A well logging data completion method and device based on a generative deep learning model

Abstract

This invention relates to the field of oil and gas exploration and development technology, and discloses a method and apparatus for well logging data completion based on a generative deep learning model. The method progressively adds Gaussian noise to missing well logging data to be restored, obtaining noisy well logging data. The data to be restored includes known well logging data and missing data, while the noisy well logging data includes known well logging data and noisy data. Based on a created observation mask, zeros are filled into the missing well logging data to be restored and the known well logging data of the noisy well logging data, respectively, to obtain first mask data and second mask data. The first mask data and second mask data are concatenated to construct a hybrid tensor. Based on the complete logging data corresponding to the data to be restored, the hybrid tensor, and the trained generative deep learning model, the noisy well logging data is denoised to obtain the completed well logging data. This invention can effectively improve the accuracy of well logging data completion.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and in particular to a method and apparatus for completing well logging data based on a generative deep learning model. Background Technology

[0002] In the process of oil and gas exploration and development, well logging data, as geological data reflecting the physical properties of underground rocks, is the core basis for reservoir evaluation, reservoir description and geological modeling.

[0003] In actual operations, due to factors such as wellbore environment (e.g., enlargement, collapse), instrument malfunction, logging obstruction, or operating costs, logging data often shows missing sections or missing entire curves, affecting subsequent detailed reservoir description and geological modeling work.

[0004] Related well logging data completion methods, such as linear interpolation or statistically based multiple regression, rely on the internal correlation of well logging data, making it difficult to capture nonlinear variation characteristics under complex geological conditions, resulting in low completion accuracy. Summary of the Invention

[0005] This invention provides a well logging data completion method and apparatus based on a generative deep learning model, which addresses the shortcomings of low well logging data completion accuracy in related technologies and improves the accuracy of well logging data completion.

[0006] In a first aspect, the present invention provides a well logging data completion method based on a generative deep learning model, comprising:

[0007] Gaussian noise is gradually added to the missing logging data to be restored, resulting in noisy logging data; wherein, the logging data to be restored includes known logging data and the missing logging data, and the noisy logging data includes the known logging data and noise data; Based on the created observation mask, 0 is filled into the missing logging data of the logging data to be restored and the known logging data of the noisy logging data, respectively, to obtain the first mask data and the second mask data; The first mask data and the second mask data are concatenated to construct a hybrid tensor; Based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model, the noisy logging data is denoised to obtain complete logging data.

[0008] Optionally, the step of denoising the noisy logging data based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model to obtain complete logging data includes: Obtain the depth sequence corresponding to the well logging data to be restored and the noise-added time step sequence corresponding to the Gaussian noise; Based on the logging data to be restored, the depth sequence, and the complete logging data, a comprehensive condition tensor is constructed; and each time step in the noisy time step sequence is encoded to obtain the feature data corresponding to each time step, and the whole is used as the feature data of the time step. The time step feature data, the hybrid tensor, and the comprehensive conditional tensor are input into the trained generative deep learning model to perform noise prediction, thereby obtaining noise prediction data. The noisy logging data is denoised based on the noise prediction data to obtain complete logging data.

[0009] Optionally, constructing a comprehensive condition tensor based on the logging data to be restored, the depth sequence, and the complete logging data includes: Determine the category identifier corresponding to the logging curve in the logging data to be restored; The complete logging data is subjected to feature projection to obtain target feature data; the depth sequence is encoded to obtain corresponding depth encoded data; and the corresponding curve category feature data is determined according to the category identifier. The target feature data, the depth encoding data, and the curve category feature data are concatenated to obtain the comprehensive conditional tensor.

[0010] Optionally, the step of performing feature projection on the complete logging data to obtain target feature data includes: The complete logging data is input into a convolutional layer so that the convolutional layer maps the complete logging data into target feature data with a set feature dimension, thereby achieving feature dimension compression and alignment and extracting nonlinear combination features between different logging parameters in the complete logging data; The depth sequence contains each depth from the logging data to be restored or the complete logging data, arranged sequentially; the encoding of the depth sequence to obtain corresponding depth-coded data includes: A fixed-position encoding strategy based on trigonometric functions and an embedded dimension index are used to encode each depth in the depth sequence to obtain the corresponding depth-encoded data. The step of determining the corresponding curve category feature data based on the category identifier includes: The category identifier is input into the constructed embedding layer so that the embedding layer can find the curve category feature data corresponding to the category identifier in the created category feature dataset.

[0011] Optionally, the generative deep learning model includes a first processing layer, a residual network, and a second processing layer; The step of inputting the time-step feature data, the hybrid tensor, and the comprehensive conditional tensor into a trained generative deep learning model for noise prediction, to obtain noise prediction data, includes: The hybrid tensor is input into the first processing layer for convolution and activation to obtain the processed data; The processed data, the integrated condition tensor, and the time step feature data are input into the residual network for deep feature extraction to obtain feature extraction data. The extracted feature data is input into the second processing layer for convolution and activation to obtain the noise prediction data.

[0012] Optionally, the residual network includes n stacked residual layers and a fusion module disposed between two adjacent residual layers, where n is an integer greater than 1; The process involves inputting the processed data, the comprehensive conditional tensor, and the time-step feature data into the residual network for deep feature extraction, resulting in feature extraction data, including: The processed data, the comprehensive condition tensor, and the time step feature data are input into the first residual layer for feature extraction to obtain the first output data and the second output data of the first residual layer; the first output data of the first residual layer and the processed data are input into the first fusion module for addition to obtain the first added data; The first summed data, the comprehensive condition tensor, and the time step feature data are input into the second residual layer for feature extraction to obtain the first output data and the second output data of the second residual layer; the first output data of the second residual layer and the first summed data are input into the second fusion module for addition to obtain the second summed data; The second summed data, the comprehensive condition tensor, and the time step feature data are input into the third residual layer for feature extraction until the second output data of the last residual layer is obtained. The second output data from each residual layer are summed to obtain the target sum data; The target sum data is numerically scaled to obtain the feature extraction data.

[0013] Optionally, each residual layer includes a depth direction feature extraction module and a curve correlation feature extraction module; The first output data of the previous residual layer, the comprehensive condition tensor, and the time step feature data are input into the current residual layer for feature extraction, resulting in the first and second output data of the current residual layer, including: The first output data of the previous residual layer, the comprehensive condition tensor, and the time step feature data are input into the depth direction feature extraction module for feature extraction to obtain depth direction feature data. The depth-direction feature data is input into the curve-related feature extraction module for feature extraction to obtain the first output data and the second output data of the current residual layer.

