Dipole dual-layer casing well imaging method based on inception module

Abstract

The application provides a dipole double-layer casing well imaging method based on an Inception module, forward modeling is performed on a real double-layer casing well model, a dipole excitation mode is adopted, different sizes of double-layer casing well array logging signals are obtained according to the forward model, neural network parameters based on the Inception module are set, and the obtained characteristic signal array is input into a neural network containing the Inception module for training and verification, the training result of the neural network is saved, and the network structure of the neural network containing the Inception module and the model trained by the optimal parameters are used for carrying out longitudinal wave and transverse wave slowness inversion and imaging on the double-layer casing well. The application establishes the relationship between the characteristic signal and the real longitudinal wave and transverse wave slowness graph of the dipole double-layer casing by using the neural network containing the Inception module, can realize fast and high-resolution imaging, and effectively improves the robustness of the neural network.

Description

Technical Field

[0001] This invention relates to the technical field of acoustic logging data processing for double-cased wells, and particularly to a dipole double-cased well imaging method based on the Inception module. Background Technology

[0002] Cementing quality inspection is of great significance for the production life of oil and gas wells and oil and gas exploration. With the continuous development of petroleum exploration technology, deep wells and ultra-deep wells are increasingly using double-casing or multi-casing completion methods. Therefore, cementing quality inspection of double-casing or multi-casing wells is crucial. Traditional sonic and ultrasonic logging technologies mainly assess the interface cementation and cement sheath health of single-casing wells. However, sonic wave propagation in multi-casing wells is more complex. Therefore, developing effective cementing quality inspection technologies for double-casing wells is a major technical challenge currently facing the sonic logging industry.

[0003] Currently, wellbore imaging methods mainly include: (1) ultrasonic pulse echo imaging method, but this method can only image the geometry of the external cemented layer and formation, which limits the detection performance of the annulus. (2) reverse time migration imaging method, but this method is highly dependent on migration artifacts, migration noise and velocity models, and has disadvantages such as amplitude imbalance and difficulty in identifying effective energy when imaging deep formations. (3) full waveform inversion imaging method, which requires a relatively accurate initial model to obtain a relatively accurate solution, and has a large amount of computation and long computation time, so it needs to be optimized and accelerated. In addition, the application of neural networks in casing wells is mainly for prediction and interpretation, and there is little research on imaging.

[0004] In summary, existing methods for inversion imaging involve large computational loads, rely on initial models, and cannot evaluate all cementation conditions in double-cased wells. Summary of the Invention

[0005] To address the imaging problem of double-cased wells, this invention proposes a dipole double-cased well imaging method based on the Inception module. A forward model of the double-cased well is established, and array acoustic logging signals of double-cased wells of different sizes are obtained based on the forward model. The array acoustic logging signals are used as feature signals and then applied to a neural network containing the Inception module to achieve rapid two-dimensional imaging of P-waves and S-waves in double-cased wells of different sizes.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: a dipole double-casing well imaging method based on the Inception module, comprising the following steps:

[0007] Step 1: Perform forward modeling on the real double-cased well model. Use dipole excitation to obtain array logging signals under different working conditions and well sizes, and obtain P-wave slowness maps and S-wave slowness maps of the real double-cased well model under different well sizes.

[0008] Step 2: Preprocess the array logging signals obtained in Step 1 to form a feature signal matrix, sample the feature signal matrix to form a dataset, and divide the dataset into training set, validation set and test set according to different purposes;

[0009] Step 3: Adjust the network structure of the neural network containing the Inception module. Use the P-wave and S-wave sluggishness maps obtained in Step 1 and the feature signal matrix obtained in Step 2 as samples to train and validate the neural network containing the Inception module. Train the neural network containing the Inception module using the training set of the dataset, and test the network structure of the trained neural network containing the Inception module using the test set of the dataset. Set the activation function and loss function, set and test the coherence parameters of the neural network containing the Inception module to adjust and seek the optimal parameters. Use the network structure of the neural network containing the Inception module and the model trained with the optimal parameters to perform P-wave and S-wave sluggishness inversion and imaging for the double-cased well.

