Logging curve reconstruction method based on ticc clustering and markov prior

Abstract

The application discloses a well logging curve reconstruction method based on TICC clustering and Markov prior, first, a first-stage GRU network is trained by using conventional well logging curves to obtain initial reconstruction values of target well logging curves; then, a TICC local state model is trained by using real target well logging curves of a test well, and a TICC state transition matrix is counted; then, a continuous and reasonable TICC state path along a depth direction is obtained by Viterbi decoding according to TICC state likelihood of the first-stage reconstruction curves and in combination with a Markov transition matrix; on this basis, a TICC state conditional curve value Gaussian prior and a TICC state conditional Markov residual prior are respectively constructed; finally, a second-stage GRU network predicts normalized residuals according to the above-mentioned prior features, and the residuals are added back to the first-stage reconstruction values to obtain final target well logging curve reconstruction results. The application can improve the accuracy, continuity and geological rationality of the final reconstruction curves without using real target well logging curves of a test well.

Description

Technical Field

[0001] This invention relates to the field of well logging curve reconstruction technology, and in particular to a well logging curve reconstruction method based on TICC clustering and Markov prior. Background Technology

[0002] Well logging records are crucial tools for understanding subsurface rock formations and exploring oil resources, providing a wealth of stratigraphic information. Different types of logging curves reflect formation characteristics through acoustic, electrical, radiometric, and volumetric properties of the rock. A complete set of logging curves is essential for detailed geological interpretation and reservoir parameter evaluation. However, in actual well logging operations and historical data processing, factors such as logging costs, wellbore conditions, instrument malfunctions, differences in measurement schemes, and incomplete data preservation can lead to partial, complete, or even missing sections of acoustic, resistivity, density, neutron, and other logging curves. Furthermore, inconsistencies in logging parameters between different wells often hinder the direct and unified interpretation and comprehensive application of multi-well data. Therefore, reconstructing missing curves using existing logging curves is of great significance for improving the completeness and utilization efficiency of well logging data.

[0003] Existing well logging curve reconstruction methods mainly utilize regression analysis, support vector machines, random forests, neural networks, and deep learning models to establish a nonlinear mapping relationship between known well logging curves and target well logging curves. These methods can improve reconstruction accuracy to some extent, but they do not adequately utilize the local dynamic changes and continuous evolution patterns along the depth direction. This leads to problems such as local jumps, excessive smoothing, and insufficient layer continuity in reconstruction results, especially in well sections with abrupt lithological changes, heterogeneous reservoir enhancement, and thin interbedded layers.

[0004] Furthermore, well logging curves typically exhibit different local evolution states along the depth direction, such as stability, rise, fall, fluctuation, and abrupt changes. Under different local states, the numerical distribution, fluctuation amplitude, trend of change, and direction and amplitude of the prediction residuals of the well logging curves often show significant differences. Even if the curve values ​​or local morphology are similar within certain depth windows, different state paths and contexts may correspond to different reasonable value ranges and error evolution patterns. Therefore, relying solely on conventional regression models or numerical mapping relationships at a single depth point is insufficient to fully characterize the local evolution patterns, state transition relationships, and vertical continuous variation characteristics of well logging curves. Summary of the Invention

[0005] To address the aforementioned problems, this invention aims to provide a well logging curve reconstruction method based on TICC clustering and Markov prior.

[0006] The technical solution of the present invention is as follows: A well logging curve reconstruction method based on TICC clustering and Markov prior includes the following steps: S1: Acquire logging data and preprocess the logging data; the logging data includes known logging curves and target logging curves; S2: Construct and train the first-stage GRU network. The first-stage GRU network takes the sequence samples of the known logging curve along the depth direction as input and outputs the initial reconstruction value of the target logging curve. S3: Taking the actual target logging curve of the training well as the object, a local logging curve subsequence is constructed using a sliding window. The local logging curve subsequence is clustered using the TICC method to identify K types of local evolution mode states and construct a TICC state model. S4: Based on the TICC state model, the actual target logging curve of the training well is divided into states to obtain a TICC state sequence arranged along the depth direction; the number of TICC state transitions between adjacent depth points is counted and normalized to construct a Markov state transition probability matrix. S5: Calculate the matching cost between the initial reconstructed value of the target logging curve and each state in the TICC state model. Combined with the Markov state transition probability matrix, obtain the optimal TICC state path that is continuous along the depth direction through Viterbi decoding, and determine the TICC category at each depth point. S6: Group the actual target logging curves of the training wells according to the TICC category, and calculate the mean and standard deviation of the target logging curve values ​​under each TICC category to construct the Gaussian prior of the TICC state condition curve values. S7: Calculate the residual between the initial reconstructed value of the target logging curve and the actual target logging curve, and construct the TICC state condition Markov residual prior. S8: The initial reconstructed value of the target logging curve, the TICC category at the current depth point, the Gaussian prior of the TICC state condition curve value, and the Markov residual prior of the TICC state condition are spliced ​​together to form the input features of the second-stage GRU network. S9: Construct a second-stage GRU network, predict residual 2 using the input features, add residual 2 back to the initial reconstruction value of the target logging curve, and obtain the final target logging curve reconstruction result.

