AI large model construction method and reconstruction method for well logging curve reconstruction, device, electronic equipment and storage medium

Abstract

The present application relates to a geophysical logging technology field, and is an AI large model construction method and reconstruction method and device for logging curve reconstruction, electronic equipment and storage medium, including an LB-DALR model, a specific curve trend semantic embedding preprocessing unit acquires a numerical matrix and a semantic embedding vector, a feature extraction and fusion unit extracts numerical features from a numerical path, extracts trend semantic features from a semantic path, and obtains multi-scale fusion features after fusion; and a decoding reconstruction unit reconstructs a logging curve. The present application solves the problem of large deviation of reconstructed curves when a single modal model is used in the process of geological logging curve reconstruction, and when complex geological logging data and data disturbed by a harsh downhole environment are faced, and numerical features and trend semantic features are extracted from a numerical path and a semantic path running in parallel, and the synergistic effect of a CrossModal multi-head cross attention mechanism and a FusionGate self-correlation dynamic gating realizes high-precision reconstruction.

Description

Technical Field

[0001] This invention relates to the field of geophysical logging technology, specifically to an AI large-scale model construction method, reconstruction method, device, electronic equipment, and storage medium for logging curve reconstruction. Background Technology

[0002] Geophysical logging is the "retina" of oil and gas exploration and development. It evaluates reservoir hydrocarbon potential, porosity, and permeability by measuring the physical properties of formation rocks (such as electrical, acoustic, and radioactive properties). The completeness of logging data directly affects the discovery and assessment of oil and gas reservoirs. Among these, spontaneous potential (SP), natural gamma ray (GR), acoustic transit time (AC), compensated neutron (CNL), lithological density (DEN), and deep lateral resistivity (RT) are core logging parameters. However, in actual drilling and logging operations, the downhole environment is extremely harsh (high temperature, high pressure, wellbore collapse), often leading to missing data or signal distortion in some logging curves. Specifically, this includes:

[0003] (1) Failure of reservoir lateral correlation: If key control wells lack characteristic curves such as natural gamma (GR), geological experts will find it difficult to identify oil and gas layers and may mistakenly classify non-target layers as effective reservoirs, resulting in geological stratification errors.

[0004] (2) Oil and gas reservoir reserve assessment bias: If deep lateral resistivity (RT) data is missing near the oil-water interface, it will be impossible to effectively distinguish between low-resistivity oil layers and water layers, which directly leads to the bias in the initial geological reserves (OOIP) calculation.

[0005] Existing well logging curve reconstruction methods include mathematical interpolation, empirical modeling, multivariate fitting, and machine learning-based methods. While mathematical interpolation can achieve data continuity and resampling, improving interpretation accuracy, it is sensitive to outliers and noise, and performs poorly in areas of drastic local changes. Empirical modeling relies on manual experience to select logging curves from nearby wells to fit the current well's data, placing high demands on neighboring wells and making operation difficult. Multivariate fitting uses nonlinear fitting based on data from points at the same depth, ignoring the complex and highly heterogeneous nature of underground formations, greatly simplifying the actual formation conditions and thus resulting in poor prediction performance, failing to meet the accuracy requirements of well logging interpretation and reservoir characterization. Machine learning-based methods are mostly single-modal models represented by CNNs, RNNs, and the standard Transformer, for example:

[0006] Existing patent document CN120539818A discloses a method for reconstructing acoustic logging curves using a bidirectional recurrent neural network. The reconstruction method includes: Step S1: acquiring data and performing sequence segmentation; Step S2: normalizing the logging curves; Step S3: superimposing and training bidirectional recurrent neural network blocks; Step S4: model testing and parameter saving; Step S5: reading the model parameter file, reloading the model, importing the logging data to be reconstructed, generating a predicted curve, and saving and exporting the results. This method employs a bidirectional recurrent neural network algorithm based on deep learning in artificial intelligence, effectively capturing the bidirectional depth sequence information contained in the logging curves, improving the accuracy of curve reconstruction, reconstructing undetected, missing, and low-quality curves, and fusing multi-curve information to enhance the sensitivity of the reconstructed acoustic curves. It improves the oilfield logging database and has the advantages of low cost, high efficiency, and high accuracy.

[0007] Existing patent document two, publication number CN121299769B, discloses a multi-scale bi-branch neural network logging curve reconstruction method based on seismic constraints. This method includes reading in-well seismic and stratigraphic data, layering and facies zone delineation of target segments based on 3D seismic interpretation, constructing a pyramid-shaped multi-scale training set sample by fusing multiple curves and well-side seismic trace information, establishing a residual-Transformer bi-branch neural network training model with a composite loss function as the core, employing a three-stage progressive training strategy, and reconstructing missing segments of the target well and evaluating the results. This method maintains high stability and robustness even under complex heterogeneous reservoir conditions; by integrating geological layering and facies zone delineation interpretation information on the basis of seismic data constraints, the reconstruction results are more consistent with geological laws; the obtained reconstructed curves provide reliable support for the reinterpretation of old data and the full utilization of incomplete logging data.

[0008] Machine learning-based methods still suffer from insurmountable logical flaws and performance bottlenecks when dealing with complex geological nonlinear mapping problems, specifically including:

[0009] The poor perception of geological sedimentation and lithological abrupt changes by a single numerical mode: Existing reconstruction models mainly rely on a single numerical mode. Limited by the local receptive field (CNN) or the finite time step memory (LSTM), such models often fall into the trap of "local numerical overfitting." They can capture local numerical information but ignore the global stratigraphic sedimentary patterns in the well logging curves, as well as the adverse operations and recording interruptions that may occur in actual operation. The resulting reconstruction curves are of low quality and may even show irregular oscillations in the predicted curves.

[0010] The vulnerability of non-stationary, non-Gaussian distributed data: Well logging data are typically nonlinear, non-stationary curve sequences accompanied by abrupt changes. Abrupt changes in formation properties at stratigraphic interfaces lead to a severe long-tailed distribution in the data distribution. Conventional models typically use mean squared error (MSE) as the loss function, which mathematically assumes that the error follows a Gaussian distribution, forcing the model to tend to output a conservative "average" to reduce the overall loss. As a result, the model experiences severe "peak shaving and valley filling" at key outlier points, directly affecting the accuracy of reservoir calculations and leading to misjudgments of reservoirs or large deviations in reserve assessments.