[0014] Optionally, the depth direction feature extraction module includes a splicing submodule, a first fusion submodule, a Transformer encoder, a first sub-convolution module, a second sub-convolution module, and a second fusion submodule; The first output data of the previous residual layer, the comprehensive conditional tensor, and the time step feature data are input into the depth direction feature extraction module for feature extraction to obtain depth direction feature data, including: Each feature data in the time step feature data is input into the splicing submodule for splicing to obtain spliced ​​feature data; The first output data of the previous residual layer and the spliced ​​feature data are input into the first fusion submodule for addition to obtain the first fused data; The first fused data is input into the Transformer encoder to perform self-attention calculation in the time dimension, thereby obtaining the first encoded data; The first encoded data is input into the first sub-convolution module for convolution to obtain the first sub-convolution data; the comprehensive condition tensor is input into the second sub-convolution module for convolution to obtain the second sub-convolution data. The first sub-convolutional data and the second sub-convolutional data are input into the second fusion submodule and added together to obtain the second fused data, which is used as the depth direction feature data.

[0015] Optionally, the curve-related feature extraction module includes a third sub-convolution module, a fourth sub-convolution module, and a fifth sub-convolution module; The step of inputting the depth-direction feature data into the curve-related feature extraction module for feature extraction to obtain the first output data and the second output data of the current residual layer includes: The depth-direction feature data is input into the third sub-convolution module for convolution to obtain the third sub-convolution data; The third sub-convolutional data is input into the fourth and fifth sub-convolutional modules respectively for convolution to obtain the fourth and fifth sub-convolutional data, which are then used as the first and second output data of the current residual layer, respectively.

[0016] Secondly, the present invention provides a well logging data completion device based on a generative deep learning model, comprising: A noise-adding unit is used to progressively add Gaussian noise to the missing logging data in the logging data to be restored, to obtain noisy logging data; wherein, the logging data to be restored includes known logging data and the missing logging data, and the noisy logging data includes the known logging data and noise data; A masking unit is used to fill 0 in the missing logging data of the logging data to be restored and the known logging data of the noisy logging data, respectively, based on the created observation mask, to obtain the first masking data and the second masking data; A splicing unit is used to splice the first mask data and the second mask data to construct a hybrid tensor; The denoising unit is used to denoise the noisy logging data based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model, so as to obtain complete logging data.

[0017] This invention provides a well logging data completion method and apparatus based on a generative deep learning model. It progressively adds Gaussian noise to missing well logging data to obtain noisy well logging data. The data to be restored includes known well logging data and missing data, while the noisy well logging data includes both known data and noise data. Based on a created observation mask, zeros are filled into both the missing and known well logging data to obtain a first mask and a second mask. The first and second mask data are then concatenated to construct a hybrid tensor. Based on the complete logging data corresponding to the data to be restored, the hybrid tensor, and the trained generative deep learning model, the noisy well logging data is denoised to obtain the completed well logging data. This invention constructs a hybrid tensor based on the data to be restored and the noisy well logging data. Using the hybrid tensor, complete logging data, and the generative deep learning model, it achieves well logging data completion, effectively improving the accuracy of well logging data completion.

[0018] This invention innovatively integrates continuous and stable engineering logging data with geological logging data, solving the problem of information silos when relying solely on geological logging data for data completion.

[0019] The generation mechanism based on the diffusion model of this invention can generate well logging curves that conform to geological laws and have high-frequency details, thus avoiding the smoothing effect of traditional regression methods.

[0020] This invention is highly versatile, and the framework can be extended to complete various types of well logging curves, such as acoustic and density logging. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a well logging data completion method based on a generative deep learning model, provided as an embodiment of the present invention; Figure 2 This is a schematic diagram of data flow in a generative deep learning model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of data flow in a residual layer provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the forward and reverse sampling process of a diffusion model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the forward and reverse sampling process of a diffusion model that integrates logging information, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of a well logging data completion device based on a generative deep learning model, provided in an embodiment of the present invention. Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0024] While deep learning methods or conventional fully connected networks have improved generation performance to some extent, most of them employ a discriminative approach of using logging data to supplement existing logging data (i.e., input A predicts output B), meaning they use existing logging curves to predict missing curves. This specific input-to-specific-output model has a fatal weakness: when the reference curve used as input is also missing, or when multiple curves are missing simultaneously across the entire well section, the model will completely fail due to a lack of input features. Compared to discriminative models, which fail completely when input curves are missing, generative models, by learning the inherent structure and distribution patterns of the data itself, can infer and reconstruct high-fidelity curves that conform to geological characteristics from noise, effectively solving the problem of prediction failure due to missing data. How to break down the barriers between engineering data and geological data, and utilize engineering logging data that is inexpensive to acquire, has good continuity, and is less prone to loss, to solve the problem of missing geological logging data, is a key technical challenge that related technologies need to address.

[0025] This embodiment aims to address the shortcomings of related technologies, such as their inability to predict complex nonlinear relationships and missing input reference curves, their failure to handle large, continuous gaps in the data, and their inability to fully utilize logging engineering data. This embodiment can infer formation properties using readily available engineering data, achieving high-precision reconstruction of missing logging curves.

[0026] The following is combined with Figures 1-5 This invention describes a well logging data completion method based on a generative deep learning model.

[0027] like Figure 1 As shown, this embodiment proposes a first method for completing well logging data based on a generative deep learning model. This method may include the following steps: S101. Gaussian noise is gradually added to the missing logging data to be restored to obtain noisy logging data; wherein, the logging data to be restored includes known logging data and missing logging data, and the noisy logging data includes known logging data and noise data.

[0028] The logging data to be restored includes logging parameter values ​​that vary with depth, showing a correspondence between depth and logging parameter values. The logging data to be restored can be a matrix containing multiple rows and columns, with each row including a depth and the corresponding logging parameter values.

[0029] Specifically, known logging data refers to logging data with known depths and logging parameter values. Missing logging data refers to logging data with missing depths and / or logging parameter values; the missing portion is recorded as a null value.

[0030] Specifically, in this embodiment, Gaussian noise can be gradually added to the missing logging data to be restored, resulting in noisy logging data. The final data obtained by gradually adding Gaussian noise to the missing logging data is the noisy data.

[0031] S102. Based on the created observation mask, fill the missing logging data and the known logging data of the noisy logging data into the logging data to be restored, respectively, to obtain the first mask data and the second mask data.