[0007] Preferably, in step S1, the preprocessing includes data cleaning and standardization.

[0008] Preferably, in step S2, when training the first-stage GRU network, the dataset is divided using a leave-one-well validation method, where one well is used as a test well and the rest are used as training wells.

[0009] Preferably, the training well is divided into training and validation segments according to depth, and an early stopping strategy is used for training control of the GRU network in the first stage.

[0010] Preferably, in step S3, when performing clustering, each cluster is characterized by an inverse covariance matrix with a Toeplitz structure; each TICC state corresponds to a class of local evolution patterns.

[0011] Preferably, the local evolutionary mode includes any one or more of the following: stable, rising, falling, fluctuating, and abrupt.

[0012] Preferably, in step S5, the matching cost is calculated using the following formula: (1) In the formula: Let represent the matching cost of the t-th local reconstruction window belonging to the k-th category; p() represents the conditional probability density; This represents the t-th local window constructed from the reconstructed curve; =k indicates that the t-th local window is divided into the k-th category; The initialization formula for Viterbi decoding is: (2) In the formula: This represents the initial cumulative cost when the first local window is in the k-th category; This represents the matching cost between the first local window and the k-th category; This represents the initial probability of the k-th category; The recursive formula is: (3) In the formula: This represents the minimum cumulative cost when the t-th depth window is in the k-th category; This represents the matching cost between the t-th local window and the k-th category; This represents the cumulative cost when the (t-1)th local window is in the i-th category; This represents the Markov state transition probability of class i to k; The corresponding optimal previous state is: (4) In the formula: This represents the optimal previous state when the current state k is reached; The optimal TICC state path is obtained by backtracking.

[0013] Preferably, in step S7, when constructing the TICC state-conditional Markov residual prior, within each TICC category, the residual is divided into several ordered residual states, the number of transitions between residual states between adjacent depth points is counted and normalized, and a TICC state-conditional Markov residual transition matrix is ​​constructed; based on the TICC category of the current depth point and the residual state of the previous depth point, the expected residual correction is extracted from the TICC state-conditional Markov residual transition matrix.

[0014] Preferably, when dividing the residual, the residual is divided into 5 ordered residual states according to the quantiles of the training well residuals.

[0015] Preferably, in step S9, when the residual is greater than 0, it indicates that the initial reconstruction value of the target logging curve is too small and needs to be corrected upward; when the residual is less than 0, it indicates that the initial reconstruction value of the target logging curve is too large and needs to be corrected downward.

[0016] The beneficial effects of this invention are: This invention can utilize the local dynamic changes and continuous evolution of logging curves along the depth direction to avoid problems such as local jumps in reconstruction results, excessive smoothing, and insufficient continuity of layers in well sections with abrupt lithological changes, reservoir heterogeneous enhancement, and thin interbedded development. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the well logging curve reconstruction method based on TICC clustering and Markov prior of the present invention. Figure 2 This is a schematic diagram of the GRU network structure; Figure 3 This is a schematic diagram of the TICC category classification results of a training well in a specific embodiment; Figure 4 This is a schematic diagram of the shear acoustic reconstruction results of the target well in Example 1; Figure 5 This is a schematic diagram of the compensated neutron reconstruction results of the target well in Example 2; Figure 6 This is a schematic diagram of the natural gamma reconstruction results of the target well in Example 3. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and technical features described in this application can be combined with each other. It should also be pointed out that, unless otherwise indicated, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terms "comprising" or "including" and similar words used in this invention refer to elements or objects preceding the word that encompass the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0020] like Figure 1 As shown, this invention provides a well logging curve reconstruction method based on TICC clustering and Markov prior, comprising the following steps: S1: Acquire logging data and preprocess the logging data; the logging data includes known logging curves and target logging curves.

