Stratum parameter inversion method of inversion neural network based on asymmetric down-sampling

Abstract

The invention discloses a stratum parameter inversion method of an inversion neural network based on asymmetric down-sampling, and relates to a while-drilling azimuth electromagnetic wave data processing technology, and the method comprises the steps: enabling adjacent one-dimensional stratum samples to form a locally continuous two-dimensional stratum sample; carrying out nonlinear compression on R response and G response in the logging response data, and then standardizing to obtain preprocessed logging electromagnetic response data; the preprocessed logging electromagnetic response data serve as input data of an inversion neural network, the corresponding two-dimensional stratum samples serve as labels, and training of the inversion neural network is completed; the inversion neural network adopts a network topology structure of an encoder and a decoder, feature maps of the same size are fused through jump connection, feature maps of different sizes are subjected to horizontal single-dimension down-sampling and up-sampling through a down-sampling module based on maximum pooling and an up-sampling module based on transpose convolution, vertical dimension information is reserved, and the feature maps of different sizes are subjected to horizontal single-dimension down-sampling and up-sampling. And high-precision real-time stratum parameter inversion can be realized.

Description

Technical Field

[0001] The present invention belongs to a technology for processing azimuthal electromagnetic wave data while drilling, and in particular relates to a technology for inverting formation parameters based on an asymmetric down-sampling inversion neural network. Background Art

[0002] Traditional logging while drilling (LWD) formation parameter inversion methods rely on fully connected neural networks or symmetric encoder-decoder structures (such as U-Net), which have the following problems:

[0003] Single feature extraction: Traditional convolution kernels are difficult to capture local details, regional context, and vertical correlation features simultaneously;

[0004] Resolution loss: Symmetric downsampling operations (such as 2×2 pooling) lead to vertical resolution loss, affecting the recognition of thin layer boundaries;

[0005] Insufficient physical continuity: Randomly generated training samples lack the continuous variation characteristics of geological parameters, resulting in limited model generalization ability.

[0006] While UNet++ improves multi-scale feature fusion capabilities through dense skip connections, it still cannot effectively address the feature dilution problem in high-dimensional electromagnetic logging data. Therefore, an improved solution combining physically constrained sample generation with network architecture optimization is urgently needed. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide an inversion method for improving the inversion performance of formation parameters.

[0008] The technical solution adopted by the present invention to solve the above technical problems is a formation parameter inversion method based on an inversion neural network with asymmetric downsampling, comprising the following steps:

[0009] Formation sample generation steps: Determine the formation model, and among the N generated one-dimensional formation samples, form a locally continuous two-dimensional formation sample with every M adjacent one-dimensional formation samples. Each column of the two-dimensional formation sample is a single-point one-dimensional formation sample. The continuity change between the M adjacent one-dimensional formation samples in the two-dimensional formation sample reflects the physical constraints of the gradual change of geological parameters. The formation parameters of the one-dimensional formation sample include formation resistivity and formation boundary.

[0010] Well logging electromagnetic response data preprocessing steps: N one-dimensional formation samples are forward-modeled to obtain well logging electromagnetic response data, the R response and G response in the well logging response data are nonlinearly compressed, and the well logging electromagnetic response after nonlinear compression is standardized to make the data feature distribution more uniform; the R response is a response that is strongly correlated with resistivity, and the G response is a response that is strongly correlated with formation boundaries;

[0011] Inversion neural network training steps: using the pre-processed well logging electromagnetic response data as input data and the two-dimensional formation samples corresponding to the pre-processed well logging electromagnetic response data as labels to complete the inversion neural network training;

[0012] The inversion neural network adopts the network topology of encoder and decoder, and the feature maps of the same size are fused through jump connections. The feature maps of different sizes are horizontally downsampled and upsampled by a downsampling module based on maximum pooling and an upsampling module based on transposed convolution, while retaining the vertical dimension information.

