Electrical imaging logging completion method and device and computer equipment

Abstract

The application provides an electrical imaging well logging completion method and device and computer equipment, and belongs to the technical field of well logging. The method comprises creating a mask of a current well logging image according to a known region and a missing region in the current well logging image; performing pixel inference and iterative completion of the missing region on the well logging image and the mask based on an image completion model to generate a well logging image after completion of the missing region, wherein the image completion model is obtained by performing training processing on a constructed neural network for performing pixel feature extraction and pixel dependency capture based on the well logging image and the mask. The method performs iterative completion of the missing region in the well logging image based on the image completion model, performs mask updating once in each iteration, gradually optimizes the image repair result through multiple rounds of operation, so that the image completion model is suitable for completion of the well logging image with multiple missing scales, and the well logging image after completion has high accuracy and no artifacts.

Description

Technical Field

[0001] This application belongs to the field of well logging image processing technology, specifically relating to an electrical imaging well logging completion method, an electrical imaging well logging completion device, a machine-readable storage medium, and a computer device. Background Technology

[0002] Electrical imaging logging technology, first proposed by Schlumberger in the 1980s, is currently an important technique for geological research and reservoir assessment. Due to instrument limitations, the measured images often contain regular gaps in the vertical direction, resulting in incomplete logging images. The coverage of logging images varies depending on the instrument and the diameter of the well. For example, Formation MicroScanner Image (FMI) images have approximately 80% coverage, Earth Imager images approximately 65%, OMRI images approximately 57%, and OBMI images only about 32%. Currently, many methods exist to complete the missing bands in logging images, but these methods are generally ineffective when dealing with large areas of missing high-dipping fractures. Given the unique and complex nature of logging images, effectively combining relevant geological knowledge to complete logging images is a crucial problem that urgently needs to be overcome. Summary of the Invention

[0003] The purpose of this application is to provide an electrical imaging logging completion method, an electrical imaging logging completion device, a machine-readable storage medium, and a computer device to overcome the technical problem in the prior art that the completion effect of the method for completing blank strips in logging images is poor when dealing with large-area, high-angle fracture missing areas.

[0004] To achieve the above objectives, the first aspect of this application provides an electrical imaging logging completion method, comprising: Create a mask for the current well logging image based on the known and missing regions in the current well logging image; The image completion model performs pixel inference and iterative completion of missing regions on the logging image and the mask to generate a logging image with missing regions completed. The image completion model is obtained by training a neural network based on the logging image and the mask to extract pixel features and capture pixel dependencies. During iterative completion, the logging image and its mask are updated based on the new known regions obtained after pixel filling by the image completion model in the current iteration step, so as to indicate the effective token region for feature aggregation in the attention layer of the neural network in the next iteration step.

[0005] In a specific embodiment of this application, the training process includes: pre-training the neural network using complete images; and training the pre-trained neural network using historical well logging images.

[0006] In a specific embodiment of this application, the complete image includes both natural images and seismic images.

[0007] In a specific embodiment of this application, pre-training the neural network using a complete image includes: Based on the neural network, inference is performed on the continuous structural features in natural images until the inference result meets the preset requirements, thus obtaining a staged neural network. Based on a staged neural network, inference is performed on the continuous structural and texture features in seismic images until the inference results meet the preset requirements, thus obtaining a pre-trained neural network.

[0008] In a specific embodiment of this application, the neural network includes a first sub-neural network for performing downsampling, a second sub-neural network connected to the first sub-neural network for filling missing regions, a third sub-neural network connected to the second sub-neural network for performing upsampling, and a fourth sub-neural network connected to the third sub-neural network for at least completing local features. The second sub-neural network includes the attention layer and is connected to a mask update module. The fourth sub-neural network takes the well logging image, the mask, and the image obtained after upsampling by the third sub-neural network as input to the first sub-neural network.

[0009] In a specific embodiment of this application, the second sub-neural network includes multiple Transformer layers connected in series. Each Transformer layer includes multiple Transformer blocks, and each Transformer block includes an input module, an encoder module, a decoder module, and an output module, wherein: The input module includes an embedding layer but does not include position encoding; The encoder module is composed of multiple encoder layers stacked together. Each encoder layer consists of two sub-layer connection structures. The first sub-layer connection structure includes a multi-head self-attention layer and a feature stitching layer connected to the multi-head self-attention layer. The second sub-layer connection structure includes a feedforward fully connected sub-layer and a feature stitching layer connected to the feedforward fully connected sub-layer. The decoder module is composed of multiple decoder layers stacked together. Each decoder layer consists of three sub-layer connection structures. The first sub-layer connection structure includes a masked multi-head self-attention layer and a feature splicing layer connected to the masked multi-head self-attention layer. The second sub-layer connection structure includes a multi-head attention layer and a feature splicing layer connected to the multi-head attention layer. The third sub-layer connection structure includes a feedforward fully connected sub-layer and a feature splicing layer connected to the feedforward fully connected sub-layer.

[0010] In a specific embodiment of this application, the fourth sub-neural network is a U-Net network.

[0011] In a specific embodiment of this application, during iterative completion, the logging image and its mask are updated based on the new known regions obtained after pixel filling by the image completion model in the current iteration step, to indicate the effective token regions for feature aggregation in the attention layer of the neural network in the next iteration step, including: In the current iteration step, the first sub-neural network is used to downsample the logging image and its mask, the second sub-neural network is used to fill the missing region pixels of the token converted from the feature map obtained after downsampling, the third sub-neural network is used to upsample the output image obtained after filling the missing region pixels of the second sub-neural network, and the fourth sub-neural network is used to at least perform local feature completion on the feature map obtained after upsampling. The current logging image and its mask are updated based on the known regions in the feature map obtained after upsampling. The updated mask is used to indicate the effective token region for feature aggregation in the attention layer of the second sub-neural network in the next iteration step. Then, the next iteration step is entered until all the missing regions in the logging image are filled with pixels and the iteration stops.

[0012] In a specific embodiment of this application, creating a mask for the current well logging image based on known and missing regions in the current well logging image includes: The known and missing regions in the current well logging image are each represented by a single pixel value to obtain a binary mask with the same size as the well logging image.

[0013] In a specific embodiment of this application, the known and missing regions in the current well logging image are each represented by a pixel value to obtain a binary mask with the same size as the well logging image, including: The known regions in the well logging images segmented using a preset segmentation threshold are subjected to hole filling processing, which includes at least expansion or corrosion processing. The well logging image after binarization and hole filling processing is represented by a single pixel value for both the known and missing regions in the current well logging image, thus obtaining the mask for the current well logging image.

