A method and system for predictive safety control oriented to well logging curve reconstruction

Abstract

The present application belongs to the technical field of predictive safety control, and particularly relates to a predictive safety control method and system for well logging curve reconstruction, which is divided into a training verification stage and a deployment control stage; the training verification stage performs full-process well-level data isolation, deployable feature unified construction, regression model training, prediction uncertainty monotonic calibration, performance knowledge base and risk judgment threshold construction; the deployment control stage completes target well input distribution difference detection and predictable grading judgment based on the performance knowledge base, and grades and controls the reconstructed curve output according to the judgment result: the reconstructed value output is blocked when it is low predictable or the input distribution is out of range, and the well logging reconstructed curve with a calibrated uncertainty interval is generated when it is medium / high predictable and the distribution is matched.

Description

Technical Field

[0001] This invention belongs to the field of pre-prediction safety control technology, and in particular relates to a pre-prediction safety control method and system for well logging curve reconstruction. Background Technology

[0002] In oil and gas exploration and development, key logging curves such as sonic transit time, neutron porosity, and bulk density are often missing. The industry commonly uses machine learning regression models to reconstruct missing curves based on existing logging curves. However, existing reconstruction schemes lack strict well-level data isolation constraints. Training, validation, and testing samples are only randomly divided according to data points, allowing samples from the same well to simultaneously flow into multiple datasets. This results in significant leakage of information from the same well during statistical parameters and model training, leading to inflated model validation metrics that fail to accurately reflect the model's generalization ability in unfamiliar target wells. Furthermore, conventional modeling introduces ground truth features derived from the target curve to be reconstructed. These features cannot be calculated in actual deployed wells with missing curves, resulting in an inconsistency between the model training and deployment feature spaces, rendering the reconstruction algorithm impractical for implementation.

[0003] Traditional logging curve reconstruction processes lack pre-prediction risk and safety management mechanisms. On the one hand, the uncertainty of the original prediction output by the model lacks a unified calibration standard, and the actual coverage of the prediction interval is far lower than the preset confidence level, making it impossible to quantify the credibility of the reconstruction results. On the other hand, the lack of a standardized performance knowledge base and input distribution difference judgment logic makes it impossible to quantitatively identify the geological data distribution deviation between the target well and the training well, as well as low-predictability conditions. Low-reliability reconstruction data is still directly output to downstream geological interpretation and reservoir evaluation modules, which can easily cause systematic deviations in reservoir parameter calculation and oil and gas layer identification, posing significant data risks to exploration and development decisions. Summary of the Invention

[0004] In view of the aforementioned problems, embodiments of the present invention provide a pre-prediction safety control method for well logging curve reconstruction, the method comprising: It is divided into a training and verification phase and a deployment and control phase; The training and validation phase includes full-process well-level data isolation, unified construction of deployable features, regression model training, monotonic calibration of prediction uncertainty, and construction of performance knowledge base and risk judgment threshold. The performance knowledge base is built and associated with the well-level isolation verification results. The base stores historical well-level isolation verification performance indicators, monotonic calibration function parameters, input feature distribution parameters, distribution difference thresholds, predictability grading thresholds and supporting output control strategies. Based on the real cross-well generalization ability obtained from well isolation verification, it realizes pre-prediction risk management. The deployment control phase sets up a pre-prediction safety gate mechanism. Before calling the regression prediction model to perform reconstruction calculation, the performance knowledge base is first retrieved to complete the target well input distribution difference detection and predictability classification. Reconstruction prediction is only allowed to start when the target well input distribution is medium / high predictable and the input distribution matches and meets the preset safety judgment conditions. Based on the safety gate determination results, the output of the reconstruction curve is controlled in a graded manner: when the reconstructed value is low or outside the input distribution, the reconstruction value output is blocked; when the reconstructed value is medium / high and the distribution matches, a logging reconstruction curve with a calibration uncertainty interval is generated.