[0010] The array logging signal mentioned in step one is wave train data including P-wave, S-wave, formation bending wave, inner casing wave component, and outer casing wave component.

[0011] The real double-casing well model includes: water, inner steel pipe, inner cement, outer steel pipe, outer cement, and formation. The water, inner steel pipe, inner cement, outer steel pipe, outer cement, and formation are all cylinders. The water, inner steel pipe, inner cement, outer steel pipe, outer cement, and formation are arranged concentrically from the inside to the outside to form a multi-layered columnar layered medium.

[0012] Step one involves collecting array logging signals for different wellbore sizes under different operating conditions:

[0013] S11: The wave field equation is obtained from the set of elastic dynamics equations;

[0014] S12: Set the boundary conditions for a realistic double-casing well model to obtain the fluid acoustic wave equation:

[0015]

[0016]

[0017] Where, ρ f p represents fluid density. fc represents the fluid sound pressure. f V represents the fluid modulus. f This is a vector representation of fluid velocity;

[0018] When dealing with the boundary conditions of a fluid-solid interface, if the radial displacement is continuous, the sound pressure on both sides of the fluid-solid interface is equal, the sound pressure in the fluid is equal to the negative radial stress, and the shear stress is zero, then:

[0019] v fr =v r

[0020] p f =p

[0021] -p f =τ rr

[0022] 0 = τ rz

[0023] Among them, v r p is the radial velocity component. f τ represents the fluid sound pressure. rr τ represents the radial principal stress. rz V represents the shear stress along the z-axis on the cylindrical surface. fr This refers to the radial velocity component of the fluid, and p refers to the sound pressure.

[0024] S13: Introduce the sound source term into the fluid sound wave equation:

[0025]

[0026] Where τ represents stress, λ represents Lamé constant, v represents velocity vector, t represents time, and ψ(t) represents the operational expression related to the sound source term. The original function of the operational expression is the second derivative of the Blackman-Harris window function (BHW) with a center frequency of 10kHz.

[0027] S14: Substitute the fluid acoustic wave equation into the wave field equation, and obtain the array logging signal of the double-cased well by solving the wave field equation substituted into the fluid acoustic wave equation; collect array logging data of double-cased wells of different sizes, and the array logging signal is wave train data including P-wave, S-wave and inner and outer casing waves and formation bending wave components.

[0028] The method for obtaining P-wave and S-wave slowness maps under different wellbore sizes in step one is as follows: By changing the sizes of water, inner steel pipe, inner cement, outer steel pipe, and outer cement, double-cased well models of different sizes are obtained. The P-wave and S-wave information of different media are known. A P-wave slowness matrix is ​​formed from the P-wave information of each medium in the double-cased well model of different sizes and the size information of the double-cased well model, which is the true P-wave slowness map. A S-wave slowness matrix is ​​formed from the S-wave information of each medium in the double-cased well model of different sizes and the size information of the double-cased well model, which is the true S-wave slowness map.

[0029] The method for forming feature signals and sampling feature signals to form a dataset in step two is as follows: the array logging signal size for different wellbore sizes is A×B, where A represents the number of sensors and B represents the number of data points contained in each sensor. The first N data points in each array logging signal are taken as a sample, and M samples are randomly sampled to form a dataset, where N is a positive integer.

[0030] Step two describes the method of dividing the dataset into training, validation, and test sets according to different purposes. This involves randomly selecting 60% of the feature signals from the feature signal matrix as the training set; randomly selecting 20% ​​of the feature signals from the feature signal matrix as the validation set; and using the remaining 20% ​​of the feature signals as the test set.