[0014] In a specific embodiment of this application, the mask update strategy is as follows: Mark all tokens within the known region of the output image obtained after upsampling by the third sub-neural network in the current iteration step as valid tokens; The missing region in the logging image of the first sub-neural network in the previous iteration step is used as the window sliding region. The preset window is slid in the output image obtained after upsampling by the third sub-neural network. If there is at least one valid token in the current window, all tokens in the current window are updated to valid tokens. Otherwise, all tokens in the current window are still invalid tokens after updating. After the window sliding is completed, the valid tokens and invalid tokens are represented by a pixel value to obtain the mask for the next iteration step.

[0015] In a specific embodiment of this application, the logging image is a planar unfolded view of an electrical imaging obtained by a rotating electrical imaging logging tool inside the wellbore. The left and right boundaries of the planar unfolded view are electrical scanning data at the same wellbore location. During pixel inference, the continuity of the texture and / or continuity structure of the left and right boundaries of the logging image after the missing region is filled is constrained.

[0016] In a specific embodiment of this application, the continuity of the texture and / or continuity structure of the left and right boundaries of the well logging image after constraining the missing region completion includes: After the last iteration step, the left and right boundaries of the output image obtained by filling the missing region pixels of the second sub-neural network are masked to obtain a new image after the left and right boundaries are occluded. The new image is segmented from the middle position, and the segmented right region is moved to the left of the segmented left region or the segmented left region is moved to the right of the segmented right region to obtain a reconstructed image after the left and right boundaries are joined. The reconstructed image is input into the third sub-neural network for upsampling, and the feature map obtained after upsampling is supplemented with at least local features using the fourth sub-neural network.

[0017] In a specific embodiment of this application, there are skip connections between the first sub-neural network and the sub-layers with the same image resolution in the third sub-neural network.

[0018] In a specific embodiment of this application, the neural network further includes a 1x1 convolution before the first sub-neural network and a 1x1 convolution between the third sub-neural network and the fourth sub-neural network.

[0019] A second aspect of this application provides an electrical imaging logging completion device, comprising: The mask creation module is used to create a mask for the current well logging image based on the known and missing regions in the current well logging image; The training module is used to train the constructed neural network that performs pixel feature extraction and pixel dependency capture based on well logging images and masks. The completion module is used to perform pixel inference and iterative completion of missing regions on the well logging image and the mask based on the image completion model, and generate a well logging image with missing regions completed. During iterative completion, the well logging image and its mask are updated according to the new known regions obtained after pixel filling by the image completion model in the current iteration step, so as to indicate the effective token region for feature aggregation in the attention layer of the neural network in the next iteration step.

[0020] In a specific embodiment of this application, the neural network includes a first sub-neural network for performing downsampling, a second sub-neural network connected to the first sub-neural network for filling missing regions, a third sub-neural network connected to the second sub-neural network for performing upsampling, and a fourth sub-neural network connected to the third sub-neural network for at least completing local features. The second sub-neural network includes the attention layer and is connected to a mask update module. The fourth sub-neural network takes the well logging image, the mask, and the image obtained after upsampling by the third sub-neural network as input to the first sub-neural network.

[0021] In a specific embodiment of this application, during iterative completion, the logging image and its mask are updated based on the new known regions obtained after pixel filling by the image completion model in the current iteration step, to indicate the effective token regions for feature aggregation in the attention layer of the neural network in the next iteration step, including: In the current iteration step, the first sub-neural network is used to downsample the logging image and its mask, the second sub-neural network is used to fill the missing region pixels of the token converted from the feature map obtained after downsampling, the third sub-neural network is used to upsample the output image obtained after filling the missing region pixels of the second sub-neural network, and the fourth sub-neural network is used to at least perform local feature completion on the feature map obtained after upsampling. The current logging image and its mask are updated based on the known regions in the feature map obtained after upsampling. The updated mask is used to indicate the effective token region for feature aggregation in the attention layer of the second sub-neural network in the next iteration step. Then, the next iteration step is entered until all the missing regions in the logging image are filled with pixels and the iteration stops.

[0022] A third aspect of this application provides a computer device, comprising: The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the electrical imaging logging completion method according to the first aspect of this application.

[0023] A fourth aspect of this application provides a machine-readable storage medium storing instructions for causing a machine to perform the electrical imaging logging completion method according to a first aspect of this application.

[0024] The above technical solution uses an image completion model to iteratively complete missing regions in well logging images. The multi-iteration completion strategy is a higher-level strategy, in which a mask update is performed in each iteration. Through multiple rounds of operation, the image restoration result is gradually optimized, making the image completion model suitable for completing well logging images with missing dimensions at multiple scales. The completed well logging image has high accuracy and no artifacts, and it can achieve good completion results, especially for well logging images with missing dimensions at large scales. Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0025] The accompanying drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the following detailed description to explain the embodiments of this application, but do not constitute a limitation on the embodiments of this application. In the drawings: Figure 1 A flowchart illustrating an electrical imaging logging completion method according to an embodiment of this application is shown schematically. Figure 2 This schematically illustrates a technical roadmap according to an embodiment of the present application; Figure 3a This illustration schematically shows a mask update diagram according to an embodiment of the present application; Figure 3b This illustration schematically shows another mask update diagram according to an embodiment of this application; Figure 4 A network architecture diagram according to an embodiment of this application is illustrated schematically; Figure 5 The illustration schematically shows a first type of original FMI logging image and its completed image according to an embodiment of this application, wherein, Figure 5 Part (a) is the first type of raw FMI logging image. Figure 5 Part (b) is the completed image; Figure 6 This illustration schematically shows a second type of original FMI logging image and its completed image according to an embodiment of this application, wherein... Figure 6 Part (a) shows the second type of original FMI logging image. Figure 6 Part (b) is the completed image; Figure 7 The illustration schematically shows a first type of original Earth Imager logging image and its completed image according to an embodiment of this application, wherein, Figure 7 Part (a) is the first type of raw Earth Imager logging image. Figure 7 Part (b) is the completed image; Figure 8 The illustration schematically shows a second type of original Earth Imager logging image and its completed image according to an embodiment of this application, wherein, Figure 8 Part (a) shows the second type of original Earth Imager logging image. Figure 8 Part (b) is the completed image; Figure 9 The illustration schematically shows a first type of original OBMI logging image and its completed image according to an embodiment of this application, wherein, Figure 9 Part (a) shows the first type of raw OBMI logging image. Figure 9 Part (b) is the completed image; Figure 10 The illustration schematically shows a second type of original OBMI logging image and its completed image according to an embodiment of this application, wherein, Figure 10 Part (a) shows the second type of original OBMI logging image. Figure 10 Part (b) is the completed image; Figure 11 A schematic block diagram illustrating the components of an electrical imaging logging completion device according to an embodiment of this application is shown. Figure 12 A schematic block diagram of a computer device according to an embodiment of this application is shown. Detailed Implementation

[0026] The specific implementation methods of the embodiments of this application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific implementation methods described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application.

