A deep learning inversion method combined with multiscale decomposition

Abstract

The present application relates to the technical field of electromagnetic wave resistivity logging, and particularly provides a deep learning inversion method combined with multi-scale decomposition. The method comprises the following steps: obtaining 9 components of an electromagnetic response signal and constructing a data set; adopting a multi-scale decomposition-local feature extraction-global sequence modeling architecture to construct an inversion model of a maximum overlap discrete wavelet transform (MODWT)-time series convolution network (TCN)-long short-term memory network (LSTM); training the inversion model and obtaining an analysis result. The method can more effectively extract local non-stationary features near the formation boundary, and improves the inversion accuracy and stability of the boundary region in a high-deviation well or a horizontal well.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic resistivity logging technology, and in particular to a deep learning inversion method that combines multi-scale decomposition. Background Technology

[0002] Electromagnetic resistivity logging while drilling (EMR) technology, with its advantages of large detection depth and good azimuth detection capability, has been widely used in geological steering and formation evaluation of highly deviated and horizontal wells. During the logging process, in order to obtain formation resistivity and boundary distance information from the EMR response, it is necessary to invert the logging data. Currently, the mainstream traditional method for inverting EMR data is the iterative inversion method. This method repeatedly calls the forward simulation model and uses gradient descent algorithms to continuously correct the formation model parameters to approximate the actual formation model, which suffers from computational time consumption. Furthermore, the inversion results are highly dependent on the selection of the initial formation model parameters. If the initial model parameters deviate significantly from the actual formation parameters, the inversion process is prone to getting trapped in local optima, failing to obtain reliable formation parameter results.

[0003] In recent years, deep learning technology has shown greater advantages in nonlinear fitting capabilities and has been gradually applied to the field of logging-while-drilling inversion, effectively solving the problems of traditional methods being time-consuming and prone to getting trapped in local optima. Because lithological and electrical parameters change significantly at formation interfaces, the logging-while-drilling electromagnetic resistivity response exhibits non-stationary abrupt changes. Traditional deep learning models focus more on extracting global features from the logging response sequence, but are insufficient in capturing local abrupt changes at formation boundaries, resulting in poor resistivity and boundary distance inversion accuracy near formation interfaces. Summary of the Invention

[0004] In view of this, the present invention provides a deep learning inversion method that combines multi-scale decomposition, which can more effectively extract local non-stationary features near formation boundaries and improve the inversion accuracy and stability of the boundary region of highly deviated wells or horizontal wells.

[0005] In a first aspect, the present invention provides a deep learning inversion method combining multi-scale decomposition, the method comprising:

[0006] Step 1: Obtain the nine components of the electromagnetic response signal and construct a dataset; Step 2: Based on Step 1, adopt the multi-scale decomposition-local feature extraction-global sequence modeling architecture to construct the inversion model of maximum overlap discrete wavelet transform MODWT-temporal convolutional network TCN-long short-term memory network LSTM. Step 3: Train the inversion model and obtain the analysis results.

[0007] Optionally, step 1 includes: Construct a horizontally layered three-layered stratigraphic model. In isotropic strata, the well inclination angle is... The stratigraphic angle was set to 60°~89°, the formation resistivity to 1~100Ω·m, and the target layer thickness to 4~7m. In each stratigraphic model, the sampling interval was 0.2m. In the constructed stratigraphic model... The resistivity of the upper layer, The resistivity of the current layer, The resistivity of the lower layer, This represents the distance between the drilling instrument and the upper boundary. The distance between the drilling instrument and the lower boundary is denoted as , where the drilling instrument is a three-dimensional holographic drilling azimuth electromagnetic resistivity instrument. In the instrument coordinate system Given a three-directional unit magnetic moment transmitting and three-directional unit magnetic moment receiving antenna system, the received signal voltage has nine eigenvalues, which can be represented using a component matrix: ; in, express When emitted in a unit magnetic moment direction, The eigenvalue of the voltage signal received by the unit magnetic moment in the direction; the nine components in the component matrix contain the voltage signals of all transmit and receive combinations in the three axes in three-dimensional space, and have the formation information in all directions in the formation space, which can be used as the basic components of the drilling instrument signal. In obtaining the formation response, based on the forward modeling theory of electromagnetic resistivity during drilling, an analytical method is used for forward modeling simulation to calculate... , , , , Five logging responses were analyzed at operating frequencies of 100kHz, 400kHz, and 2MHz. Indicates the amplitude ratio of anisotropic signals. Indicates anisotropic signal due to phase difference. Indicates the amplitude ratio compared to geological signals. Indicates phase difference geological signals, Represents the resistivity signal; The definitions are as follows: ; ; ; in, express When emitted in a unit magnetic moment direction, The eigenvalue of the voltage signal received per unit magnetic moment in the direction; This indicates the calculation of the phase value of the signal; The sliding window method is used to reconstruct the samples, with a window size of 15. This means that each input sample contains well logging response data for the current measuring point and its 14 previous historical measuring points, and the data dimension of a single input sample is 15. 15 represents the sequence length and the 15-dimensional logging response features corresponding to a single depth sampling point, respectively, to obtain depth sequence samples; these depth sequence samples are then used as input to a deep learning model, corresponding to the upper resistivity in the formation model. Current layer resistivity lower layer resistivity Distance between the drilling instrument and the upper boundary Distance of the drilling instrument from the lower boundary As output labels, to quantitatively evaluate the model's inversion performance and generalization ability, the dataset is divided into training set, validation set, and test set in a ratio of 7:2:1.

