Electromagnetic exploration fault identification method and device based on deep learning

Abstract

The invention relates to the field of geophysical electromagnetic exploration intelligent identification, in particular to an electromagnetic exploration fault identification method based on deep learning, which comprises the following steps: constructing a fault identification deep learning model; obtaining an existing resistivity inversion profile data set to train a deep learning model; and taking the actual resistivity inversion section as input, and outputting a fault distribution result by using the trained deep learning model. According to the method, the theoretical fault model resistivity inversion profile and the actual fault model resistivity inversion profile jointly form the training set for training the deep learning model based on the image feature data driving of the electromagnetic exploration profile, so that the accuracy and efficiency of the deep learning model in fault automatic identification are remarkably improved. The invention further provides a calculation device based on the fault identification method.

Description

Technical Field

[0001] The present invention relates to the field of intelligent identification of geophysical electromagnetic exploration, and in particular to a method and device for electromagnetic exploration fault identification based on deep learning. Background Art

[0002] Electromagnetic exploration has been widely used in the fields of oil and gas, mineral and geothermal resource exploration, engineering geological exploration, and geological disaster prevention due to its convenience, low cost, deep detection, wide range, and high accuracy. However, the research on electromagnetic exploration is mainly focused on electromagnetic exploration methods, electromagnetic forward inversion algorithms, and electrical interface imaging, but the interpretation of faults is less involved. However, accurate identification of fault locations is crucial in both geological resource exploration and engineering geological exploration. For example, in the field of oil and gas exploration, faults can not only serve as channels for oil and gas migration, but can also laterally block oil and gas to form structural oil and gas reservoirs. Therefore, fault identification is a basic task in electromagnetic data interpretation.

[0003] At present, the conventional fault interpretation method of electromagnetic exploration is to manually identify faults through frequency-pseudo-section diagrams, electromagnetic inversion profiles and geological knowledge. This method is difficult to identify faults, has a long period of time and relies on subjective consciousness, and requires extremely high prior knowledge of people. At the same time, since the electrical difference on both sides of the fault is mainly reflected by the different conductivity of the strata on both sides, and the conductivity difference between the strata on both sides is very small or the fault distance is small, the electrical difference characteristics presented on the profile are not obvious, and it is difficult for the human eye to identify the fault position. Therefore, finding a method for electromagnetic exploration data fault identification that does not rely on human factors and is accurate and effective has become an urgent problem to be solved. Summary of the invention

[0004] According to one aspect of the present application, a method for electromagnetic exploration fault identification based on deep learning is provided, including: constructing a deep learning model for fault identification; acquiring an existing resistivity inversion profile data set to train the deep learning model; taking the actual resistivity inversion profile as input, and using the trained deep learning model to output the fault distribution result.

[0005] As a preferred implementation, the existing resistivity inversion profile data set includes a theoretical fault model resistivity inversion profile data set and an actual fault model resistivity inversion profile data set.

[0006] The resistivity inversion profile data set of the theoretical fault model is obtained by the following steps:

[0007] A plurality of undulating stratum resistivity models are established according to the Cauchy distribution function; a straight line is fitted to any two points of each of the undulating stratum resistivity models that are not at the same height, and the data in its domain is moved up / down with the straight line as the boundary to form a fault; a plurality of undulating stratum resistivity models with formed faults are stored as theoretical fault models, and the left and right boundaries of the fault are marked as fault labels; the theoretical fault model is forward simulated to obtain forward apparent resistivity data; the forward apparent resistivity data is inverted to output resistivity inversion profile data of the theoretical fault model.

[0008] As a preferred implementation, the deep learning model is a Res-UNet network model improved based on the UNet network.

