Gravity gradient data potential field conversion method and system based on U-net deep neural network multi-task learning

Abstract

The invention discloses a gravity gradient data potential field conversion method and system based on U-net deep neural network multi-task learning, and relates to the technical field of gravity field data conversion, and the method comprises the following steps: constructing a geologic model body of an underground model, carrying out the forward modeling of the geologic model body, and constructing a component data set; training a U-net multi-task network by using the component data set to obtain a data conversion model; and inputting actual gravity data into the data conversion model to obtain converted gravity data. According to the method, gradient component conversion can be rapidly carried out, and the gravitational field map of a complex geologic body model can be predicted. Therefore, the network is strong in generalization.

Description

Technical Field

[0001] The present invention belongs to the technical field of gravity field data conversion, and specifically relates to a gravity gradient data potential field conversion method and system based on multi-task learning of a U-net deep neural network. Background Technique

[0002] Gravity exploration has become an important technical means for solving geological problems such as oil and gas, minerals, and regional geotectonic research due to its advantages of low economic cost, wide detection range, easy implementation, and short application cycle. People obtain the underground spatial distribution and density amplitude distribution range of the target geological structure by measuring the gravity anomaly generated by the geological body with uneven density inside the earth. In recent years, with the continuous improvement of the detection accuracy and efficiency of gravity exploration instruments, the obtained high-precision gravity gradient data can be effectively used to improve the density imaging of gravity exploration.

[0003] How to convert the gravity data g of relative Bouguer vertical gravity z into high-precision gradient gravity data is of crucial importance. Agarwal (Agarwal BNP, Lal T. Calculation of the second vertical derivative of gravity field[J]. Pure and Applied Geophysics, 1969, 76(1):5-16. DOI:10.1007 / BF00877833) obtained the gravity gradient data of other components based on the Laplace equation. However, this method still requires the gradient data in one direction to obtain other components. Along with the innovation of modern computer technology and digital signal analysis theory, many derivative calculation methods in the spatial domain have been developed, such as the least squares method, spline function interpolation method, Taylor series expansion method, etc. However, these methods have a greater impact on the results of noisy data. Geophysicists have tried to convert the Bouguer vertical gravity data g zTransform to the frequency domain, then apply the corresponding derivative transformation operator, and finally inverse transform the result back to the spatial domain to achieve the conversion between potential field data (Hou Chongchu, Liu Xiufang. Conversion of gravity and magnetic potential fields using two-dimensional Fourier transform [J]. Journal of Beijing Normal University: Natural Science Edition, 1978(2):16); Ma Guoqing (Ma Guoqing, Huang Danian, Du Xiaojuan, etc. Application of Hartley transform in the calculation of potential field (gravity, magnetic) anomaly derivatives [J]. Journal of Jilin University: Earth Science Edition, 2014(1):8. DOI: 10.13278 / j.cnki.jjuese.201401301) adopted the Hartley transform in the calculation of potential field data derivatives and successfully completed the identification of the magnetic anomaly boundary in the Zhurihe area. However, the above methods are all carried out in the frequency domain. Due to the influence of the sampling interval size of potential field data, there will be burr phenomena on the boundary of potential field data and the changes are drastic. At the same time, since the wavenumber domain derivative transformation operator is equivalent to a high-pass filter, it is easily affected by high-frequency interference, resulting in an unstable calculation process, and it must be corrected to ensure the calculation accuracy. Tai Zhenhua (Tai Zhenhua. Research and application of high-precision processing methods for potential field data [D]. Jilin University, 2016) adopted the idea of the iterative method and calculated the corresponding filtering operator using the Taylor expansion. This method effectively enhanced the high-frequency signal and suppressed the low-frequency signal. However, this method using the iterative method leads to slow calculation efficiency. If there is noise in the data, it will cause errors in the conversion of high-order potential field data.)