[0013] Test steps: The trained inversion neural network is used to receive input real-time logging electromagnetic response data, output predicted formation parameters, and complete the inversion of actual formation parameters.

[0014] The neural inversion method of the present invention improves the formation parameter inversion performance through the following innovations:

[0015] The inversion neural network adopts a multi-scale dynamic feature fusion module and an asymmetric downsampling strategy: in the multi-scale dynamic feature fusion module, the outputs of each branch are weighted and fused, and the weight parameters are adaptively optimized through back propagation; horizontal single-dimensional downsampling is adopted to compress only the horizontal resolution and retain the vertical dimension information; the decoder uses transposed convolution for upsampling to ensure that the feature map sizes are strictly aligned.

[0016] Physically constrained sample generation: Generate continuous stratigraphic samples based on a Markov chain perturbation strategy to simulate the gradual variation of geological parameters; every M 1D samples are spliced into 2D input to enhance the convolutional neural network's ability to extract spatial features.

[0017] Hybrid feature compression transformation: Nonlinear compression is performed on the R response and G response in the logging response data to solve the problem of uneven data distribution. After standardization, the feature distribution is more uniform, which improves the model's sensitivity to weak signals.

[0018] The beneficial effects of the present invention are that, based on the improved network structure, a physical constraint sample generation strategy, a hybrid feature compression change and an LWD formation parameter inversion network are integrated, and the electromagnetic logging inversion accuracy and efficiency are improved through multi-scale feature fusion and asymmetric downsampling strategy, which solves the problem of high-precision inversion of the formation resistivity and formation boundary position around the logging instrument through the while-drilling azimuthal electromagnetic wave logging response during the logging while drilling process; it can achieve high-precision real-time formation parameter inversion and provide reliable geological guidance for logging decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the three-layer stratigraphic model;

[0020] Figure 2 Generate flow charts for physical constraint samples;

[0021] Figure 3 This is the R-BInvNet network architecture diagram;

[0022] Figure 4 This is a stratigraphic model diagram of the structure;

[0023] Figure 5 This is the inversion effect diagram. DETAILED DESCRIPTION

[0024] Steps for setting up the formation model:

[0025] The formation model targeted by the present invention is as follows Figure 1 As shown, the formation model is set to a three-layer formation model, including the upper layer R1, the middle layer R2 and the lower layer R3. The logging while drilling instrument is placed in the middle layer, and the instrument center coordinate z is always 0, that is, z = 0. The boundary between the middle layer R2 and the upper layer R1 is the upper boundary, and the upper boundary position is represented by B up , that is, z=B up The boundary between the middle layer R2 and the lower layer R3 is the lower boundary, and the position of the lower boundary is represented by B dn , that is, z=B dn .

[0026] The generation method of training samples for formation parameter inversion neural network is based on Markov chain perturbation strategy, which is used to generate continuous formation samples that simulate the gradual change characteristics of geological parameters. The process is as follows: Figure 2 As shown:

[0027] First, set the number of samples N and the sample parameter range. In this embodiment, the range of the formation boundary is set to -6 meters to 6 meters, and the range of the three formation resistivities is set to 1~100Ω / m. The i-th formation sample i includes the formation resistivity R i and boundary position B i,up and B i,dn At this time, the data of the formation sample is one-dimensional.

[0028] Then, in order to make more full use of the performance of the convolutional neural network to extract edge features, the sample needs to be changed from one-dimensional to two-dimensional. Considering the continuity of the formation, the present invention proposes to transform a part of adjacent one-dimensional formation samples into a locally continuous two-dimensional formation sample. The number of single-point samples included in the two-dimensional formation sample is set according to actual needs. In an embodiment, every M = 8 adjacent one-dimensional formation samples form a locally continuous two-dimensional formation sample. Each column of the two-dimensional formation sample is a single-point one-dimensional formation sample, and each row is a formation value of 8 adjacent formation samples.