[0011] The transformative opportunity brought by Large Language Models (LLM): The emergence of large language models (such as DeepSeek and the GPT series) provides a completely new paradigm for solving the above-mentioned problems. LLM not only learns the grammar of human language, but also internalizes general knowledge logic, including geology and physics, through massive text pre-training. Through appropriate Prompt engineering, well logging data are transformed into textual descriptions containing row numbers, numerical sequences, and rates of change. Like a seasoned geologist, LLM can numerically examine the trend and shape of curves, identifying geological sedimentary patterns that are difficult for single-modal models to capture.

[0012] In summary, the existing technology lacks a multimodal perception architecture that can not only accurately extract data details from logging curves through numerical paths, but also grasp the sedimentary patterns of formations by utilizing the trend information of the curves, and eliminate the semantic gap between different modes through dynamic interaction mechanisms, so as to achieve high-precision and highly interpretable dynamic reconstruction of logging curves. Summary of the Invention

[0013] This invention provides an AI large model construction method, reconstruction method, device, electronic device and storage medium for well logging curve reconstruction, which overcomes the shortcomings of the above-mentioned prior art. It can effectively solve the problem of large curve deviation in the reconstruction process of existing single-modal models when facing complex geological well logging data and data interfered with by harsh downhole environments.

[0014] One of the technical solutions of this invention is achieved through the following measures: a method for constructing a large AI model for well logging curve reconstruction, comprising:

[0015] Multiple samples were acquired and divided into training and testing sets according to a ratio. Each sample included the original logging data sequence of historical wells and the label information of the corresponding logging curves.

[0016] The LB-DALR model is trained using a training set. Training ends when the training stopping condition is met, resulting in a well logging curve reconstruction model. The LB-DALR model includes:

[0017] The curve trend semantic embedding preprocessing unit preprocesses the original logging data sequence to obtain the corresponding numerical matrix and transforms the original logging data sequence into a semantic embedding vector that reflects the overall trend of the logging curve.

[0018] The feature extraction and fusion unit is equipped with a numerical path and a semantic path. The numerical path introduces a TransformerEncoder to extract numerical features from the numerical matrix. The semantic path is equipped with a semantic alignment mechanism based on a large language model to obtain trend semantic features rich in curve trend information. The numerical features and trend semantic features are then fused to obtain multi-scale fused features.

[0019] The decoding and reconstruction unit introduces a Transformer Decoder to reconstruct logging curves based on multi-scale fusion features;

[0020] The trained logging curve reconstruction model was tested using a test set, the model parameters were optimized, and a logging curve reconstruction model that meets the test evaluation requirements was output.

[0021] The following are further optimizations and / or improvements to the above-mentioned technical solution:

[0022] The aforementioned curve trend semantic embedding preprocessing unit includes:

[0023] Load the original well logging data sequence, segment it using an overlapping sliding window strategy, and generate a numerical matrix. Each type of well logging data was independently normalized to eliminate the magnitude differences between different geological parameters. , R is the number of input feature channels, R is the set of real numbers, and L is the window length.

[0024] The Trends indicator within the window will index the row number range. The normalized raw logging data sequence and trend index Trends are filled into the preset Prompt template to obtain the text prompt words. The trend index Trends is obtained by calculating the difference between the mean at the end of the window data and the mean at the beginning.

[0025] The text prompts are input into the pre-trained DeepSeek model, and the hidden state vectors of the last layer of the DeepSeek model at the corresponding token positions are extracted as semantic embedding vectors that reflect the overall trend of the logging curve.

[0026] The aforementioned feature extraction and fusion unit includes:

[0027] The dual-path collaborative representation module includes:

[0028] (1) Numerical path TS-Path, including:

[0029] Decoupling normalization: using the standard normalization module to normalize the input numerical matrix. Standardize;

[0030] Feature dimension mapping: Normalized numerical features are mapped to a preset hidden layer dimension by a linear layer to represent the original number of feature channels.

[0031] The encoding logic uses a Transformer Encoder network structure to receive the numerical matrix after feature dimension mapping, capture local numerical information features, and output numerical features.

[0032] (2) Semantic Path (LLM-Path), including:

[0033] Semantic source introduction: Carrying and transmitting semantic embedding vectors generated by the DeepSeek model;

[0034] Based on the semantic alignment mechanism of the large language model, the semantic embedding vector and numerical features are aligned through cascaded mapping and spatial transpose, and the aligned semantic feature tensor is output.

[0035] Curve trend semantic representation generation: The aligned semantic feature tensor is fed into the encoder to generate trend semantic features rich in curve trend information.

[0036] The cross-modal interactive alignment and autocorrelation dynamic gating fusion module fuses numerical features and trend semantic features to obtain multi-scale fused features.

[0037] The aforementioned semantic alignment mechanisms based on large language models include:

[0038] Dimensionality reduction: Using a dimensionality reduction operator composed of a multilayer perceptron, the semantic embedding vector is nonlinearly mapped to the target dimension;

[0039] Transpose and Projection: Perform a spatial transpose operation on the dimensionality-reduced semantic embedding vector and use a projector to map it to a unified hidden layer dimension to generate an aligned semantic feature tensor.

[0040] The aforementioned cross-modal interaction alignment and autocorrelation dynamic gating fusion module includes:

[0041] The CrossModal retrieval module uses numerical features as queries and aligned trend semantic features as keys and values. It calculates the association weights between numerical information and curve trends through multi-head cross attention to generate interactive features.

[0042] FusionGate autocorrelation dynamic gating introduces autocorrelation confidence weights. The proportion of interactive features introduced is dynamically adjusted, and weighted fusion is performed based on an adaptive fusion formula to obtain multi-scale fused features, where:

[0043]

[0044] in, For numerical features, For interactive features; For linear layers; Flatten is the flattening operation; This is the Sigmoid activation function.