[0009] As another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0010] As another aspect of the present application, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] The deep learning-based electromagnetic exploration data fault identification method provided by the present invention is driven by the image feature data of the electromagnetic exploration profile. The theoretical fault model resistivity inversion profile and the actual fault model resistivity inversion profile are combined into a training set for training the deep learning model, which significantly improves the accuracy and efficiency of the deep learning model in automatic fault identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flow chart of a method for identifying faults in electromagnetic exploration based on deep learning according to an embodiment of the present application is shown;

[0014] Figure 2 The horizontal formation resistivity model of the embodiment of the present application is shown;

[0015] Figure 3 The resistivity models of undulating formations of different forms according to the embodiments of the present application are shown;

[0016] Figure 4 shows different numbers of fault resistivity models and their labels according to an embodiment of the present application;

[0017] Figure 5 The forward modeling data of a certain acquisition point of the theoretical fault model of the embodiment of the present application is shown;

[0018] Figure 6 The synthetic fault resistivity model and its inversion profile of the embodiment of the present application are shown;

[0019] Figure 7 A schematic diagram of the Res-UNet network model of an embodiment of the present application is shown;

[0020] Figure 8 The actual resistivity inversion profile of the embodiment of the present application is shown;

[0021] Fig. 9 The figure shows the fusion diagram of the actual resistivity inversion section and the real fault label of the embodiment of the present application;

[0022] Fig.10 A fusion diagram of the actual resistivity inversion profile and the fault identification result of the embodiment of the present application is shown. DETAILED DESCRIPTION

[0023] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and is not a limitation of the scope of protection, applicability or examples set forth in the claims. The function and arrangement of the elements discussed can be changed without departing from the scope of protection of the present disclosure. Various examples can omit, replace or add various processes or components as needed. For example, the described method can be performed in an order different from the described order, and various steps can be added, omitted or combined. In addition, the features described relative to some examples can also be combined in other examples.

[0024] It should be noted that the references to "one embodiment", "embodiment", "some embodiments" and the like in the specification indicate that the described embodiments may include specific features, structures or characteristics, but not every embodiment may include the specific features, structures or characteristics. In addition, such expressions do not necessarily refer to the same embodiment. In addition, when specific features, structures or characteristics are described in conjunction with an embodiment, it should be within the knowledge of technicians in the relevant field to implement such features, structures or characteristics in conjunction with other embodiments that are explicitly or not explicitly described.

[0025] Embodiments of the electromagnetic exploration fault identification method and device based on deep learning according to the present invention will now be described with reference to the accompanying drawings.

[0026] A fault identification method for electromagnetic exploration based on deep learning, such as Figure 1 As shown, it includes: building a deep learning model for fault identification; obtaining an existing resistivity inversion profile data set to train the deep learning model; taking the actual resistivity inversion profile as input, and using the trained deep learning model to output the fault distribution results.

[0027] As a preferred embodiment, the existing resistivity inversion profile data set includes a theoretical fault model resistivity inversion profile data set and an actual fault model resistivity inversion profile data set.

[0028] The theoretical fault model resistivity inversion profile data set is obtained by the following steps: establishing a plurality of undulating stratum resistivity models according to the Cauchy distribution function; fitting straight lines with any two points of each of the undulating stratum resistivity models that are not at the same height, and moving the data in its domain up / down with the straight line as the boundary to form a fault; storing a plurality of undulating stratum resistivity models with formed faults as theoretical fault models, and marking the left and right boundaries of the faults as fault labels; forward modeling the theoretical fault models to obtain forward apparent resistivity data; inverting the forward apparent resistivity data to output theoretical fault model resistivity inversion profile data.

[0029] Creating theoretical fault models and fault labels with various shapes and in line with geological characteristics is the key to efficient and accurate fault identification. The following describes in detail the steps of obtaining the resistivity inversion profile data set of the theoretical fault model.

[0030] Creating a resistivity model of an undulating formation

[0031] First, multiple multi-layer horizontal formation resistivity models are established. In order to increase the diversity of the models, the formation thickness, number of formation layers and resistivity value of each model can be randomly set according to the needs. Figure 2 Based on the horizontal formation resistivity model, the Cauchy distribution function is used to simulate the lateral undulation of the actual formation. The function expression is as follows:

[0032]

[0033] Where: A is the height amplitude parameter for controlling the undulating formation, x0 is the position parameter for controlling the center point of the undulating formation, and γ is the range parameter for controlling the undulating formation;