[0032] Specifically, this embodiment can be based on the logging data to be restored. Create observation mask The observation data to be restored Multiplying with the observation mask fills in 0s in the missing logging data of the logging data to be restored, and retains data values ​​in the known logging data of the logging data to be restored, thus obtaining the first mask data after processing the logging data to be restored.

[0033] Specifically, in this embodiment, the observation mask can be subtracted from 1. The corresponding missing mask is obtained. Noisy logging data With missing mask Multiplication is performed to fill the known logging data with 0s and retain the data values ​​in the noisy logging data, thus obtaining the second mask data after processing the noisy logging data.

[0034] S103. Concatenate the first mask data and the second mask data to construct a hybrid tensor.

[0035] It should be noted that the first mask data and the second mask data are matrix data with the same number of rows and columns.

[0036] Specifically, in this embodiment, the first mask data and the second mask data can be horizontally concatenated, and the concatenated data can be determined as a hybrid tensor.

[0037] S104. Based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model, the noisy logging data is denoised to obtain the complete logging data.

[0038] It should be noted that, through research and experimentation, the inventors have concluded that well logging data, as engineering data acquired in real time during drilling, contains a wealth of information, including drilling time, total hydrocarbons, chromatographic analysis, drilling pressure, and rotation speed. While these engineering parameters are primarily used to monitor drilling safety and make preliminary judgments about oil and gas shows, they are inherently physically related to formation lithology and physical properties. For example, the drilling speed often reflects the hardness (lithology) of the formation, and changes in gas logging values ​​directly indicate the hydrocarbon content of the reservoir. However, related technologies often overlook this deep connection, treating engineering data and geological data separately. From a physical perspective, both are objective reflections of the same formation properties (such as lithology, porosity, and fluid properties) under different physical fields and engineering environments. This inherent consistency of data origin and physical coupling provides a solid theoretical basis for using continuous and stable engineering logging data to infer and reconstruct missing geological logging data.

[0039] The complete logging data corresponding to the logging data to be restored refers to complete logging data collected from the same well. The depth distribution in the logging data to be restored and the complete logging data can be the same. The logging parameter values ​​corresponding to different depths in the complete logging data are all known, with no missing parts.

[0040] Specifically, this embodiment can use complete logging data, hybrid tensors, and a trained generative deep learning model to denoise noisy logging data, thereby completing the missing data in the logging data to be restored and obtaining complete logging data.

[0041] Optionally, step S104 includes: Obtain the depth sequence corresponding to the logging data to be restored and the time step sequence with added Gaussian noise; Based on the logging data to be restored, the depth sequence, and the complete logging data, a comprehensive condition tensor is constructed; and each time step in the noisy time step sequence is encoded to obtain the feature data corresponding to each time step, and the whole is used as the time step feature data. The time step feature data, the hybrid tensor, and the comprehensive conditional tensor are input into the trained generative deep learning model to perform noise prediction, and noise prediction data is obtained. Noise-inducing logging data is processed based on noise prediction data to obtain complete logging data.

[0042] The depth sequence includes each depth in the logging data to be restored, and the depths in the depth sequence can be arranged in ascending order.

[0043] It is understood that Gaussian noise is gradually added to the missing logging data to be restored in this embodiment. The noisy time step sequence corresponding to the Gaussian noise includes each time step for which noise is added, and the time steps in the noisy time step sequence can be arranged in ascending order. Specifically, in this embodiment, each time step in the noisy time step sequence can be encoded separately to obtain the feature data corresponding to each time step, and the whole sequence can be used as the time step feature data.

[0044] Optionally, the above-mentioned comprehensive condition tensor is constructed based on the logging data to be restored, the depth sequence, and the complete logging data, including: Determine the category identifier corresponding to the logging curve in the logging data to be restored; Feature projection is performed on the complete logging data to obtain target feature data; depth sequences are encoded to obtain corresponding depth encoded data; and corresponding curve category feature data are determined based on category identifiers. By concatenating the target feature data, deep encoding data, and curve category feature data, a comprehensive conditional tensor is obtained.

[0045] Optionally, the above-mentioned feature projection of complete logging data to obtain target feature data includes: The complete logging data is input into the convolutional layer so that the convolutional layer maps the complete logging data into target feature data with a set feature dimension, thereby achieving feature dimension compression and alignment and extracting nonlinear combination features between different logging parameters in the complete logging data; Optionally, the depth sequence contains each depth from the logging data to be restored or the complete logging data, arranged sequentially; the above-mentioned encoding of the depth sequence to obtain the corresponding depth-coded data includes: A fixed-position encoding strategy based on trigonometric functions and an embedded dimension index are used to encode each depth in the depth sequence to obtain the corresponding depth-encoded data. Optionally, the curve category feature data determined above based on the category identifier includes: The category identifier is input into the constructed embedding layer so that the embedding layer can look up the curve category feature data corresponding to the category identifier in the created category feature dataset.

[0046] like Figure 2 As shown, this embodiment can perform well logging engineering feature projection. Let the complete well logging data be... .in, For batch size, The number of original logging parameters. This is the length of the depth sequence. This embodiment can utilize a convolutional layer as a feature projector to project the original feature dimensions. Mapping to a fixed-dimensional latent space dimension This operation preserves the time / depth dimension. While keeping the parameters unchanged, feature dimension compression and alignment were achieved, and nonlinear combination features between different logging parameters were extracted.

[0047] like Figure 2 As shown, this embodiment can perform depth location encoding. To enable the model to perceive the sequential order of logging data along the depth direction, we adopted a fixed location encoding strategy based on sine and cosine functions. For each depth index location... and embedded dimension index The formula for calculating its encoded value is: ; ; It should be noted that this encoding method can generate periodic and unique relative position information, enabling the model to capture the trend characteristics of well logging curves as they change with depth.

[0048] like Figure 2 As shown, this embodiment can perform well logging curve category embedding. To distinguish well logging curves with different physical properties (such as acoustic AC, gamma GR, resistivity RT, etc.), this embodiment assigns a unique integer index ID to each type of curve. A learnable embedding layer, i.e., a parameter matrix, is constructed. ( (This represents the number of curve types). The model can find and output the corresponding feature vector based on the input ID index. This vector implicitly contains the differences in the physical properties of different logging curves, guiding the model to generate specific types of curve data.