[0008] Optionally, step 2 includes: First, the original well logging response is decomposed into multiple scales using the Maximum Overlap Discrete Wavelet Transform (MODWT) to separate high-frequency components containing abrupt changes in formation interfaces, i.e., the multi-scale depth sequence. Second, TCN is introduced to extract features from the multi-scale depth sequence. The multi-scale depth sequence is input into the TCN module, and local features are captured synchronously using the causal dilatation convolution mechanism to obtain serialized features. Then, LSTM is used to model the serialized features and learn the nonlinear mapping relationship between formation resistivity and boundary distance. Finally, the inversion results of upper, middle, and lower layer resistivity and upper and lower boundary distances are output through a fully connected layer.

[0009] Optionally, MODWT is introduced as a data preprocessor to perform multi-scale decomposition of the original logging response, aiming to effectively separate the boundary abrupt features in the mixed signal from the background trend and provide response features for the subsequent network; MODWT uses a normalized low-pass scale filter. and high-pass wavelet filter ,in, , The filter length is the low-pass scale filter of the DWT. and high-pass wavelet filter Scaling and It satisfies: ; And satisfy: , ; By performing cyclic filtering on the discrete signal, and ensuring that the length of the decomposition coefficients at each level remains consistent with the original signal without filtering or downsampling, multi-scale fine decomposition is achieved; let the original logging response sequence of length N be... ,in For sampling point index; As the low-frequency approximation coefficients of the 0th layer, for the ... Layer decomposition, ,in If the maximum decomposition level is , then the th Low-frequency approximation coefficients of layer MODWT and high frequency detail coefficient By examining the first The approximation coefficients of the layer are obtained by filtering, and the expression is: ; ; in, For sequence length; The cyclic translation operation for taking the modulus; Using the first Layered periodic low-frequency scale filters and the Layered periodic high-pass wavelet filter Cyclic cross-correlation operations are performed with the low-frequency approximation coefficients and high-frequency detail coefficients, respectively, to reconstruct the decomposed coefficients into low-frequency approximation components with the same length and physical location as the original sequence. and high-frequency detail components The reconstructed expression is: ; ; go through After layer MODWT decomposition and multi-resolution reconstruction, the original well logging response sequence can be represented as the first layer. The sum of the low-frequency approximation components of each layer and the high-frequency detail components of each layer: .

[0010] Optionally, TCN employs a causal dilated convolution structure, the expression of which is: ; in, for Output at all times For the first Each convolutional kernel weight, Input a sequence of historical moments. Used to ensure the causality of sequences, The kernel size is [size]. The coefficient of thermal expansion; During feature extraction, the dilated convolution mechanism in TCN exponentially increases the receptive field without adding extra parameters. The convolution kernels are spaced apart in the sequence. For skip sampling, the receptive field calculation expression is: ; in, To feel the size of the wild, For the number of network layers, For the first The coefficient of thermal expansion of the layer; TCN introduces residual connections. Each residual module contains two layers of causal dilated convolutions and combines ReLU activation function and Dropout regularization to prevent overfitting. At the same time, the residual module adds the input to the output of the convolutional layer through cross-layer connections. When the feature dimensions of the input and output sequences are inconsistent, 1×1 convolution is used for dimension matching.

[0011] Optionally, an LSTM model can be introduced after TCN, utilizing its gating mechanism to capture the variation features in multi-scale depth sequences and achieve the mapping from high-dimensional feature space to formation physical parameters. The LSTM model includes an input gate, a forget gate, an output gate, and an internal state to store long-term memory, and its expression is as follows: ; in, , , , , , These represent the forget gate, input gate, candidate cell state, current cell state, and output gate, respectively. Output of the hidden state at any given time; , , , These represent the weight matrices corresponding to the forget gate, input gate, cell state, and output gate, respectively. , , , These represent the bias terms of the corresponding gating units; , These represent the sigmoid and hyperbolic tangent activation functions, respectively. First of all, At time 1, the LSTM model extracts features based on TCN as input. and the hidden state of the previous moment Calculate the forget gate to control The first step is to determine the degree of information retention in the input gate; secondly, the input gate controls the degree of writing of the current input information, and new feature information is generated from the candidate cell state; after generating the new feature information, the cell state is updated to fuse the current information and historical information; finally, the output gate is based on the updated information. generate The output at the current time step is fed into the fully connected layer to achieve the final mapping of resistivity and boundary distance, and it also participates in the calculation of the next time step.