[0031] The neural network containing the Inception module described in step three includes an input layer, a convolutional layer, a max pooling layer, an upper block module, a lower block module, an average pooling layer, a ReLU fully connected layer, a Dropout layer, a Sigmoid fully connected layer, and an output layer connected in sequence. Both the upper and lower block modules include an upper Inception module and a lower Inception module connected in sequence. The max pooling layer is connected to the upper Inception module in the upper block module, and the average pooling layer is connected to the lower Inception module in the lower block module. The neural network containing the Inception module has 1000 iterations and a batch size of 12.

[0032] Both the upper and lower Inception modules include a 1×1 convolutional kernel, a 3×3 convolutional kernel, a 5×5 convolutional kernel, and a 3×3 max pooling layer. The inputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers are connected to the max pooling layer or, through the Concatenate module, to the outputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers of the previous Inception module. The outputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers are connected to the average pooling layer or the inputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers of the next Inception module.

[0033] The activation function described in step three is the Sigmoid function, the loss function is the mean squared error function, and the optimization algorithm is the Adam algorithm.

[0034] The beneficial effects of this invention are as follows: This invention utilizes a neural network containing an Inception module to image the P-wave and S-wave slowness of dipole double-cased wells. By performing forward modeling based on a real double-cased well model, array acoustic logging signals at different sizes are obtained. Array logging signals of different well sizes under different operating conditions are acquired, and P-wave slowness maps and S-wave slowness maps under different real double-cased well model operating conditions are obtained. The use of dipole excitation avoids the defect that monopoles cannot excite S-waves in soft formations. Direct use of array logging signals does not require additional signal processing. The method is simple to operate and highly accurate. Secondly, this invention utilizes a neural network containing an Inception module to rapidly image array logging signals. The neural network establishes the relationship between the array logging signals and the P-wave and S-wave slowness models of double-cased wells, enabling rapid, high-resolution imaging. The method used in this invention does not require an initial model and can converge to a relatively accurate result. The activation function of the neural network is the Sigmoid function, as its output range corresponds to (0,1), which corresponds to the slowness value of the P-wave or S-wave after the slowness map is normalized. This invention can identify and image different formations and cement information, and can also locate and image micro-rings at the casing-cement interface and the cement-formation interface, exhibiting strong universality and wide applicability. The neural network algorithm of this invention is simple, and the network architecture is easy to build. Combining different convolutional kernels not only increases the receptive field but also improves the robustness of the neural network. This method can achieve rapid imaging of P-wave and S-wave slowness models of dipole logging data under complex working conditions, providing a foundation for cementing quality inspection of multi-cased wells. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the present invention.

[0037] Figure 2 This is a schematic diagram of a double-cased well with a transducer module and a receiver array in this invention.

[0038] Figure 3 The following are the double-cased well model and its received signals in this invention: (a) the double-cased well model of actual size, and (b) the array logging signal corresponding to (a).

[0039] Figure 4 This is a diagram of the neural network structure containing the Inception module in this invention.

[0040] Figure 5 Comparison of P-wave slowness models for double-cased wells in this invention: (a) P-wave slowness model of a real double-cased well, (b) P-wave slowness model of a double-cased well obtained by neural network inversion, and (c) Slowness section profile.

[0041] Figure 6 The following is a comparison of the shear wave slowness models for double-cased wells in this invention: (a) actual shear wave slowness model of a double-cased well, (b) shear wave slowness model of a double-cased well obtained by neural network inversion, and (c) slowness section profile. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] like Figure 1As shown, this invention proposes a dipole-based double-cased well imaging method based on the Inception module. First, using dipole excitation, double-cased well models of different sizes are established based on the actual wellbore dimensions. Second, logging signals from double-cased well arrays of different sizes are obtained based on the forward model. Next, the parameters of a neural network based on the Inception module are set, and the obtained feature signal array is used as a sample input to the neural network containing the Inception module for training and validation. The training results of the neural network, i.e., the P-wave and S-wave slowness inversion results of the double-cased well, are saved, thereby enabling imaging of the double-cased well. The specific steps are as follows:

[0044] Step 1: Perform forward modeling using a real double-cased well model. Use dipole excitation to obtain array logging signals for different wellbore sizes under different operating conditions. Obtain P-wave slowness maps and S-wave slowness maps for real double-cased well models with different wellbore sizes under different operating conditions.