[0049] like Figure 2 As shown, this embodiment can concatenate the target feature data, depth encoded data, and curve category feature data obtained after processing the complete logging data, depth sequence, and logging curve category identifier respectively to obtain a comprehensive condition tensor, thereby realizing the construction of multi-source conditions and injecting them into the generative deep learning model for noise data prediction.

[0050] Optional, such as Figure 2 As shown, the generative deep learning model includes a first processing layer, a residual network, and a second processing layer. The above-mentioned time-step feature data, mixture tensor, and comprehensive conditional tensor are input into a trained generative deep learning model for noise prediction, resulting in noise prediction data, including: The hybrid tensor is input into the first processing layer for convolution and activation to obtain the processed data; The processed data, the integrated condition tensor, and the time step feature data are input into the residual network for deep feature extraction to obtain the feature extraction data. The extracted feature data is input into the second processing layer for convolution and activation to obtain noise prediction data.

[0051] Optionally, the residual network described above includes n stacked residual layers and a fusion module positioned between two adjacent residual layers, where n is an integer greater than 1. In this case, the processed data, the comprehensive conditional tensor, and the time-step feature data are input into the residual network for deep feature extraction to obtain feature extraction data, including: The processed data, the comprehensive condition tensor, and the time step feature data are input into the first residual layer for feature extraction to obtain the first output data and the second output data of the first residual layer; the first output data of the first residual layer and the processed data are input into the first fusion module for addition to obtain the first added data; The first summed data, the comprehensive condition tensor, and the time step feature data are input into the second residual layer for feature extraction to obtain the first output data and the second output data of the second residual layer; the first output data of the second residual layer and the first summed data are input into the second fusion module for addition to obtain the second summed data; The second summed data, the comprehensive condition tensor, and the time step feature data are input into the third residual layer for feature extraction until the second output data of the last residual layer is obtained. The second output data from each residual layer is summed to obtain the target summed data. Numerical scaling is applied to the target summation data to obtain feature extraction data.

[0052] It should be noted that the core processing unit of the denoising network is the residual layer based on the Transformer mechanism. Given that well logging data possesses both vertical sequentiality (depth) and horizontal feature correlation (between curves), the residual layer incorporates two cascaded processing modules: a depth-direction feature extraction module and a curve correlation feature extraction module. The depth-direction feature extraction module utilizes the Transformer encoder layer and is specifically designed to capture global dependencies and formation structure features along the depth direction. The input features are first self-attention calculated in the time dimension, and the output serves as the input for the next module. The curve correlation feature extraction module follows immediately after the depth-direction feature extraction module and is specifically designed to capture the physical property correlations and coupling relationships between different well logging curves at the same depth point.

[0053] Optional, such as Figure 3 As shown, each residual layer contains a depth direction feature extraction module and a curve correlation feature extraction module; The first output data, the synthesis condition tensor, and the time step feature data of the previous residual layer are input into the current residual layer for feature extraction, resulting in the first and second output data of the current residual layer, including: The first output data of the previous residual layer, the comprehensive condition tensor and the time step feature data are input into the depth direction feature extraction module for feature extraction to obtain the depth direction feature data. The depth-direction feature data is input into the curve correlation feature extraction module for feature extraction, and the first output data and the second output data of the current residual layer are obtained.

[0054] Optional, such as Figure 3 As shown, the depth direction feature extraction module includes a splicing submodule, a first fusion submodule, a Transformer encoder, a first sub-convolution module, a second sub-convolution module, and a second fusion submodule. The first output data of the previous residual layer, the comprehensive condition tensor, and the time step feature data are input into the depth-direction feature extraction module for feature extraction, resulting in depth-direction feature data, including: Each feature data in the time step feature data is input into the splicing submodule for splicing to obtain spliced ​​feature data; The first output data of the previous residual layer and the spliced ​​feature data are input into the first fusion submodule for addition to obtain the first fused data; The first fused data is input into the Transformer encoder to perform self-attention calculation in the time dimension, and the first encoded data is obtained. The first encoded data is input into the first sub-convolution module for convolution to obtain the first sub-convolution data; the comprehensive condition tensor is input into the second sub-convolution module for convolution to obtain the second sub-convolution data. The first and second sub-convolutional data are input into the second fusion submodule and added together to obtain the second fused data, which is used as the depth-direction feature data.

[0055] Optional, such as Figure 3 As shown, the curve-related feature extraction module includes a third sub-convolution module, a fourth sub-convolution module, and a fifth sub-convolution module. The aforementioned input of depth-direction feature data into the curve correlation feature extraction module for feature extraction yields the first and second output data of the current residual layer, including: The depth-oriented feature data is input into the third sub-convolution module for convolution to obtain the third sub-convolution data; The third sub-convolution data is input into the fourth and fifth sub-convolution modules respectively for convolution to obtain the fourth and fifth sub-convolution data, which are then used as the first and second output data of the current residual layer, respectively.

[0056] It should be noted that, as Figure 2 and Figure 3As shown, the overall data flow of the network follows a topology of backbone residual progression-branch jump aggregation, specifically including the following three key stages: Input Transformation Stage: Hybrid Tensors First, convolution and activation are performed on the input projection layer, mapping the original channel dimension to a high-dimensional hidden layer feature dimension. Thus, the initial hidden state is obtained. .

[0057] Interlayer progression stage (backbone pathway): Included in the model The stacked residual layers. For the first... One residual layer ( Its input data includes the output features of the previous layer. Global condition tensor and the embedding vector at the current time step The output of the previous layer becomes the input of the next layer. After internal temporal / feature attention processing and gated fusion, the layer outputs updated residual features. This feature is directly used as the first The input from one layer enters the next level of processing, forming a hierarchical feature extraction.

[0058] Jump output phase (branch path): Within each residual block, in addition to the output being passed to the next layer of the trunk... In addition, a skip connection feature was also isolated in parallel. This feature does not participate in subsequent calculations on the backbone network, but is directly sent to the aggregation nodes at the end of the network for processing.

[0059] Specifically, regarding the reconstruction process of well logging data, in this embodiment, after cascaded feature extraction of the time and feature dimensions through the residual layer, the model completes the noise prediction data through the following steps. Prediction and denoising restoration of well logging data: Multi-scale feature aggregation: Instead of directly using the output of the last residual block for prediction, the model employs a skip connection aggregation strategy. This involves aggregating all skip connection features from the outputs of the n residual blocks. While maintaining consistency in the channel dimension, perform element-wise summation and divide by . Perform numerical scaling.