[0012] Optionally, step 3 includes: In the data preprocessing stage, the db2 wavelet basis function is used to perform a two-layer MODWT decomposition on the input sequence. The low-frequency approximation component A2 of the second layer, the high-frequency detail component D1 of the first layer, and the high-frequency detail component D2 of the second layer are concatenated along the feature dimension to form a multi-scale deep sequence, which is used as the input to the TCN network. In the network architecture, the TCN module contains four residual blocks. To balance the extraction capability of local features and computational efficiency, the convolution kernel size is uniformly set to 2, and the dilation coefficients are set to 1, 2, 4, and 8 respectively to maximize the temporal receptive field. Combined with the padding strategy, the causality is ensured to be the same length as the sequence. The serialized features output by TCN are input to LSTM in a time-by-time form, and the feature vector of each time step is used as the input of LSTM at the corresponding time step. LSTM adopts a three-layer unidirectional structure, with the number of hidden layer units set to 128, 64, and 64 respectively. After LSTM sequence modeling, the hidden state output of the last time step of LSTM is extracted, and then passed through two fully connected layers to finally output the resistivity of the upper, middle, and lower layers and the distance between the upper and lower boundaries. The training uses mean squared error (MSE) as the loss function to quantify the deviation between the predicted and true values; the Adam optimizer is used for parameter updates to balance convergence speed and stability; the batch size is set to 256, the learning rate is set to 0.0008, and the number of training epochs is set to 300. After the model training was completed, a five-layer formation model was selected, and the well inclination angle was set to 80° to simulate the working conditions of a highly deviated well or a horizontal well. The resistivity of each layer was 63Ω·m, 22Ω·m, 10Ω·m, 36Ω·m, and 5Ω·m, respectively. The thicknesses of the three middle layers to be inverted were 4m, 6m, and 5m, respectively. The five-layer formation model was inverted using three models: LSTM, TCN-LSTM, and MODWT-TCN-LSTM. A region-specific analysis strategy based on boundary distance is introduced, defining the distance from a sample to the nearest boundary as... ,when If the current sample is in a boundary region, it is classified as a non-boundary region; otherwise, it is classified as a non-boundary region. The three models are applied to the test set, and the average relative error of resistivity and the average absolute error of boundary distance are calculated for different regions.

[0013] In a second aspect, embodiments of the present invention provide a computer-readable storage medium comprising a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to execute the deep learning inversion method combining multi-scale decomposition in the first aspect or any possible implementation thereof.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions that, when executed by the device, cause the device to perform a deep learning inversion method combining multi-scale decomposition in the first aspect or any possible implementation of the first aspect.

[0015] The technical solution provided by this invention includes acquiring nine components of the electromagnetic response signal and constructing a dataset; employing a multi-scale decomposition-local feature extraction-global sequence modeling architecture to construct an inversion model of maximum overlap discrete wavelet transform (MODWT), temporal convolutional network (TCN), and long short-term memory network (LSTM); training the inversion model and obtaining analysis results. This method can more effectively extract local non-stationary features near formation boundaries, improving the inversion accuracy and stability of the boundary region of highly deviated wells or horizontal wells. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments 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.

[0017] Figure 1 A flowchart of a deep learning inversion method combining multi-scale decomposition provided in an embodiment of the present invention; Figure 2 A schematic diagram of a stratigraphic model provided in an embodiment of the present invention; Figure 3 A schematic diagram of the MODWT-TCN-LSTM inversion model provided in an embodiment of the present invention; Figure 4 A schematic diagram of the upper resistivity inversion results provided in an embodiment of the present invention; Figure 5 A schematic diagram of the current layer resistivity inversion results provided in an embodiment of the present invention; Figure 6 A schematic diagram of the lower-level resistivity inversion results provided in an embodiment of the present invention; Figure 7 A schematic diagram of the upper boundary distance inversion results provided in an embodiment of the present invention; Figure 8 A schematic diagram of the lower boundary distance inversion results provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0019] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0020] Figure 1 The flowchart of the deep learning inversion method combining multi-scale decomposition provided in the embodiments of the present invention is as follows: Figure 1 As shown, the method includes: Step 1: Obtain the nine components of the electromagnetic response signal and construct a dataset.