[0045] The real double-casing well model is mainly divided into six parts: drilling fluid, inner steel pipe, inner cement sheath, outer steel pipe, outer cement sheath, and formation. The radii of the drilling fluid, inner steel pipe, inner cement sheath, outer steel pipe, and outer cement sheath are all randomly generated based on the actual wellbore size. Figure 2 The image shows the established double-casing well model. The P-wave velocities or S-wave velocities of the drilling fluid, inner casing, inner cement sheath, outer casing, outer cement sheath, and formation are converted into P-wave deceleration or S-wave deceleration. Since deceleration is the reciprocal of velocity, the formula for calculating deceleration is used... Calculations can obtain P-wave or S-wave slowness maps for double-cased wells of different sizes, where v cement The longitudinal wave velocity or transverse wave velocity of cement, s cement This indicates the longitudinal wave slowness value or the transverse wave slowness value of cement. The specific method is as follows:

[0046] S11: The wave field equation is obtained from the elastic dynamics equations:

[0047] Specifically, for a homogeneous isotropic elastic medium with a certain density ρ and Lamé constant, the elastic dynamics equations under multipole source action are as follows:

[0048]

[0049] Among them, v r ,v θ ,v z These represent the radial, circumferential, and axial velocity components, respectively; τ rr ,τ θθ ,τ zz These represent the principal stresses in the radial, circumferential, and axial directions, respectively; τ rθτ represents the shear force along the circumference on a cylindrical surface. rz τ represents the shear stress along the z-axis on the cylindrical surface. θz Let λ represent the shear force along the z-axis in a plane containing the z-axis, λ and μ represent the Lamé constant, ρ represent the density, and t represent time.

[0050] Therefore, the entire wave field satisfies the following wave field equation:

[0051]

[0052] S12: Set the boundary conditions for a realistic double-casing well model to obtain the fluid acoustic wave equation:

[0053] Specifically, the equation for sound waves in a fluid is as follows:

[0054]

[0055]

[0056] Where, ρ f This indicates fluid density, with units of kg / m³. 3 ;p f c represents the sound pressure of a fluid, with the unit being Pa; f This represents the fluid modulus, with units of Pa; v f This is a vector representation of fluid velocity.

[0057] Typically, when dealing with boundary conditions at fluid-solid interfaces, it is assumed that radial displacement is continuous, sound pressure on both sides of the fluid-solid interface is equal, and the sound pressure in the fluid is equal to the negative radial stress and the shear stress is zero, as shown below:

[0058]

[0059] S13: Introduce the sound source term into the fluid sound wave equation:

[0060]

[0061] Where ψ(t) represents the operational expression related to the sound source term, and the second derivative of the Blackman-Harris window (BHW) with a center frequency of 10kHz is chosen as the source function. The BHW window function is defined as follows:

[0062]

[0063] Where ω i =2πi / period, i = 0, 1, 2, 3.

[0064] S14: Substitute the fluid acoustic wave equation into the wave field equation, and solve the wave field equation substituted into the fluid acoustic wave equation to obtain the array logging signal of the double-cased well; collect array logging data of double-cased wells of different sizes. The array logging signal is wave train data including components such as longitudinal waves, transverse waves, inner and outer casing waves, and formation bending waves.