[0021] In one specific embodiment, the preprocessing includes data cleaning and standardization. During data cleaning, the integrity of the data is checked to ensure that there are no large areas of missing or anomalies within the depth range, and missing values ​​are filled in. Additionally, during preprocessing, correlation analysis can be performed on existing logging curves to obtain correlation coefficients with the reconstructed curves, selecting curves that are more relevant to the target variable, thereby reducing interference from noise and redundant information. In this embodiment, the standardization process can eliminate the impact of differences in dimensions and numerical ranges between different logging curves on model training.

[0022] In one specific embodiment, the logging data includes natural gamma (GR), DTC (Difference Transmission Time), compensated density (DEN), compensated neutron (CNL), and DTS (Difference Transmission Time). Optionally, when the target logging curve is natural gamma, the known logging curve includes DTC, compensated density, and compensated neutron; when the target logging curve is compensated neutron, the known logging curve includes DTC, natural gamma, and compensated density; when the target logging curve is DTS, the known logging curve includes DTC, natural gamma, compensated neutron, and compensated density.

[0023] S2: Construct and train the first-stage GRU network. The first-stage GRU network takes the sequence samples of the known logging curves along the depth direction as input and outputs the initial reconstruction value of the target logging curve.

[0024] In a specific embodiment, when training the first-stage GRU network, the dataset is divided using a leave-one-well validation method, where one well is used as a test well and the rest are used as training wells. The training wells are divided into training and validation segments according to depth, and an early stopping strategy (if the validation loss does not decrease within several consecutive iterations, training is stopped) is used to control the training of the first-stage GRU network.

[0025] In the above embodiments, dividing the training well into training and validation segments according to depth order ensures that the model training, validation, and testing processes conform to the actual characteristics of continuous changes in logging curves along depth, and avoids depth information leakage caused by random division. Employing an early stopping strategy for training control prevents the model from overfitting to the training segment.

[0026] In a specific embodiment, when training the first-stage GRU network, depth sequence samples are constructed using a sliding window method based on the known continuous variation characteristics of well logging curves along the depth direction, and these samples are input into the first-stage GRU. Let the window length be D and the input feature dimension be F, then a single-well sample can be represented as: (5) In the formula: X represents the sequence sample tensor input to the first-stage GRU network; R represents the real number field; B represents the number of sequence samples contained in a training batch; In the first stage of the GRU network, the input sequence is passed step by step along the depth direction, and historical information is preserved and filtered through update gates and reset gates. For example... Figure 2 As shown, GRU can simultaneously utilize the logging features at the current depth point and the hidden state passed down from the previous depth point, thereby learning the continuous vertical variation of the logging curve. For the t-th depth position, the input feature vector is denoted as x. t The hidden state at the previous depth position is denoted as h. t-1 The hidden state at the current depth position is denoted as h. t GRU controls the fusion process of historical information and current input information through update gates, reset gates, and candidate hidden states.

[0027] The update gate controls how much historical information from the previous depth position is retained in the current depth position. Its calculation formula is: (6) In the formula: σ represents the update gate vector at the t-th depth position; σ represents the Sigmoid activation function. This represents the weight matrix from the input features to the update gate; This represents the cyclic weight matrix from the hidden state at the previous depth position to the update gate; This represents the bias vector of the updated gate; The reset gate controls how much historical information from the previous depth position needs to be forgotten when calculating the current candidate hidden state. Its calculation formula is as follows: (7) In the formula: Represents the reset gate vector at the t-th depth position; This represents the weight matrix from the input features to the reset gate; This represents the cyclic weight matrix from the hidden state at the previous depth position to the reset gate; This represents the bias vector of the reset gate; Under the control of the update gate and reset gate, GRU further calculates the candidate hidden states at the current depth position: (8) In the formula: Let represent the candidate hidden state at depth t; tanh represents the hyperbolic tangent activation function. This represents the weight matrix from the input features to the candidate hidden state; Represents the cyclic weight matrix; The bias vector representing the candidate hidden state; This represents the Hadamard element-wise product; Finally, GRU merges the hidden state of the previous depth position with the current candidate hidden state to obtain the hidden state of the current depth position: (9) In the formula: This represents the current hidden state at depth t; The initial reconstructed value of the target logging curve is obtained through the first-stage GRU network.

[0028] S3: Using the actual target logging curves of the training well as the object, a local logging curve subsequence is constructed using a sliding window. The TICC method is used to cluster the local logging curve subsequences, identify K types of local evolution mode states, and construct a TICC state model.