[0021] In this embodiment of the invention, step 1 includes: Construct a horizontally layered three-layered stratigraphic model. In isotropic strata, the well inclination angle is... The formation angle is set to 60°~89°, the formation resistivity to 1~100Ω·m, and the target layer thickness to 4~7m. In each formation model, the sampling interval is 0.2m to obtain the instrument's complete logging response under that model. Figure 2 As shown, in the constructed stratigraphic model The resistivity of the upper layer, The resistivity of the current layer, The resistivity of the lower layer, This represents the distance between the drilling instrument and the upper boundary. The distance between the drilling instrument and the lower boundary is denoted as , where the drilling instrument is a three-dimensional holographic drilling azimuth electromagnetic resistivity instrument. In the instrument coordinate system Given a three-directional unit magnetic moment transmitting and three-directional unit magnetic moment receiving antenna system, the received signal voltage has nine eigenvalues, which can be represented using a component matrix: ; in, express When emitted in a unit magnetic moment direction, The eigenvalue of the voltage signal received by the unit magnetic moment in the direction; the nine components in the component matrix contain the voltage signals of all transmit and receive combinations in the three axes in three-dimensional space, and have the formation information in all directions in the formation space, which can be used as the basic components of the drilling instrument signal. In obtaining the formation response, based on the forward modeling theory of electromagnetic resistivity during drilling, an analytical method is used for forward modeling simulation to calculate... , , , , Five logging responses were analyzed at operating frequencies of 100kHz, 400kHz, and 2MHz. Indicates the amplitude ratio of anisotropic signals. Indicates anisotropic signal due to phase difference. Indicates the amplitude ratio compared to geological signals. Indicates phase difference geological signals, Represents the resistivity signal; The definitions are as follows: ; ; ; in, express When emitted in a unit magnetic moment direction, The eigenvalue of the voltage signal received per unit magnetic moment in the direction; This indicates the calculation of the phase value of the signal; Since each operating frequency includes the aforementioned five types of logging responses, each depth sampling point contains a total of 15 dimensions of logging response features. Because logging-while-drilling response data is typical depth sequence data, to effectively utilize current and historical logging point information, this invention employs a sliding window method to reconstruct samples. The window size is 15, meaning each input sample contains logging response data for the current logging point and its previous 14 historical logging points, with a data dimension of 15 for a single input sample. 15 represents the sequence length and the 15-dimensional logging response features corresponding to a single depth sampling point, respectively, to obtain depth sequence samples; these depth sequence samples are then used as input to a deep learning model, corresponding to the upper resistivity in the formation model. Current layer resistivity lower layer resistivity Distance between the drilling instrument and the upper boundary Distance of the drilling instrument from the lower boundary As output labels, to quantitatively evaluate the model's inversion performance and generalization ability, the dataset is divided into training set, validation set, and test set in a ratio of 7:2:1.

[0022] Step 2: Based on Step 1, adopt the multi-scale decomposition-local feature extraction-global sequence modeling architecture to construct the inversion model of maximum overlap discrete wavelet transform MODWT-temporal convolutional network TCN-long short-term memory network LSTM.

[0023] In this embodiment of the invention, step 2 includes: To address the issue of poor inversion performance of traditional methods near stratigraphic boundaries, this invention constructs a MODWT-TCN-LSTM inversion model, such as... Figure 3 As shown, firstly, the original well logging response is decomposed into multiple scales using the Maximum Overlap Discrete Wavelet Transform (MODWT) to separate high-frequency components containing abrupt changes in formation interfaces, i.e., the multi-scale depth sequence. Secondly, TCN is introduced to extract features from the multi-scale depth sequence. The multi-scale depth sequence is input into the TCN module, and local features are captured synchronously using the causal dilatation convolution mechanism to obtain serialized features. Then, LSTM is used to model the serialized features and learn the nonlinear mapping relationship between formation resistivity and boundary distance. Finally, the inversion results of upper, middle, and lower layer resistivity and upper and lower boundary distances are output through a fully connected layer.

[0024] In this embodiment of the invention, the electromagnetic resistivity logging response during drilling is a non-stationary signal that varies with depth. When electrical changes occur at the formation interface, the logging response will exhibit violent high-frequency fluctuations near the boundary, while regions far from the boundary will show a relatively stable low-frequency trend. Therefore, this invention introduces MODWT as a data preprocessor to perform multi-scale decomposition of the original logging response, aiming to effectively separate the boundary abrupt change features in the mixed signal from the background trend, providing response features for subsequent networks. MODWT uses a normalized low-pass scale filter. and high-pass wavelet filter ,in, , The filter length; the two filters pass through a traditional DWT low-pass scale filter. and high-pass wavelet filter Scaling yields the following result: ; And satisfy: , ; By performing cyclic filtering on the discrete signal, and ensuring that the length of the decomposition coefficients at each level remains consistent with the original signal without filtering or downsampling, multi-scale fine decomposition is achieved; let the original logging response sequence of length N be... ,in For sampling point index; As the low-frequency approximation coefficients of the 0th layer, for the ... Layer decomposition, ,in If the maximum decomposition level is , then the th Low-frequency approximation coefficients of layer MODWT and high frequency detail coefficient By examining the first The approximation coefficients of the layer are obtained by filtering, and the expression is: ; ; in, For sequence length; The cyclic translation operation for taking the modulus; Because MODWT eliminates the downsampling operation of traditional DWT during decomposition, its decomposition results possess translation invariance. Compared to traditional DWT, MODWT effectively avoids the problem of drastic changes in wavelet coefficients caused by small translations in the logging response sequence; simultaneously, the reconstructed components at each scale maintain a strict equal-length and positional correspondence with the original sequence, which is more conducive to preserving and extracting local abrupt changes near the boundaries. Using the first... Layered periodic low-frequency scale filters and the Layered periodic high-pass wavelet filter Cyclic cross-correlation operations are performed with the low-frequency approximation coefficients and high-frequency detail coefficients, respectively, to reconstruct the decomposed coefficients into low-frequency approximation components with the same length and physical location as the original sequence. and high-frequency detail components The reconstructed expression is: ; ; go through After layer MODWT decomposition and multi-resolution reconstruction, the original well logging response sequence can be represented as the first layer. The sum of the low-frequency approximation components of each layer and the high-frequency detail components of each layer: .