[0005] Preferably, the training and verification phase includes the following steps: Step S1, multi-well logging data acquisition and depth alignment, constructing a dataset carrying well identifiers, and distinguishing between reference wells with complete target curves and target wells with missing target curves; Step S2: Establish a full-process well-level data isolation mechanism. The dataset is divided using well identifiers as the smallest unit of division. Group cross-validation or leave-one-well cross-validation is used to divide the dataset. Various preprocessing and feature standardization statistical parameters are calculated independently only by training wells. After the parameters are frozen, they are applied to the verification / test wells separately to prevent cross-well information leakage. Step S3: Based on the available input curves of the target well, deployable features are uniformly constructed. The features are obtained only from the input curves and their depth window statistics and curve coupling transformation quantities, without containing the true value of the target curve to be reconstructed and its derived quantities. Training and deployment adopt completely consistent feature calculation rules. Step S4: Train the regression model using deployable features under well isolation constraints, obtain the original prediction uncertainty and prediction residuals based on the validation samples, and freeze the model and feature processing parameters after training and store them in the performance knowledge base. Step S5: Construct a calibration function with monotonic constraints using well isolation verification residuals, solve the post-calibration uncertainty based on the original prediction uncertainty, and generate a calibration prediction interval with a specified coverage level by combining the verification residual quantiles. Step S6a: During the training phase, a performance knowledge base is built, pre-storing target curve types, input curve combinations, well-level isolation verification indicators, monotonic calibration function parameters, input feature distribution parameters, distribution difference thresholds, predictability grading thresholds, and corresponding output control strategies.

[0006] Preferably, when constructing deployable features in step S3, at least one of the following is calculated within the depth window: mean, standard deviation, range, first / second difference, and local slope of the input curve; at least one of the following is calculated: ratio, difference, correlation, and mutual information among multiple curves; and a logarithmic transformation is performed on the RT curve as a derived feature. The normalized statistics within each cross-validation iteration are calculated only from the current training wells, without introducing validation set data, further eliminating information leakage between iterations.

[0007] Preferably, the monotonic calibration function in step S5 is solved by order-preserving regression combined with PAVA adjacent violation constraint algorithm, and the constraint condition is that the greater the uncertainty of the original prediction, the higher the upper limit of the residual after calibration is monotonically non-decreasing. The interval half-width adopts or , where q α and c α All are determined by the well-level isolation verification residuals during the training and verification phase.

[0008] Preferably, the predictability grading threshold is determined through nested cross-validation: Inner layer optimization set based on well isolation verification R 2 The thresholds are divided into high, medium, and low levels, and the outer validation set verifies the stability of the threshold levels. R 2 Values ​​above the high threshold are considered highly predictable, values ​​between the two thresholds are considered moderately predictable, and values ​​below the low threshold are considered lowly predictable.

[0009] Preferably, the deployment control stage includes the following steps: Step S6b, construct deployable features for the target well, query the performance knowledge base using the target curve type and input curve combination as joint retrieval keys, and obtain the reference feature distribution mean vector μ, covariance matrix Σ, regularization parameter λ, and distribution difference threshold τ; Calculate the regularized squared Mahalanobis distance between the target well characteristics and the reference distribution: ; in, ; If the proportion of sampling points exceeding the threshold or the length of consecutive well segments exceeding the threshold exceeds the preset limit, the corresponding well / well segment is determined to be an out-of-distribution sample. Step S7: Based on the predictable level and the out-of-distribution judgment result, output in stages: for low predictability or out-of-distribution scenarios, generate blocking signals, set the reconstructed values ​​to empty or invalid masks, and prohibit downstream output; When the prediction is medium to high and the distribution matches, the frozen regression model is invoked to generate reconstructed values ​​and the original uncertainty, and then the calibration prediction interval is output through the monotonic calibration function. Medium predictability outputs reconstructed data with risk labels, while high predictability outputs standard reconstructed data without risk labels.

[0010] Preferably, when the performance knowledge base contains multiple matching entries, the entry that best matches the geological range, sampling scale, feature definition, and deployment conditions of the training well is selected as the query result.