[0004] In recent years, with the rapid development of deep learning, people train through a large number of data sets to construct implicit relationships between various data. Therefore, the present invention adopts the U-net deep neural network. By inputting the g z component data, the g zz 、g xx 、g yy component data can be obtained quickly. Considering the differences in the g zz 、g xx 、g yy component data, a multi-task learning method is adopted, and respective corresponding network channels are established. At the same time, the Laplace regularization term is added for physical constraint learning to construct the physical relationship of the three-component data. The theoretical results are calculated by constructing an underground model and compared and verified with the method of the present invention.) Summary of the Invention

[0005] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0006] A method for potential field conversion of gravity gradient data based on multi-task learning of U-net deep neural network, comprising the following steps:

[0007] Construct a geological model body for the subsurface model, and perform forward modeling on the geological model body to construct a component dataset;

[0008] Train a U-net multi-task network using the component dataset to obtain a data conversion model.

[0009] Input actual gravity data into the data conversion model to obtain converted gravity data.

[0010] Preferably, the method for constructing the geological model body includes:

[0011] Construct the geological model body in the shape of a rectangular parallelepiped. Obtain the characteristics of the rectangular parallelepiped: the length, width, and height of the rectangular parallelepiped and the position of the spatial center point of the rectangular parallelepiped .

[0012] Use the normal distribution function to construct the length, width, and height of the rectangular parallelepiped :

[0013] ,

[0014] where f(t) represents the normal distribution function, σ represents one-third of the maximum range of the model space, and μ represents one-fifth of the maximum range of the model space.

[0015] Use the uniform distribution function to construct the position of the spatial center point of the rectangular parallelepiped :

[0016] ,

[0017] where f(s) represents the uniform distribution function, a represents the minimum range of the model space, and b represents the maximum range of the model space.

[0018] Preferably, the method for obtaining the component dataset through forward modeling includes:

[0019] ,

[0020] ,

[0021] ,

[0022] ,

[0023] ,

[0024] where g z represents the gravity data of the geological model body, g zz 、g xx 、g yyRepresent the three components of gravity data, G represents the universal gravitational constant, ρ represents the residual density, and r ijk represents the distance between the observation point and the spatial center point of the geological model body, and i, j, k represent the vertex coordinates of the geological model body, x i represents the difference between the abscissa of the observation point and the abscissa of the underground model, y j represents the difference between the ordinate of the observation point and the ordinate of the underground model, z k represents the difference between the depth coordinate of the observation point and the depth coordinate of the underground model.

[0025] Preferably, the method for training the U-net multi-task network includes:

[0026] Construct the U-net multi-task network, and the U-net multi-task network includes: a number of 3×3 convolutional kernels, a number of max-pooling layers and a number of upsampling layers.

[0027] Taking the gravity data g z as the input, the components g zz 、g xx and g yy as the output, and constructing a loss function with the Laplace equations of the three components as physical constraints, training the U-net multi-task network to obtain the data conversion model.

[0028] Preferably, the loss function is:

[0029] ,

[0030] where loss represents the loss function, σ1 and σ2 represent the weights of the loss function, and l gzz 、l gxx 、l gyy represent the label loss functions corresponding to the three gradient components, represents the Laplace equation.

[0031] The present invention also provides a gravity gradient data potential field conversion system based on U-net deep neural network multi-task learning. The conversion system applies the conversion method described in any one of the above, and includes: a data set construction module, a model training module and a data conversion module.

[0032] The data set construction module is used to construct a geological model body of the underground model and perform forward modeling on the geological model body to construct a component data set.

[0033] The model training module uses the component data set to train the U-net multi-task network to obtain a data conversion model.

[0034] The data conversion module is used to input actual gravity data into the data conversion model to obtain converted gravity data.

[0035] Preferably, in the dataset construction module, the process of constructing the geological model body includes:

[0036] Construct the geological model body in the shape of a rectangular parallelepiped. Obtain the characteristics of the rectangular parallelepiped: the length, width and height of the rectangular parallelepiped and the position of the spatial center point of the rectangular parallelepiped .