[0029] Formation sample generation steps:

[0030] When generating the i-th formation sample, first judge whether the remainder of the current i-th sample divided by 8 is 1. If so, generate an independent new sample; otherwise, add perturbations to the previous sample to generate a continuous new sample. This can ensure the continuity of every 8 continuously generated samples.

[0031] If so, to avoid overfitting and ensure the diversity of the generated two-dimensional formation samples, a random sample is generated within the preset sample range; otherwise, a Markov chain perturbation strategy under physical constraints is adopted, and the value of sample i will be perturbed based on the i-1-th sample. The perturbation range △d of the formation boundary satisfies 0.05 < |△d| < 0.2, so that adjacent one-dimensional samples have the smallest change to avoid ineffective fine-tuning and will not be too large to exceed the normal formation change range; the perturbation range △R of the resistivity satisfies 0.02 < |△R| < 0.1 (the resistivity takes the logarithmic value), simulating the physical property gradual change caused by underground fluid migration. And the upper boundary coordinate B i,up and the lower boundary coordinate B i,dn satisfy B i,up > B i,dn + △d min where the minimum spacing △d min = 0.5, which can prevent the upper and lower boundaries from crossing and reduce the occurrence of singular values during the forward modeling process.

[0032] Based on the constructed samples, call the electromagnetic forward modeling program. The transmitting antenna is equivalent to an electromagnetic dipole source, and the formation is a layered medium. The logging electromagnetic response data corresponding to the given formation sample, frequency, and source distance are calculated through the one-dimensional Green's function; the calculated logging electromagnetic response data is input and stored as the observation data of the inversion network.

[0033] Continuously generate samples until the set scale N is reached. In an embodiment, N = 400000 is adopted to ensure the sufficiency of the training data set. When the sample size does not reach the threshold i < N, then continue to update i = i + 1, start the i-th formation sample generation step, and end the sample generation until i = N is satisfied.

[0034] Well logging electromagnetic response data preprocessing steps:

[0035] The present invention pre-processes the observation data input into the neural network to make the data distribution in the data set more reasonable, so that the neural network can better fit the target, accelerate convergence, and improve training efficiency and effect. In this embodiment, the electromagnetic response data corresponding to the well logging of a formation sample is composed of 8 sub-responses. , It is a real number domain, among which 4 sub-responses have a strong correlation with resistivity and are called R responses. The other 4 sub-responses have a strong correlation with the formation boundary.

[0036] The nonlinear compression processing method for R response is:

[0037] ;

[0038] is the jth sub-response with a strong correlation with resistivity in the logging electromagnetic response data of the i-th formation sample, After nonlinear compression ; To preset the first fine-tuning parameter, in the embodiment ;

[0039] The nonlinear compression processing method for G response is:

[0040] ;

[0041] The jth sub-response in the logging electromagnetic response data of the i-th formation sample that has a strong correlation with the formation boundary, After nonlinear compression ; ln is the natural logarithm, α presets the second fine-tuning parameter, and in the embodiment, α=0.7.

[0042] Merge the R response and G response of all formation samples after nonlinear compression processing to obtain the logging electromagnetic response data set after nonlinear compression. ;

[0043] Then the data set To perform standardization:

[0044] Calculation Dataset The mean of the data distribution for each seed response in and standard deviation :

[0045] ;

[0046] ;

[0047] in, For the dataset The response of row i and column j in Well logging electromagnetic response dataset The mean of the data distribution in row j, Well logging electromagnetic response dataset The standard deviation of the data distribution in row j;

[0048] Reuse and And the first fine-tuning parameter Complete the final data preprocessing and obtain the preprocessed logging electromagnetic response data set :

[0049] ;

[0050] for The sub-response in the i-th row and j-th column is used as the observation data input of the inversion neural network, and the formation resistivity and boundary in the corresponding formation sample are normalized and used as the label of the inversion neural network.