[0029] In one specific embodiment, during clustering, each cluster is characterized by an inverse covariance matrix with a Toeplitz structure; each TICC state corresponds to a local evolutionary pattern. Optionally, the local evolutionary pattern includes any one or more of stationary, rising, falling, fluctuating, and abrupt changes.

[0030] In the above embodiments, the inverse covariance Θ is represented in the following Toeplitz form: (10) Where 'a' represents distinct elements in the matrix. This process can be viewed as well-logging curve subsequences within the same type of local evolution model having similar local correlation structures, i.e., the DTS variation relationship between adjacent or similar depth points remains relatively stable within this type of model.

[0031] For a dataset with K clusters, the goal is to solve for the K inverse covariances Θ = Θ1, ... Θ. K One per cluster. TICC uses a method similar to expectation maximization to solve it, expressed as: (11) In the formula: Θ represents the set of inverse covariance matrices to be learned; T represents the matrix space that satisfies the Toeplitz structure constraints; P represents the set of cluster partitions for local window samples; K represents the number of categories; The regularization parameter matrix represents the degree of sparsity of the inverse covariance matrix; This represents the observed feature vector formed by the t-th local DTS window; This represents the set of local window samples that have been assigned to the i-th category; Represents a local window Log-likelihood of the Gaussian graph model corresponding to the i-th category; This represents a penalty parameter that controls the continuity of states between adjacent windows; It is an indicator function used to check whether adjacent attribute observations are assigned to the same cluster.

[0032] S4: Based on the TICC state model, the actual target logging curve of the training well is divided into states to obtain a TICC state sequence arranged along the depth direction; the number of TICC state transitions between adjacent depth points is counted and normalized to construct a Markov state transition probability matrix.

[0033] S5: Calculate the matching cost between the initial reconstructed value of the target logging curve and each state in the TICC state model. Combined with the Markov state transition probability matrix, obtain the optimal TICC state path that is continuous along the depth direction through Viterbi decoding, and determine the TICC category at each depth point.

[0034] In one specific embodiment, the matching cost is calculated using the following formula: (1) In the formula: Let $\mathbf{t}$ represent the matching cost of the $t$-th local reconstruction window belonging to the $k$-th category. The smaller the value, the better the local reconstruction window matches the k-th category; p() represents the conditional probability density. This represents the t-th local window constructed from the reconstructed curve; =k indicates that the t-th local window is divided into the k-th category; The initialization formula for Viterbi decoding is: (2) In the formula: This represents the initial cumulative cost when the first local window is in the k-th category; This represents the matching cost between the first local window and the k-th category; This represents the initial probability of the k-th category; The recursive formula is: (3) In the formula: This represents the minimum cumulative cost when the t-th depth window is in the k-th category; This represents the matching cost between the t-th local window and the k-th category; This represents the cumulative cost when the (t-1)th local window is in the i-th category; This represents the Markov state transition probability of class i to k; The corresponding optimal previous state is: (4) In the formula: This represents the optimal previous state when the current state k is reached; The optimal TICC state path is obtained by backtracking.

[0035] In one specific embodiment, the target logging curve is a shear sonic curve, and the classification result of the actual shear sonic curve of a training well is as follows: Figure 3 As shown.

[0036] S6: Group the actual target logging curves of the training wells according to the TICC category, and calculate the mean and standard deviation of the target logging curve values ​​under each TICC category to construct the Gaussian prior of the TICC state condition curve values.

[0037] In a specific embodiment, the mean and the standard deviation are calculated using the following formulas: (12) (13) In the formula: The mean; Standard deviation; This represents the set of curve values ​​in the training well that are classified into the k-th category; Represents a set The number of samples included; Represents a set One of the true values.

[0038] S7: Calculate the residual between the initial reconstructed value of the target logging curve and the actual target logging curve, and construct the TICC state condition Markov residual prior.

[0039] In a specific embodiment, when constructing the TICC state-conditional Markov residual prior, within each TICC category, residual 1 is divided into several ordered residual states. The number of transitions between residual states between adjacent depth points is counted and normalized to construct the TICC state-conditional Markov residual transition matrix. Based on the TICC category of the current depth point and the residual state of the previous depth point, the expected residual correction is extracted from the TICC state-conditional Markov residual transition matrix. Optionally, when dividing residual 1, it is divided into 5 ordered residual states based on the quantiles of the training well residuals.