[0011] A pre-prediction safety control system for well logging curve reconstruction includes: a data acquisition module, a well-level data isolation module, a feature construction module, a performance knowledge base module, a regression prediction module, an uncertainty calibration module, a predictability classification and safety gate determination module, a risk classification output module, a memory, and a processor. The processor schedules each module to run in two phases: training and verification, and deployment and control. The performance knowledge base module is used to store historical well-level isolation verification performance, monotonic calibration function parameters, input feature distribution parameters, distribution difference threshold, predictability grading threshold and supporting output control strategies constructed from well-level isolation verification results. During the deployment phase, the predictability classification and safety gate determination module is run first to complete the safety gate verification before prediction: This module retrieves data from the performance knowledge base module and sequentially completes the target well input distribution difference detection and predictability classification determination. Only after the verification is passed will the regression prediction module and uncertainty calibration module be started to calculate the reconstructed curve and the calibration prediction interval. The risk classification output module outputs data for normal channels, risk-marked channels, and blocked channels based on the safety door determination results. In scenarios with low predictability or outside the input distribution, the risk classification output module generates a blocking signal, sets the reconstructed value to empty or adds an invalid mask, and prohibits the output of the reconstructed result downstream.

[0012] Preferably, the system is applicable to multiple oil and gas exploration blocks and can safely reconstruct missing logging curves such as sonic transit time, neutron porosity, and bulk density. Based on the predictability classification and safety gate determination module and the performance knowledge base module, the system can complete the prediction risk identification and graded output control of target wells across structural units and geological blocks.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the pre-prediction safety control method for well logging curve reconstruction.

[0014] Based on the above, a full-process well-level data isolation mechanism is used to isolate training, validation, and test datasets by well identifier. Various statistical parameters are generated independently from training wells and frozen for reuse, completely eliminating the problem of information leakage within the same well. Well-level isolation validation metrics can objectively characterize the model's cross-well generalization ability. At the same time, the generation rules for deployable features are uniformly constrained, avoiding the participation of the ground truth value of the target curve to be reconstructed and its derivatives in feature calculation throughout the process. This ensures that the feature space is completely consistent between the training and deployment stages, solving the defect of the model's inability to be deployed from the bottom layer. Combined with a monotonic calibration function built based on well isolation residuals, the actual coverage of the nominal 95% prediction interval is increased from 59.8% to 89.3%, achieving accurate quantification of prediction uncertainty and significantly improving the credibility of the reconstruction results.

[0015] This invention establishes a standardized performance knowledge base and introduces squared Mahalanobis distance to detect input distribution differences. It combines nested cross-validation to divide the predictability into three levels: high, medium, and low, forming a closed loop for pre-prediction safety judgment. For low predictability and out-of-distribution well sections, it automatically generates blocking signals to shield unreliable reconstructed values ​​from flowing downstream. For medium predictability, it adds risk indicators to the output, and for high predictability, it outputs standard and effective reconstructed curves. The hierarchical control output mode can accurately avoid reservoir evaluation deviations caused by geological data offsets and insufficient model prediction capabilities, realizing pre-control of risks in the entire process of logging curve reconstruction. It is suitable for the safety reconstruction needs of missing curves in oil and gas wells in multiple blocks and across structural units. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention, illustrating the relationship between the training and verification phase and the deployment control phase. The training and verification phase establishes a well-level data isolation mechanism, a set of deployable features, a regression prediction model, a monotonic calibration function, and a performance knowledge base. The deployment control phase first queries the performance knowledge base based on the available input curves of the target well and calculates the input distribution differences, performs predictability grading and out-of-distribution determination, and then executes prohibited output, risk flag output, or normal output based on the determination results.

[0017] Figure 2 This is a schematic diagram of the well-level data isolation mechanism of the present invention, showing that the training well set, validation well set, and test well set are isolated from each other with well identifiers as the smallest dividing unit; data cleaning statistics, standardization parameters, feature statistics parameters, and hyperparameter search are determined only by the training well set and applied independently to the validation well set and the test well set, and the validation well set and the test well set do not participate in the fitting of the above parameters.