[0045] The aforementioned decoding and reconstruction unit includes:

[0046] The Transformer Decoder module is used to smoothly decode the fused multi-scale features;

[0047] The high-dimensional features are compressed back into the one-dimensional target space by the output projection layer to obtain the normalized prediction value;

[0048] Execute an independent inverse normalization strategy: Utilize the label information statistics of the samples in the training set, and output the reconstructed well logging curves through a linear transformation, where the linear transformation process is as follows:

[0049]

[0050] in, The final output of the denormalization is the reconstructed logging curve; These are normalized predicted values; and All are statistical quantities. The mean, The standard deviation is denoted as .

[0051] The second technical solution of the present invention is achieved through the following measures: a well logging curve reconstruction method, comprising:

[0052] Obtain the logging data sequence of the target well;

[0053] The logging data sequence of the target well is output to the logging curve reconstruction model to obtain the reconstructed logging curve. The logging curve reconstruction model is constructed using the AI ​​large model construction method for logging curve reconstruction.

[0054] The third technical solution of the present invention is achieved through the following measures: an AI large model construction device for well logging curve reconstruction, comprising:

[0055] The sample acquisition module acquires multiple samples and divides them into training and test sets according to a ratio. Each sample includes the original logging data sequence of historical wells and the label information of the corresponding logging curve.

[0056] The training module trains the pre-defined LB-DALR model using the training set. Training terminates when the training stopping condition is met, resulting in a well logging curve reconstruction model. The LB-DALR model includes:

[0057] The curve trend semantic embedding preprocessing unit preprocesses the original logging data sequence to obtain the corresponding numerical matrix and transforms the original logging data sequence into a semantic embedding vector that reflects the overall trend of the logging curve.

[0058] The feature extraction and fusion unit is equipped with a numerical path and a semantic path. The numerical path introduces a TransformerEncoder to extract numerical features from the numerical matrix. The semantic path is equipped with a semantic alignment mechanism based on a large language model to obtain trend semantic features rich in curve trend information. The numerical features and trend semantic features are then fused to obtain multi-scale fused features.

[0059] The decoding and reconstruction unit introduces a Transformer Decoder to reconstruct logging curves based on multi-scale fusion features;

[0060] The testing module uses a test set to test the trained well logging curve reconstruction model, optimize the model parameters, and output a well logging curve reconstruction model that meets the test evaluation requirements.

[0061] The fourth technical solution of the present invention is achieved through the following measures: a well logging curve reconstruction device, comprising:

[0062] The data acquisition unit acquires the logging data sequence of the target well;

[0063] The reconstruction unit outputs the logging data sequence of the target well to the logging curve reconstruction model to obtain the reconstructed logging curve. The logging curve reconstruction model is constructed using an AI large model construction method for logging curve reconstruction.

[0064] The fifth technical solution of the present invention is achieved through the following measures: an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the computer program is loaded and executed by the processor to implement the steps in the AI ​​large model construction method or the well logging curve reconstruction method for well logging curve reconstruction.

[0065] The sixth technical solution of the present invention is achieved by the following measures: a storage medium storing a computer program that can be read by a computer, the computer program being configured to execute steps in the AI ​​large model construction method or the well logging curve reconstruction method at runtime.

[0066] This invention addresses the problem of large curve deviations in the reconstruction of geological logging curves using single-modal models when dealing with complex geological logging data and data affected by harsh downhole environments. Specifically, it includes:

[0067] Multi-scale feature extraction, based on the DeepSeek model, extracts curve values ​​and semantic embedding vectors reflecting the overall trend of well logging curves using a curve trend semantic embedding preprocessing module. Through the dual-path collaborative representation module (HDP-CA), parallel numerical path (TS-Path) and semantic path (LLM-Path) are proposed. TS-Path employs a Transformer Encoder structure, focusing on capturing local numerical features from the normalized well logging numerical sequence. LLM-Path, on the other hand, collaboratively represents local details and curve trends in well logging data, overcoming the limitation of single-modality models in simultaneously considering local numerical information and global curve trends. Furthermore, the semantic alignment mechanism based on a large language model (LB-TSMA) solves the dimensionality alignment problem between the natural language representation of the large language model and the original numerical representation, projecting the high-dimensional trend semantic vectors extracted by DeepSeek onto a feature space consistent with the numerical features, providing a foundation for accurate trend fitting.

[0068] High-precision reconstruction, through the CrossModal multi-head cross-attention mechanism, aligns numerical information with curve trend semantics. This allows the model to "correct" errors caused by adverse environments or noise interference by using trend semantics generated by LLM, avoiding the irregular oscillations common in traditional regression models and outputting a more stable and physically logical reconstructed curve.

[0069] Strong robustness: The synergistic effect of the CrossModal multi-head attention mechanism and the FusionGate autocorrelation dynamic gating significantly improves the model's performance under complex perturbation conditions, and it has strong environmental adaptability. Attached Figure Description

[0070] Appendix Figure 1 This is a schematic diagram of the implementation environment provided for an embodiment of the present invention.

[0071] Appendix Figure 2 This is a schematic diagram of the AI ​​large model construction method for well logging curve reconstruction provided in an embodiment of the present invention.

[0072] Appendix Figure 3 A schematic diagram of the network structure of the curve trend semantic embedding preprocessing unit provided in an embodiment of the present invention.

[0073] Appendix Figure 4 This is a schematic diagram of the network structure of the dual-path collaborative representation module provided in an embodiment of the present invention.

[0074] Appendix Figure 5 This is a schematic diagram of the network structure of the cross-modal interactive alignment and autocorrelation dynamic gating fusion module and the decoding and reconstruction unit provided in an embodiment of the present invention.

[0075] Appendix Figure 6 This is a schematic diagram of the logging curve reconstruction method provided in an embodiment of the present invention.

[0076] Appendix Figure 7 A visualization diagram showing the comparison between the actual and predicted values ​​of the mh136 blind well provided in an embodiment of the present invention.

[0077] Appendix Figure 8 The test well diagram of the mh136 blind well provided in the embodiment of the present invention.