[0040] In the above embodiment, for the k-th type of TICC state, let the residual state of the (i-1)-th logging curve be a, and the residual state of the ith logging curve be b. Then, the conditional probability of transitioning from state a to state b is defined as: (14) In the formula: This indicates the number of times a data point is transferred from category a to category b. This represents the total number of transitions from category a to all candidate states.

[0041] This yields the Markov transition probability matrix for the well logging curve interval: (15) in, .

[0042] Each TICC state corresponds to a residual state transition matrix, which describes the continuous evolution of the first-stage prediction residual along the depth direction under the same local curve morphology. This method can distinguish the differences in error direction and error magnitude under different curve morphologies.

[0043] S8: The initial reconstructed value of the target logging curve, the TICC category at the current depth point, the Gaussian prior of the TICC state condition curve value, and the Markov residual prior of the TICC state condition are spliced ​​together to form the input features of the second-stage GRU network.

[0044] S9: Construct a second-stage GRU network, predict residual 2 using the input features, add residual 2 back to the initial reconstruction value of the target logging curve, and obtain the final target logging curve reconstruction result.

[0045] In one specific embodiment, when the residual is greater than 0, it indicates that the initial reconstruction value of the target logging curve is too small and needs to be corrected upward; when the residual is less than 0, it indicates that the initial reconstruction value of the target logging curve is too large and needs to be corrected downward.

[0046] In one specific embodiment, the well logging curve reconstruction method based on TICC clustering and Markov prior of the present invention further includes a step of evaluating and analyzing the final reconstruction result. Optionally, indicators such as root mean square error, mean absolute error, and coefficient of determination are used to evaluate the final target well logging curve reconstruction result, and the model effect is analyzed by comparing the actual curve and the reconstructed curve.

[0047] In this invention, a technical approach of "first-stage main trend reconstruction + TICC-Markov state prior + second-stage standardized residual correction" is adopted. First, the first-stage GRU network is trained using known logging curves to obtain the initial reconstruction value of the target curve. Then, the TICC local state model is trained using the real target logging curve of the training well, and the TICC state transition matrix is ​​statistically analyzed. Next, based on the TICC state likelihood of the first-stage reconstructed curve, the Viterbi decoding is used to obtain a continuous and reasonable TICC state path along the depth direction. On this basis, Gaussian priors for TICC state condition curve values ​​and Markov residual priors for TICC state conditions are constructed respectively. Finally, the second-stage GRU network predicts the standardized residuals based on the above prior features, and the residuals are added back to the first-stage reconstruction value to obtain the final target logging curve reconstruction result.

[0048] In this way, the second-stage GRU network no longer relies solely on the numerical mapping relationship between the known logging curve and the target logging curve. Instead, it further integrates the initial reconstruction results from the first stage and the Markov residual transfer prior between the TICC local evolution mode categories, thereby improving the continuity, stability, and geological rationality of the final target logging curve. This allows the present invention to improve the accuracy, continuity, and geological rationality of the final reconstructed curve without using the actual target logging curve from the test well, addressing the problem in existing target logging curve reconstruction models of insufficient utilization of local state, residual direction, and vertical continuity constraints.

[0049] Example 1 Taking a target well in a certain study area as an example, its shear acoustic waveform is reconstructed using the well logging curve reconstruction method based on TICC clustering and Markov prior described in this invention. In this embodiment, the actual shear acoustic waveform curve and the reconstructed shear acoustic waveform curve of the target well are as follows: Figure 4 As shown. From Figure 4As can be seen, the present invention can accurately reconstruct the shear acoustic curve of the target well using only four known logging curves: longitudinal wave time difference, natural gamma, compensated neutron, and compensated density.

[0050] Example 2 Unlike Example 1, this example uses compensated neutrons as the target logging curve, and employs the P-wave travel time, natural gamma, and compensated density as the known logging curves. The actual compensated neutron curve and the reconstructed compensated neutron curve of the target well are as follows: Figure 5 As shown. From Figure 5 As can be seen, the present invention can accurately reconstruct the compensated neutron curve of the target well using only three known logging curves: P-wave time difference, natural gamma, and compensated density.

[0051] Example 3 Unlike Example 1, this example uses natural gamma ray as the target logging curve and employs the P-wave travel time, compensated density, and compensated neutrons as the known logging curves. The true natural gamma ray curve and the reconstructed natural gamma ray curve of the target well are as follows: Figure 6 As shown. From Figure 6 As can be seen, the present invention can accurately reconstruct the natural gamma curve of the target well using only three known logging curves: longitudinal wave time difference, compensated density, and compensated neutron.