[0018] Figure 3 This is a schematic diagram of the deployable feature construction of the present invention. It shows that the actual usable input curve combination of the target well can be one or more of, for example, GR, RT, and CAL. The deployable feature vector is formed by multi-scale window statistics, curve coupling, geological prior proxy, and vertical variation calculation. The true value of the target curve is only used as a label in the training and verification stage of the reference well, and is not used as an input feature in the calculation during the deployment stage of the target well.

[0019] Figure 4 This is a schematic diagram of uncertainty calibration and graded output control of the present invention. It shows that a monotonic calibration function is established based on the original uncertainty and verification error data generated by well-level isolation verification. Based on the well-level isolation verification performance knowledge base and the target well deployable feature distribution, the input distribution difference is calculated, predictability is graded and out-of-distribution is determined. Based on the determination result, prohibited output, risk flag output or normal output is executed.

[0020] Figure 5This diagram illustrates the pre-prediction safety control system module for well logging curve reconstruction according to the present invention. It shows the data acquisition module, well-level data isolation module, feature construction module, regression prediction module, uncertainty calibration module, predictability grading and out-of-distribution determination module, risk grading output module, and the data flow, parameter reading flow, and control flow between the memory and processor. The memory stores model parameters, calibration functions, performance knowledge base, μ, Σ, τ, λ, grading thresholds, and output control strategies. The processor reads from the memory and controls the operation of the corresponding functional modules 501-507. The predictability grading and out-of-distribution determination module performs a pre-prediction safety gate determination before calling the regression prediction module. The regression prediction module and uncertainty calibration module are only called when the target curve is at a medium or high predictability level and out-of-distribution determination has not been triggered. When the predictability is low or out of distribution, the risk grading output module generates a blocking signal, sets the reconstruction value field to a null value or an invalid mask, sets the output status field to a blocked state, and prohibits the writing of the target curve reconstruction value to the downstream data processing interface.

[0021] In the diagram: 101 represents multi-well data acquisition and depth alignment; 102 represents well-level data isolation; 103 represents deployable feature construction; 104 represents training the regression model; 105 represents establishing a monotonic calibration function; 106 represents establishing a performance knowledge base, grading thresholds, and distribution difference thresholds; 107 represents constructing deployable features for the target well and determining the target curve type and actual input combination; 108 represents querying the performance knowledge base; 109 represents calculating input distribution difference; 110 represents predictability grading and out-of-distribution determination; 111 represents prohibiting output and generating blocking signals; 112 represents risk indicator output; and 113 represents normal output.

[0022] 201 represents the training well set; 202 represents the validation well set; 203 represents the test well set; 204 represents the cleaning statistics; 205 represents the standardized parameters; 206 represents the characteristic statistical parameters; 207 represents the hyperparameter search and model parameter fitting; 208 represents the validation / test application; 209 represents the well set mutual exclusion constraint; 210 represents the validation / test results without leakage in the same well.

[0023] 301 represents the actual available input curve combination for the target well; 302 represents the GR curve; 303 represents the RT curve; 304 represents the CAL curve; 305 represents the multi-scale window statistical features; 306 represents the curve coupling features; 307 represents the geological prior proxy features; 308 represents the vertical variation features; 309 represents the deployable feature vector formed according to the same definition; 310 represents the true value of the target curve of the reference well; 311 represents the model training and well-level isolation verification; 312 represents the deployable features of the target well; 313 represents the input for knowledge base query, distribution difference calculation, and regression prediction after passing the safety gate.

[0024] 401 represents the well-level isolation validation regression prediction set; 402 represents the original prediction uncertainty; 403 represents the validation prediction residual calibration data pair; 404 represents the monotonic calibration function; 405 represents the calibration function and residual quantiles; 406 represents the performance knowledge base; 407 represents the target well deployability characteristics; 408 represents the input distribution difference calculation; 409 represents the predictability classification and out-of-distribution determination; 410 represents the prohibited output and blocking signal; 411 represents the risk indicator output; 412 represents the normal output.