[0078] Appendix Figure 9 A schematic diagram of the structure of an AI large model construction device for well logging curve reconstruction provided in an embodiment of the present invention.

[0079] Appendix Figure 10 This is a schematic diagram of the logging curve reconstruction device provided in an embodiment of the present invention. Detailed Implementation

[0080] The present invention is not limited to the following embodiments, and the specific implementation can be determined according to the technical solution of the present invention and the actual situation.

[0081] Those skilled in the art will understand that, unless otherwise stated, in the embodiments of the present invention, a "module" or "unit" refers to a computer program or part of a computer program with a predetermined function, which works together with other related parts to achieve a predetermined goal, and can be implemented, wholly or partially, using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0082] This invention provides a well logging curve reconstruction method, comprising: acquiring a well logging data sequence of a target well; outputting the well logging data sequence of the target well into a well logging curve reconstruction model to obtain a reconstructed well logging curve. The construction of the well logging curve reconstruction model includes acquiring multiple samples and dividing them into a training set and a test set according to a ratio. Each sample includes the original well logging data sequence of historical wells and the label information of the corresponding well logging curve. The training set is used to train a preset LB-DALR model. When the training stopping condition is met, the training ends to obtain the well logging curve reconstruction model. The LB-DALR model includes: a curve trend semantic embedding preprocessing unit, which preprocesses the original well logging data sequence to obtain the corresponding numerical matrix and transforms the original well logging data sequence into a semantic embedding vector reflecting the overall trend of the well logging curve; and a feature extraction and fusion unit, which sets up a numerical path and a semantic path, with the numerical path introduced by a Transformer. The Encoder extracts numerical features from the numerical matrix; the semantic path setting uses a semantic alignment mechanism based on a large language model to obtain trend semantic features rich in curve trend information; and it fuses numerical features and trend semantic features to obtain multi-scale fusion features; the decoding and reconstruction unit introduces a TransformerDecoder to reconstruct the logging curve based on the multi-scale fusion features; the trained logging curve reconstruction model is tested using a test set, the model parameters of the logging curve reconstruction model are optimized, and a logging curve reconstruction model that meets the test evaluation requirements is output.

[0083] The method provided in this embodiment of the invention may involve artificial intelligence (AI) technology and may be implemented based on artificial intelligence technology, such as using deep learning to train a corresponding model using samples.

[0084] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence.

[0085] Deep learning (DL) specifically refers to machine learning based on deep neural network models and methods. It has developed from statistical machine learning, artificial neural network algorithms, and other algorithms, combined with the advancements in big data and computing power. The most important technical feature of deep learning is its ability to automatically extract features.

[0086] The aforementioned machine learning and deep learning typically include techniques such as neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0087] In deep learning, the loss function is used to predict the target value by comparing the predicted value with the target value. This is done by updating the weight vector of each layer of the neural network based on the difference between the two values ​​(usually with an initialization process before the first update, where parameters are pre-configured for each layer) until the network can predict the target value or a value very close to it. Therefore, deep learning requires pre-defining "how to compare the difference between the predicted value and the target value," which is the loss function.

[0088] As attached Figure 1 The diagram illustrates an implementation environment provided by an embodiment of the present invention. This implementation environment may include: training equipment and usage equipment.

[0089] Both the training equipment and the equipment used are computer devices; optionally, the computer device is a terminal device, such as a mobile phone, tablet computer, PC (Personal Computer) or other electronic devices; or, the computer device is a server, which can be a single server, a server cluster composed of multiple servers, or a cloud computing service center. This embodiment of the invention does not limit this.

[0090] Training equipment refers to computer equipment capable of training and learning neural networks. Optionally, the training equipment has the ability to acquire neural networks and train and learn them according to application requirements. For example, the training equipment acquires neural networks from other devices through a network and then trains them with training samples according to application requirements, so that the neural network has the ability to reconstruct well logging curves. Optionally, the training equipment has the ability to build neural networks. It can build neural networks itself according to application requirements and then train and learn them. For example, in order to realize well logging curve reconstruction based on well logging data, the training equipment builds its own neural network and then trains and learns it with samples according to application requirements.

[0091] The device used refers to a computer device that has the requirement to use a neural network. Optionally, the device uses a neural network from other devices through the network according to the application requirements. For example, if the device has the requirement to reconstruct well logging curves, it can use the network to obtain a neural network that has been trained and learned to reconstruct well logging curves from other devices, and use the neural network to reconstruct well logging curves.

[0092] Based on this, the technical solution of the present invention will be described and explained below with reference to several examples.

[0093] Example 1: As shown in the attached document Figure 2 As shown, this invention discloses a method for constructing a large AI model for well logging curve reconstruction, including:

[0094] Step S110: Obtain multiple samples and divide them into training set and test set according to the ratio. Each sample includes the original logging data sequence of historical wells and the label information of the corresponding logging curve.

[0095] Step S120: Train the preset LB-DALR model using the training set. When the training stopping condition is met, end the training to obtain the well logging curve reconstruction model, wherein the LB-DALR model includes:

[0096] The curve trend semantic embedding preprocessing unit preprocesses the original logging data sequence to obtain the corresponding numerical matrix and transforms the original logging data sequence into a semantic embedding vector that reflects the overall trend of the logging curve.

[0097] The feature extraction and fusion unit is equipped with a numerical path and a semantic path. The numerical path introduces a Transformer Encoder to extract numerical features from the numerical matrix. The semantic path is equipped with a semantic alignment mechanism based on a large language model to obtain trend semantic features rich in curve trend information. The numerical features and trend semantic features are then fused to obtain multi-scale fused features.

[0098] The decoding and reconstruction unit introduces a Transformer Decoder to reconstruct logging curves based on multi-scale fusion features.

[0099] Step S130: Test the trained logging curve reconstruction model using the test set, optimize the model parameters of the logging curve reconstruction model, and output a logging curve reconstruction model that meets the test evaluation requirements.

[0100] The training stopping conditions mentioned above may include the loss function and the maximum number of iterations. The loss function can be set as needed.