[0034] When the undulating stratum resistivity model is established according to the Cauchy distribution function from the horizontal stratum resistivity model, different stratum undulation forms are constructed by adjusting different amplitude parameters A, scale parameters γ and position parameters x0. First, the point number of the horizontal stratum resistivity model is used as the independent variable x, and then the independent variable x is substituted into the Cauchy distribution function f(x,x0,γ) to calculate the function value corresponding to each point number, and the calculated function value is uniformly rounded to obtain F(x,x0,γ), and then the column data corresponding to each point number is moved up or down as a whole by F(x,x0,γ) units. The data that overflows the array boundary during the movement is discarded. Since the units of movement of each column data are different and determined by the Cauchy distribution function, the stratum model will form a lateral undulation form similar to the Cauchy distribution function. In this process, different amplitude parameters A, scale parameters γ and position parameters x0 can be set to construct different stratum undulation forms, thereby improving the diversity of stratum models. Some models such as Figure 3 shown.

[0035] Create theoretical fault resistivity model and label

[0036] Multiple straight lines are fitted with each of the undulating formation resistivity models to form multiple faults, and different types of faults are formed by adjusting the slope and movement distance of the fitted straight lines. In a specific embodiment, a Cartesian rectangular coordinate system is established with the lower left corner vertex of the undulating formation resistivity model as the coordinate origin, and a point is randomly selected on the upper and lower boundaries of the model, and the coordinates of the two points A1 (x1, y1) and B1 (x2, y2) are recorded, and the coordinates of the two points are fitted to the straight line equation:

[0037]

[0038] The resistivity model of the undulating strata is divided into two areas, left and right, with a straight line as the boundary. One of the areas is selected, and the data of the area is moved upward or downward by n units as a whole. During the movement, the data that overflows the model boundary is discarded. The strata in the moved area and the unmoved area will be discontinuous in the horizontal direction, thus forming a fault structure, such as Figure 4 As shown in a; at the same time, a zero-value model with grayscale values ​​of all 0 and the same pixel size as the undulating stratum resistivity model (for example, 512*512) is generated. In the zero-value model, the boundary between the left and right areas is marked with a grayscale value of 255 to form a fault label.

[0039] On the basis of a single fault resistivity model, select point groups at different positions again, fit the straight line equations with different slopes and different positions, and repeatedly move the data in the left and right areas of the new straight line equation to form multiple fault system models. After generating a fault according to the above steps, randomly select a set of point coordinates (A2, B2) again, and the coordinates of at least one point in (A2, B2) are different from the coordinates of the first selected point (A1, B1). Use the second selected points to fit a new straight line and repeat the above steps to generate the second fault, and mark the position of the second fault in the label model, such as Figure 4 b. Repeat the above steps to generate multiple faults, such as Figure 4 c. In this process, the slope of the straight line can be controlled according to the distance between the horizontal axis coordinates of points A and B, thereby controlling the inclination of the fault. The fault distance is controlled according to the value n of the data moving up and down in the selected area. The larger the moving value n, the larger the fault distance. In practical applications, multiple normal fault models, reverse fault models, and combined normal and reverse fault models can be generated in a targeted manner in combination with the actual geological characteristics of the study area.

[0040] After the fault structure is formed, in order to reduce the jagged features formed by the sudden change of formation resistivity on both sides of the fault, the fault model needs to be smoothed. The smoothing method used in the present invention is Gaussian smoothing, and the formula is as follows:

[0041]

[0042] Where:

[0043] G(i,j) is the value of the Gaussian kernel, which represents the weighting coefficient of adjacent pixels in the image;

[0044] I(x+i,y+j) represents the pixel value at (x+i,y+j) in the fault model;

[0045] k is half the size of the Gaussian kernel.