[0027] If the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0028] In actual production, due to the special design of imaging logging tools and the different wellbore diameters, the logging images obtained after scanning along the wellbore wall contain varying degrees of data loss, with OBMI images showing a pixel loss rate as high as 70%. Preliminary research revealed that existing methods for completing blank strips in logging images are not ideal in multi-scale missing logging images. For example, artifacts exist after completion, and the completion effect on geological texture features and continuous structural features such as fractures, bedding, and unconformities is poor. One existing technology uses only the U-Net network for blank strip completion in logging images. This is because the U-Net network fully utilizes the local repair capabilities of CNNs; the local connections of neurons ensure that the U-Net network can focus on a specific local area, thus better capturing local features. However, its completion effect on continuous structural features needs improvement.

[0029] To overcome the above-mentioned shortcomings, this application provides a neural network-based electrical imaging logging completion method. This method completes logging images based on an iterative completion strategy that can handle multi-scale missing data, and specifically includes the following: Create a mask for the current well logging image based on the known and missing regions in the current well logging image; The image completion model performs pixel inference and iterative completion of missing regions on the current well logging image and its mask, generating a well logging image with the missing regions completed. The image completion model is obtained by training a constructed neural network based on the well logging image and mask to extract pixel features and capture pixel dependencies. During iterative completion, the logging image and its mask are updated based on the new known regions obtained after pixel filling by the image completion model in the current iteration step, so as to indicate the effective token region for feature aggregation in the attention layer of the neural network in the next iteration step.

[0030] It's important to understand that a mask is a binary image of the same size as the logging image, used to indicate the area on the logging image to be manipulated. Binarization means representing the known and missing areas in the logging image with a single pixel value, which can be 0 and 1 respectively. The known area is relative to the missing area. For example, the area corresponding to the micro-scan data from an electrical imaging logging tool is the known area.

[0031] As described in the above embodiment, a multi-iterative completion strategy (also known as multi-incremental completion) is used, in which the logging image and mask are updated in each iteration. Mask update is a dynamic propagation mechanism applied to the attention layer. The mask is used to indicate whether the token is valid and is automatically updated during forward propagation. With the continuous updating of the mask, it ensures that the missing area can participate in subsequent calculations. The multi-iterative completion strategy is a higher-level strategy. It performs a mask update in each iteration. Through multiple rounds of operations, it gradually optimizes the image restoration result, directly affecting the coherence and detail quality of the final restoration. This compensates for the inconsistency that mask update may cause when dealing with large-area missing areas or complex scenes. It can be seen that mask update itself is not a means to directly improve restoration quality. Multi-iterative completion is the strategy that ultimately and directly improves restoration quality. In summary, the above embodiment, through the multi-iterative completion strategy, enables the image restoration model to have high-quality restoration capabilities for logging images missing at multiple scales, thereby improving the accuracy of electrical imaging logging restoration and making it suitable for logging images missing at various scales.

[0032] Figure 1 A flowchart illustrating an electrical imaging logging completion method according to an embodiment of this application is shown schematically. Figure 1 As shown, the electrical imaging logging completion method may include steps 102 to 104.

[0033] Step 102: Create a mask for the current well logging image based on the known and missing regions in the current well logging image.

[0034] A mask is a type of binary image. As an example, a well logging image mask can be obtained by representing the known and missing regions of the current well logging image with a single pixel value, resulting in a binary mask of the same size as the well logging image. The pixel values ​​representing the known regions and the missing regions can be 0 and 1, respectively. In a binary image, 0 represents black and 1 represents white.

[0035] To clearly identify known and missing regions in a well logging image, threshold segmentation can be used. In practical applications, the segmented well logging image may still contain holes that need to be filled. These holes can be filled using dilation or erosion methods. For example, as another example, the mask for the well logging image can be obtained as follows: Segment the known and missing regions in the current well logging image using a preset segmentation threshold; perform hole-filling processing on the segmented known regions in the well logging image, including at least dilation or erosion processing; binarize the well logging image after hole filling to obtain the mask for the current well logging image.

[0036] Step 104: Based on the image completion model, perform pixel inference and iterative completion of the missing regions of the current logging image and its mask to generate a logging image with the missing regions completed.

[0037] Specifically, in this application, the image completion model is obtained by training a neural network based on well logging images and masks, performing pixel feature extraction and pixel dependency capture. During iterative completion, the well logging image and its mask are updated according to the new known regions obtained after pixel filling by the neural network in the current iteration step, indicating the effective token regions for feature aggregation in the attention layer of the neural network in the next iteration step. As mentioned above, this application demonstrates that: missing regions of the well logging image are repaired through a multi-iteration completion strategy, and the mask is dynamically updated according to the new known regions obtained in the current iteration step at each iteration.

[0038] As an optional embodiment of this application, the training process includes: pre-training the neural network using complete images; and training the pre-trained neural network using historical logging images.

[0039] In this application, a complete image indicates that the image has no missing regions.

[0040] It is important to understand that this application obtains the image completion model based on a combination of pre-training and retraining. Because, to adapt to the well logging image completion task, a new neural network is constructed to perform pixel feature extraction and pixel dependency capture, clarifying the difference between this modeling approach and that combined with transfer learning and model fine-tuning. The neural network performing pixel feature extraction and pixel dependency capture can be a neural network containing multiple feature extraction layers and multiple attention layers.

[0041] As an optional embodiment of this application, the complete image includes natural images and seismic images. Seismic images refer to seismic images obtained through seismic exploration in oil and gas exploration and development. Natural images are those that are not natural images. In this application, natural images refer to images captured by a camera based on its photosensitive unit, while non-natural images refer to images converted from other means, such as radiographic imaging, functional imaging, magnetic resonance imaging, and ultrasound imaging.

[0042] As a preferred option, combined Figure 2 As shown, the neural network is pre-trained based on the complete image in the following manner: Based on neural networks, inference is performed on the continuous structural features in natural images until the inference results meet the preset requirements, thus obtaining a staged neural network. Based on a staged neural network, inference is performed on the continuous structural and texture features in seismic images until the inference results meet the preset requirements, thus obtaining a pre-trained neural network.

[0043] As is known, in order to ensure that the inference results meet the preset requirements, the training effect is usually constrained by defining a loss function.