[0037] Use the normal distribution function to construct the length, width and height of the rectangular parallelepiped :

[0038] ,

[0039] where f(t) represents the normal distribution function, σ represents one-third of the maximum range of the model space, and μ represents one-fifth of the maximum range of the model space.

[0040] Use the uniform distribution function to construct the position of the spatial center point of the rectangular parallelepiped :

[0041] ,

[0042] where f(s) represents the uniform distribution function, a represents the minimum range of the model space, and b represents the maximum range of the model space.

[0043] Preferably, in the dataset construction module, the process of obtaining the component dataset through forward modeling includes:

[0044] ,

[0045] ,

[0046] ,

[0047] ,

[0048] ,

[0049] where g z represents the gravity data of the geological model body, g zz , g xx , g yy represent the three components of the gravity data, G represents the gravitational constant, ρ represents the residual density, r ijk represents the distance between the observation point and the spatial center point of the geological model body, i, j, k represent the vertex coordinates of the geological model body, x i represents the difference between the abscissa of the observation point and the abscissa of the underground model, y jIt represents the difference between the ordinate of the observation point and the ordinate of the subsurface model, z k It represents the difference between the depth coordinate of the observation point and the depth coordinate of the subsurface model.

[0050] Preferably, the working process of the model training module includes:

[0051] Construct the U-net multi-task network, which includes: several 3×3 convolutional kernels, several max-pooling layers and several upsampling layers.

[0052] Using the gravity data g z as the input, and the components g zz , g xx and g yy as the output, and constructing a loss function with the Laplace equations of the three components as physical constraints, training the U-net multi-task network to obtain the data conversion model.

[0053] Preferably, the loss function is:

[0054] ,

[0055] where loss represents the loss function, σ1 and σ2 represent the weights of the loss function, and l gzz , l gxx , l gyy represent the label loss functions corresponding to the three gradient components, represents the Laplace equation.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] The present invention uses the deep learning network U-net to analyze geophysical exploration data. By establishing the relationship between the input data g z and the output data g zz , g xx , g yy components, and at the same time adopting the multi-task learning mode to establish each adapted network structure, and considering that the output three-component data satisfies the Laplace equation, so a Laplace regularization term is added as a physical constraint for learning. The data conversion of the present invention has the characteristics of higher accuracy and faster speed. At the same time, the network model has more physical properties, which can not only quickly perform gradient component conversion, but also predict the gravity field map of complex geological body models. This fully shows that the network has strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] To more clearly illustrate the technical solution of the present invention, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.

[0059] Figure 1 It is a schematic flowchart of the method according to an embodiment of the present invention.

[0060] Figure 2 It is a schematic diagram of the geological model body according to an embodiment of the present invention.

[0061] Figure 3 It is a schematic structural diagram of the U-net multi-task network according to an embodiment of the present invention.

[0062] Figure 4 It is a schematic diagram of the Laplace equation of the three gradient components according to an embodiment of the present invention.

[0063] Figure 5 It is the loss function and its accuracy curve according to an embodiment of the present invention.

[0064] Figure 6 They are three separated geological rectangular bodies according to an embodiment of the present invention. Among them, (a) is the true g zz component, (b) is the true g xx component, (c) is the true g yy component, (d) is the predicted g zz component, (e) is the predicted g xx component, (f) is the predicted g yy component.

[0065] Figure 7 They are three coupled geological rectangular bodies according to an embodiment of the present invention. Among them, (a) is the true g zz component, (b) is the true g xx component, (c) is the true g yy component, (d) is the predicted g zz component, (e) is the predicted g xx component, (f) is the predicted g yy component. Detailed implementation manners

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0067] To make the above objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0068] Embodiment 1

[0069] In this embodiment, as Figure 1 shown, the gravity gradient data potential field conversion method based on multi-task learning of the U-net deep neural network includes the following steps:

[0070] S1. Construct a geological model body of the underground model, and perform forward modeling on the geological model body to construct a component data set.