[0046] Forward calculation

[0047] The above fault resistivity model is forward calculated to simulate the apparent resistivity data collected in the actual field work of electromagnetic exploration. Taking the wide-area electromagnetic sounding method as an example, the current wide-area electromagnetic sounding working methods include EE x Law, EE z Different working modes define different wide-area apparent resistivities. x The method is used as an example, and the wide-area apparent resistivity is defined as:

[0048]

[0049] Where: is the measured voltage, I is the current intensity of the dipole source, r is the length of the radial vector r, k is the wave number, ω is the angular frequency, μ is the magnetic permeability, ε is the dielectric constant, For EE x The electromagnetic effect function of wide-area electromagnetic sounding, is the angle between the x-axis and the radial vector r, dL is the dipole moment, To measure the distance between electrodes.

[0050] Since the apparent resistivity ρ a The expression of is an implicit expression for the unknown resistivity ρ, so the present invention uses a computer iteration method to solve it:

[0051] First, arbitrarily select the initial value ρ0 and substitute ρ0 into the electromagnetic effect function Then, the calculated value is substituted into the definition equation of wide-area apparent resistivity to obtain the first apparent resistivity value ρ1. Then, the following conditions are used to determine whether the apparent resistivity accuracy meets the requirements:

[0052]

[0053] Where μ is the expected accuracy. If the expected accuracy is not met, substitute ρ1 into the new apparent resistivity ρ2 and iterate again to determine whether ρ2 meets the requirement. Repeat this process until the solved ρ n Meet the accuracy requirements.

[0054] The generated theoretical fault models are calculated using the wide-area apparent resistivity calculation method one by one, and appropriate iteration accuracy is selected according to the requirements. The wide-area apparent resistivity of the theoretical model is obtained point by point through forward modeling. The forward modeled apparent resistivity of a point in a model is shown in the attached figure. Figure 5 shown.

[0055] Inversion calculation

[0056] The theoretical fault resistivity model is forward calculated and the apparent resistivity is obtained as the initial model for inversion calculation. The linear equations for electromagnetic inversion calculation are:

[0057]

[0058] Where: si is the observed value of apparent resistivity corresponding to the ith frequency (i.e., the apparent resistivity data obtained by forward calculation of the theoretical fault model), ρ ci is the forward modeled apparent resistivity corresponding to the ith frequency, Δm j is the modification amount of the j-th prediction model, nm is the number of model parameters, and J is the Jacobian matrix.

[0059] Then use the singular value decomposition method to solve the above equation to get the modification value Δm of the prediction model m, and the new prediction model parameter is m+Δm. Use the new prediction model parameters to forward calculate the apparent resistivity ρ ci (m+Δm), the error between the forward calculated apparent resistivity and the observed apparent resistivity is calculated. If the error meets the accuracy requirement, the prediction model is the inversion result. Otherwise, ρ ci (m+Δm) is substituted into the equation again and iterated until the apparent resistivity calculated by the new prediction model meets the accuracy requirements. The inversion section that meets the accuracy requirements is used as the characteristic map of the theoretical fault model and used as the training set of the neural network.

[0060] The generated forward model is used as the initial model for inversion, and the inversion profiles are obtained through iterative inversion. For example, 2000 theoretical fault model resistivity inversion profiles can be obtained. The inversion profile of a certain model is shown in the attached figure. Figure 6 shown.

[0061] In a preferred embodiment, the actual fault model resistivity inversion profile data set is obtained by the following steps:

[0062] A uniform window size (e.g., 512*512) is used to intercept the fault from the actual electromagnetic inversion profile, and the left and right boundaries of the fault are marked as fault labels. In order to increase the diversity of the fault model training set, a batch of wide-area electromagnetic inversion profiles obtained by processing actual field data are added to the training set, and the actual profiles are artificially labeled with faults; similarly, in the fault label, the fault position is marked with a gray value of 255, and the gray values ​​of the remaining positions are set to 0. Considering that the actual inversion profile is too large and the shape of each profile is different, it is not suitable for direct training of deep learning neural networks. Therefore, a window with the same size as the theoretical model is set, and the window is randomly moved on the actual inversion profile. The data in the window on the profile is clipped, and the position parameters of the window are recorded at the same time. The label map is clipped at the same position on the fault label of the resistivity inversion profile.