[0060] Output Projection and Reconstruction: The aggregated feature tensor is then passed through an output projection module. This module contains two convolutional layers and an intermediate... The activation function maps the high-dimensional hidden layer features back to the original logging data channel dimension, thus outputting the final prediction result.

[0061] Denoising and Restoration: The model outputs the predicted result, which represents the noise component that needs to be removed at the current time step. At each step of the inference phase, the predicted noise is extracted from the current noisy data using the backsampling formula of the diffusion model. By subtracting from the middle value, the noise-free logging curve reconstruction value is gradually restored. .

[0062] This embodiment improves the accuracy of well logging data completion in the following ways: Hybrid input tensor construction: The observed conditional channel is fused with the noisy target channel, enabling the network to directly reference real physical constraint information and improve completion accuracy.

[0063] Multi-source condition injection: By using well logging data and other engineering information as prior conditions, the model's ability to perceive geological features is further enhanced.

[0064] Residual networks effectively handle the time and feature dimensions of data, improving the model's ability to process the inherent complex relationships in well logging data.

[0065] This embodiment can be applied to scenarios such as well logging data preprocessing, old well review, and geological modeling data preparation in oil and gas field exploration and development. It is particularly suitable for repairing large sections of distorted or missing well logging curves caused by poor wellbore conditions.

[0066] The well logging data completion method based on a generative deep learning model proposed in this embodiment can progressively add Gaussian noise to the missing well logging data to be restored, obtaining noisy well logging data. The well logging data to be restored includes known well logging data and missing well logging data, while the noisy well logging data includes known well logging data and noisy data. Based on a created observation mask, zeros are filled into the missing well logging data to be restored and the known well logging data of the noisy well logging data, respectively, to obtain first mask data and second mask data. The first mask data and the second mask data are concatenated to construct a hybrid tensor. Based on the complete logging data corresponding to the well logging data to be restored, the hybrid tensor, and the trained generative deep learning model, the noisy well logging data is denoised to obtain the completed well logging data. This embodiment constructs a hybrid tensor based on the well logging data to be restored and the noisy well logging data, and uses the hybrid tensor, complete logging data, and generative deep learning model to achieve well logging data completion, effectively improving the accuracy of well logging data completion.

[0067] In the second well logging data completion method based on a generative deep learning model proposed in this embodiment, in order to construct and train the deep learning-based generative model, this embodiment first needs to construct a high-quality paired sample set containing geological well logging data and engineering well logging data as the basis for model training. This dataset consists of two core data parts: geological well logging data, which serves as the model reconstruction target, and engineering well logging data, which serves as the model condition guide. Both types of data are physically continuous sequence data distributed along the well depth direction, with strict spatial correspondence.

[0068] In the data preprocessing stage, this embodiment uses depth indexing as a benchmark to perform depth alignment and standardization on two types of data with potentially inconsistent sampling intervals, forming one-to-one corresponding depth sequence pairs. This data structure ensures that at any depth point, the instantaneous response of drilling engineering parameters can be accurately mapped to the corresponding geological attributes.

[0069] This depth-aligned sequence data structure also addresses the limitations of relying solely on the internal correlations of well logging data for completion. This embodiment embeds well logging data strongly correlated with formation lithology and physical properties into the model as additional prior conditions to guide the denoising process of the diffusion model. The well logging data embedding process can be considered a key step in model preprocessing and conditional fusion: first, feature projection is performed. Let the original well logging data be... By using convolutional layers, the logging features are mapped from the original dimension to a high-dimensional latent space, resulting in projected features. This operation not only achieved dimensional alignment but also extracted the local nonlinear correlations between logging parameters.

[0070] This is followed by multi-dimensional feature fusion. To enable the model to perceive depth location information and logging curve category information, this embodiment can project the logging features... Concatenate with auxiliary features to construct a comprehensive conditional tensor These auxiliary features include temporal / depth embeddings of depth index maps using sinusoidal positional encoding, and leveraging learnable... The layer embeds the target features mapped to each logging curve ID. These auxiliary features include: first, depth location encoding, which is a vector sequence generated using sine and cosine functions. It maps discrete depth indices to fixed-dimensional real-value vectors. Specifically, using sine and cosine trigonometric formulas, each discrete depth index value is encoded into a fixed-length real-value vector to preserve the sequential position information of the logging data along the depth direction. Secondly, there is the well logging curve category embedding, which is a learnable feature vector lookup table. This embodiment can assign unique integer index IDs to different types of well logging curves, such as acoustic, density, and resistivity curves. Through this embedding layer, discrete IDs are mapped to a continuous, fixed-dimensional feature vector. This vector is continuously updated during training, ultimately learning a latent feature representation capable of distinguishing the physical properties of different well logging curves.

[0071] Finally, in the condition injection phase, the fused condition tensor It is fed into the denoising network. Within the network's residual blocks, a dedicated conditional projection layer... After transformation, the features of the intermediate layer are applied in the form of addition or gating mechanisms. Within the residual blocks of the network, convolutional layers further transform... The channel dimension is adjusted to match that of the intermediate layer features, and then incorporated into the intermediate layer features in an element-wise addition manner. The fused features are further processed non-linearly through a gated activation unit, that is, the sigmoid function is used to control the gate value of information flow, the Tanh function is used to extract feature content, and the two are multiplied to obtain the output.

[0072] In this way, logging data is no longer a simple input splicing, but a kind of global contextual information that dynamically adjusts the network's prediction of missing logging data, ensuring that the generated logging curves are consistent with the physical laws of the drilling operation response.

[0073] It should be noted that the training phase includes a forward process. The forward process is a Markov chain with fixed parameters, the purpose of which is to distribute the raw logging data. Gaussian noise is gradually added until the data becomes a standard Gaussian distribution where each component is independent. During this process, only geological logging data is processed. Noise-adding operations were performed on the engineering logging data. Maintain the original state as a stable reference for the subsequent reverse process.

[0074] like Figure 4 As shown, assuming This represents the initial complete logging data; the forward process is performed through... Noise is gradually added at each time step to generate a sequence. At any time step The status of the data Depends only on the state of the previous time step Its conditional probability distribution is expressed as:

[0075] in, These are predefined noise variance scheduling parameters, typically varying with time steps. The increase is linear or nonlinear. It is an identity matrix.