[0065] By changing the dimensions of water, inner steel pipe, inner cement, outer steel pipe, and outer cement, double-cased well models of different sizes were obtained, as shown in Table 1. The P-wave and S-wave information of different media is known. A P-wave slowness matrix, i.e., the true P-wave slowness map, is formed by combining the P-wave information of each medium layer and the size information of the double-cased well model for each size. Similarly, a S-wave slowness matrix, i.e., the true S-wave slowness map, is formed by combining the S-wave information of each medium layer and the size information of the double-cased well model for each size. In the S-wave slowness map, since water has no S-wave information, its S-wave velocity is assumed to be 800 m / s. This parameter is only used to illustrate the slowness model of the wellbore and has no meaning for the S-wave information of water.

[0066] Table 1 Material properties of double-casing wells

[0067]

[0068] Step 2: Preprocess the double-casing logging signals obtained in Step 1. Divide the array acoustic signal length to ensure that the dimensions of all array logging signals are consistent, thus forming a feature signal. The size of the array acoustic signal for double-casing well models of different sizes is A×B, where A represents the number of sensors and B represents the number of data points contained in each sensor. Take the first N data points from each array logging signal as a sample, i.e., the sample size is A×N (N≤B), and randomly sample M samples to form a dataset, where N is a positive integer. Sample the feature signals to form the dataset, and divide the dataset into different parts according to different purposes. Randomly select 60% as the training set for training; randomly select 20% of the data for error estimation and parameter tuning during the training process, i.e., the validation set; and finally, use the last 20% of the data to test the performance of the neural network, i.e., the test set.

[0069] Figure 3 The image shows a dimensional model of a double-cased well and its received signals. Figure 3 (a) shows a cross-sectional view of a double-cased well at actual dimensions. Figure 3 (b) indicates Figure 3 (a) The array logging signal corresponding to the model mainly includes the received signals of 24 sensors, each sensor containing 2440 data points.

[0070] This invention performs forward modeling based on a real double-cased well model to obtain array acoustic logging signals of different sizes. It collects array logging signals of different well sizes under different operating conditions and obtains P-wave and S-wave slowness maps under different real double-cased well model operating conditions. The dipole excitation method can avoid the defect that monopoles cannot excite S-waves in soft formations. The array logging signals can be used directly without additional signal processing methods, making the operation simple and the accuracy high.

[0071] Step 3: Use the feature signals from Step 2 and the actual longitudinal or transverse wave slowness maps of the dipole double-layer sleeve from Step 1 as samples to train and validate the neural network containing the Inception module, and obtain the inversion results. Figure 4 The diagram shows the neural network structure containing the Inception module, which includes a sequentially connected input layer, convolutional layer, max pooling layer, upper block module, lower block module, average pooling layer, ReLU fully connected layer, Dropout layer, Sigmoid fully connected layer, and output layer. Both the upper and lower block modules contain sequentially connected upper and lower Inception modules. The max pooling layer is connected to the upper Inception module within the upper block module, and the average pooling layer is connected to the lower Inception module within the lower block module. The neural network containing the Inception module iterates 1000 times, with a batch size of 12. The input layer receives logging signals from dipole double-casing wells, and the output layer outputs the P-wave or S-wave inversion results for the double-casing well. The specific method is as follows:

[0072] The main function of convolutional layers is to extract features from the input data. A convolutional layer contains multiple convolutional kernels, each of which traverses the input data and performs a convolution operation using sparse connections. The output feature map is then obtained through a non-linear mapping using an activation function. The specific calculation process is as follows:

[0073]

[0074] in, This represents the output obtained after the p-th convolutional kernel in layer b undergoes a convolution operation. and Let these represent the weights and biases of the p-th convolutional kernel in layer b, respectively. This represents the local input feature value of the b-th layer performing the q-th convolution operation, where * represents the convolution operation and f represents the non-linear activation function.

[0075] Max pooling and average pooling layers primarily downsample the feature sequences obtained from convolution operations by calculating local maxima or averages, respectively. This reduces network parameters while filtering out some redundant information. Their calculation formulas are as follows:

[0076]

[0077] Where down indicates downsampling, which typically uses max pooling and average pooling; and This represents the weights and biases of the pooling layer.