[0063] The actual fault resistivity profiles and their labels are stored and divided into a training set and a test set according to a preset ratio. In order to improve the generalization ability of the network model, reduce the risk of overfitting of the network model and improve the robustness of the model, it is necessary to perform data enhancement operations on the actual model training set data to increase the diversity of the data set. The data enhancement method for the actual fault resistivity features and their labels of the present invention includes but is not limited to flipping (left and right, up and down), rotating, shifting, cropping, deforming, scaling, and blurring.

[0064] In actual EE xIn the inversion profiles obtained by wide-area electromagnetic exploration, the fault positions are manually marked, and then the inversion profiles and fault labels are randomly moved and cut using the above-mentioned random movement method. The profiles obtained by cutting and having the same size as the theoretical model are then subjected to data enhancement operations such as left-right flipping, up-down flipping, and blurring to generate 1,000 actual fault models.

[0065] In a preferred embodiment, the deep learning model is a Res-UNet network model improved based on the UNet network, and the model structure is as follows: Figure 7 As shown in the figure. Based on the UNet network, the Res-UNet network model replaces the standard convolutional layers of the encoder and decoder with convolutional blocks containing residual modules. Through the residual module, the network can pass the gradient from the output layer to the shallower layer, and at the same time change the learning target of the network to the residual, which will effectively alleviate the problem of network gradient disappearance and improve the convergence speed of the network.

[0066] Replace the ReLU activation function after some convolutional layers of the network encoder and decoder with the LeakyReLU activation function. The LeakyReLU activation function helps maintain the flow of gradients in the network by retaining a smaller slope in the negative interval so that the gradient does not disappear due to negative input values. The expression of the LeakyReLU function is as follows:

[0067]

[0068] Wherein: α represents the slope of the non-zero interval, and the present invention takes 0.01.

[0069] The output layer of the network uses a sigmoid activation function to obtain a probability distribution image of the predicted fault.

[0070] At the same time, since the number of fault samples in the fault label is far less than the number of non-fault samples, this leads to an imbalance in the number of samples in the label, which affects the performance of the network and makes the fault recognition effect poor. Therefore, the Focal Loss function is further used to solve the problem of sample imbalance. The expression of the Focal Loss function is as follows:

[0071]

[0072] Where: N represents the number of samples, y i Represents the true sample label value, p i Represents the predicted value, α represents the category balance factor, and γ represents the focus factor. Based on the Focal Local function, the category balance factor α is used to suppress the imbalance in the number of fault and non-fault samples, and the focus factor γ is used to control the imbalance in the number of simple / easy-to-separate samples, thereby reducing the loss contribution of easy-to-separate samples and improving the model training efficiency.

[0073] In a preferred embodiment, the theoretical model and the actual model are combined into a training set for training the Res-UNet network. The electromagnetic exploration resistivity profile is input into the trained network model to obtain a fault probability distribution map. Specifically, a total of 2000 theoretical fault inversion models and 1000 actual fault resistivity inversion models are generated to form a training set to train the Res-UNet network. The training is iteratively trained until the loss value tends to converge. The model with stable iterative convergence is used as the optimal model for fault identification. The typical number of iterations is, for example, 150 times. A certain actual wide-area electromagnetic exploration inversion profile is converted into a grayscale image, and then the grayscale image is input into the optimal network model for fault identification. Figure 8-10 The figures show the actual resistivity inversion profile, the fusion diagram of the actual resistivity inversion profile and the real fault label, and the fusion diagram of the fault identification result using this method and the actual resistivity inversion profile. Fig.10 It can be seen that the method of the present invention identifies both large and small faults in the electromagnetic profile with high recognition accuracy and no misidentification. The identified faults have clear morphology, strong continuity and obvious detail features.

[0074] According to one embodiment, a computer-readable storage medium is provided, on which a program code is stored, and when the program code is executed by a processor, the processor is enabled to perform various operations and functions in the present specification in combination with the various embodiments of the present application. Specifically, a system or device equipped with a readable storage medium can be provided, on which a software program code that implements the functions of any of the above embodiments is stored, and a computer or processor of the system or device is enabled to read and execute instructions stored in the readable storage medium.