[0071] The method for constructing the geological model body includes: constructing a geological model body in the shape of a rectangular body. Obtain the characteristics of the rectangular body: the length, width, and height of the rectangular body and the position of the spatial center point of the rectangular body . Use the normal distribution function to construct the length, width, and height of the rectangular body . Use the uniform distribution function to construct the position of the spatial center point of the rectangular body .

[0072] In this embodiment, since both the gravity data and the gravity gradient data are abnormal data generated by geological objects with uneven underground density distribution. Therefore, the first step is to construct a geological model body of the underground model. In this embodiment, a rectangular body is used as the geological model body, as Figure 2 shown. The rectangular body has 6 characteristics, namely the length, width, and height of the rectangular body , and its spatial center point position . Use the probability function method to construct 5000 sets of geological model bodies for its 6 characteristics. Among them, the length, width, and height of the rectangular body are constructed using the normal distribution function:

[0073] ,

[0074] where f(t) represents the normal distribution function, σ represents one-third of the maximum range of the model space, and μ represents one-fifth of the maximum range of the model space. And the spatial center point position is constructed using the uniform distribution function:

[0075] ,

[0076] where f(s) represents the uniform distribution function, a represents the minimum range of the model space, and b represents the maximum range of the model space.

[0077] After determining the geological model body, it is necessary to perform forward modeling on the geological model body to obtain a component data set through forward modeling:

[0078] ,

[0079] ,

[0080] ,

[0081] ,

[0082] ,

[0083] Among them, g z represents the gravity data of the geological model body, g zz , g xx , g yy represent the three components of the gravity data, G represents the universal gravitational constant, ρ represents the residual density, r ijk represents the distance between the observation point and the spatial center point of the geological model body, i, j, k represent the vertex coordinates of the geological model body, x i represents the difference between the abscissa of the observation point and the abscissa of the underground model, y j represents the difference between the ordinate of the observation point and the ordinate of the underground model, z k represents the difference between the depth coordinate of the observation point and the depth coordinate of the underground model.

[0084] According to the above method, a total of 5000 sets of data sets are established, of which 80% are used for the training set and 20% are used for the test set: Two separated anomalies are randomly established in the geological model body, and the generated g z forward modeling field map is obtained, that is, the total anomaly generated by the underground anomaly on the observation surface. At the same time, the g zz , g xx , g yy three-component forward modeling field maps are obtained. The g yy component mainly describes the vertical boundary of the anomaly body, the g xx component mainly describes the horizontal boundary of the anomaly body, and the g zz component is the derivative of g z in the z direction, mainly depicting the boundary of the model. Therefore, observing the g zz component in this embodiment can clearly distinguish the boundary position of the model. Since the sizes of the four component field maps are all (64, 64), there will be no information loss caused by upsampling and downsampling of the model in the present invention.

[0085] S2. Use the component data set to train the U-net multi-task network to obtain a data conversion model.

[0086] The method for training a U-net multi-task network includes: constructing a U-net multi-task network, where the U-net multi-task network includes: several 3×3 convolutional kernels, several max pooling layers, and several upsampling layers; using the gravity data g z as the input, and the components g zz , g xx and g yy as the output, and constructing a loss function with the Laplace equations of the three components as the physical constraints, and training the U-net multi-task network to obtain a data conversion model.

[0087] In this embodiment, considering that the morphologies and amplitudes of the three gradient components (g zz , g xx , g yy ) are different, a U-net multi-task network structure is used as the initial network training model. Among them, the input end is the g z component, and the output end is the g zz , g xx , g yy components. The network structure is as shown in Figure 3 , including: several 3×3 convolutional kernels, several max pooling layers, and several upsampling layers. The working process of this network includes: the input end is the field map of the g z component, passing through a 3×3 convolutional kernel for dimensionality increase operation to become data features with 64 channels. Then, max pooling layer downsampling is performed, that is, the size of the input feature map becomes half of the original, as shown in the following formula:

[0088] ,

[0089] ,

[0090] ,

[0091] where, F X is a matrix of size H×W, then the output F YIts size is H / 2×W / 2; then the channel dimension is increased to 128, and so on, until the final number of channels becomes 512 channels and the feature map size becomes H / 8×W / 8. This part is the encoder of the network, whose main purpose is to extract the features of the input data. Each step of such an operation depends on the output feature information of the previous step. It has a powerful representation ability and can capture the complex mapping relationship between input and output. After encoding, the extracted features need to be decoded. The decoder has strong flexibility, is suitable for input and output sequences of various lengths, and can handle different types of sequence conversion tasks. The upsampling in the middle part mainly converts the encoded feature matrix into data of size H×W. Before passing through upsampling, the features extracted from each layer of the original encoding part also need to be concatenated. The purpose of this is to prevent information loss during the learning process of the network, and thus the method of concatenation is used to retain the original information. The convolutional operations in the last three parts of the encoder are based on a multi-task network structure. By inputting g z it is found that the forms and amplitudes of g zz 、g xx 、g yy at the output end are all different. Therefore, convolutional channels that conform to each feature need to be established. In addition, since the high-order components of gravity satisfy the Laplace equation, that is, the sum of g zz 、g xx 、g yy components is equal to zero, as shown in Figure 4 it can be found that the sum of each element of the three gradient components is equal to zero. Therefore, the Laplace method is used as the physical constraint in the loss function, and the loss function is:

[0092] ,

[0093] where loss represents the loss function, σ1 and σ2 represent the weights of the loss function, and l gzz 、l gxx 、l gyy represent the label loss functions corresponding to the three gradient components, represents the Laplace equation. Since the above three components represent the second-order derivatives of the gravity potential, and at the same time the gravity potential is a conservative force field, the sum of the corresponding elements of the even-order derivatives is 0 (as shown in Figure 4 ), so in this embodiment, is added to the traditional deep learning loss function as a physical constraint.

[0094] S3. Input the actual gravity data into the data conversion model to obtain the converted gravity data.

[0095] Embodiment 2

[0096] In this embodiment, the effectiveness of the present invention will be verified.

[0097] Figure 5 The loss function and its accuracy in Embodiment 1 are shown. As can be seen from the figure, the loss function in this network structure converges well, and there is no overfitting state. From the accuracy curve, the accuracy of the validation set can reach more than 95%. This shows that the method we provided has good effects to a certain extent.

[0098] In Figure 6 , three abnormal geological model bodies with different densities and burial depths are set, and the theoretical Bouguer gravity anomaly is calculated and put into the network for prediction. Among them Figure 6 (a), Figure 6 (b) and Figure 6 (c) are three high-order gradient components calculated theoretically. Figure 6 (d), Figure 6 (e) and Figure 6 (f) are the prediction results of the present invention. It can be clearly found that three abnormal bodies can be found in the predicted g zz component, and the boundaries are clearly depicted. The g xx component and the g yy component depict the lateral and longitudinal boundaries of the geological body respectively. At the same time, the predicted three-component results are all consistent with the theoretically calculated results.

[0099] In Figure 7 , three abnormal geological model bodies with different densities and burial depths are set, but the three abnormal rectangular blocks are coupled together, and the theoretical Bouguer gravity anomaly is calculated and put into the network for prediction. Among them Figure 7 (a), Figure 7 (b) and Figure 7 (c) are three high-order gradient components calculated theoretically. Figure 7 (d), Figure 7 (e) and Figure 7 (f) are the prediction results of the present invention. It can be clearly found that the predicted three-component results are all consistent with the theoretically calculated results. Therefore, our method can not only quickly perform gradient component conversion, but also predict the gravity field map of complex geological body models. This fully shows that the network has strong generalization ability.

[0100] Embodiment 3

[0101] In this embodiment, a gravity gradient data potential field conversion system based on multi-task learning of U-net deep neural network is characterized by including: a data set construction module, a model training module and a data conversion module.