[0101] This invention discloses an AI large-scale model construction method for well logging curve reconstruction. The pre-set LB-DALR model core framework is based on a dual-path collaborative network of "numerical path-semantic path". In the numerical path, the original well logging data sequence is preprocessed to obtain the corresponding numerical matrix, and a Transformer Encoder is introduced to focus on capturing local numerical features from the numerical matrix and outputting the corresponding numerical features. In the semantic path, the original well logging data sequence is transformed into a semantic embedding vector reflecting the overall trend of the well logging curve, and a semantic alignment mechanism based on a large language model is set to obtain trend semantic features rich in curve trend information. This realizes the collaborative representation of local detail information and curve trend in well logging data, overcoming the defect of single-modality models that cannot take into account both local numerical information and global curve trend. Furthermore, the numerical features and trend semantic features are fused to obtain multi-scale fusion features, realizing deep interaction and balance of dual-modal features, providing a foundation for high-precision reconstruction of well logging curves.

[0102] Example 2: This embodiment of the invention is a further optimization of the above embodiments, wherein the LB-DALR model includes:

[0103] (a) Curve Trend Semantic Embedding Preprocessing Unit (GTE), as shown in the appendix Figure 3 As shown, it includes:

[0104] Load the original well logging data sequence, segment it using an overlapping sliding window strategy, and generate a numerical matrix. Each type of well logging data was independently normalized to eliminate the magnitude differences between different geological parameters. , The input feature channel is R, which is the set of real numbers, and L is the window length. The window length and step size of the overlapping sliding window can be set according to the requirements. For example, if the window length is set to L=24 and the step size is 1, then the window length L=24 means that each window contains 24 consecutive sampling points of data features (such as natural gamma GR, acoustic time difference AC, density DEN, etc.).

[0105] The Trends indicator within the window will index the row number range. Normalized raw well logging data sequence The Trends indicator is entered into the preset Prompt template to obtain the text prompt words;

[0106] The Trends indicator is obtained by calculating the difference between the mean at the end of the window and the mean at the beginning of the window. If the value is positive, it indicates an upward trend; if it is negative, it indicates a downward trend. The absolute value reflects the degree of change.

[0107] Among them, the default prompt template can be: "You are given 24 consecutive log values ​​from row 1 to row 24. Analyze the overall trend and describe it briefly using appropriate petrophysical or geologicalterminology. Input format: Row 1:value1 Row 2: value2, ..., valuen. Output format: Trend:[concise trenddescription]".

[0108] The text prompts are input into a pre-trained DeepSeek model, and the hidden state vectors of the last layer of the DeepSeek model at the corresponding token positions are extracted as semantic embedding vectors reflecting the overall trend of the logging curve. , R is the set of real numbers, specifically:

[0109] This extracts the hidden states vector output at the end of the sequence (Token) of Layer 28 in the Transformer architecture of the DeepSeek model, rather than the final generated text token. This vector... It not only encodes the statistical features of the numerical sequence, but also enriches the semantic representation of the curve trend internalized by the model.

[0110] Furthermore, the above preprocessing process can also employ atomic storage and breakpoint resumption, specifically:

[0111] Given that well logging data is typically in the terabyte (TB) range, the generated embedding vectors are persistently managed using H5 format. To ensure the robustness of data generation, the system employs an atomic write mechanism: Atomic write: Each batch of embeddings is first written to a temporary file (.tmp), and only after passing integrity verification is it renamed to a formal file (.h5). Intelligent resume: If processing is interrupted, upon restart, an index verification algorithm automatically compares the already generated H5 files, skipping processed windows, thus achieving efficient breakpoint resume for large-scale pre-generation tasks.

[0112] (ii) Feature extraction and fusion unit, including:

[0113] (1) Dual-path collaborative representation module (HDP-CA), as shown in the appendix Figure 4 As shown, it includes:

[0114] Numerical paths (TS-Path) include:

[0115] Decoupling normalization: using the standard normalization module to normalize the input numerical matrix. Standardization is performed. This operation calculates the mean and standard deviation of each physical channel along the batch dimension to ensure that the input tensor follows a standard Gaussian distribution, while not introducing statistical information from the target column to be reconstructed;

[0116] Feature dimension mapping: Normalized numerical features are processed through a linear layer to map the number of original feature channels. Mapped to a preset hidden dimension This step transforms the raw sensed values ​​into a high-dimensional abstract representation that can be extracted by a deep network;

[0117] The encoding logic employs a Transformer Encoder network structure to receive the numerical matrix after feature dimension mapping, capture local numerical information features, and output numerical features. The Transformer Encoder network structure includes a linear attention Transformer, which consists of multiple stacked self-attention blocks. R is the set of real numbers, B is the batch size, and L is the window length, i.e., the sequence length.

[0118] Semantic path LLM-Path, including:

[0119] Semantic source introduction: Carrying and transmitting semantic embedding vectors generated by the DeepSeek model ;

[0120] Trend Co-representation: Unlike TS-Path, which focuses on instantaneous fluctuations, LLM-Path provides guidance on "what trend the entire curve should follow"; this structured separation design can overcome the shortcomings of single numerical models that are prone to deviating from the true evolution law when faced with severe data noise or complex strata;

[0121] A semantic alignment mechanism based on large language models (LB-TSMA) is used to address the relationship between natural language representation (1536-dimensional hidden layer vectors) and trend numerical representation (…). To address the dimensional differences between matrices, this invention designs a dedicated LB-TSMA mechanism. This mechanism, through cascaded mapping and spatial transpose, "unfolds" static global features into serialized features with spatiotemporal structure, including:

[0122] Dimensionality reduction: Using a dimensionality reduction operator composed of a multilayer perceptron, semantic embedding vectors are transformed. (i.e., high-dimensional semantic vectors (1536 dimensions)) are non-linearly mapped to the target dimension;

[0123] Among them, the target dimension, which is the same as the length of the curve sequence, can be set to L=24;

[0124] The mathematical mapping logic is as follows:

[0125]

[0126] in, This is the semantic embedding vector after dimensionality reduction; For Gaussian error linear units, there is an excitation function; This is a linear transformation layer used to reduce the dimension of features from... Projected to the target dimension L=24.