[0044] To enable the neural network to infer missing pixels, and considering that the image features in well logging images mainly include texture features and continuous structural features (such as fractures, bedding, and unconformities), the training objective is to enable the neural network to infer both continuous structural and texture features. Since natural images are readily available and cover various scenes, the diverse scenes help the network learn diverse features, which is beneficial for many downstream tasks. Therefore, in the preferred embodiment described above, the neural network is first trained using natural images to improve its ability to infer continuous structural features, for example, with the goal of recovering some bedding features. However, because natural images typically do not contain geological structures commonly found in well logging images, the neural network trained with natural images lacks the ability to infer geological structures such as fractures and bedding, leading to the introduction of numerous unrealistic color spots. To eliminate these unrealistic color spots, the neural network trained with natural images is then pre-trained using seismic images. Seismic images are readily available, allowing the neural network to learn relevant geological knowledge, thus enabling the trained neural network to infer both continuous structural and texture features.

[0045] In summary, by pre-training with natural images first and then with seismic images, the pre-trained neural network acquires better reasoning capabilities for continuous structural and textural features, thereby improving the accuracy of filling in missing well logging areas.

[0046] The neural network constructed in the above embodiment performs pixel feature extraction and pixel dependency capture. In an improved embodiment of this application, a neural network framework is proposed. This framework includes a first sub-neural network for downsampling, a second sub-neural network connected to the first sub-neural network for filling missing regions, a third sub-neural network connected to the second sub-neural network for upsampling, and a fourth sub-neural network connected to the third sub-neural network for at least completing local features. The second sub-neural network includes an attention layer and is connected to a mask update module. The fourth sub-neural network takes the well logging image and mask input to the first sub-neural network, and the image obtained after upsampling by the third sub-neural network, as input. That is, from shallow to deep, the network framework includes the first, second, third, and fourth sub-neural networks connected sequentially. The first sub-neural network is the input end of the network architecture, and the data input to the network architecture includes the well logging image and its mask. The data input to the fourth sub-neural network includes the well logging image and its mask input to the first sub-neural network, and the image obtained after upsampling by the third sub-neural network.

[0047] Based on the above network architecture, the mask update strategy can be expressed as: Mark all tokens within the known region of the output image obtained after upsampling by the third sub-neural network in the current iteration step as valid tokens; The missing region in the logging image of the first sub-neural network in the previous iteration step is used as the window sliding region. The preset window is slid in the output image obtained after upsampling by the third sub-neural network. The sliding region is the region corresponding to the aforementioned window sliding region. If there is at least one valid token in the current window, all tokens in the current window are updated to valid tokens. Otherwise, all tokens in the current window are still invalid tokens after updating. After the window sliding is completed, the valid tokens and invalid tokens are represented by a pixel value to obtain the mask for the next iteration step.

[0048] It should be understood that the area during sliding is the area corresponding to the aforementioned window sliding area. Since the logging image input to the first sub-neural network and the image output by the third sub-neural network have the same size, the correspondence means, for example, that after taking a corner point of the image as the origin, they are within the same coordinate range within the image.

[0049] For example, such as Figure 3a and Figure 3bAs shown, the mask is used to indicate whether a token is valid and is automatically updated during forward propagation. The mask update strategy is similar to partial convolution; as the mask is continuously updated, invalid regions (missing regions in the logging image) are gradually marked as valid, and eventually all tokens will be marked as valid tokens. Figure 3a and Figure 3b In the code, U represents mask update. If there is a valid token of 1 in the current 2×2 window (dashed box), all tokens in the window will be updated to be valid tokens. If all tokens in the window are invalid tokens of 0, all tokens in the window will remain invalid tokens after the update.

[0050] As an example, the first sub-neural network is a multi-layer convolutional downsampling module, which is the first part of the network architecture. It is used for token extraction and can reduce the computational complexity and memory overhead of subsequent networks. The second sub-neural network is a Transformer layer, which realizes the interaction of remote information of the image at different resolutions, and is the second part of the network architecture. The third sub-neural network is a multi-layer convolutional upsampling module, which is used to restore the tokens processed by the Transformer layer to the original image size (the size of the input well logging image), and is the third part of the network architecture. The fourth sub-neural network is a high-frequency detail (local feature) completion module, which is used to refine the image, that is, to complete the local features, and is the fourth part of the network architecture.

[0051] Based on the above network architecture, during iterative completion, the logging image and its mask are updated according to the new known regions obtained after pixel filling by the image completion model in the current iteration step. This indicates the effective token regions for feature aggregation in the attention layer of the neural network in the next iteration step, including: In the current iteration step, the first sub-neural network is used to downsample the logging image and its mask, the second sub-neural network is used to fill the missing pixels of the token converted from the feature map obtained after downsampling, the third sub-neural network is used to upsample the output image obtained after filling the missing pixels of the second sub-neural network, and the fourth sub-neural network is used to at least complete the local features of the feature map obtained after upsampling. The current logging image and its mask are updated based on the known regions in the feature map obtained after upsampling. The updated mask is used to indicate the effective token region for feature aggregation in the attention layer of the second sub-neural network in the next iteration step. Then, the iteration continues until all missing regions in the logging image are filled with pixels.

[0052] To overcome the discontinuity in some landmark bedding areas remaining in well logging images after a single completion, the above embodiments employ a multi-iterative completion strategy to gradually repair large-area missing well logging images, inferring pixel values ​​in the missing areas using known image information. Each iteration further refines and corrects the missing areas of the image based on the previous completion, reducing the difference between the completed area and the surrounding original area, thus improving the overall continuity of the completed well logging image. Each generation generates a complete image, selecting only the missing pixels around the known areas as the prediction result for this iteration. After multiple iterations, a complete well logging image after multiple completions can be obtained.

[0053] For example, in an improved embodiment of this application, the Transformer Block within the second sub-neural network is improved as follows to better adapt to the task of completing large-scale missing well logging images: 1) the LN layer and positional embedding in the original Transformer structure are removed; 2) the residual structure is replaced with concatenation. The original Transformer achieves good completion results when completing small-scale missing well logging images, but when the well logging image is missing at a large scale, the training process of the original Transformer becomes unstable and may experience gradient explosion. This is because when the missing region is large, the number of invalid tokens increases, and the LN layer in the original Transformer greatly amplifies the invalid tokens, leading to training instability. The removal of positional embedding stems from the fact that early convolutional operations utilize local perception, considering the relative positional relationship between pixels and their neighbors during computation, thus implicitly encoding positional information. For example, stacking multiple 3×3 convolutions expands the receptive field, gradually capturing larger contextual and spatial information, making explicit positional embedding unnecessary. In image processing, using convolution to provide positional information captures spatial information. Furthermore, the residual structure in the original Transformer makes the network more inclined to learn high-frequency information, but the large number of invalid tokens means that the features extracted by the network cannot provide sufficient low-frequency information, making it difficult for subsequent network layers to learn high-frequency information, which also makes network training difficult. Therefore, concatenation is used to replace the residual structure. Combining these two improvements enhances training efficiency and the accuracy of Transformer blocks in completing multi-scale well logging images.