[0102] The dataset construction module is used to construct the geological model body of the underground model and perform forward modeling on the geological model body to construct the component dataset.

[0103] In the dataset construction module, the process of constructing the geological model body includes: constructing a geological model body in the shape of a rectangular prism. Obtaining the characteristics of the rectangular prism: the length, width, and height of the rectangular prism and the position of the spatial center point of the rectangular prism . Using the normal distribution function to construct the length, width, and height of the rectangular prism :

[0104] ,

[0105] where f(t) represents the normal distribution function, σ represents one-third of the maximum range of the model space, and μ represents one-fifth of the maximum range of the model space. Using the uniform distribution function to construct the position of the spatial center point of the rectangular prism :

[0106] ,

[0107] where f(s) represents the uniform distribution function, a represents the minimum range of the model space, and b represents the maximum range of the model space.

[0108] In the dataset construction module, the process of obtaining the component dataset through forward modeling includes:

[0109] ,

[0110] ,

[0111] ,

[0112] ,

[0113] ,

[0114] where g z represents the gravity data of the geological model body, g zz , g xx , g yy represent the three components of the gravity data, G represents the gravitational constant, ρ represents the residual density, r ijk represents the distance between the observation point and the spatial center point of the geological model body, i, j, k represent the vertex coordinates of the geological model body, x i represents the difference between the abscissa of the observation point and the abscissa of the underground model, y j represents the difference between the ordinate of the observation point and the ordinate of the underground model, z k represents the difference between the depth coordinate of the observation point and the depth coordinate of the underground model.

[0115] The model training module trains the U-net multi-task network using the component data sets to obtain a data conversion model.

[0116] The working process of the model training module includes: constructing a U-net multi-task network, which includes: a number of 3×3 convolutional kernels, a number of max pooling layers and a number of upsampling layers. Using the gravity data g z as the input, and the components g zz , g xx and g yy as the outputs, and constructing a loss function with the Laplace equations of the three components as physical constraints to train the U-net multi-task network to obtain a data conversion model.

[0117] It is characterized in that the loss function is:

[0118] ,

[0119] where loss represents the loss function, σ1 and σ2 represent the weights of the loss function, and l gzz , l gxx , l gyy represent the label loss functions corresponding to the three gradient components, represents the Laplace equation.

[0120] The data conversion module is used to input the actual gravity data into the data conversion model to obtain the converted gravity data.

[0121] The above-described embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A gravity gradient data potential field conversion method based on multi-task learning of U-net deep neural network, characterized in that Including the following steps: Construct a geological model body of the underground model, and perform forward modeling on the geological model body to construct a component data set; Use the component data set to train the U-net multi-task network to obtain a data conversion model; Input the actual gravity data into the data conversion model to obtain the converted gravity data.

2. The gravity gradient data potential field conversion method based on multi-task learning of U-net deep neural network according to claim 1, characterized in that The method for constructing the geological model body includes: Construct the geological model body in the shape of a rectangular body; obtain the characteristics of the rectangular body: the length, width and height of the rectangular body and the position of the spatial center point of the rectangular body ; Construct the length, width, and height of the rectangular body using the normal distribution function : ， where f(t) represents the normal distribution function, σ represents one-third of the maximum range of the model space, and μ represents one-fifth of the maximum range of the model space; Construct the spatial center point position of the rectangular body by using the uniform distribution function : ， where f(s) represents the uniform distribution function, a represents the minimum range of the model space, and b represents the maximum range of the model space.

3. The gravity gradient data potential field conversion method based on multi-task learning of U-net deep neural network according to claim 1, characterized in that The method for obtaining the component data set through forward modeling includes: ， ， ， ， ， where, g z represents the gravity data of the geological model body, g zz , g xx , g yy represent the three components of the gravity data, G represents the universal gravitational constant, ρ represents the residual density, r ijk represents the distance between the observation point and the spatial center point of the geological model body, i, j, k represent the vertex coordinates of the geological model body, x i represents the difference between the abscissa of the observation point and the abscissa of the underground model, y j represents the difference between the ordinate of the observation point and the ordinate of the underground model, z k represents the difference between the depth coordinate of the observation point and the depth coordinate of the underground model.