[0051] Inversion neural network training steps:

[0052] First, an inversion neural network is constructed. The improvement of the formation parameter inversion method of the present invention is based on the structure of multi-scale dynamic feature fusion and asymmetric downsampling.

[0053] Establish as Figure 3 The inversion neural network model of the present invention shown is called the resistivity-boundary inversion neural network R-BInvNet, which is improved on the existing UNet++ and includes a 3×3 convolution block Convolution (3×3), a multi-scale dynamic feature fusion Multi-Feature Fusion module, an upsampling module based on transposed convolution, a downsampling module based on maximum pooling MaxPool, and a 1×4 convolution block Convolution (1×4). The downsampling module based on maximum pooling retains significant features and reduces the amount of data by taking the maximum value in the local area. The present invention uses a downsampling module based on maximum pooling to adopt horizontal single-dimensional downsampling, only compressing the horizontal resolution and retaining the vertical dimension information. The upsampling module based on transposed convolution realizes the upsampling convolution through inverse operation. The present invention uses an upsampling module based on transposed convolution to perform horizontal single-dimensional upsampling to ensure that the feature map sizes are strictly aligned.

[0054] The network topology of the R-BInvNet embodiment is an encoder and decoder structure, and feature maps of the same size are fused through skip connections:

[0055] Encoder path:

[0056] Input layer: receives the 8×8 single-channel feature map of the input;

[0057] First stage: The 3×3 convolution block converts the 8×8 feature map from 1 channel to 32 channels. The 32-channel 8×8 feature map is fused through the multi-scale dynamic feature fusion module to obtain the 32-channel 8×8 first-level fusion feature. The first-level fusion feature is downsampled through the maximum pooling with a pooling kernel of 1×2 and then output as a 32-channel 8×4 feature map to the second stage.

[0058] Second stage: The 3×3 convolution block converts the 32-channel 8×4 feature map from 32 channels to 64 channels. The 64-channel feature map is fused through the multi-scale dynamic feature fusion module to obtain the 64-channel 8×4 second-level first fusion feature. The second-level first fusion feature is downsampled through the maximum pooling with a pooling kernel of 1×2 and then output as a 64-channel 8×2 feature map to the third stage.

[0059] The third level: The 3×3 convolution block converts the 64-channel 8×2 feature map from 64 channels to 128-channel 8×2 feature map. The 128-channel 8×2 feature map is fused by the multi-scale dynamic feature fusion module to obtain the 128-channel 8×2 third-level feature.

[0060] Decoder path:

[0061] Primary upsampling: The 64-channel 8×4 second-level first fusion feature is converted from a 64-channel to a 32-channel 8×8 feature map through transposed convolution. It is then concatenated with the 32-channel 8×8 first-level first fusion feature through skip connection. The concatenated feature is passed through a 3×3 convolution block and the multi-scale dynamic feature fusion module to obtain a 32+32-channel 8×8 first-level second fusion feature.

[0062] Secondary upsampling: The 128-channel 8×2 third-level features are converted from 128 channels to 64 channels through transposed convolution, and then spliced with the second-level first fusion features through jump connection. The spliced features are passed through a 3×3 convolution block and a multi-scale dynamic feature fusion module to obtain the 64+64-channel 8×4 second-level second fusion features;

[0063] Final upsampling: The 64+64-channel 8×4 second-level second fusion features are converted from 64 channels to 32 channels through transposed convolution, and then spliced with the first-level first fusion features and the first-level second fusion features through skip connections. The spliced features are passed through a 3×3 convolution block and a multi-scale dynamic feature fusion module to obtain a 32+32+32-channel 8×8 first-level third fusion feature.

[0064] Output layer: The 32+32+32 channel 8×8 first-level third fusion feature is converted into a 32-channel 8×8 feature map through a 3×3 convolution block, and finally passes through a 1×4 convolution block Convolution (1×4) to output the final fusion feature of a single channel 8×5 target size.