[0025] In this embodiment of the invention, TCN adopts a causal dilated convolution structure. Causality ensures that the model only uses data from current and historical measurement points and cannot use future data, thus meeting the engineering requirements of logging while drilling. Its expression is: ; in, for Output at all times For the first Each convolutional kernel weight, Input a sequence of historical moments. Used to ensure the causality of sequences, The kernel size is [size]. The coefficient of thermal expansion; During feature extraction, the dilated convolution mechanism in TCN exponentially increases the receptive field without adding extra parameters, enabling the model to process long sequences. Each convolutional layer can still process local data within the receptive field, allowing it to capture local features and short-term trends. The convolutional kernels are spaced apart within the sequence. Skip sampling, for example: when the expansion coefficient At this time, there is no gap between the convolution kernels; the convolution calculation is performed directly on the current measurement point and the immediately adjacent historical measurement points; when the dilation coefficient... At this time, there is a measurement point interval between the convolution kernels, and convolution calculation is performed once every other measurement point. This causal dilation convolution structure allows the model to exponentially increase the receptive field without adding additional network parameters. The receptive field calculation expression is: ; in, To feel the size of the wild, For the number of network layers, For the first The coefficient of thermal expansion of the layer; To overcome the vanishing gradient problem in deep network training, TCN introduces residual connections. Each residual module contains two layers of causal dilated convolutions, combined with ReLU activation and Dropout regularization to prevent overfitting. Simultaneously, the residual modules add the input to the output of the convolutional layer through cross-layer connections. When the feature dimensions of the input and output sequences are inconsistent, a 1 / 2 convolutional layer is used. 1. Convolutional layers are used for dimension matching to ensure smooth feature fusion. The introduction of residual modules not only accelerates model convergence but also effectively preserves key non-stationary fluctuation features at the formation interface extracted by MODWT.

[0026] In this embodiment of the invention, the feature sequence extracted by the TCN module contains rich information on local deep dependencies and abrupt changes in formation boundaries. Since the electromagnetic logging-while-drilling (EMD) data is essentially sequence data that varies with the depth of the logging point, this invention introduces an LSTM model after the TCN, utilizing its gating mechanism to capture the changing features in the multi-scale depth sequence and achieve the mapping from a high-dimensional feature space to formation physical parameters. The LSTM model includes an input gate, a forget gate, an output gate, and an internal state to store long-term memory; its expression is: ; in, , , , , , These represent the forget gate, input gate, candidate cell state, current cell state, and output gate, respectively. Output of the hidden state at any given time; , , , These represent the weight matrices corresponding to the forget gate, input gate, cell state, and output gate, respectively. , , , These represent the bias terms of the corresponding gating units; , These represent the sigmoid and hyperbolic tangent activation functions, respectively. First of all, At time 1, the LSTM model extracts features based on TCN as input. and the hidden state of the previous moment Calculate the forget gate to control The degree to which historical information is retained allows the model to selectively retain or discard historical information; secondly, the input gate controls the degree to which current input information is written, and candidate cell states generate new feature information; after generating new feature information, the cell state is updated to fuse current and historical information; finally, the output gate is based on the updated information. generate The output at the current time step is fed into the fully connected layer to achieve the final mapping of resistivity and boundary distance, and it also participates in the calculation of the next time step.

[0027] Step 3: Train the inversion model and obtain the analysis results.

[0028] In this embodiment of the invention, step 3 includes: In the data preprocessing stage, the db2 wavelet basis function is used to perform a two-layer MODWT decomposition on the input sequence. The low-frequency approximation component A2 of the second layer, the high-frequency detail component D1 of the first layer, and the high-frequency detail component D2 of the second layer are concatenated along the feature dimension to form a multi-scale deep sequence, which is used as the input to the TCN network. In the network architecture, the TCN module contains four residual blocks. To balance the extraction capability of local features and computational efficiency, the convolution kernel size is uniformly set to 2, and the dilation coefficients are set to 1, 2, 4, and 8 respectively to maximize the temporal receptive field. Combined with the padding strategy, the causality is ensured to be the same length as the sequence. The serialized features output by TCN are input to LSTM in a time-by-time form, and the feature vector of each time step is used as the input of LSTM at the corresponding time step. LSTM adopts a three-layer unidirectional structure, with the number of hidden layer units set to 128, 64, and 64 respectively. After LSTM sequence modeling, the hidden state output of the last time step of LSTM is extracted, and then passed through two fully connected layers to finally output the resistivity of the upper, middle, and lower layers and the distance between the upper and lower boundaries. The training uses mean squared error (MSE) as the loss function to quantify the deviation between the predicted and true values; the Adam optimizer is used for parameter updates to balance convergence speed and stability; the batch size is set to 256, the learning rate is set to 0.0008, and the number of training epochs is set to 300. After model training was completed, a five-layer formation model was selected, with a well inclination angle set to 80° to simulate the working conditions of highly deviated or horizontal wells. The resistivity of each layer was 63 Ω·m, 22 Ω·m, 10 Ω·m, 36 Ω·m, and 5 Ω·m, respectively. The thicknesses of the three middle layers to be inverted were 4m, 6m, and 5m, respectively. The five-layer formation model was inverted using three models: LSTM, TCN-LSTM, and MODWT-TCN-LSTM. The inversion results are shown below. Figures 4 to 8 As shown; To further evaluate the inversion capability of the proposed MODWT-TCN-LSTM model near formation boundaries in highly deviated or horizontal wells, a regional analysis strategy based on boundary distance is introduced, defining the distance from the sample to the nearest boundary as... ,when When the current sample is in the boundary region, it is divided into the non-boundary region; otherwise, it is divided into the non-boundary region. The three models are applied to the test set, and the average relative error of resistivity and the average absolute error of boundary distance in different regions are statistically analyzed. The results are shown in Table 1.