4. The gravity gradient data potential field conversion method based on multi-task learning of U-net deep neural network according to claim 3, characterized in that The method for training the U-net multi-task network includes: Construct the U-net multi-task network, which includes: a number of 3×3 convolutional kernels, a number of max pooling layers, and a number of upsampling layers; With the gravity data g z as the input, the components g zz , g xx and g yy as the output, and using the Laplace equations of the three components as physical constraints to construct a loss function, training the U-net multi-task network to obtain the data conversion model.

5. The gravity gradient data potential field conversion method based on multi-task learning of U-net deep neural network according to claim 4, characterized in that The loss function is: ， where loss represents the loss function, σ1 and σ2 represent the weights of the loss function, l gzz 、l gxx 、l gyy represent the label loss functions corresponding to three gradient components, represents the Laplace equation.

6. A gravity gradient data potential field conversion system based on multi-task learning of the U-net deep neural network, the conversion system applying the conversion method according to any one of claims 1-5, characterized in that, Including: A data set construction module, a model training module, and a data conversion module; The data set construction module is used to construct a geological model body of the underground model, and perform forward modeling on the geological model body to construct a component data set; The model training module uses the component data set to train the U-net multi-task network to obtain a data conversion model; The data conversion module is used to input the actual gravity data into the data conversion model to obtain the converted gravity data.

7. The gravity gradient data potential field conversion system based on multi-task learning of the U-net deep neural network according to claim 6, characterized in that In the data set construction module, the process of constructing the geological model body includes: Construct the geological model body in the shape of a rectangular body; obtain the characteristics of the rectangular body: the length, width and height of the rectangular body and the position of the spatial center point of the rectangular body ; Construct the length, width, and height of the rectangular body using the normal distribution function : ， where f(t) represents the normal distribution function, σ represents one-third of the maximum range of the model space, and μ represents one-fifth of the maximum range of the model space; Construct the spatial center point position of the rectangular body by using the uniform distribution function : ， where f(s) represents the uniform distribution function, a represents the minimum range of the model space, and b represents the maximum range of the model space.

8. The gravity gradient data potential field conversion system based on multi-task learning of U-net deep neural network according to claim 6, characterized in that, In the data set construction module, the process of obtaining the component data set through forward modeling includes: ， ， ， ， ， Among them, g z represents the gravity data of the geological model body, g zz , g xx , g yy represent the three components of the gravity data, G represents the gravitational constant, ρ represents the residual density, r ijk represents the distance between the observation point and the spatial center point of the geological model body, i, j, k represent the vertex coordinates of the geological model body, x i represents the difference between the abscissa of the observation point and the abscissa of the underground model, y j represents the difference between the ordinate of the observation point and the ordinate of the underground model, z k represents the difference between the depth coordinate of the observation point and the depth coordinate of the underground model.

9. The gravity gradient data potential field conversion system based on multi-task learning of U-net deep neural network according to claim 8, characterized in that, The work process of the model training module includes: Construct the U-net multi-task network, which includes: a number of 3×3 convolutional kernels, a number of max pooling layers, and a number of upsampling layers; With gravity data g z as the input, components g zz , g xx and g yy as the output, and using the Laplace equations of the three components as physical constraints to construct a loss function, training the U-net multi-task network to obtain the data conversion model.

10. The gravity gradient data potential field conversion system based on multi-task learning of U-net deep neural network according to claim 9, characterized in that, The loss function is: ， where loss represents the loss function, σ1 and σ2 represent the weights of the loss function, and l gzz 、l gxx 、l gyy represent the label loss functions corresponding to three gradient components, represents the Laplace equation.

Images