer).

[0100] In this embodiment, the denoising prediction network model used includes five convolutional layers in the fully convolutional network layer, and the encoder-decoder includes two down-sampling modules and two up-sampling modules.

[0101] Step 4: Train the denoising prediction network model using noisy data;

[0102] Specifically, the training and testing sets are input into the denoising prediction network model for training and testing, including the following steps:

[0103] Step 4.1: Preprocess the training set data to obtain the preprocessed dataset;

[0104] Before inputting noisy data into the network, the data needs to be normalized. Taking a set of noisy data as an example, it contains... and They need to be normalized separately;

[0105] Step 4.2: Input the preprocessed dataset into the denoising prediction network model for training and testing;

[0106] During training and testing, the loss function The calculation formula is as follows:

[0107] (6)

[0108] In the formula, n is a matrix. x The number of elements, For the denoising result of the output, For normalized noisy data, the training process of the denoising prediction network model is equivalent to obtaining the desired result. Network parameters;

[0109] Step 5: Evaluate the effectiveness of the trained denoising prediction network model until a qualified denoising prediction network model is obtained.

[0110] First, the trained denoising prediction network model is initially evaluated for effectiveness using a loss function. Then, it is validated a second time using multiple sets of theoretical model data until a qualified denoising prediction network model is obtained.

[0111] Specifically:

[0112] After the denoising prediction network model is trained, the effectiveness of the obtained model is verified using a test set. A model that meets the conditions is a qualified model.

[0113] After each training session, the test set is preprocessed in step 4.1, and the loss function in step 4.2 is used to verify the effectiveness of the model. If the loss value of a certain test is greater than that of the previous test, the previous training result is regarded as an effective denoising prediction network model.

[0114] In this embodiment, when the denoising prediction network model meets the conditions, the model is trained 50 times, and the final model loss value is 0.0019.

[0115] Subsequently, the denoising network prediction model after the preliminary evaluation is secondarily verified using a plurality of groups of simulation data of the pre-designed, known noise and real data relationship of the theoretical model, to determine whether the denoising result meets the processing requirements of subsequent boundary identification calculation;

[0116] If the denoising result meets the standard, it is considered as an effective denoising result; if it does not meet the standard, the data set size needs to be adjusted and the denoising model needs to be retrained to adapt to the actual data, until a qualified denoising prediction network model is obtained.

[0117] In the secondary verification process of the denoising prediction network model, the abnormal data generated by the theoretical model is used to denoise the trained denoising prediction network model under the condition of adding not less than 5% Gaussian noise. Since the noisy data is normalized before input, it needs to be de-normalized to obtain the de-normalized result when evaluating the denoising result.

[0118] In order to obtain a result that is more in line with the requirements of subsequent boundary identification processing, maximizes noise and retains the original data characteristics, an evaluation index MSIM is introduced for judgment. The index combines the numerical accuracy term and the feature similarity term SSIM, and its calculation formula is as follows:

[0119] (7)

[0120] The closer the value of the index is to 1, the better the denoising effect of the model. The condition whether the trained denoising prediction network model meets the standard needs to be evaluated by denoising a plurality of groups of data.

[0121] In this embodiment, 10 groups of model-generated abnormal data are used, and 1% to 7% Gaussian noise is added to the data. The MSIM value of the model in the test is greater than 0.9, so 0.9 is set as the lower limit of the denoising result in this embodiment. Results less than 0.9 are considered invalid.

[0122] In addition, a theoretical model is used to verify the denoising effect and the new evaluation index MSIM. The theoretical model and its data image are shown in Figure 2 1%, 3% and 5% Gaussian noise is added to the data, and the denoising method of the application, median filtering and low-pass filtering are used for denoising calculation. The denoising results and related parameter values are shown in Figures 3 to 5 , Table 1, Table 2, Table 3 and Table 4.

[0123] Table 1 Comparison of denoising effects of different methods under different noise levels (MSE)

[0124] Noise level of the denoising method The invention Median filtering Low-pass filtering 1% 0.0309 0.2814 0.1585 3% 0.0591 1.8261 1.2136 5% 0.1552 4.5871 3.4869

[0125] Table 2 Comparison of denoising effects of different methods under different noise levels (SNR)​

[0126] Noise level of the denoising method The invention Median filtering Low-pass filtering 1% 49.8968 37.2523 34.7571 3% 42.4349 28.4105 26.6360 5% 37.3395 23.8268 22.6358

[0127] Table 3 Comparison of denoising effects of different methods under different noise levels (SSIM)