[0065] The multi-scale dynamic feature fusion module includes a local detail branch, a regional context branch, and a vertical correlation branch, and fuses multi-scale features through dynamic weights. Its structure is a parallel multi-branch convolution, with three branches:

[0066] The feature extraction of the local detail branch obtains the first branch feature : Use a standard convolution with a convolution kernel of 3×3 without expansion, dilation=1, and padding=1 on each side. A smaller convolution kernel is beneficial for extracting high-frequency features.

[0067] The feature extraction of the regional context branch obtains the second branch feature : Using a 3×3 dilated convolution kernel, each element will be separated by one eigenvalue, and the dilation coefficient is 2, that is, dilation=2, which will expand the receptive field to 5×5 and capture low-frequency features;

[0068] The feature extraction of the vertical correlation branch obtains the third branch feature : Using 1×3 strip convolution, the horizontal convolution is enhanced to capture the internal correlation of the logging electromagnetic response corresponding to each 2D formation sample;

[0069] Then the three branch features are dynamically fused through 1×1 convolution Each feature Perform weighted fusion, formula:

[0070] ;

[0071] in is the weight parameter of the lth branch, which can be adaptively optimized by the network according to back propagation.

[0072] The R-BInvNet network of the present invention adopts an asymmetric up-and-down sampling strategy between different scales. In view of the vertical continuity of LWD data, R-BInvNet abandons the traditional 2×2 pooling and adopts horizontal single-dimensional downsampling. The height of the convolution kernel is 1 and the width is 2, kernel size = (1, 2), and two adjacent eigenvalues are fused in the width direction, while the height direction remains unchanged, that is, only a horizontally adjacent 1×2 area of the input is covered each time; the step size in the height direction is 1 without jump, and the step size in the width direction is 2, stride = (1, 2), and the output size in the width direction is halved to achieve downsampling, while the height direction remains unchanged. This strategy retains the vertical dimension information, avoids the loss of shallow details, and provides support for cross-layer feature recovery of the decoder. The corresponding upsampling stage uses 1×2 transposed convolution to ensure that the feature map size is strictly aligned with the input.

[0073] Both network training and testing were accelerated using an NVIDIA GeForce RTX 4070 Ti SUPER GPU in the PyTorch framework. The loss function was set to:

[0074] ;

[0075] in is the mean square error loss, , is the i-th stratum sample as the label, is the i-th predicted stratum sample output by the inversion neural network , is the L2 norm, is the R-BInvNet network weight parameter set, is the L1 regularization coefficient, used to prevent overfitting. During training, the Adam optimizer was used for parameter update, with an initial learning rate of 0.001, a batch size of 128, and 200 epochs of training.

[0076] Test steps:

[0077] The trained R-BInvNet network is used to receive input real-time logging electromagnetic response data, output formation parameters, and complete the inversion of actual formation parameters.

[0078] The embodiment is constructed as follows Figure 4 The formation model shown in the figure is calculated by forward modeling to obtain electromagnetic response, and then inverted by trained R-BInvNet. The horizontal and vertical coordinates are in meters, representing width and depth respectively, and the grayscale is the resistivity expressed in logarithms. Figure 5As shown in the figure, where the horizontal and vertical axes are in meters, representing width and depth, respectively, and the grayscale represents the resistivity expressed logarithmically. This demonstrates that the embodiment can achieve high-precision real-time inversion. In actual projects, after receiving electromagnetic response data from logging instruments, they are input into the trained R-BInvNet. The network outputs the formation resistivity and boundary locations, rendering the formation image in real time.