[0025] 501 represents the data acquisition module; 502 represents the well-level data isolation module; 503 represents the feature construction module; 504 represents the regression prediction module; 505 represents the uncertainty calibration module; 506 represents the predictability classification and out-of-distribution determination module; 507 represents the risk classification output module; 508 represents the memory; and 509 represents the processor. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings.

[0027] Well-level data isolation refers to ensuring that data from the same well does not appear simultaneously in the training set and validation / test set throughout the entire process of data preprocessing, feature calculation, model training, and validation evaluation, using the well identifier as the smallest dividing unit. Furthermore, the statistical parameters and model parameters of the training set do not introduce information from the validation or test set. This well-level data isolation is achieved through GroupKFold or Leave-One-Well-Out cross-validation.

[0028] Deployable features: These are features that can be computed during both the model training and actual deployment phases. Deployable features are calculated solely from the actual available input curves of the target well, their depth window statistics, inter-curve interactions, and depth changes. They do not include the true value of the target curve to be reconstructed or any quantities derived from the true value of the target curve to be reconstructed.

[0029] This invention provides a pre-prediction safety control method for well logging curve reconstruction, characterized by including a training and verification phase and a deployment control phase: Step S1: Multi-well logging data acquisition and depth alignment. Logging data from multiple wells are acquired, and depth alignment is performed according to a preset depth sampling interval to establish a dataset including a well identifier field. Wells with the target curve to be reconstructed are used as reference wells, and at least one well with a missing target curve is used as the target well.

[0030] Step S2: Establish a full-process well-level data isolation mechanism. Using the well identifier field as the smallest unit, establish a well-level data isolation mechanism that runs through the entire process of data preprocessing, feature calculation, model training, and validation evaluation: Within each training iteration, missing value imputation parameters, outlier truncation thresholds, standardized mean and standard deviation, depth window statistical parameters, and feature selection parameters are calculated only based on the training well data; these parameters obtained from the training wells are frozen and applied independently to validation and test wells, without being mixed with data from the validation or test wells for calculation; cross-validation is performed on a well-by-well basis to ensure no information leakage occurs between training and validation wells, or between training and test wells, in the stages of statistical parameter calculation, feature standardization, model training, and hyperparameter selection. The well-by-well cross-validation can be implemented through group cross-validation or leave-one-well cross-validation.

[0031] Step S3: Based on the actual available input curve combinations for the target well, construct a deployable feature set with the same input combinations in the reference wells. First, determine the input curve list based on the actual available input curve combinations for the target well, and select the same input curve combinations from the reference wells. Align and handle missing values ​​for the input curves at the same depth sampling interval, and perform standardization based on the statistical parameters determined by the training wells. Calculate at least one of the following for the input curves within one or more preset depth windows: mean, standard deviation, range, local slope, first-order difference, or second-order difference. Calculate at least one of the following for different input curves: ratio, difference, correlation, or mutual information measure. The logarithmic transformation of the RT is used as a derived feature of the RT, not as an additional input curve. The feature set does not contain the true value of the target curve to be reconstructed or any quantity derived from it. The same feature definition is used simultaneously in the training and deployment phases to ensure consistency of the feature space between the training and deployment phases.

[0032] Step S4: Based on the well-level data isolation established in Step S2, at least one regression prediction model is trained in the reference well data, using the deployable feature set constructed in Step S3 as the input matrix and the true values ​​of the corresponding target curves to be reconstructed in the reference wells as training labels. Within each training iteration, model fitting and hyperparameter selection are performed based on the training well data, and validation prediction values, original prediction uncertainties, and prediction residuals are output on the validation wells. After training and validation are completed, the final model parameters, feature processing parameters, and model number are frozen and written to the performance knowledge base or stored in association with the performance knowledge base. This step belongs to the training and validation stage and does not generate target curve reconstruction values ​​that can be used downstream on the target wells. The deployment control stage only uses the frozen model parameters, and the true values ​​of the target curves must not be used in the target well deployment stage.