[0052] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A well logging curve reconstruction method based on TICC clustering and Markov prior, characterized in that, Includes the following steps: S1: Acquire logging data and preprocess the logging data; the logging data includes known logging curves and target logging curves; S2: Construct and train the first-stage GRU network. The first-stage GRU network takes the sequence samples of the known logging curve along the depth direction as input and outputs the initial reconstruction value of the target logging curve. S3: Taking the actual target logging curve of the training well as the object, a local logging curve subsequence is constructed using a sliding window. The local logging curve subsequence is clustered using the TICC method to identify K types of local evolution mode states and construct a TICC state model. S4: Based on the TICC state model, the actual target logging curve of the training well is divided into states to obtain a TICC state sequence arranged along the depth direction; the number of TICC state transitions between adjacent depth points is counted and normalized to construct a Markov state transition probability matrix. S5: Calculate the matching cost between the initial reconstructed value of the target logging curve and each state in the TICC state model. Combined with the Markov state transition probability matrix, obtain the optimal TICC state path along the depth direction through Viterbi decoding, and determine the TICC category at each depth point. S6: Group the actual target logging curves of the training wells according to the TICC category, and calculate the mean and standard deviation of the target logging curve values ​​under each TICC category to construct the Gaussian prior of the TICC state condition curve values. S7: Calculate the residual between the initial reconstructed value of the target logging curve and the actual target logging curve, and construct the TICC state condition Markov residual prior. S8: The initial reconstructed value of the target logging curve, the TICC category at the current depth point, the Gaussian prior of the TICC state condition curve value, and the Markov residual prior of the TICC state condition are spliced ​​together to form the input features of the second-stage GRU network. S9: Construct a second-stage GRU network, predict residual 2 using the input features, add residual 2 back to the initial reconstruction value of the target logging curve, and obtain the final target logging curve reconstruction result.

2. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 1, characterized in that, In step S1, the preprocessing includes data cleaning and standardization.

3. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 1, characterized in that, In step S2, when training the first-stage GRU network, the dataset is divided using a leave-one-well validation method, where one well is used as a test well and the rest are used as training wells.

4. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 3, characterized in that, The training well is divided into training and validation segments according to depth, and an early stopping strategy is used for the training control of the first stage of the GRU network.

5. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 1, characterized in that, In step S3, during clustering, each cluster is represented by an inverse covariance matrix with a Toeplitz structure; each TICC state corresponds to a local evolutionary pattern.

6. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 5, characterized in that, The local evolutionary pattern includes any one or more of the following: stable, rising, falling, fluctuating, and abrupt.

7. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 1, characterized in that, In step S5, the matching cost is calculated using the following formula: （1） In the formula: Let represent the matching cost of the t-th local reconstruction window belonging to the k-th category; p() represents the conditional probability density; This represents the t-th local window constructed from the reconstructed curve; =k indicates that the t-th local window is divided into the k-th category; The initialization formula for Viterbi decoding is: （2） In the formula: This represents the initial cumulative cost when the first local window is in the k-th category; This represents the matching cost between the first local window and the k-th category; This represents the initial probability of the k-th category; The recursive formula is: （3） In the formula: This represents the minimum cumulative cost when the t-th depth window is in the k-th category; This represents the matching cost between the t-th local window and the k-th category; This represents the cumulative cost when the (t-1)th local window is in the i-th category; This represents the Markov state transition probability of class i to k; The corresponding optimal previous state is: （4） In the formula: This represents the optimal previous state when the current state k is reached; The optimal TICC state path is obtained by backtracking.

8. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 1, characterized in that, In step S7, when constructing the TICC state-conditional Markov residual prior, within each TICC category, the residual is divided into several ordered residual states. The number of transitions between residual states between adjacent depth points is counted and normalized to construct the TICC state-conditional Markov residual transition matrix. Based on the TICC category of the current depth point and the residual state of the previous depth point, the expected residual correction is extracted from the TICC state-conditional Markov residual transition matrix.

9. The well logging curve reconstruction method based on TICC clustering and Markov prior as described in claim 8, characterized in that, When dividing the residual, the residual is divided into 5 ordered residual states based on the quantiles of the training well residuals.

10. The well logging curve reconstruction method based on TICC clustering and Markov prior according to any one of claims 1-9, characterized in that, In step S9, when the residual is greater than 0, it indicates that the initial reconstruction value of the target logging curve is too small and needs to be corrected upward; when the residual is less than 0, it indicates that the initial reconstruction value of the target logging curve is too large and needs to be corrected downward.

Images