[0076] Based on the superposition property of Gaussian distribution, it can be derived that for any time step... Data Regarding the raw data The closed-form solution. Let ,and Then the marginal distribution of the forward process can be represented as:

[0077] This indicates that this embodiment can achieve the desired result by sampling standard Gaussian noise. Directly obtain noise data at any given time:

[0078] along with , ,at this time It approximately follows a standard Gaussian distribution. .

[0079] This embodiment also includes a backward process. The backward process is the core of the diffusion model's data generation, designed to extract data from pure noise. The original data was recovered by gradually removing noise. Due to the true inverse transformation distribution This requires the distribution information of the entire dataset, which is usually difficult to calculate directly. Therefore, this embodiment can use a deep neural network (with parameters of...). To approximate this inverse conditional distribution: ; in, and These are the mean and variance of the neural network predictions, respectively. To simplify training, the variance is usually fixed. For constants (e.g.) or ), and transform the problem of predicting the mean into the problem of predicting noise.

[0080] Specifically, the network Trained to predict in Noise components added at any time The mean from the previous time step can be reconstructed using the predicted noise and a denoising formula: ; Thus achieve from arrive Sampling.

[0081] Regarding the model training method, this embodiment addresses the common problem of large, continuous missing segments in well logging curves by designing an interval masking strategy during the training phase. Unlike random point missing segments, this strategy randomly selects a continuous segment of length from the training samples for masking, forcing the model to learn to predict the content of the missing segment using contextual information and logging condition information, thereby improving the model's robustness to large data loss in practical applications.

[0082] In the inference phase, the data generation process in this embodiment for well logging curve completion is not completely random, but rather based on known observation data to recover missing data. Therefore, this embodiment employs a conditional diffusion probability model. Within this framework, it is crucial to clearly distinguish the existence states of two types of data: well logging engineering data, as accompanying records of the drilling process, typically exists continuously throughout the entire process and serves as a strong global condition guiding the inference process; while well logging geological data, as the target to be completed, may contain some known segments or be completely missing across the entire well section. The model constructed in this embodiment is capable of performing inference in both of these scenarios. Accordingly, the target well logging data is defined. It consists of two parts: the observed part and missing parts The objective of this embodiment is to utilize the existing logging data and the data that may exist. Together, they serve as conditional information to generate data that conforms to the distribution of geological data. .

[0083] Within this framework, the reverse generation process is modified to be subject to a conditional probability distribution controlled by multi-source data: ; in, Indicates the missing part is in The noise state at any given time. The neural network's current input is a mixture conditional tensor, which contains not only the current noisy data. and time step It also explicitly includes observational data. And complete logging engineering data The optimization objective of the model is to minimize the mean square error between the predicted noise and the actual noise. ; In the actual reasoning process, in order to ensure the consistency between the generated data and the known observation data, this embodiment will forcibly reset the values ​​of the known parts to the observed values ​​after each denoising step. Noisy version or use directly (Depending on the specific sampling strategy), thereby guiding the generation process to converge toward the real data manifold.

[0084] like Figure 5As shown, regarding the reasoning method, the missing value imputation process begins with the definition of the data state. In this embodiment, the condition mask is determined by marking the missing components as 0. And generate conditional observation logs filled with 0s. Subsequently, the filling process begins from the final diffusion step. T To begin, the initial, fully noisy target is obtained by filling in the missing values ​​with Gaussian noise. After that, the model was... T Step-by-step reverse denoising: The trained denoising model progressively predicts and eliminates introduced noise, thereby restoring the data. Crucially, after each denoising step, data consistency correction is performed: the denoising results are corrected using known, non-missing data. Specifically, the model retains the true observations for known segments while only updating the missing segment values ​​generated by the denoising model. This ensures that the final completion result seamlessly integrates with the observed data, achieving reliable missing value prediction.

[0085] This embodiment innovatively integrates continuous and stable engineering logging data with geological logging data, solving the information silo problem when relying solely on geological logging data for supplementation.

[0086] The generation mechanism based on the diffusion model in this embodiment can produce well logging curves that conform to geological laws and have high-frequency details, thus avoiding the smoothing effect of traditional regression methods.

[0087] This embodiment is highly versatile, and the framework can be extended to complete various types of logging curves, such as acoustic and density logging.

[0088] like Figure 6 As shown, this embodiment proposes a well logging data completion device based on a generative deep learning model. The device includes: The noise-adding unit 101 is used to progressively add Gaussian noise to the missing logging data of the logging data to be restored, so as to obtain noisy logging data; wherein, the logging data to be restored includes known logging data and missing logging data, and the noisy logging data includes known logging data and noise data; Masking unit 102 is used to fill 0 in the missing logging data of the logging data to be restored and the known logging data of the noisy logging data based on the created observation mask, respectively, to obtain the first mask data and the second mask data; The splicing unit 103 is used to splice the first mask data and the second mask data to construct a hybrid tensor; The denoising unit 104 is used to denoise the noisy logging data based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model, so as to obtain the complete logging data.

[0089] It should be noted that the processing procedures of the noise-adding unit 101, the masking unit 102, the splicing unit 103, and the noise-reducing unit 104, and their beneficial effects, can be found in the following references: Figure 1 Steps S101 to S104 in the process will not be described again.

[0090] Optionally, the noise reduction unit 104 is also used for: Obtain the depth sequence corresponding to the logging data to be restored and the time step sequence with added Gaussian noise; Based on the logging data to be restored, the depth sequence, and the complete logging data, a comprehensive condition tensor is constructed; and each time step in the noisy time step sequence is encoded to obtain the feature data corresponding to each time step, and the whole is used as the time step feature data. The time step feature data, the hybrid tensor, and the comprehensive conditional tensor are input into the trained generative deep learning model to perform noise prediction, and noise prediction data is obtained. Noise-inducing logging data is processed based on noise prediction data to obtain complete logging data.

[0091] Optionally, the noise reduction unit 104 is also used for: Determine the category identifier corresponding to the logging curve in the logging data to be restored; Feature projection is performed on the complete logging data to obtain target feature data; depth sequences are encoded to obtain corresponding depth encoded data; and corresponding curve category feature data are determined based on category identifiers. By concatenating the target feature data, deep encoding data, and curve category feature data, a comprehensive conditional tensor is obtained.