[0029] Table 1 Comparison of inversion errors in boundary and non-boundary regions ; In this embodiment of the invention, near the formation boundary, due to abrupt changes in electrical parameters, the corresponding well logging response typically exhibits significant non-stationary characteristics, thus presenting a higher inversion difficulty compared to non-boundary regions. As shown in Table 1, the LSTM model shows significantly higher errors in resistivity and boundary distance in the boundary region compared to the non-boundary region, indicating that the traditional LSTM has limited ability to capture local abrupt changes near the formation. In contrast, TCN-LSTM enhances local feature extraction capabilities through causal dilatation convolution, reducing the boundary region error to some extent. Further introduction of MODWT further enhances the error improvement in the boundary region. Specifically, the average relative error of resistivity in the boundary region is reduced by 7.1987 percentage points compared to LSTM, and the average absolute error of boundary distance is reduced by 0.2106 m. In the non-boundary region, the average relative error of resistivity is reduced by 3.9041 percentage points compared to LSTM, and the average absolute error of boundary distance is reduced by 0.1271 m. Compared to non-boundary regions, MODWT-TCN-LSTM shows a more significant improvement in error in boundary regions, indicating that the proposed model can more effectively extract local non-stationary abrupt change features near stratigraphic boundaries. This is mainly due to MODWT's ability to separate multi-scale high-frequency boundary information and TCN's ability to extract local features, thereby improving the inversion accuracy and stability in complex boundary regions.

[0030] To further evaluate the overall inversion performance of different models, the relative error percentage of resistivity and the absolute error percentage of boundary distance were statistically analyzed in the inversion results of the test set, and the results are shown in Table 2.

[0031] Table 2 Percentage of Inversion Parameters ; Compared with LSTM and TCN-LSTM, MODWT-TCN-LSTM shows a higher proportion of small errors in all inversion parameters, while the proportion of large error samples is also reduced. The inversion results near the boundary in Table 1 show that the proposed model can not only improve the inversion capability near the boundary, but also has good stability in the overall stratigraphic parameter inversion.

[0032] In resistivity inversion, MODWT-TCN-LSTM achieved optimal results for samples with errors less than 10%, with the sample proportions for upper, current, and lower resistivity layers reaching 96.2%, 97.6%, and 89.6%, respectively, representing improvements of 7.6, 3.9, and 18.4 percentage points compared to the basic LSTM model. In boundary distance inversion, MODWT-TCN-LSTM achieved a sample proportion of 77.4% within an absolute error range of less than 0.2 m, an improvement of 20.5 percentage points compared to LSTM.

[0033] This invention addresses the problem of insufficient inversion accuracy near formation boundaries in electromagnetic resistivity logging while drilling (EMR) in highly deviated or horizontal wells. It constructs a MODWT-TCN-LSTM inversion model that integrates MODWT, TCN, and LSTM. Results show that the proposed model can effectively separate high-frequency non-stationary abrupt changes near formation boundaries using MODWT, and combines TCN and LSTM to achieve local feature extraction and depth sequence modeling, thereby improving the inversion accuracy of resistivity and boundary distance in the boundary region. Furthermore, it exhibits a higher proportion of small-error samples on the overall test set, indicating that the model also has good performance in terms of overall inversion accuracy and stability, and can be used as a method for inversion in EMR logging while drilling in highly deviated or horizontal wells.

[0034] The technical solution provided by this invention includes acquiring nine components of the electromagnetic response signal and constructing a dataset; employing a multi-scale decomposition-local feature extraction-global sequence modeling architecture to construct an inversion model of maximum overlap discrete wavelet transform (MODWT), temporal convolutional network (TCN), and long short-term memory network (LSTM); training the inversion model and obtaining analysis results. This method can more effectively extract local non-stationary features near formation boundaries, improving the inversion accuracy and stability of the boundary region of highly deviated wells or horizontal wells.

[0035] The various steps in the embodiments of the present invention can be performed by an electronic device. This electronic device includes, but is not limited to, tablet computers, portable PCs, and desktop computers.

[0036] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is running, it controls the electronic device containing the computer-readable storage medium to execute the above-described embodiment of the deep learning inversion method combined with multi-scale decomposition.

[0037] Figure 9 A schematic diagram of an electronic device provided in an embodiment of the present invention, such as... Figure 9 As shown, the electronic device 21 includes a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211. When the computer program 213 is executed by the processor 211, it implements the deep learning inversion method combined with multi-scale decomposition in the embodiment. To avoid repetition, it will not be described in detail here.