[0078] The Inception module consists of an upper Inception module and a lower Inception module connected in sequence. Both the upper and lower Inception modules include a 1×1 convolutional kernel, a 3×3 convolutional kernel, a 5×5 convolutional kernel, and a 3×3 max pooling layer. The inputs of the 1×1 convolutional kernel, 3×3 convolutional kernel, 5×5 convolutional kernel, and 3×3 max pooling layer are connected to the max pooling layer or, through the Concatenate module, to the outputs of the 1×1 convolutional kernel, 3×3 convolutional kernel, 5×5 convolutional kernel, and 3×3 max pooling layer of the previous Inception module. The outputs of the 1×1 convolutional kernel, 3×3 convolutional kernel, 5×5 convolutional kernel, and 3×3 max pooling layer are connected to the average pooling layer or the inputs of the 1×1 convolutional kernel, 3×3 convolutional kernel, 5×5 convolutional kernel, and 3×3 max pooling layer of the next Inception module. Specifically, in the upper-level block module, the inputs of the 1×1, 3×3, and 5×5 convolutional kernels and the 3×3 max-pooling layer of the upper-level Inception module are connected to the max-pooling layer. The outputs of the 1×1, 3×3, 5×5, and 3×3 convolutional kernels and the 3×3 max-pooling layer of the upper-level Inception module are connected to the inputs of the 1×1, 3×3, 5×5, and 3×3 max-pooling layers of the lower-level Inception module in the upper-level block module. The outputs of the 1×1, 3×3, 5×5, and 3×3 max-pooling layers of the lower-level Inception module are connected to the lower-level block module. The upper-middle Inception module's 1×1, 3×3, 5×5, and 3×3 max-pooling layers are connected to their inputs. The lower-level block module's upper-middle ...

[0079] Using 1×1 convolutions reduces the number of input channels, lowering computational complexity and introducing non-linear transformations. 3×3 and 5×5 convolutional kernels, on the other hand, extract features over a larger receptive field. The Inception module effectively increases network width by using filters and pooling operations of different sizes within the same layer to collect different features from the input of the previous layer. Combining different convolutional kernels not only increases the receptive field but also improves the robustness of the neural network.

[0080] After multiple convolution and pooling operations, the extracted features are typically planarized and connected through two fully connected layers. These fully connected layers, serving as the final output layer, map the learned features to the corresponding labels on the input data. In the case of inversion slowness model prediction, this classifier maps the extracted features to the corresponding labels on the input data, serving as the final output of the neural network, thus yielding different inversion slowness models for double-cased wells.

[0081] A fully connected layer has L layers besides the input layer, where layer l has m nodes and layer l-1 has n nodes. During the forward propagation, the output of layer l can be written as:

[0082]

[0083] Among them, c l It is the output of the l-th layer, z l W is the inactive output of layer l. l It represents the weights between layer (l-1) and layer l, d l It is the deviation of the lth layer.

[0084] Using the Sigmoid (sigmoid growth function) as the activation function guarantees that the output value is within the range [0,1]. Its expression is as follows:

[0085]

[0086] Where × represents the output of the previous layer of the neural network.

[0087] The loss function uses the mean squared error function:

[0088]

[0089] Z a S represents the P-wave or S-wave slowness map of a one-dimensional real double-cased well. a This represents the predicted P-wave or S-wave slowness map obtained from neural network inversion. The optimization algorithm for the loss function is the Adam algorithm (Adaptive Moment Estimation Optimization Algorithm).

[0090] Figure 5 To test the comparison results of the P-wave slowness model, Figure 5 (a) represents the true P-wave slowness model. Figure 5 (b) represents the P-wave slowness model obtained through neural network inversion. Figure 5(c) shows the slow-wave cross-sectional view. It can be seen that the inversion results are very similar to the real model, including the identification and characterization of relatively small-sized media information such as the inner casing, inner cement sheath, outer casing, and outer cement sheath. From the slow-wave cross-sectional view, it can be seen that the cross-sectional curves of the real model and the inversion results are very similar, indicating that this method can image the P-wave slow-wave model of dipole double-cased wells and the imaging results are good.