[0092] Optionally, the noise reduction unit 104 is also used for: The complete logging data is input into the convolutional layer so that the convolutional layer maps the complete logging data into target feature data with a set feature dimension, thereby achieving feature dimension compression and alignment and extracting nonlinear combination features between different logging parameters in the complete logging data; The depth sequence sequentially lists each depth from the logging data to be restored or the complete logging data; the denoising unit 104 is also used for: A fixed-position encoding strategy based on trigonometric functions and an embedded dimension index are used to encode each depth in the depth sequence to obtain the corresponding depth-encoded data. The noise reduction unit 104 is also used for: The category identifier is input into the constructed embedding layer so that the embedding layer can look up the curve category feature data corresponding to the category identifier in the created category feature dataset.

[0093] Optionally, the generative deep learning model includes a first processing layer, a residual network, and a second processing layer; The noise reduction unit 104 is also used for: The hybrid tensor is input into the first processing layer for convolution and activation to obtain the processed data; The processed data, the integrated condition tensor, and the time step feature data are input into the residual network for deep feature extraction to obtain the feature extraction data. The extracted feature data is input into the second processing layer for convolution and activation to obtain noise prediction data.

[0094] Optionally, the residual network includes n stacked residual layers and a fusion module set between two adjacent residual layers, where n is an integer greater than 1; The noise reduction unit 104 is also used for: The processed data, the comprehensive condition tensor, and the time step feature data are input into the first residual layer for feature extraction to obtain the first output data and the second output data of the first residual layer; the first output data of the first residual layer and the processed data are input into the first fusion module for addition to obtain the first added data; The first summed data, the comprehensive condition tensor, and the time step feature data are input into the second residual layer for feature extraction to obtain the first output data and the second output data of the second residual layer; the first output data of the second residual layer and the first summed data are input into the second fusion module for addition to obtain the second summed data; The second summed data, the comprehensive condition tensor, and the time step feature data are input into the third residual layer for feature extraction until the second output data of the last residual layer is obtained. The second output data from each residual layer is summed to obtain the target summed data. Numerical scaling is applied to the target summation data to obtain feature extraction data.

[0095] Optionally, each residual layer includes a depth-direction feature extraction module and a curve-related feature extraction module; The first output data, the comprehensive condition tensor, and the time step feature data of the previous residual layer are input into the current residual layer for feature extraction, resulting in the first and second output data of the current residual layer, which are set as follows: The first output data of the previous residual layer, the comprehensive condition tensor and the time step feature data are input into the depth direction feature extraction module for feature extraction to obtain the depth direction feature data. The depth-direction feature data is input into the curve correlation feature extraction module for feature extraction, and the first output data and the second output data of the current residual layer are obtained.

[0096] Optionally, the depth-direction feature extraction module includes a splicing submodule, a first fusion submodule, a Transformer encoder, a first sub-convolution module, a second sub-convolution module, and a second fusion submodule; The first output data of the previous residual layer, the comprehensive condition tensor, and the time step feature data are input into the depth direction feature extraction module for feature extraction, resulting in depth direction feature data, which is set as follows: Each feature data in the time step feature data is input into the splicing submodule for splicing to obtain spliced ​​feature data; The first output data of the previous residual layer and the spliced ​​feature data are input into the first fusion submodule for addition to obtain the first fused data; The first fused data is input into the Transformer encoder to perform self-attention calculation in the time dimension, and the first encoded data is obtained. The first encoded data is input into the first sub-convolution module for convolution to obtain the first sub-convolution data; the comprehensive condition tensor is input into the second sub-convolution module for convolution to obtain the second sub-convolution data. The first and second sub-convolutional data are input into the second fusion submodule and added together to obtain the second fused data, which is used as the depth-direction feature data.

[0097] Optionally, the curve-related feature extraction module includes a third sub-convolution module, a fourth sub-convolution module, and a fifth sub-convolution module; The depth-direction feature data is input into the curve correlation feature extraction module for feature extraction, resulting in the first and second output data of the current residual layer, which are set as follows: The depth-oriented feature data is input into the third sub-convolution module for convolution to obtain the third sub-convolution data; The third sub-convolution data is input into the fourth and fifth sub-convolution modules respectively for convolution to obtain the fourth and fifth sub-convolution data, which are then used as the first and second output data of the current residual layer, respectively.

[0098] The well logging data completion device based on a generative deep learning model proposed in this embodiment can progressively add Gaussian noise to the missing well logging data to be restored, obtaining noisy well logging data. The well logging data to be restored includes known well logging data and missing well logging data, while the noisy well logging data includes known well logging data and noisy data. Based on a created observation mask, zeros are filled into the missing well logging data to be restored and the known well logging data of the noisy well logging data, respectively, to obtain first mask data and second mask data. The first mask data and second mask data are concatenated to construct a hybrid tensor. Based on the complete logging data corresponding to the well logging data to be restored, the hybrid tensor, and the trained generative deep learning model, the noisy well logging data is denoised to obtain the completed well logging data. This embodiment constructs a hybrid tensor based on the well logging data to be restored and the noisy logging data, and uses the hybrid tensor, complete logging data, and generative deep learning model to achieve well logging data completion, which can effectively improve the accuracy of well logging data completion.

[0099] In this embodiment, the well logging data completion device based on a generative deep learning model is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0100] This invention also provides a computer device having the above-described features. Figure 6 The image shows a well logging data completion device based on a generative deep learning model.

[0101] Please see Figure 7 The present invention provides a schematic diagram of the structure of a computer device according to an optional embodiment. The computer device includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some optional embodiments, multiple processors and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 7 Take a processor 10 as an example.

[0102] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0103] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0104] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function. The data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0105] Memory 20 may include volatile memory, such as random access memory. Memory may also include non-volatile memory, such as flash memory, hard disk, or solid-state drive. Memory 20 may also include combinations of the above types of memory.

[0106] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0107] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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; and these 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 the present invention.

Claims

1. A well logging data completion method based on a generative deep learning model, characterized in that, include: Gaussian noise is gradually added to the missing logging data to be restored, resulting in noisy logging data; wherein, the logging data to be restored includes known logging data and the missing logging data, and the noisy logging data includes the known logging data and noise data; Based on the created observation mask, 0 is filled into the missing logging data of the logging data to be restored and the known logging data of the noisy logging data, respectively, to obtain the first mask data and the second mask data; The first mask data and the second mask data are concatenated to construct a hybrid tensor; Based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model, the noisy logging data is denoised to obtain complete logging data.