[0127] Transpose and Projection: Performing a spatial transpose operation on the dimensionality-reduced semantic embedding vector Then, a projector is used to map them to a uniform hidden layer dimension, generating an aligned semantic feature tensor. Specifically:

[0128] Spatial transpose: The dimension of the reduced vector is transposed from (B, L) to (B, L, 1); this operation rotates the semantic information from the "feature attribute" dimension to the "spatial step" dimension, establishing a preliminary trend correspondence.

[0129] Channel-aligned projection: Using the llm_projector module, a single semantic dimension is mapped to a unified hidden dimension through a linear layer (the hidden dimension is consistent with the numerical feature and can be set to...). );

[0130]

[0131] in, For the aligned semantic feature tensor; Linear layer; This is a transpose operation; This is the semantic embedding vector after dimensionality reduction.

[0132] This step transforms each short text description into a set of dense feature maps with the same width and length, eliminating mathematical computational barriers between different modalities.

[0133] Curve trend semantic representation generation: Aligned semantic feature tensor The data is fed into the encoder llm_encoder to generate trend semantic features rich in curve trend information. .

[0134] It should be noted that, due to It originates from the curvilinear trend semantics of a large language model. At this stage, the encoder learns the distribution intensity of these "semantic units" on 24 sampling points through a self-attention mechanism, thereby generating trend semantic features rich in curvilinear trend information. .

[0135] In this embodiment, the Dual-Path Collaborative Representation Module (HDP-CA) is configured with a parallel numerical path (TS-Path) and a semantic path (LLM-Path). The TS-Path employs a Transformer Encoder structure, focusing on capturing local numerical features from the normalized well logging numerical sequences. The LLM-Path introduces a pre-trained DeepSeek language model as a semantic engine to extract the trend semantics of the curves. This design achieves collaborative representation of local details and curve trends in well logging data, overcoming the limitation of single-modality models that cannot simultaneously consider both local numerical information and global curve trends.

[0136] Furthermore, this embodiment utilizes a semantic alignment mechanism based on large language models (LB-TSMA) to address the dimensionality differences between natural language representations and numerical representations. It designs a cascaded mapping based on Dimension Reducer, Transpose, and LLM Projector to project high-dimensional (1536-dimensional) semantic vectors onto a "step-channel" space consistent with the numerical features. This mechanism aligns the discrete, unstructured trend text semantics with the original data modality.

[0137] (2) Cross-modal interactive alignment and autocorrelation dynamic gating fusion module (CMA&GF module), as shown in the appendix. Figure 5 As shown, it includes:

[0138] CrossModal retrieval module: based on numerical features For querying, use aligned trend semantic features. Using key and value as keys, multi-head cross-attention is used to calculate the weights of the numerical information-curve trend association, generating interaction features. ;

[0139]

[0140] in, For attention mechanisms; Numerical features The corresponding query matrix; For aligned trend semantic features The corresponding key matrix; For aligned trend semantic features The corresponding value matrix; The Softmax function normalizes the attention scores into a probability distribution, representing the attention weight of each query position to different key positions. To query the projection matrix; The key projection matrix; The projection matrix is ​​a value. The dimension for the key / query.

[0141] The formula calculates each depth step. The matching weight with the trend of this curve segment. For example, when numerical features... When a sharp change in resistivity is detected at a certain depth point, the model will automatically use an attention mechanism to focus on the trend semantic features. The system retrieves the corresponding curve trend semantics and injects the most relevant semantic features into the current point, thereby achieving accurate alignment between numerical information and trend semantics.

[0142] The CrossModal search described above aims to dynamically retrieve the most matching curve trend context, using the current actual numerical characteristics as anchors.

[0143] FusionGate Autocorrelation Dynamic Gating: Considering the different reconstruction difficulties in different geological strata, this invention introduces the FusionGate mechanism to dynamically adjust the proportion of interactive features introduced. Specifically:

[0144] Autocorrelation confidence calculation: This involves calculating the numerical features... Perform Flatten flattening and linear layer mapping to assess the completeness and volatility of the current numerical sequence and generate a global confidence scalar.

[0145] Dynamic weight generation: The global confidence scalar is mapped to the (0, 1) interval using the Sigmoid activation function to obtain the autocorrelation confidence weights. This allows for the dynamic adjustment of the contribution of semantic paths to the final reconstruction result;

[0146]

[0147] in, For linear layers; Flatten is the flattening operation; Use the Sigmoid activation function;

[0148] Dynamic complementarity: Weighted fusion is performed based on an adaptive fusion formula to obtain multi-scale fusion features. .

[0149]

[0150] in, For interactive features.

[0151] In the above formula, numerical characteristics The fundamental residual terms are always retained, ensuring the physical accuracy of the reconstructed curve. Weights The model then adapts to the sequence content. In stable depositional phases or areas with missing / noisy data, the model tends to increase... To fill and smooth out the curve trend semantics of LLM, the model is introduced to incorporate more LLM curve trend semantics; in thin interlayers or abrupt interface segments, the model tends to reduce... The values ​​are optimized to suppress excessive smoothing and prioritize the retention of numerical information captured by TS-Path. This adaptive fusion formula achieves adaptive complementarity of dual-modal features, emphasizing trend guidance in stable data segments and distinguishing lithological abrupt changes and anomalous noise in abrupt segments, accurately restoring curve information and realizing the transformation from single-modal curve reconstruction to dual-modal reconstruction.

[0152] (III) Decoding and Reconstruction Unit: Introducing the Transformer Decoder to reconstruct well logging curves based on multi-scale fusion characteristics; including:

[0153] The Transformer Decoder is used to process the fused multi-scale features. Perform smooth decoding;

[0154] The high-dimensional features are compressed back into the one-dimensional target space by the output projection layer to obtain the normalized prediction value;

[0155] Implement an independent inverse normalization strategy: utilize the label information statistics (mean) of the samples in the training set. with standard deviation The reconstructed logging curves are output through a linear transformation, the process of which is shown below:

[0156]

[0157] in, The final output of the denormalization is the reconstructed logging curve; These are normalized predicted values; All are statistical quantities. The mean, The standard deviation is denoted as .