[0091] Figure 6 To test the comparison results of the shear wave slowness model, Figure 6 (a) represents the true shear wave slowness model. Figure 6 (b) represents the transverse wave slowness model obtained through neural network inversion. Figure 6 (c) shows the slowness profile. The water slowness information is only for illustrating the double-casing well model; its slowness value is meaningless. It can be seen that for shear wave slowness inversion, the inversion results match the actual model well in terms of shear wave slowness values, downhole structure, and geological interfaces. The slowness profile shows that the cross-sectional curves of the actual model and the inversion results are basically consistent, indicating that this method can also image the shear wave slowness model of dipole double-casing wells with good imaging results.

[0092] This invention uses a deep learning method based on a neural network containing an Inception module to perform rapid P-wave and S-wave imaging of the input feature signals of a dipole double-cased well. By using a neural network containing an Inception module to establish the relationship between the feature signals and the actual P-wave and S-wave slowness maps of the dipole double-cased well, rapid and high-resolution imaging can be achieved. Furthermore, combining different convolutional kernels with this neural network can not only increase the receptive field but also improve the robustness of the neural network.

[0093] This invention can perform slow-motion P-wave and S-wave imaging of dipole double-cased wells with different cement and formation parameters. It can also locate and image micro-rings at the casing-cement interface and the cement-formation interface, demonstrating strong versatility and wide applicability. The neural network algorithm of this invention, including the Inception module, is simple and easy to build. Furthermore, combining different convolutional kernels can not only increase the receptive field but also improve the robustness of the neural network. This method can achieve rapid P-wave and S-wave imaging of dipole double-cased well logging data under complex operating conditions, providing a foundation for cementing quality inspection of double-cased wells.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dipole double-casing well imaging method based on the Inception module, characterized in that, Includes the following steps: Step 1: Perform forward modeling on the real double-cased well model. Use dipole excitation to obtain array logging signals under different working conditions and well sizes, and obtain P-wave slowness maps and S-wave slowness maps of the real double-cased well model under different well sizes. Step one describes obtaining array logging signals for different wellbore sizes under different operating conditions: S11: The wave field equation is obtained from the set of elastic dynamics equations; S12: Set the boundary conditions for a realistic double-casing well model to obtain the fluid acoustic wave equation: ； ； in, Indicates fluid density; Indicates fluid sound pressure. Indicates fluid modulus. This is a vector representation of fluid velocity; When dealing with the boundary conditions of a fluid-solid interface, if the radial displacement is continuous, the sound pressure on both sides of the fluid-solid interface is equal, the sound pressure in the fluid is equal to the negative radial stress, and the shear stress is zero, then: ； in, For radial velocity components, Indicates fluid sound pressure. Indicates radial principal stress. Indicates along the cylindrical surface Shear stress in the axial direction, It refers to the radial velocity component of the fluid. This refers to sound pressure level; S13: Introduce the sound source term into the fluid sound wave equation: ； in, Indicates stress, Represents Lamé constant, Let t represent the velocity vector and t represent time. This represents the operational expression related to the sound source term. The original function of the operational expression is the second derivative of the Blackman-Harris window function with a center frequency of 10 kHz. S14: Substitute the fluid acoustic wave equation into the wave field equation, and obtain the array logging signal of the double-cased well by solving the wave field equation substituted into the fluid acoustic wave equation; collect array logging data of double-cased wells of different sizes. The array logging signal is wave train data including P-wave, S-wave and inner and outer casing waves and formation bending wave components. Step 2: Preprocess the array logging signals obtained in Step 1 to form a feature signal matrix, sample the feature signal matrix to form a dataset, and divide the dataset into training set, validation set and test set according to different purposes; Step 3: Adjust the network structure of the neural network containing the Inception module. Use the P-wave and S-wave sluggishness maps obtained in Step 1 and the feature signal matrix obtained in Step 2 as samples to train and validate the neural network containing the Inception module: Train the neural network containing the Inception module using the training set of the dataset, and test the network structure of the trained neural network containing the Inception module using the test set of the dataset. Set the activation function and loss function, set and test the coherence parameters of the neural network containing the Inception module to adjust and seek the optimal parameters. Use the network structure of the neural network containing the Inception module and the model trained with the optimal parameters to perform P-wave and S-wave sluggishness inversion and imaging for the double-cased well. The real double-casing well model includes: water, inner steel pipe, inner cement, outer steel pipe, outer cement, and formation. The water, inner steel pipe, inner cement, outer steel pipe, outer cement, and formation are all cylinders. The water, inner steel pipe, inner cement, outer steel pipe, outer cement, and formation are arranged concentrically from the inside to the outside to form a multi-layered columnar layered medium. The method for obtaining P-wave and S-wave slowness maps under different wellbore sizes in step one is as follows: By changing the sizes of water, inner steel pipe, inner cement, outer steel pipe, and outer cement, double-cased well models of different sizes are obtained. The P-wave and S-wave information of different media are known. A P-wave slowness matrix is ​​formed from the P-wave information of each medium in the double-cased well model of different sizes and the size information of the double-cased well model, which is the true P-wave slowness map. A S-wave slowness matrix is ​​formed from the S-wave information of each medium in the double-cased well model of different sizes and the size information of the double-cased well model, which is the true S-wave slowness map.