[0033] Step S5: Establish a monotonic calibration function based on the well-level isolation verification residuals. In well-level isolation verification, let u be the original prediction uncertainty of the i-th verification sample.i The predicted value is The actual value is y i The predicted residual is This forms the calibration data pair {(u i ,e i Monotonic calibration function f cal It can be represented as: The constraint condition is when hour, For the target well sample x, its original prediction uncertainty is u(x), and the calibrated uncertainty is... The calibration prediction interval can be expressed as: Where H(x) is the half-width of the interval, given by u cal (x) is determined together with the empirical quantile of the well-level isolation verification residual. In one implementation, or , where q α and c α All values ​​are determined by the well-level isolation verification residuals from the training and verification phase. The interval half-width is determined by the post-calibration uncertainty output by the monotonic mapping function and the empirical quantile of the well-level isolation verification residuals. It is used to balance coverage and interval width under a given business risk level, forming a nominal 95% calibration prediction interval or a calibration prediction interval with a preset coverage level.

[0034] Step S6a involves establishing a performance knowledge base, predictability grading threshold, and distribution difference threshold during the training and validation phase. The performance knowledge base records at least the following: target curve type, input curve combination, training well set, validation well set, well-level isolation validation R², RMSE, MAE, calibration prediction interval coverage, monotonic calibration function number or parameters, input feature mean vector μ, input feature covariance matrix Σ, distribution difference threshold τ, regularization parameter λ, predictability grading results, and output control strategy.

[0035] Step S6b: In the deployment control phase, deployable features are first constructed for the target well. Based on the target curve type and the actual available input curves for the target well, the performance knowledge base is queried to obtain the corresponding historical well-level isolation verification performance, grading thresholds, distribution parameters, and output control strategies. The squared Mahalanobis distance statistic between the input features and the reference distribution at each depth sampling point or continuous depth window of the target well is calculated. Where μ is the mean vector of input features from the training well or reference well, Σ is the covariance matrix of input features from the training well or reference well, and τ is the distribution difference threshold; Σ reg =Σ+λI, where λ is a non-negative regularization parameter determined by the training well data; when Σ is invertible and the condition number does not exceed a preset threshold, Σ reg Σ can be used; when Σ is irreversible or the condition number exceeds a preset threshold, Σ is adopted.reg =Σ+λI or Moore-Penrose generalized inverse calculation of the square Mahalanobis distance statistic. When the proportion of sampling points exceeding the threshold τ is greater than the preset proportion, or the length of consecutive well segments exceeding the threshold exceeds the preset length, the target well or corresponding well segment is determined as an out-of-distribution sample. If the target curve is at a low predictability level or triggers out-of-distribution determination, a blocking signal is generated, the reconstructed value field is set to null or invalid mask, the output status field is set to blocking status, and the reconstructed value is prohibited from entering the downstream module; if the target curve is at a medium or high predictability level and does not trigger out-of-distribution determination, the regression prediction model frozen in step S4 is called, the deployable features of the target well are input into the regression prediction model, and the target curve reconstruction value and original prediction uncertainty are generated; then the monotonic calibration function established in step S5 is called to generate the calibration prediction interval. In step S7, the normal channel, risk indicator channel, or blocking signal are output according to the high, medium, and low levels and out-of-distribution determination results, respectively.

[0036] In a preferred embodiment, for features in the deployable feature set that involve cross-well statistical normalization or cross-well statistical parameters, within each training layer of cross-validation, the statistics (mean and standard deviation) required for the cross-well statistical normalization are independently calculated using only the data from the wells included in the current training layer, and these statistics are applied independently to the corresponding validation layer without introducing statistical information from the validation layer or the full dataset. This measure, in conjunction with the well-level isolation mechanism in step S2, further eliminates potential information leakage between cross-validation layers.

[0037] The query rules for the performance knowledge base are as follows: the system uses a combination of the target curve type and the actual available input curves of the target well as a joint search key to query knowledge base entries that match this joint search key. When multiple candidate entries exist, entries whose geological range, sampling scale, and feature definitions are consistent with the deployment conditions of the target well are prioritized. The query results must return at least the historical well-level isolation verification performance, calibration function, grading threshold, μ, Σ, τ, and corresponding output control strategy for use by the pre-prediction safety gate.