[0054] As an example, the second sub-neural network includes at least one Transformer layer, and each Transformer layer includes at least one Transformer block in which the above-mentioned improvement is applied, namely: the improved Transformer block includes an input module, an encoder module, a decoder module, and an output module, wherein: The input module includes an embedding layer but does not include position encoding; The encoder module is composed of multiple encoder layers stacked together. Each encoder layer consists of two sub-layer connection structures. The first sub-layer connection structure includes a multi-head self-attention layer and a feature stitching layer connected to the multi-head self-attention layer. The second sub-layer connection structure includes a feedforward fully connected sub-layer and a feature stitching layer connected to the feedforward fully connected sub-layer. The decoder module consists of multiple stacked decoder layers. Each decoder layer consists of three sub-layer connection structures. The first sub-layer connection structure includes a masked multi-head self-attention layer and a feature splicing layer connected to the masked multi-head self-attention layer. The second sub-layer connection structure includes a multi-head attention layer and a feature splicing layer connected to the multi-head attention layer. The third sub-layer connection structure includes a feedforward fully connected sub-layer and a feature splicing layer connected to the feedforward fully connected sub-layer.

[0055] In the example above, the feature splicing layer is a serial structure that replaces the residual structure in the original Transformer.

[0056] As another improved embodiment of this application, since the logging image is a planar unfolded view of the electrical imaging obtained by rotating the electrical imaging logging tool inside the wellbore, and the left and right boundaries of the planar unfolded view are the electrical scanning data of the same wellbore position, this improved embodiment introduces the "cylinder" constraint of the logging image, that is: when the image completion model is inferring pixels, it constrains the continuity of the texture and / or continuity structure of the left and right boundaries of the logging image after the missing area is completed.

[0057] Well logging images are acquired using an electrical imaging logging tool on an existing well. A well is a cylindrical wellbore. By rotating and descending along the wellbore wall, the electrical imaging logging tool collects formation information from the wellbore surface. In actual production, to visually represent the logging data, the final data is usually unfolded into a planar image. The unfolded image is essentially a planar unfolding of a cylindrical surface image. In a complete cylindrical surface image, the 0-degree and 360-degree positions are identical. Therefore, after unfolding the cylindrical surface image into a planar image, the left and right boundaries of the planar image correspond to the 0-degree and 360-degree positions of the cylindrical surface image, respectively, representing a continuous representation of the same location. The well logging image completed using an image completion model reflects a continuous 360-degree image of the wellbore. Ideally, the completed well logging image should have seamless textural transitions at its left and right boundaries, meaning it can be stitched together to form a complete cylindrical surface image. Therefore, by introducing cylindrical constraints, the continuity of the texture and / or continuity structure of the left and right boundaries in the completed well logging image is improved, which is more in line with actual geological understanding, that is, a more natural and continuous well logging image is obtained.

[0058] In an optional embodiment of this application, the continuity of texture and / or continuity structure of the left and right boundaries in the well logging image after the missing region is filled is achieved in the following way: After the last iteration step, the left and right boundaries of the output image obtained by filling the missing region pixels of the second sub-neural network are masked to obtain a new image after the left and right boundaries are occluded. The new image is segmented from the middle position, and the segmented right region is moved to the left of the segmented left region or the segmented left region is moved to the right of the segmented right region to obtain a reconstructed image after the left and right boundaries are joined. The reconstructed image is input into the third sub-neural network for upsampling, and the feature map obtained after upsampling is supplemented with at least local features using the fourth sub-neural network.

[0059] In the above optional embodiments, by joining the originally separate left and right boundaries, it helps the subsequent fourth sub-neural network to utilize the contextual information of the left and right sides during the repair process, thereby making the consistency of the texture and / or continuity structure better after the occluded parts of the left and right boundaries are filled.

[0060] As another improved embodiment of this application, there are skip connections between sub-layers with the same image resolution in the first sub-neural network and the third sub-neural network. Through skip connections, the feature information obtained by downsampling the sub-layers of the first sub-neural network is passed to the corresponding sub-layers of the third sub-neural network, thereby making the network easier to optimize. The term "corresponding" refers to having the same image resolution.

[0061] Based on the above embodiments, in a specific application of this application, the network architecture of the neural network is as follows: Figure 4 As shown, the network architecture includes CNN convolutional downsampling, Transformer Layer, CNN convolutional upsampling, and U-Net network, with network depth increasing from shallow to deep. The first sub-neural network is a multi-layer convolutional downsampling module, which is the first part of the network architecture and is used for token extraction. Figure 4 The downsampling module shown in the image first performs a 1x1 convolution on the incomplete well logging image and its mask input. Figure 4 As shown by the red arrow, the number of image channels is increased to 180 through a single convolutional layer. Then, three convolutions with a stride of 2 are used to downsample the feature map from (w, h) to (w / 8, h / 8). The downsampling is as follows: Figure 4 As shown by the gray arrows, the feature map changes during the downsampling process are as follows: [180, 512, 512], [180, 256, 256], [180, 128, 128], [180, 64, 64]. By using CNN convolutions to combine local inductive priors, the neural network can focus on extracting features from local regions of the logging image in the early stages, while using surrounding information to fill holes, laying the foundation for more complex visual processing tasks later. The second sub-neural network is a Transformer layer, consisting of five sequentially connected Transformer layers. The Transformer layers realize the interaction of remote information of the image at different resolutions, while the mask is updated synchronously (e.g., ...). Figure 4 The Mask update module (shown) guides attention calculation, and the Transformer layer is the second part of the network architecture. The feature map output from the first part of the network architecture is converted into tokens and then fed into the Transformer layer. The number of tokens in the five Transformer layers are w / 8×h / 8, w / 16×h / 16, w / 32×h / 32, w / 16×h / 16, and w / 8×h / 8, respectively. The increase or decrease in the number of tokens is achieved through convolution. Figure 4The T shown represents the Transformer layer. The numbers 4096, 1024, 256, 1024, and 4096 below T represent the number of tokens in each Transformer layer. Furthermore, the number of Transformer blocks in the five Transformer layers are 2, 3, 4, 3, and 2, respectively. The Transformer blocks are improvements to the original Transformer structure, including the removal of LN and positional encoding, and the use of concatenation to replace the residual structure, as mentioned earlier. The third sub-neural network is a multi-layer convolutional upsampling module, used to restore the tokens processed by the Transformer layer to the original image size (the size of the input well logging image). It is the third part of the network architecture. It first restores the feature map to the original image size through three upsampling operations. The upsampling process is as follows... Figure 4 As shown by the blue arrow, the feature map changes sequentially during the upsampling process: [180,64,64], [180,128,128], [180,256,256], [180,512,512]. This is followed by a convolutional layer, as shown... Figure 4As shown by the red arrow, the image dimensions are restored to the original image dimensions. A skip connection is used between the upsampling layer in the third part and the corresponding downsampling layer in the first part. The fourth sub-neural network is a high-frequency detail (local feature) completion module. In this application, the U-Net network is used for refined image processing, i.e., completing local features. This is the fourth part of the network architecture. After the first three parts, we have achieved the overall structure repair of the image. The fourth part completes the high-frequency details of the image based on the repair of the first three stages. The U-Net network consists of an encoder part and a decoder part. The U-Net network fully utilizes the local repair capability of CNNs. The local connections of neurons ensure that the network can focus on a certain local region and better capture local features. Compared with Transformer, convolution has higher computational efficiency. The computational complexity of CNN is linearly related to the input size, while the self-attention mechanism of Transformer makes its computational complexity proportional to the square of the input size. The fourth part takes the mask corresponding to the missing region, the original incomplete well logging image, and the image completed by the third part as input. The encoder first increases the image dimension to 64 using a single convolutional layer, then downsamples the image size from (h, w) to (32 / h, 32 / w) using five convolutional layers. The number of feature dimensions doubles after each convolution, ultimately increasing from 64 to 512 dimensions. The decoder design is symmetrical to the encoder. It first upsamples the feature map size from (32 / h, 32 / w) to (h, w) using five convolutional layers, halving the number of feature dimensions after each convolution, ultimately reducing from 512 to 64 dimensions. Finally, a single convolutional layer restores the feature dimension from 64 back to the original image dimension. The U-Net network uses skip connections between the encoder and decoder, allowing information from the downsampled layers to be passed to the upsampled layers, making the network easier to optimize.