[0128] Noise level of the denoising method The invention Median filtering Low-pass filtering 1% 0.9996 0.9932 0.9915 3% 0.9986 0.9491 0.9389 5% 0.9959 0.8726 0.8501

[0129] Table 4 Comparison of denoising effects of different methods under different noise levels (MSIM)

[0130] Noise level of the denoising method The invention Median filtering Low-pass filtering 1% 0.9848 0.8867 0.9273 3% 0.9713 0.6514 0.6953 5% 0.9307 0.5257 0.5364

[0131] Tables 1 to 4 show the results of three parameters under different noise levels, proving that the model meets the requirements of an effective model. Moreover, under any noise level, the method achieves the best denoising effect. In addition, the variation law of the newly proposed MSIM index is consistent with the above three other evaluation indexes, verifying the reliability of the evaluation index. Through Figures 3 to 5 It can be seen that the denoising method has clear denoising result boundaries and does not change the original data characteristics, proving that it has good denoising effect.

[0132] Specifically, the embodiment also uses the magnetic anomaly data of Macheng iron ore area in Macheng Town, Luanan County, Hebei Province, China for denoising application, and the data used and the denoising result are as shown in Figure 6 The data scale of the used area is 64x64, and the unit length in the x direction and the y direction is 50m. By using the denoising method proposed in the application, the denoising calculation of the data of Macheng iron ore area is realized, and the obtained denoising result suppresses noise and obtains more accurate data result.

[0133] Step 6: input the noisy data into the qualified denoising prediction network model to obtain the denoising result;

[0134] Specifically, the noisy data is input into the qualified denoising prediction network model to calculate the noise weighting matrix; the noise weighting matrix and the noisy data are input into the network to obtain the denoising matrix and process the noisy data to obtain the final denoising result;

[0135] The noise weighting matrix can be obtained after the noisy data is input into the full convolution network , the noise weighting matrix and the noisy data are input into Net to obtain the denoising matrix , the sum of the denoising matrix and the noisy data is calculated to obtain the final denoising result , and the specific calculation formula is as follows:

[0136] (8)

[0137] (9)

[0138] (10)

[0139] In the formula, represents the activation function, and L represents the number of fully convolutional layers used.

[0140] Step 7: Use the denoising results to identify boundaries, obtain the correlation coefficient matrix, and then determine the location of the ore body and the structure.

[0141] Specifically, after applying the vertical guide method and the Theta diagram method to the obtained denoising results, the correlation coefficient of the boundary identification results of the two methods is calculated, and the location of the ore body and structure is determined by the correlation coefficient.

[0142] Step 7.1: Apply the vertical derivative method and the Theta plot method to the obtained denoising results to obtain the vertical derivative. Theta plot parameters The calculation formulas are as follows:

[0143]

[0144]

[0145] Step 7.2: [The sentence is incomplete and requires more context to translate accurately.] and Calculate the correlation coefficient matrix The calculation formula is as follows:

[0146]

[0147] In the formula, A represents the number of matrix elements.

[0148] Step 7.3: Determine the location of the ore body and the structure using the correlation coefficient matrix.

[0149] The correlation coefficient matrix can be obtained through the above steps. Because the boundary portion of the anomaly projected onto the horizontal plane in the vertical derivative method is a low-value area, while the Theta diagram method is the opposite, and when there are undulations on the surface of the anomaly, the boundary of the anomaly projected onto the horizontal plane obtained by the vertical derivative method and the Theta diagram method is inaccurate, the present invention can determine the ore body boundary more accurately by using the correlation coefficient method.

[0150] In this embodiment, the measured data of the Macheng area are further processed, and the processing result is as follows: Figure 7As shown, by the boundary recognition result, it can be seen that, due to the southwest inclination of the ore body, the anomaly body circled by the Theta diagram method is not accurate in the horizontal plane projection at the lower left of the figure, and after calculation by the correlation coefficient method, the determined projection region position is clear, which provides a basis for subsequent data processing and interpretation.

[0151] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present application.