2. The dipole double-casing well imaging method based on the Inception module according to claim 1, characterized in that, The method for assembling feature signals and sampling them to form a dataset, as described in step two, is as follows: the array logging signal size for different wellbore sizes is... Where A represents the number of sensors and B represents the number of data points contained in each sensor. The first N data points in each array logging signal are taken as a sample, and M samples are randomly selected to form a dataset, where N is a positive integer.

3. The dipole double-casing well imaging method based on the Inception module according to claim 2, characterized in that, Step two describes the method of dividing the dataset into training, validation, and test sets according to different purposes by randomly selecting 60% of the feature signals from the feature signal matrix as the training set. 20% of the feature signals in the feature signal matrix are randomly selected as the validation set; the remaining 20% ​​of the feature signals are selected as the test set.

4. The dipole double-casing well imaging method based on the Inception module according to claim 3, characterized in that, The neural network containing the Inception module described in step three includes an input layer, a convolutional layer, a max pooling layer, an upper block module, a lower block module, an average pooling layer, a ReLU fully connected layer, a Dropout layer, a Sigmoid fully connected layer, and an output layer connected in sequence. Both the upper and lower block modules include an upper Inception module and a lower Inception module connected in sequence. The max pooling layer is connected to the upper Inception module in the upper block module, and the average pooling layer is connected to the lower Inception module in the lower block module. The neural network containing the Inception module has 1000 iterations and a batch size of 12.

5. The dipole double-casing well imaging method based on the Inception module according to claim 4, characterized in that, Both the upper and lower Inception modules include a 1×1 convolutional kernel, a 3×3 convolutional kernel, a 5×5 convolutional kernel, and a 3×3 max pooling layer. The inputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers are connected to the max pooling layer or, through the Concatenate module, to the outputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers of the previous Inception module. The outputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers are connected to the average pooling layer or the inputs of the 1×1, 3×3, 5×5, and 3×3 max pooling layers of the next Inception module.

6. The dipole double-casing well longitudinal and transverse wave imaging method based on the Inception module according to any one of claims 3-5, characterized in that, The activation function described in step three is the Sigmoid function, the loss function is the mean squared error function, and the optimization algorithm is the Adam algorithm.

Images