[0062] Furthermore, the training process for the neural network with the above architecture mainly includes steps A1 and A2.

[0063] Step A1, pre-training based on natural and seismic images, specifically refers to: Because the total amount of well logging data is relatively small, we used a large number of complete natural image datasets, Places365-Standard, along with seismic images for two rounds of pre-training. The Places365-Standard dataset is widely used in computer vision; it contains approximately 1.8 million images covering 365 scene categories. Using it for early model pre-training allows the model to learn image features from various scenes. The diverse scenes help the model learn diverse features, which is beneficial for many downstream tasks. The seismic data comes from our collected 3D seismic data. By systematically deploying a large number of seismic detectors on the surface or seabed and generating seismic waves, these waves penetrate underground strata and are reflected in different ways by different strata before being recorded by the detectors. After noise reduction, velocity correction, and other processing, 3D seismic volume data is finally formed. By analyzing this data, we can obtain detailed 3D views of underground strata; in addition, this data can also be used to study geological issues such as stratigraphic structure and faults. The seismic images used during pre-training are cut from 3D seismic data volumes. Pre-training is performed on this seismic dataset. Since seismic images themselves contain a wealth of geological information, they help the model learn geological information related to stratigraphic structure. After the first round of pre-training using only the Places365-Standard dataset, a staged neural network (such as...) is obtained. Figure 2 (As shown) This method is used for completing well logging images. While it can recover some bedding features of the image as a whole, it also introduces a large number of unrealistic color spots. A second round of pre-training using complete seismic data is performed on a staged neural network to obtain a pre-trained neural network. Compared to the completion effect of the staged neural network, the pre-trained neural network can eliminate unrealistic color spots.

[0064] The loss function used in the above pre-training process is Perceptual Loss, Adversarial Loss, and a regularization term for the discriminator. Given the original image... I in Complete the image as follows I out The corresponding mask is M. The perceptual loss includes content loss, ensuring that the image generated by the neural network retains the original content. A pre-trained VGG-19 is used to project the image into a high-level feature space for loss calculation. Content loss. The calculation formula is as follows: (Formula 1); In Formula 1 above, content loss The computation uses the 15th layer of VGG-19 (the last convolutional layer of the fourth convolutional block) for feature representation. C, H, W These represent the number of channels, height, and width of the image, respectively. This represents a function that projects an image into a high-dimensional feature space using the 15th layer of a pre-trained VGG-19 network.

[0065] The formula for calculating adversarial losses is as follows: (Formula 2); In formula two above, L G This represents the adversarial loss of the generator. G(z) It is the output of the generator. D(x) It is the discriminator for the input image x For true estimates, D ( G ( z )) is the discriminator's estimate that the generated data is true, and z represents the random noise input to the generator.

[0066] The formula for calculating the regularization term of the discriminator is as follows: (Formula 3); In formula three above, R 1 represents the regularization term of the discriminator. p data ( x ) represents the distribution of the real data, ||∇ D (x)|| represents the discriminator output relative to its input. x The gradient norm.

[0067] Based on this, the total loss function during the pre-training phase is calculated using the following formula: (Formula 4); In Formula 4 above, λ represents the regularization weight, η represents the content loss weight, λ can take the value 10, and η can take the value 0.1.

[0068] Step A2 involves fine-tuning the hyperparameters within the neural network using incomplete logging images.

[0069] We collected publicly available seafloor formation logging data from the IODP (International Ocean Discovery Program), including FMI, OMRI, OBMI, and Earth Imager logging data. The FMI image coverage is approximately 80%, Earth Imager approximately 65%, OMRI approximately 57%, and OBMI approximately 32%. In this application, FMI images were used as training data. A training dataset was created using FMI images with approximately 80% coverage, and a subset of FMI images were selected for the training set, which contains wellbore information for a total length of 3241m. First, the FMI images were completed using methods such as CIFLOG logging interpretation software or other methods, for example, using generative adversarial networks. The model pre-training used a large number of complete natural and seismic images. During model fine-tuning, to ensure consistency between the pre- and post-training data, relatively complete data without significant inconsistencies was provided to the model. Because perceptual loss was chosen in the loss function calculation, which calculates the difference between features in the feature space of the true image before and after the calculation, the feature space distribution will be altered if the complete image is not used during the fine-tuning stage. The experiment used the completed image to create a 512×512 dataset, and used the original FMI logging images to create corresponding mask labels. These were then used in the subsequent loss function calculation by multiplying the image by the mask labels to calculate part of the loss.

[0070] Using seismic data to retrain the pre-trained neural network for well logging image completion resulted in poor bedding continuity. Therefore, fine-tuning with more well logging data was performed to further adapt the neural network to the well logging image completion task. Specifically, approximately 60,000 well logging images containing missing regions were used to fine-tune the pre-trained neural network. During fine-tuning, the loss function was modified to include partial loss to incorporate losses from known regions, avoiding the introduction of inaccurate information. The fine-tuned neural network achieved smoother and more continuous well logging image completion overall, demonstrating improved accuracy and reliability of the image completion model.