Claims

1. A formation parameter inversion method based on an asymmetric downsampling inversion neural network, characterized in that: The following steps are involved: Formation sample generation steps: Determine the formation model, and among the N generated one-dimensional formation samples, form a locally continuous two-dimensional formation sample with every M adjacent one-dimensional formation samples. Each column of the two-dimensional formation sample is a single-point one-dimensional formation sample. The continuity change between the M adjacent one-dimensional formation samples in the two-dimensional formation sample reflects the physical constraints of the gradual change of geological parameters. The formation parameters of the one-dimensional formation sample include formation resistivity and formation boundary. Well logging electromagnetic response data preprocessing steps: N one-dimensional formation samples are forward-modeled to obtain well logging electromagnetic response data, the R response and G response in the well logging response data are nonlinearly compressed, and the well logging electromagnetic response after nonlinear compression is standardized to make the data feature distribution more uniform; the R response is a response that is strongly correlated with resistivity, and the G response is a response that is strongly correlated with formation boundaries; Inversion neural network training steps: using the pre-processed well logging electromagnetic response data as input data and the two-dimensional formation samples corresponding to the pre-processed well logging electromagnetic response data as labels to complete the inversion neural network training; The inversion neural network adopts the network topology of encoder and decoder, and the feature maps of the same size are fused through jump connections. The feature maps of different sizes are horizontally downsampled and upsampled by a downsampling module based on maximum pooling and an upsampling module based on transposed convolution, while retaining the vertical dimension information. Test steps: The trained inversion neural network is used to receive input real-time logging electromagnetic response data, output predicted formation parameters, and complete the inversion of actual formation parameters.

2. The formation parameter inversion method based on the asymmetric downsampling inversion neural network according to claim 1, characterized in that: The formation model is a three-layer formation model, including an upper layer, a middle layer and a lower layer. The logging while drilling instrument is placed in the middle layer, and the center coordinate of the instrument is always 0; The formation parameters include the formation resistivity of the upper layer, the middle layer and the lower layer, and the upper and lower formation boundaries. The upper formation boundary is the boundary between the upper layer and the middle layer, and the lower formation boundary is the boundary between the middle layer and the lower layer.

3. The formation parameter inversion method based on the asymmetric downsampling inversion neural network according to claim 2, characterized in that: Among every M adjacent one-dimensional formation samples, one one-dimensional formation sample is randomly generated, and the other M-1 one-dimensional formation samples are obtained by adding disturbance to the randomly generated one-dimensional formation sample.

4. The method according to claim 3, wherein: In the formation sample generation step, when generating the i-th one-dimensional formation sample, first determine whether the remainder of the current sample number i divided by M is 1. If so, generate an independent new sample. Otherwise, add disturbance to the i-1-th one-dimensional formation sample to generate a continuous new sample. Among them, the disturbance range △d of the formation boundary satisfies 0.05<|△d|<0.2, the disturbance range △R of the logarithmic resistivity satisfies 0.02<|△R|<0.1, and the upper boundary coordinate B of the formation is 0. i,up and the coordinates of the lower boundary of the stratum B i,dn Satisfy B i,up >B i,dn +△d min , where the minimum spacing △d min =0.

5.

5. The formation parameter inversion method based on the asymmetric downsampling inversion neural network according to claim 1, characterized in that: The nonlinear compression of the R response and G response in the logging response data is performed as follows: The nonlinear compression processing method for R response is: ； is the jth sub-response with strong correlation with resistivity in the logging electromagnetic response data of the i-th one-dimensional formation sample, After nonlinear compression ; is to preset the first fine-tuning parameter; The nonlinear compression processing method for G response is: ； The jth sub-response in the logging electromagnetic response data of the i-th one-dimensional formation sample that has a strong correlation with the formation boundary, After nonlinear compression ; ln is the natural logarithm, and α is the second fine-tuning parameter; Merge the R response and G response of all one-dimensional formation samples after nonlinear compression processing to obtain the logging electromagnetic response data set after nonlinear compression. .

6. The formation parameter inversion method based on the asymmetric downsampling inversion neural network according to claim 5, characterized in that: The normalization process of the logging electromagnetic response after nonlinear compression is as follows: Calculation Dataset The mean of the data distribution for each seed response in and standard deviation : ； ； in, For the dataset The response of row i and column j in Well logging electromagnetic response dataset The mean of the data distribution in row j, Well logging electromagnetic response dataset The standard deviation of the data distribution in row j; Reuse and And the first fine-tuning parameter Complete the final data preprocessing and obtain the preprocessed logging electromagnetic response data set : ； for The response in row i and column j.