[0158] The above can be utilized The interface pre-calculates the mean and variance from the label information of the samples in the training set.

[0159] Example 3: As shown in the attached document Figure 6 As shown, this embodiment of the invention discloses a method for reconstructing well logging curves, including:

[0160] Step S210: Obtain the logging data sequence of the target well;

[0161] Step S220: Output the logging data sequence of the target well to the logging curve reconstruction model to obtain the reconstructed logging curve. The logging curve reconstruction model is constructed using the AI ​​large model construction method for logging curve reconstruction disclosed in Examples 1 and 2.

[0162] Example 4: The well logging curve reconstruction model constructed in this invention was used in the MH136 blind well test, and the results were obtained. Figure 7 The diagram showing the comparison between the actual and predicted values ​​of the MH136 blind well is accompanied by a visual representation. Figure 8 The test well chart of the MH136 blind well shown shows that, by introducing test parameters, the AC curve can be found: R=0.87, R 2 =0.76; RT curve: R=0.91, R 2 =0.83; The test structure is excellent, so the present invention exhibits excellent fitting accuracy and physical consistency in complex lithological abrupt change sections, which meets the requirements of oil and gas exploration for the integrity of logging data and high signal-to-noise ratio, and has important application value for intelligent logging interpretation and formation evaluation.

[0163] Example 5: As shown in the attached document Figure 9 As shown, this invention discloses an AI large model construction device for well logging curve reconstruction, comprising:

[0164] The sample acquisition module acquires multiple samples and divides them into training and test sets according to a ratio. Each sample includes the original logging data sequence of historical wells and the label information of the corresponding logging curve.

[0165] The training module trains the pre-defined LB-DALR model using the training set. Training terminates when the training stopping condition is met, resulting in a well logging curve reconstruction model. The LB-DALR model includes:

[0166] The curve trend semantic embedding preprocessing unit preprocesses the original logging data sequence to obtain the corresponding numerical matrix and transforms the original logging data sequence into a semantic embedding vector that reflects the overall trend of the logging curve.

[0167] The feature extraction and fusion unit is equipped with a numerical path and a semantic path. The numerical path introduces a Transformer Encoder to extract numerical features from the numerical matrix. The semantic path is equipped with a semantic alignment mechanism based on a large language model to obtain trend semantic features rich in curve trend information. The numerical features and trend semantic features are then fused to obtain multi-scale fused features.

[0168] The decoding and reconstruction unit introduces a Transformer Decoder to reconstruct logging curves based on multi-scale fusion features.

[0169] The testing module uses a test set to test the trained well logging curve reconstruction model, optimize the model parameters, and output a well logging curve reconstruction model that meets the test evaluation requirements.

[0170] Example 6: As shown in the appendix Figure 10 As shown, an embodiment of the present invention discloses a well logging curve reconstruction device, comprising:

[0171] The data acquisition unit acquires the logging data sequence of the target well;

[0172] The reconstruction unit outputs the logging data sequence of the target well to the logging curve reconstruction model to obtain the reconstructed logging curve. The logging curve reconstruction model is constructed using an AI large model construction method for logging curve reconstruction.

[0173] Example 7: This embodiment of the invention discloses a storage medium storing a computer program that can be read by a computer. The computer program is configured to execute steps in the AI ​​large model construction method or the well logging curve reconstruction method when it runs.

[0174] The aforementioned storage media may include, but are not limited to, USB flash drives, read-only memory, portable hard drives, magnetic disks, optical disks, and other media capable of storing computer programs.

[0175] Example 8: This embodiment of the invention discloses an electronic device, including a processor and a memory. The memory stores a computer program, which is loaded and executed by the processor to implement the steps in the AI ​​large model construction method or the well logging curve reconstruction method for well logging curve reconstruction.

[0176] The processor described above can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an ASIC, an FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. It can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The memory can include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, portable hard drives, magnetic disks, or optical disks.

[0177] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented 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. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0178] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0179] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0180] The above content is only a specific embodiment of the present invention, which has strong adaptability and implementation effect. However, the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention. Therefore, equivalent changes made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for constructing a large AI model for well logging curve reconstruction, characterized in that, include: Multiple samples were acquired and divided into training and testing sets according to a ratio. Each sample included the original logging data sequence of historical wells and the label information of the corresponding logging curves. The LB-DALR model is trained using a training set. Training ends when the training stopping condition is met, resulting in a well logging curve reconstruction model. The LB-DALR model includes: The curve trend semantic embedding preprocessing unit preprocesses the original logging data sequence to obtain the corresponding numerical matrix and transforms the original logging data sequence into a semantic embedding vector that reflects the overall trend of the logging curve. The feature extraction and fusion unit includes numerical and semantic paths. The numerical path uses a TransformerEncoder to extract numerical features from the numerical matrix. The semantic path employs a semantic alignment mechanism based on a large language model to obtain trend semantic features rich in curve trend information. The numerical and trend semantic features are then fused to obtain multi-scale fused features, including: The dual-path collaborative representation module includes: (1) Numerical path TS-Path, including: Decoupling normalization: using the StandardNorm module to normalize the input numerical matrix. Standardize; Feature dimension mapping: Normalized numerical features are mapped to a preset hidden layer dimension through a linear projection layer to represent the original number of feature channels. The encoding logic uses a Transformer Encoder network structure to receive the numerical matrix after feature dimension mapping, capture local numerical information features, and output numerical features. (2) Semantic Path (LLM-Path), including: Semantic source introduction: Carrying and transmitting semantic embedding vectors generated by the DeepSeek model; Based on the semantic alignment mechanism of the large language model, the semantic embedding vector and numerical features are aligned through cascaded mapping and spatial transpose, and the aligned semantic feature tensor is output. Curve trend semantic representation generation: The aligned semantic feature tensor is fed into the encoder to generate trend semantic features rich in curve trend information. A cross-modal interactive alignment and autocorrelation dynamic gating fusion module fuses numerical features and trend semantic features to obtain multi-scale fusion features; The decoding and reconstruction unit introduces a Transformer Decoder to reconstruct logging curves based on multi-scale fusion features; The trained logging curve reconstruction model was tested using a test set, the model parameters were optimized, and a logging curve reconstruction model that meets the test evaluation requirements was output.