[0038] Electronic device 21 includes, but is not limited to, processor 211 and memory 212. Those skilled in the art will understand that... Figure 9This is merely an example of electronic device 21 and does not constitute a limitation on electronic device 21. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0039] The processor 211 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0040] The memory 212 can be an internal storage unit of the electronic device 21, such as a hard disk or RAM of the electronic device 21. The memory 212 can also be an external storage device of the electronic device 21, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the electronic device 21. Furthermore, the memory 212 can include both internal and external storage units of the electronic device 21. The memory 212 is used to store computer programs and other programs and data required by network devices. The memory 212 can also be used to temporarily store data that has been output or will be output.

[0041] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0042] 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 scope of protection of the present invention.

Claims

1. A deep learning inversion method combining multi-scale decomposition, characterized in that, The method includes: Step 1: Obtain the nine components of the electromagnetic response signal and construct a dataset; Step 2: Based on Step 1, adopt the multi-scale decomposition-local feature extraction-global sequence modeling architecture to construct the inversion model of maximum overlap discrete wavelet transform MODWT-temporal convolutional network TCN-long short-term memory network LSTM. Step 3: Train the inversion model and obtain the analysis results; Step 1 includes: Construct a horizontally layered three-layered stratigraphic model. In isotropic strata, the well inclination angle is... The stratigraphic angle was set to 60°~89°, the formation resistivity to 1~100Ω·m, and the target layer thickness to 4~7m. In each stratigraphic model, the sampling interval was 0.2m. In the constructed stratigraphic model... The resistivity of the upper layer, The resistivity of the current layer, The resistivity of the lower layer, This represents the distance between the drilling instrument and the upper boundary. The distance between the drilling instrument and the lower boundary is denoted as , where the drilling instrument is a three-dimensional holographic drilling azimuth electromagnetic resistivity instrument. In the instrument coordinate system Given a three-directional unit magnetic moment transmitting and three-directional unit magnetic moment receiving antenna system, the received signal voltage has nine eigenvalues, which can be represented using a component matrix: ； in, express When emitted in a unit magnetic moment direction, The eigenvalue of the voltage signal received by the unit magnetic moment in the direction; the nine components in the component matrix contain the voltage signals of all transmit and receive combinations in the three axes in three-dimensional space, and have the formation information in all directions in the formation space, which can be used as the basic components of the drilling instrument signal. In obtaining the formation response, based on the forward modeling theory of electromagnetic resistivity during drilling, an analytical method is used for forward modeling simulation to calculate... , , , , Five logging responses were analyzed at operating frequencies of 100kHz, 400kHz, and 2MHz. Indicates the amplitude ratio of anisotropic signals. Indicates anisotropic signal due to phase difference. Indicates the amplitude ratio compared to geological signals. Indicates phase difference geological signals, Represents the resistivity signal; The definitions are as follows: ； ； ； in, express When emitted in a unit magnetic moment direction, The eigenvalue of the voltage signal received per unit magnetic moment in the direction; This indicates the calculation of the phase value of the signal; The sliding window method is used to reconstruct the samples, with a window size of 15. This means that each input sample contains well logging response data for the current measuring point and its 14 previous historical measuring points, and the data dimension of a single input sample is 15. 15 represents the sequence length and the 15-dimensional logging response features corresponding to a single depth sampling point, respectively, to obtain depth sequence samples; these depth sequence samples are then used as input to a deep learning model, corresponding to the upper resistivity in the formation model. Current layer resistivity lower layer resistivity Distance between the drilling instrument and the upper boundary Distance of the drilling instrument from the lower boundary As output labels, to quantitatively evaluate the model's inversion performance and generalization ability, the dataset is divided into training set, validation set, and test set in a ratio of 7:2:

1. Step 2 includes: First, the original well logging response is decomposed into multiple scales using the Maximum Overlap Discrete Wavelet Transform (MODWT) to separate high-frequency components containing abrupt changes in formation interfaces, i.e., the multi-scale depth sequence. Second, TCN is introduced to extract features from the multi-scale depth sequence. The multi-scale depth sequence is input into the TCN module, and local features are captured synchronously using the causal dilatation convolution mechanism to obtain serialized features. Then, LSTM is used to model the serialized features and learn the nonlinear mapping relationship between formation resistivity and boundary distance. Finally, the inversion results of upper, middle, and lower layer resistivity and upper and lower boundary distances are output through a fully connected layer.