2. The method according to claim 1, characterized in that, The noisy logging data is denoised based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model to obtain complete logging data, including: Obtain the depth sequence corresponding to the well logging data to be restored and the noise-added time step sequence corresponding to the Gaussian noise; Based on the logging data to be restored, the depth sequence, and the complete logging data, a comprehensive condition tensor is constructed; and each time step in the noisy time step sequence is encoded to obtain the feature data corresponding to each time step, and the whole is used as the feature data of the time step. The time step feature data, the hybrid tensor, and the comprehensive conditional tensor are input into the trained generative deep learning model to perform noise prediction, thereby obtaining noise prediction data. The noisy logging data is denoised based on the noise prediction data to obtain complete logging data.

3. The method according to claim 1, characterized in that, The step of constructing a comprehensive condition tensor based on the logging data to be restored, the depth sequence, and the complete logging data includes: Determine the category identifier corresponding to the logging curve in the logging data to be restored; The complete logging data is subjected to feature projection to obtain target feature data; the depth sequence is encoded to obtain corresponding depth encoded data; and the corresponding curve category feature data is determined according to the category identifier. The target feature data, the depth encoding data, and the curve category feature data are concatenated to obtain the comprehensive conditional tensor.

4. The method according to claim 3, characterized in that, The step of performing feature projection on the complete logging data to obtain target feature data includes: The complete logging data is input into a convolutional layer so that the convolutional layer maps the complete logging data into target feature data with a set feature dimension, thereby achieving feature dimension compression and alignment and extracting nonlinear combination features between different logging parameters in the complete logging data; The depth sequence contains each depth from the logging data to be restored or the complete logging data, arranged sequentially; the encoding of the depth sequence to obtain corresponding depth-coded data includes: A fixed-position encoding strategy based on trigonometric functions and an embedded dimension index are used to encode each depth in the depth sequence to obtain the corresponding depth-encoded data. The step of determining the corresponding curve category feature data based on the category identifier includes: The category identifier is input into the constructed embedding layer so that the embedding layer can find the curve category feature data corresponding to the category identifier in the created category feature dataset.

5. The method according to claim 2, characterized in that, The generative deep learning model includes a first processing layer, a residual network, and a second processing layer. The step of inputting the time-step feature data, the hybrid tensor, and the comprehensive conditional tensor into a trained generative deep learning model for noise prediction, to obtain noise prediction data, includes: The hybrid tensor is input into the first processing layer for convolution and activation to obtain the processed data; The processed data, the integrated condition tensor, and the time step feature data are input into the residual network for deep feature extraction to obtain feature extraction data. The extracted feature data is input into the second processing layer for convolution and activation to obtain the noise prediction data.

6. The method according to claim 5, characterized in that, The residual network includes n stacked residual layers and a fusion module set between two adjacent residual layers, where n is an integer greater than 1; The process involves inputting the processed data, the comprehensive conditional tensor, and the time-step feature data into the residual network for deep feature extraction, resulting in feature extraction data, including: The processed data, the comprehensive condition tensor, and the time step feature data are input into the first residual layer for feature extraction to obtain the first output data and the second output data of the first residual layer; the first output data of the first residual layer and the processed data are input into the first fusion module for addition to obtain the first added data; The first summed data, the comprehensive condition tensor, and the time step feature data are input into the second residual layer for feature extraction to obtain the first output data and the second output data of the second residual layer; the first output data of the second residual layer and the first summed data are input into the second fusion module for addition to obtain the second summed data; The second summed data, the comprehensive condition tensor, and the time step feature data are input into the third residual layer for feature extraction until the second output data of the last residual layer is obtained. The second output data from each residual layer are summed to obtain the target sum data; The target sum data is numerically scaled to obtain the feature extraction data.

7. The method according to claim 6, characterized in that, Each residual layer contains a depth direction feature extraction module and a curve correlation feature extraction module; The first output data of the previous residual layer, the comprehensive condition tensor, and the time step feature data are input into the current residual layer for feature extraction, resulting in the first and second output data of the current residual layer, including: The first output data of the previous residual layer, the comprehensive condition tensor, and the time step feature data are input into the depth direction feature extraction module for feature extraction to obtain depth direction feature data. The depth-direction feature data is input into the curve-related feature extraction module for feature extraction to obtain the first output data and the second output data of the current residual layer.

8. The method according to claim 7, characterized in that, The depth direction feature extraction module includes a splicing submodule, a first fusion submodule, a Transformer encoder, a first sub-convolution module, a second sub-convolution module, and a second fusion submodule; The first output data of the previous residual layer, the comprehensive conditional tensor, and the time step feature data are input into the depth direction feature extraction module for feature extraction to obtain depth direction feature data, including: Each feature data in the time step feature data is input into the splicing submodule for splicing to obtain spliced ​​feature data; The first output data of the previous residual layer and the spliced ​​feature data are input into the first fusion submodule for addition to obtain the first fused data; The first fused data is input into the Transformer encoder to perform self-attention calculation in the time dimension, thereby obtaining the first encoded data; The first encoded data is input into the first sub-convolution module for convolution to obtain the first sub-convolution data; the comprehensive condition tensor is input into the second sub-convolution module for convolution to obtain the second sub-convolution data. The first sub-convolutional data and the second sub-convolutional data are input into the second fusion submodule and added together to obtain the second fused data, which is used as the depth direction feature data.

9. The method according to claim 7, characterized in that, The curve-related feature extraction module includes a third sub-convolution module, a fourth sub-convolution module, and a fifth sub-convolution module; The step of inputting the depth-direction feature data into the curve-related feature extraction module for feature extraction to obtain the first output data and the second output data of the current residual layer includes: The depth-direction feature data is input into the third sub-convolution module for convolution to obtain the third sub-convolution data; The third sub-convolutional data is input into the fourth and fifth sub-convolutional modules respectively for convolution to obtain the fourth and fifth sub-convolutional data, which are then used as the first and second output data of the current residual layer, respectively.

10. A well logging data completion device based on a generative deep learning model, characterized in that, include: A noise-adding unit is used to progressively add Gaussian noise to the missing logging data in the logging data to be restored, to obtain noisy logging data; wherein, the logging data to be restored includes known logging data and the missing logging data, and the noisy logging data includes the known logging data and noise data; A masking unit is used to fill 0 in the missing logging data of the logging data to be restored and the known logging data of the noisy logging data, respectively, based on the created observation mask, to obtain the first masking data and the second masking data; A splicing unit is used to splice the first mask data and the second mask data to construct a hybrid tensor; The denoising unit is used to denoise the noisy logging data based on the complete logging data corresponding to the logging data to be restored, the hybrid tensor, and the trained generative deep learning model, so as to obtain complete logging data.

Images