Claims

1. A boundary identification method based on intelligent gravity and magnetic field data denoising, characterized in that, It comprises the following steps: Step 1: generating an original geophysical gravity and magnetic field data set using a random combination forward method; In step 1, the random combination forward method is used to randomly generate random models of different sizes, shapes, positions, densities and magnetic susceptibilities in the determined survey area, including cuboids, spheres and random walk bodies, to calculate their gravity and magnetic field anomaly data and form the original geophysical gravity and magnetic field data set; Step 2: adding random percentage Gaussian noise to all data in the original geophysical gravity and magnetic field data set to obtain a noisy data set; Step 3: constructing a denoising prediction network model; Step 4: training the denoising prediction network model using noisy data; Step 5: evaluating the effectiveness of the trained denoising prediction network model until a qualified denoising prediction network model is obtained; Step 6: inputting noisy data into the qualified denoising prediction network model to obtain denoising results; The step 6 is that the noise-containing data is input into the denoising prediction network model to obtain a denoising result, specifically, the noise-containing data is input into the full convolution network to obtain a noise weighting matrix , the noise-containing data is input into the Net to obtain a denoising matrix , the sum of the denoising matrix and the noise-containing data is calculated to obtain a final denoising result , and a specific calculation formula is as follows: ​ ； ； ； In the formula, represents an activation function, and L represents the number of used fully convolutional layers. Step 7: using the denoising results for boundary identification to obtain a correlation coefficient matrix and determine the positions of ore bodies and structures; In step 7, the denoising results are used for boundary identification to obtain a correlation coefficient matrix and determine the positions of ore bodies and structures, which comprises the following steps: Step 7.1: Using vertical derivative method and Theta map method on the obtained denoised result, vertical derivative and Theta map method parameters The calculation formula is respectively: ； ； Step 7.2: To the obtained and The correlation coefficient matrix is calculated with the formula: ； In the formula, A is the number of matrix elements; Step 7.3: determining the positions of ore bodies and structures through the correlation coefficient matrix. 2.The boundary identification method based on intelligent gravity and magnetic field data denoising according to claim 1, characterized in that, The random model parameter setting formula is expressed as follows: First, determine the center position of the random model M, set M as a random value, and the value range conforms to the following formula: ； wherein respectively represent the coordinates of the center position of M in three directions, respectively represent the range of the observation space in the x, y, and z axis directions; If it is a random cuboid, the parameters of the random model are: ； In the formula, denote the parameters of a cuboid model, and l, w and h denote the long side, short side and height of the cuboid, respectively; If it is a random sphere, the parameters of the random model are: ； wherein denotes the parameters of the sphere model, d denotes the sphere model radius; If it is a random walk body, the parameters of the random model are: ； wherein denotes the parameter of random walk, where a and t represent the size of the edge of a single cubic voxel and the number of cubes, respectively. 3.The boundary identification method based on intelligent gravity and magnetic field data denoising according to claim 2, characterized in that, The noisy data set obtained in step 2 is: after adding random percentage Gaussian noise to the original data corresponding to the noisy data generated by adding random percentage Gaussian noise to all data in the original geophysical gravity and magnetic field data set, obtaining the noisy data set, and dividing the training set and the test set according to the preset proportion.

4. The boundary identification method based on intelligent gravity and magnetic field data denoising according to claim 3, characterized in that, The noisy data set is obtained by adding random percentage Gaussian noise with a maximum value of 1% to 7% to all data in the original geophysical gravity and magnetic field data set, and adding noise to obtain the noisy data set according to the following formula: ； wherein and denote the obtained noisy data and the data to be noisy, respectively, is a random integer between 1 and 7.

5. The boundary identification method based on intelligent gravity and magnetic field data denoising according to claim 4, characterized in that, The constructed denoising prediction network model comprises a full convolution network layer and an encoder-decoder Net; wherein the convolution layer in the full convolution network layer comprises at least 4 layers, and the structure of the Net comprises at least two down-sampling modules and two up-sampling modules; the up-sampling module comprises at least one convolution layer and one up-sampling layer, and the down-sampling module comprises at least one convolution layer and one down-sampling layer. 6.The boundary identification method based on intelligent gravity and magnetic field data denoising according to claim 5, characterized in that, Step 4 uses noisy data to train the denoising prediction network model, which comprises the following steps: Step 4.1: preprocessing the training set data to obtain a preprocessed data set; Step 4.2: inputting the preprocessed data set into the denoising prediction network model for training and testing; In the training and testing process, the loss function is calculated as follows: ； where n is the number of elements of the matrix x, is the denoised result of the output, is the normalized noisy data.

7. The boundary identification method based on intelligent gravity and magnetic field data denoising according to claim 6, characterized in that, The step 5 is to evaluate the effectiveness of the trained denoising prediction network model until a qualified denoising prediction network model is obtained, specifically: firstly, the trained denoising prediction network model is preliminarily evaluated for effectiveness by using a loss function, and then the denoising prediction network model is verified again by using multiple sets of theoretical model data until a qualified denoising prediction network model is obtained.

Images