2. The method according to claim 1, characterized in that, MODWT is introduced as a data preprocessor to perform multi-scale decomposition of the raw logging response, aiming to effectively separate the boundary abrupt features in the mixed signal from the background trend and provide response features for the subsequent network; MODWT uses a normalized low-pass scale filter. and high-pass wavelet filter ,in, , The filter length is the low-pass scale filter of the DWT. and high-pass wavelet filter Scaling and It satisfies: ； And satisfy: ， ； By performing cyclic filtering on the discrete signal, and ensuring that the length of the decomposition coefficients at each level remains consistent with the original signal without filtering or downsampling, multi-scale fine decomposition is achieved; let the original logging response sequence of length N be... ,in For sampling point index; As the low-frequency approximation coefficients of the 0th layer, for the ... Layer decomposition, ,in If the maximum decomposition level is , then the th Low-frequency approximation coefficients of layer MODWT and high frequency detail coefficient By examining the first The approximation coefficients of the layer are obtained by filtering, and the expression is: ； ； in, For sequence length; The cyclic translation operation for taking the modulus; Using the first Layered periodic low-frequency scale filters and the Layered periodic high-pass wavelet filter Cyclic cross-correlation operations are performed with the low-frequency approximation coefficients and the high-frequency detail coefficients, respectively, to reconstruct the decomposed coefficients into low-frequency approximation components with the same length and physical location as the original sequence. and high-frequency detail components The reconstructed expression is: ； ； go through After layer MODWT decomposition and multi-resolution reconstruction, the original well logging response sequence can be represented as the first layer. The sum of the low-frequency approximation components of each layer and the high-frequency detail components of each layer: 。 3. The method according to claim 1, characterized in that, TCN employs a causal dilated convolution structure, the expression of which is: ； in, for Output at all times For the first Each convolutional kernel weight, Input a sequence of historical moments. Used to ensure the causality of sequences, The kernel size is [size]. The coefficient of thermal expansion; During feature extraction, the dilated convolution mechanism in TCN exponentially increases the receptive field without adding extra parameters. The convolution kernels are spaced apart in the sequence. For skip sampling, the receptive field calculation expression is: ； in, To feel the size of the wild, For the number of network layers, For the first The coefficient of thermal expansion of the layer; TCN introduces residual connections. Each residual module contains two layers of causal dilated convolutions and combines ReLU activation function and Dropout regularization to prevent overfitting. At the same time, the residual module adds the input to the output of the convolutional layer through cross-layer connections. When the feature dimensions of the input and output sequences are inconsistent, 1×1 convolution is used for dimension matching.

4. The method according to claim 1, characterized in that, Following TCN, the LSTM model is introduced, utilizing its gating mechanism to capture the variation features in multi-scale depth sequences and achieve the mapping from high-dimensional feature space to formation physical parameters. The LSTM model includes an input gate, a forget gate, an output gate, and an internal state to store long-term memory, and its expression is: ； in, , , , , , These represent the forget gate, input gate, candidate cell state, current cell state, and output gate, respectively. Output of the hidden state at any given time; , , , These represent the weight matrices corresponding to the forget gate, input gate, cell state, and output gate, respectively. , , , These represent the bias terms of the corresponding gating units; , These represent the sigmoid and hyperbolic tangent activation functions, respectively. First of all, At time 1, the LSTM model extracts features based on TCN as input. and the hidden state of the previous moment Calculate the forget gate to control The first step is to determine the degree of information retention in the input gate; secondly, the input gate controls the degree of writing of the current input information, and new feature information is generated from the candidate cell state; after generating the new feature information, the cell state is updated to fuse the current information and historical information; finally, the output gate is based on the updated information. generate The output at the current time step is fed into the fully connected layer to achieve the final mapping of resistivity and boundary distance, and it also participates in the calculation of the next time step.

5. The method according to claim 1, characterized in that, Step 3 includes: In the data preprocessing stage, the db2 wavelet basis function is used to perform a two-layer MODWT decomposition on the input sequence. The low-frequency approximation component A2 of the second layer, the high-frequency detail component D1 of the first layer, and the high-frequency detail component D2 of the second layer are concatenated along the feature dimension to form a multi-scale deep sequence, which is used as the input to the TCN network. In the network architecture, the TCN module contains four residual blocks. To balance the extraction capability of local features and computational efficiency, the convolution kernel size is uniformly set to 2, and the dilation coefficients are set to 1, 2, 4, and 8 respectively to maximize the temporal receptive field. Combined with the padding strategy, the causality is ensured to be the same length as the sequence. The serialized features output by TCN are input to LSTM in a time-by-time form, and the feature vector of each time step is used as the input of LSTM at the corresponding time step. LSTM adopts a three-layer unidirectional structure, with the number of hidden layer units set to 128, 64, and 64 respectively. After LSTM sequence modeling, the hidden state output of the last time step of LSTM is extracted, and then passed through two fully connected layers to finally output the resistivity of the upper, middle, and lower layers and the distance between the upper and lower boundaries. The training uses mean squared error (MSE) as the loss function to quantify the deviation between the predicted and true values; the Adam optimizer is used for parameter updates to balance convergence speed and stability; the batch size is set to 256, the learning rate is set to 0.0008, and the number of training epochs is set to 300. After the model training was completed, a five-layer formation model was selected, and the well inclination angle was set to 80° to simulate the working conditions of a highly deviated well or a horizontal well. The resistivity of each layer was 63Ω·m, 22Ω·m, 10Ω·m, 36Ω·m, and 5Ω·m, respectively. The thicknesses of the three middle layers to be inverted were 4m, 6m, and 5m, respectively. The five-layer formation model was inverted using three models: LSTM, TCN-LSTM, and MODWT-TCN-LSTM. A region-specific analysis strategy based on boundary distance is introduced, defining the distance from a sample to the nearest boundary as... ,when If the current sample is in a boundary region, it is classified as a non-boundary region; otherwise, it is classified as a non-boundary region. The three models are applied to the test set, and the average relative error of resistivity and the average absolute error of boundary distance are calculated for different regions.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the deep learning inversion method combining multi-scale decomposition as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, include: One or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs including instructions that, when executed by the device, cause the device to perform the deep learning inversion method combined with multi-scale decomposition as described in any one of claims 1 to 5.

Images