7. The formation parameter inversion method based on the asymmetric downsampling inversion neural network according to claim 1, characterized in that: The network topology of the inversion neural network includes an encoder path and a decoder path; The encoder path includes: Input layer: receives the input 8×8 single-channel feature map; First stage: The 3×3 convolution block converts the 8×8 feature map from 1 channel to 32 channels. The 32-channel 8×8 feature map is fused through the multi-scale dynamic feature fusion module to obtain the 32-channel 8×8 first-level fusion feature. The first-level fusion feature is downsampled through the maximum pooling with a pooling kernel of 1×2 and then output as a 32-channel 8×4 feature map to the second stage. Second stage: The 3×3 convolution block converts the 32-channel 8×4 feature map from 32 channels to 64 channels. The 64-channel feature map is fused through the multi-scale dynamic feature fusion module to obtain the 64-channel 8×4 second-level first fusion feature. The second-level first fusion feature is downsampled through the maximum pooling with a pooling kernel of 1×2 and then output as a 64-channel 8×2 feature map to the third stage. The third level: The 3×3 convolution block converts the 64-channel 8×2 feature map from 64 channels to 128-channel 8×2 feature map. The 128-channel 8×2 feature map is fused by the multi-scale dynamic feature fusion module to obtain the 128-channel 8×2 third-level feature. Decoder path: Primary upsampling: The 64-channel 8×4 second-level first fusion feature is converted from a 64-channel to a 32-channel 8×8 feature map through transposed convolution, and then concatenated with the 32-channel 8×8 first-level first fusion feature through jump connection. The concatenated feature is passed through a 3×3 convolution block and a multi-scale dynamic feature fusion module to obtain a 32+32-channel 8×8 first-level second fusion feature. Secondary upsampling: The 128-channel 8×2 third-level features are converted from 128 channels to 64 channels through transposed convolution, and then spliced with the second-level first fusion features through jump connection. The spliced features are passed through a 3×3 convolution block and a multi-scale dynamic feature fusion module to obtain the 64+64-channel 8×4 second-level second fusion features; Final upsampling: The 64+64-channel 8×4 second-level second fusion features are converted from 64 channels to 32 channels through transposed convolution, and then spliced with the first-level first fusion features and the first-level second fusion features through skip connections. The spliced features are passed through a 3×3 convolution block and a multi-scale dynamic feature fusion module to obtain a 32+32+32-channel 8×8 first-level third fusion feature. Output layer: The third fusion feature of the first level is reduced to 32 channels through a 3×3 convolution block, and finally outputs the final fusion feature of a single channel 8×5 target size through a 1×4 convolution block.

8. The formation parameter inversion method based on the asymmetric downsampling inversion neural network according to claim 1, characterized in that: The multi-scale dynamic feature fusion module includes a parallel local detail branch for extracting high-frequency features, a regional context branch for extracting low-frequency features, and a vertical correlation branch for extracting the internal correlation of the logging electromagnetic response corresponding to each two-dimensional formation sample. Finally, the three branch features are fused through dynamic weighting. Get fusion feature output : ； in is the weight parameter of the lth branch, It is a 1×1 convolution.

9. The formation parameter inversion method based on the asymmetric downsampling inversion neural network according to claim 1, characterized in that: Feature extraction of local detail branch First branch feature Specifically, a standard convolution without expansion of the convolution kernel 3×3 is used, the expansion coefficient dilation=1, and the padding amount padding=1; Feature extraction of regional context branch Second branch features Specifically, the convolution kernel 3×3 is used, and the dilation coefficient is dilation=2; Feature extraction of vertical correlation branch Third branch feature Specifically, 1×3 strip convolution is used.

Images