[0071] Furthermore, the images used in the fine-tuning process are incomplete logging images. To avoid introducing inaccurate information during fine-tuning, partial loss is used in the fine-tuning stage. Given the original logging image as... I in The completed image is I out The corresponding mask is M. In M, 1 represents the real area and 0 represents the blank space. The mask M is multiplied by each image pixel to eliminate the influence of missing data on model fine-tuning. Specifically, the formula for calculating the loss function in the fine-tuning process is as follows: The partial content loss function is calculated using the following formula: (Formula 5); The countermeasure loss is calculated using the following formula: (Formula 6); Part of the MAE loss is calculated using the following formula: (Formula 7); In formula seven above, This indicates partial MAE loss. n Indicates the total number of pixels.

[0072] The total loss function during the fine-tuning phase is calculated using the following formula: (Formula 8); In this application, during model fine-tuning, λ is set as the regularization weight and η is the weight for partial content loss. λ can be set to 10 and η to 0.1.

[0073] The trained image completion model is used to complete electrical imaging logging in the actual work area. For example... Figures 5 to 10 As shown, FMI logging data, Earth Imager logging data, and OBMI logging data from an actual work area were used to complete the missing imaging logging bands. Simultaneously, SSIM (Structural Similarity Index), MSE (Mean Squared Error), PSNR (Peak Signal-to-Noise Ratio), FID (Frechet Inception Distance), and MAE (Mean Absolute Error) were used as evaluation metrics. SSIM considers the structural information of the image, with a value range of [-1, 1], where 1 indicates that the two images are completely identical, and -1 indicates that the two images are extremely different. FID measures the similarity of the generated image and the real image in the feature space; the lower the FID value, the closer the generated image is statistically to the real image, and the higher the value, the closer it is to the real image. Combined with... Figures 5 to 10 As shown, the well logging images completed by the image completion model constructed in this application have more reasonable texture and semantic information as a whole, while not introducing any artifacts, which makes it easier for researchers to interpret and analyze the subsurface structure based on the complete well logging images.

[0074] Corresponding to the electrical imaging logging completion method in the above embodiments, Figure 11The diagram illustrates the components of the electrical imaging logging completion device 500 provided in the embodiments of this application. For ease of explanation, only the parts relevant to the embodiments of this application are shown.

[0075] like Figure 11 As shown, the electrical imaging logging completion device 500 includes: The mask creation module 510 is used to create a mask for the current well logging image based on the known and missing regions in the current well logging image; Training module 520 is used to perform training processing on the constructed neural network based on well logging images and masks to perform pixel feature extraction and pixel dependency capture; The completion module 530 is used to perform pixel inference and iterative completion of the missing regions of the logging image and the mask based on the image completion model, and generate a logging image with the missing regions completed. During iterative completion, the logging image and its mask are updated according to the new known regions obtained after pixel filling by the image completion model in the current iteration step, so as to indicate the effective token regions for feature aggregation in the attention layer of the neural network in the next iteration step.

[0076] As an embodiment of this application, the electrical imaging logging completion device 500 can achieve the following: Figure 1 The embodiments shown are as well as other related method embodiments in this application.

[0077] The process by which each module in the electrical imaging logging completion device 500 provided in this application implements its respective function can be found in the foregoing. Figure 1 The descriptions of the embodiments shown and other related method embodiments are not repeated here.

[0078] It should be noted that the information interaction and execution process between the above modules are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Furthermore, all of the above modules can be applied to computing devices that include memory and a processor.

[0079] Figure 12 A schematic block diagram of a computer device according to an embodiment of the present application is shown. In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as shown below. Figure 12As shown in the figure, the computer device includes a processor A01, a network interface A02, a display screen A04, an input device A05, and a memory (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A06. The non-volatile storage medium A06 stores an operating system B01 and a computer program B02. The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A06. The network interface A02 is used for communication with external terminals via a network connection. When the computer program is executed by the processor A01, it implements an electrical imaging logging completion method. The display screen A04 can be a liquid crystal display (LCD) or an e-ink display. The input device A05 can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0080] Those skilled in the art will understand that Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0081] In one embodiment, the electrical imaging logging completion device 500 provided in this application can be implemented as a computer program, and the computer program can be implemented in, for example... Figure 12 The computer device shown operates on the device. The computer device's memory can store various program modules that make up the electrical imaging logging completion apparatus 500. The computer program, composed of these program modules, causes the processor to execute the steps in the electrical imaging logging completion methods of the various embodiments of this application described in this specification.

[0082] In one embodiment, this application also provides a machine-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the electrical imaging logging completion method in the above embodiments.

[0083] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0084] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0085] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for completing electrical imaging logging, characterized in that, include: Create a mask for the current well logging image based on the known and missing regions in the current well logging image; The image completion model performs pixel inference and iterative completion of missing regions on the logging image and the mask to generate a logging image with missing regions completed. The image completion model is obtained by training a neural network based on the logging image and the mask to extract pixel features and capture pixel dependencies. During iterative completion, the logging image and its mask are updated based on the new known regions obtained after pixel filling by the image completion model in the current iteration step, so as to indicate the effective token region for feature aggregation in the attention layer of the neural network in the next iteration step.

2. The electrical imaging logging completion method according to claim 1, characterized in that, The training process includes: pre-training the neural network using complete images; and training the pre-trained neural network using historical well logging images.

3. The electrical imaging logging completion method according to claim 2, characterized in that, The complete imagery includes both natural and seismic images.

4. The electrical imaging logging completion method according to claim 3, characterized in that, Pre-training the neural network using complete images includes: Based on the neural network, inference is performed on the continuous structural features in natural images until the inference result meets the preset requirements, thus obtaining a staged neural network. Based on a staged neural network, inference is performed on the continuous structural and texture features in seismic images until the inference results meet the preset requirements, thus obtaining a pre-trained neural network.

5. The electrical imaging logging completion method according to claim 1, characterized in that, The neural network includes a first sub-neural network for performing downsampling, a second sub-neural network connected to the first sub-neural network for filling missing regions, a third sub-neural network connected to the second sub-neural network for performing upsampling, and a fourth sub-neural network connected to the third sub-neural network for at least completing local features. The second sub-neural network includes the attention layer and is connected to a mask update module. The fourth sub-neural network takes the well logging image, the mask, and the image obtained after upsampling by the third sub-neural network as input to the first sub-neural network.

6. The electrical imaging logging completion method according to claim 5, characterized in that, The second sub-neural network includes multiple Transformer layers connected in series. Each Transformer layer includes multiple Transformer blocks, and each Transformer block includes an input module, an encoder module, a decoder module, and an output module, wherein: The input module includes an embedding layer but does not include position encoding; The encoder module is composed of multiple encoder layers stacked together. Each encoder layer consists of two sub-layer connection structures. The first sub-layer connection structure includes a multi-head self-attention layer and a feature stitching layer connected to the multi-head self-attention layer. The second sub-layer connection structure includes a feedforward fully connected sub-layer and a feature stitching layer connected to the feedforward fully connected sub-layer. The decoder module is composed of multiple decoder layers stacked together. Each decoder layer consists of three sub-layer connection structures. The first sub-layer connection structure includes a masked multi-head self-attention layer and a feature splicing layer connected to the masked multi-head self-attention layer. The second sub-layer connection structure includes a multi-head attention layer and a feature splicing layer connected to the multi-head attention layer. The third sub-layer connection structure includes a feedforward fully connected sub-layer and a feature splicing layer connected to the feedforward fully connected sub-layer.