[0075] The above description of the present disclosure is provided to enable any person of ordinary skill in the art to implement or use the present disclosure. Various modifications to the present disclosure will be apparent to those of ordinary skill in the art, and the general principles defined herein may be applied to other variations without departing from the scope of protection of the present disclosure. Therefore, the present disclosure is not limited to the examples and designs described herein, but is consistent with the widest range of principles and novel features disclosed herein.

[0076] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described in the present invention. Although the present invention has been described in detail in this specification with reference to the above embodiments, the present invention is not limited to the above specific implementation methods. Any technical solutions and improvements that modify or replace the present invention without departing from the spirit and scope of the invention are all included in the scope of the claims of the present invention.

Claims

1. A method for identifying faults in electromagnetic exploration based on deep learning, characterized in that: include: Build a deep learning model for fault identification; Obtain existing resistivity inversion profile datasets to train deep learning models; Taking the actual resistivity inversion profile as input, the trained deep learning model is used to output the fault distribution results.

2. The method according to claim 1, characterized in that: The existing resistivity inversion profile data set includes a theoretical fault model resistivity inversion profile data set and an actual fault model resistivity inversion profile data set.

3. The method according to claim 2, characterized in that: The resistivity inversion profile data set of the theoretical fault model is obtained by the following steps: Multiple undulating stratum resistivity models are established based on the Cauchy distribution function; Fitting straight lines to any two points of the resistivity model of each undulating formation that are not at the same height, and moving the data in the domain upward / downward with the straight line as the boundary to form a fault; storing a plurality of undulating stratum resistivity models of formed faults as theoretical fault models, and marking the left and right boundaries of the faults as fault labels; Performing forward simulation on the theoretical fault model to obtain forward apparent resistivity data; The forward modeling apparent resistivity data is inverted to output theoretical fault model resistivity inversion profile data.

4. The method according to claim 3, characterized in that: The expression of the Cauchy distribution function to simulate the lateral fluctuation of the stratum is: Where: A is the height amplitude parameter for controlling the undulating formation, x0 is the position parameter for controlling the center point of the undulating formation, and γ is the range parameter for controlling the undulating formation; Moreover, when the undulating stratum resistivity model is established by the horizontal stratum resistivity model according to the Cauchy distribution function, different stratum undulation shapes are constructed by adjusting different amplitude parameters A, scale parameters γ and position parameters x0.

5. The method according to claim 3, characterized in that: A plurality of straight lines are fitted with the resistivity models of the undulating formations to form a plurality of faults; and different types of faults are formed by adjusting the slopes and moving distances of the fitted straight lines.

6. The method according to claim 5, characterized in that: After the fault is formed, the data on both sides of the fault are smoothed.

7. The method according to claim 3, characterized in that: Wide-area electromagnetic sounding (EE) x Method for forward modeling: Where: is the measured voltage, I is the current intensity of the dipole source, r is the length of the radial vector r, k is the wave number, For EE x The electromagnetic effect function of wide-area electromagnetic sounding, is the angle between the x-axis and the radial vector r, dL is the dipole moment, To measure the distance between electrodes.

8. The method according to claim 3, characterized in that: The linear equations for inversion of the forward apparent resistivity data are: Where: si is the observed value of apparent resistivity corresponding to the ith frequency, ρ ci is the forward calculated value of apparent resistivity corresponding to the ith frequency, Δm j is the modification amount of the j-th prediction model, nm is the number of model parameters, and J is the Jacobian matrix.

9. The method according to claim 2, characterized in that: The actual fault model resistivity inversion profile data set is obtained by the following steps: A fault is cut from the actual electromagnetic method inversion section using a window of uniform size, and the left and right boundaries of the fault are marked as fault labels; A data enhancement operation is performed according to the intercepted fault, and the data enhancement operation includes but is not limited to flipping, rotating, shifting, cropping, deforming, scaling, and blurring.

10. The method according to claims 1-9, characterized in that: The deep learning model is a Res-UNet network model improved based on the UNet network, and the Focal Loss function is used as the loss function to train the deep learning model.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 10 are implemented.

12. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 10 are implemented.

Images