2. The AI ​​large model construction method for well logging curve reconstruction according to claim 1, characterized in that, The curve trend semantic embedding preprocessing unit includes: Load the original well logging data sequence, segment it using an overlapping sliding window strategy, and generate a numerical matrix. Each type of well logging data was independently normalized to eliminate the magnitude differences between different geological parameters. The number of input feature channels; The Trends indicator within the window will index the row number range. The normalized raw logging data sequence and trend index Trends are filled into the preset Prompt template to obtain the text prompt words. The trend index Trends is obtained by calculating the difference between the mean at the end of the window data and the mean at the beginning. The text prompts are input into the pre-trained DeepSeek model, and the hidden state vectors of the last layer of the DeepSeek model at the corresponding token positions are extracted as semantic embedding vectors that reflect the overall trend of the logging curve.

3. The AI ​​large model construction method for well logging curve reconstruction according to claim 1, characterized in that, Semantic alignment mechanisms based on large language models include: Dimensionality reduction: Using a dimensionality reduction operator composed of a multilayer perceptron, the semantic embedding vector is nonlinearly mapped to the target dimension; Transpose and Projection: Perform spatial transpose on the dimensionality-reduced semantic embedding vector and use a projector to map it to a unified hidden layer dimension to generate an aligned semantic feature tensor. or / and, The cross-modal interaction alignment and autocorrelation dynamic gating fusion module includes: The CrossModal retrieval module uses numerical features as queries and aligned trend semantic features as keys and values. It calculates the association weights between numerical information and curve trends through multi-head cross attention to generate interactive features. FusionGate autocorrelation dynamic gating introduces autocorrelation confidence weights. g， The proportion of interactive features introduced is dynamically adjusted, and weighted fusion is performed based on an adaptive fusion formula to obtain multi-scale fused features, where: in, For numerical features, For interactive features.

4. The AI ​​large model construction method for well logging curve reconstruction according to claim 1 or 2, characterized in that, The decoding and reconstruction unit includes: The Transformer Decoder module is used to smoothly decode the fused multi-scale features; The high-dimensional features are compressed back into the one-dimensional target space by the output projection layer to obtain the normalized prediction value; Execute an independent inverse normalization strategy: Utilize the label information statistics of the samples in the training set, and output the reconstructed well logging curves through a linear transformation, where the linear transformation process is as follows: in, The final output of the denormalization is the reconstructed logging curve; These are normalized predicted values; All are statistical quantities. The mean, The standard deviation is denoted as .

5. A method for reconstructing well logging curves, characterized in that, include: Obtain the logging data sequence of the target well; The logging data sequence of the target well is output to the logging curve reconstruction model to obtain the reconstructed logging curve, wherein the logging curve reconstruction model is constructed using the method described in any one of claims 1 to 4.

6. An AI large model construction device for well logging curve reconstruction using the method described in any one of claims 1 to 4, characterized in that, include: The sample acquisition module acquires multiple samples and divides them into training and test sets according to a ratio. Each sample includes the original logging data sequence of historical wells and the label information of the corresponding logging curve. The training module trains the pre-defined LB-DALR model using the training set. Training terminates when the training stopping condition is met, resulting in a well logging curve reconstruction model. The LB-DALR model includes: The curve trend semantic embedding preprocessing unit preprocesses the original logging data sequence to obtain the corresponding numerical matrix and transforms the original logging data sequence into a semantic embedding vector that reflects the overall trend of the logging curve. The feature extraction and fusion unit includes numerical and semantic paths. The numerical path uses a TransformerEncoder to extract numerical features from the numerical matrix. The semantic path employs a semantic alignment mechanism based on a large language model to obtain trend semantic features rich in curve trend information. The numerical and trend semantic features are then fused to obtain multi-scale fused features, including: The dual-path collaborative representation module includes: (1) Numerical path TS-Path, including: Decoupling normalization: using the StandardNorm module to normalize the input numerical matrix. Standardize; Feature dimension mapping: Normalized numerical features are mapped to a preset hidden layer dimension through a linear projection layer to represent the original number of feature channels. The encoding logic uses a Transformer Encoder network structure to receive the numerical matrix after feature dimension mapping, capture local numerical information features, and output numerical features. (2) Semantic Path (LLM-Path), including: Semantic source introduction: Carrying and transmitting semantic embedding vectors generated by the DeepSeek model; Based on the semantic alignment mechanism of the large language model, the semantic embedding vector and numerical features are aligned through cascaded mapping and spatial transpose, and the aligned semantic feature tensor is output. Curve trend semantic representation generation: The aligned semantic feature tensor is fed into the encoder to generate trend semantic features rich in curve trend information. A cross-modal interactive alignment and autocorrelation dynamic gating fusion module fuses numerical features and trend semantic features to obtain multi-scale fusion features; The decoding and reconstruction unit introduces a Transformer Decoder to reconstruct logging curves based on multi-scale fusion features; The testing module uses a test set to test the trained well logging curve reconstruction model, optimize the model parameters, and output a well logging curve reconstruction model that meets the test evaluation requirements.

7. A logging curve reconstruction device using the method described in claim 5, characterized in that, include: The data acquisition unit acquires the logging data sequence of the target well; The reconstruction unit outputs the logging data sequence of the target well to the logging curve reconstruction model to obtain the reconstructed logging curve.

8. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which is loaded and executed by the processor to implement the steps of the method as claimed in any one of claims 1 to 4 or the method as claimed in claim 5.

9. A storage medium, characterized in that, The storage medium stores a computer program that can be read by a computer, the computer program being configured to execute the steps of the method as described in any one of claims 1 to 4 or the method as described in claim 5 when it is run.

Images