7. The electrical imaging logging completion method according to claim 5, characterized in that, The fourth sub-neural network is a U-Net network.

8. The electrical imaging logging completion method according to claim 5, characterized in that, During iterative completion, the logging image and its mask are updated based on the new known regions obtained after pixel filling by the image completion model in the current iteration step. This indicates the effective token regions for feature aggregation in the attention layer of the neural network in the next iteration step, including: In the current iteration step, the first sub-neural network is used to downsample the logging image and its mask, the second sub-neural network is used to fill the missing region pixels of the token converted from the feature map obtained after downsampling, the third sub-neural network is used to upsample the output image obtained after filling the missing region pixels of the second sub-neural network, and the fourth sub-neural network is used to at least perform local feature completion on the feature map obtained after upsampling. The current logging image and its mask are updated based on the known regions in the feature map obtained after upsampling. The updated mask is used to indicate the effective token region for feature aggregation in the attention layer of the second sub-neural network in the next iteration step. Then, the next iteration step is entered until all the missing regions in the logging image are filled with pixels and the iteration stops.

9. The electrical imaging logging completion method according to claim 1, characterized in that, Create a mask for the current well logging image based on the known and missing regions in the current well logging image, including: The known and missing regions in the current well logging image are each represented by a single pixel value to obtain a binary mask with the same size as the well logging image.

10. The electrical imaging logging completion method according to claim 9, characterized in that, The known and missing regions in the current well logging image are each represented by a single pixel value to obtain a binary mask of the same size as the well logging image, including: The known regions in the well logging images segmented using a preset segmentation threshold are subjected to hole filling processing, which includes at least expansion or corrosion processing. The well logging image after binarization and hole filling processing is represented by a single pixel value for both the known and missing regions in the current well logging image, thus obtaining the mask for the current well logging image.

11. The electrical imaging logging completion method according to claim 5 or 8, characterized in that, The mask update strategy is as follows: Mark all tokens within the known region of the output image obtained after upsampling by the third sub-neural network in the current iteration step as valid tokens; The missing region in the logging image of the first sub-neural network in the previous iteration step is used as the window sliding region. The preset window is slid in the output image obtained after upsampling by the third sub-neural network. If there is at least one valid token in the current window, all tokens in the current window are updated to valid tokens. Otherwise, all tokens in the current window are still invalid tokens after updating. After the window sliding is completed, the valid tokens and invalid tokens are represented by a pixel value to obtain the mask for the next iteration step.

12. The electrical imaging logging completion method according to claim 1 or 5, characterized in that, The logging image is a planar unfolded view of the electrical imaging obtained by rotating the electrical imaging logging tool inside the wellbore. The left and right boundaries of the planar unfolded view are the electrical scanning data of the same wellbore location. During pixel inference, the continuity of the texture and / or continuity structure of the left and right boundaries of the logging image after the missing region is filled is constrained.

13. The electrical imaging logging completion method according to claim 12, characterized in that, The continuity of texture and / or continuity structure of the left and right boundaries in the well logging image after constraint missing region completion includes: After the last iteration step, the left and right boundaries of the output image obtained by filling the missing region pixels of the second sub-neural network are masked to obtain a new image after the left and right boundaries are occluded. The new image is segmented from the middle position, and the segmented right region is moved to the left of the segmented left region or the segmented left region is moved to the right of the segmented right region to obtain a reconstructed image after the left and right boundaries are joined. The reconstructed image is input into the third sub-neural network for upsampling, and the feature map obtained after upsampling is supplemented with at least local features using the fourth sub-neural network.

14. The electrical imaging logging completion method according to claim 5, characterized in that, There are skip connections between the first sub-neural network and the sub-layers with the same image resolution in the third sub-neural network.

15. The electrical imaging logging completion method according to claim 5, characterized in that, The neural network also includes a 1x1 convolution before the first sub-neural network and a 1x1 convolution between the third sub-neural network and the fourth sub-neural network.

16. An electrical imaging logging completion device, characterized in that, include: The mask creation module is used to create a mask for the current well logging image based on the known and missing regions in the current well logging image; The training module is used to train the constructed neural network that performs pixel feature extraction and pixel dependency capture based on well logging images and masks. The completion module is used to perform pixel inference and iterative completion of missing regions on the well logging image and the mask based on the image completion model, and generate a well logging image with missing regions completed. During iterative completion, the well logging image and its mask are updated according to the new known regions obtained after pixel filling by the image completion model in the current iteration step, so as to indicate the effective token region for feature aggregation in the attention layer of the neural network in the next iteration step.

17. The electrical imaging logging completion device according to claim 16, characterized in that, The neural network includes a first sub-neural network for performing downsampling, a second sub-neural network connected to the first sub-neural network for filling missing regions, a third sub-neural network connected to the second sub-neural network for performing upsampling, and a fourth sub-neural network connected to the third sub-neural network for at least completing local features. The second sub-neural network includes the attention layer and is connected to a mask update module. The fourth sub-neural network takes the well logging image, the mask, and the image obtained after upsampling by the third sub-neural network as input to the first sub-neural network.

18. The electrical imaging logging completion device according to claim 17, characterized in that, During iterative completion, the logging image and its mask are updated based on the new known regions obtained after pixel filling by the image completion model in the current iteration step. This indicates the effective token regions for feature aggregation in the attention layer of the neural network in the next iteration step, including: In the current iteration step, the first sub-neural network is used to downsample the logging image and its mask, the second sub-neural network is used to fill the missing region pixels of the token converted from the feature map obtained after downsampling, the third sub-neural network is used to upsample the output image obtained after filling the missing region pixels of the second sub-neural network, and the fourth sub-neural network is used to at least perform local feature completion on the feature map obtained after upsampling. The current logging image and its mask are updated based on the known regions in the feature map obtained after upsampling. The updated mask is used to indicate the effective token region for feature aggregation in the attention layer of the second sub-neural network in the next iteration step. Then, the next iteration step is entered until all the missing regions in the logging image are filled with pixels and the iteration stops.

19. A computer device, characterized in that, include: The memory is configured to store instructions; as well as The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the electrical imaging logging completion method according to any one of claims 1 to 15.

20. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to perform the electrical imaging logging completion method according to any one of claims 1 to 15.

Images