[0038] This invention also provides a pre-prediction safety control system for well logging curve reconstruction, including a data acquisition module, a well-level data isolation module, a feature construction module, a regression prediction module, an uncertainty calibration module, a predictability classification and out-of-distribution determination module, and a risk classification output module. During deployment, the feature construction module first outputs deployable features for the target well; the predictability classification and out-of-distribution determination module first queries the performance knowledge base and completes the pre-prediction safety gate determination; only when the target curve is at a medium or high predictability level and out-of-distribution determination is not triggered, does the regression prediction module generate reconstruction values ​​and the uncertainty calibration module generate calibration prediction intervals. The risk classification output module simultaneously receives the reconstruction values, calibration prediction intervals, predictability level, and out-of-distribution determination results, and outputs a normal data channel signal, a risk indicator channel signal, or a blocking signal; in the blocking state, the risk classification output module sets the reconstruction value field to null or an invalid mask and prohibits writing the target curve reconstruction value to the downstream data processing interface.

[0039] In one embodiment, multi-well logging data from the Dengying Formation of the Seidan System in the Sichuan Basin were used for verification.

[0040] The reference wells comprised 20 wells with 85,551 valid samples. Target curves included sonic transit time, neutron porosity, and bulk density. Input curves were selected based on the actual availability of data from the target wells. It should be understood that this embodiment is only used to illustrate the specific implementation and technical effects of the present invention and does not constitute a limitation on the scope of protection. The present invention is also applicable to other oil and gas exploration and development blocks containing multi-well logging data.

[0041] Experimental environment: Python language, Scikit-learn machine learning library, NumPy numerical computation library. Model training and validation were both divided into wells, and data cleaning, statistical parameter calculation, feature transformation, and hyperparameter selection were all limited to the corresponding training wells.

[0042] Example 1: Quantitative verification of the data leakage effect in the same well; Twenty reference wells (85,551 samples) were selected, with sonic transit time as the target curve. The random forest model parameters were: 100 trees, maximum depth 15, and minimum number of samples per leaf node 30. Under the random sample partitioning method, the coefficient of determination R² was 0.9016 (95% interval estimate of R²: [0.895, 0.908]); under the leave-one-well cross-validation method, the coefficient of determination R² was 0.4005 (95% interval estimate of R²: [0.372, 0.429]). The coefficient of determination under the random partitioning method was artificially inflated by approximately 0.501 (absolute increment) compared to the well-level isolation validation method, representing a relative overestimation of approximately 125%. The results indicate that validation results without well-level data isolation cannot reliably reflect the model's generalization ability on unseen wells.

[0043] Example 2: Predictability experiment of predicting target curves for AC, CNL and DEN using only GR, RT and CAL derived features; In a scenario where only GR, RT, and CAL are available for the target well, AC, CNL, and DEN are used as target curves. Following step S3, a deployable feature set is constructed using only GR, RT, CAL, and their derived features. With one well left open for cross-validation, the mean R² for AC is 0.1882, for CNL it is 0.0653, and for DEN it is -0.3739. This example is used to compare the cross-well predictability of different target curves under a given input combination.

[0044] Under the above conditions, acoustic transit time and neutron porosity still retain some cross-well prediction information, while the coefficient of determination R² for bulk density is negative, indicating that the model's predictive ability under these input conditions is not better than the training set mean benchmark. Therefore, predictability classification should be differentiated based on the target curve type, input curve combination, and well-level isolation verification performance knowledge base.

[0045] DEN exhibits low predictability under this given input combination, consistent with the petrophysical interpretation that conventional input curves have limited ability to resolve differences in carbonate mineral density. However, this interpretation represents a possible mechanism based on existing results and requires further verification by incorporating residual relationships such as CAL, depth, well location, target variance, AC, CNL, and logarithmic transformation derivatives of RT. This paper does not extrapolate this result to mean that DEN is unpredictable under other well sets, sampling scales, input combinations, or missing scenarios.

[0046] Example 3: Functional verification of key modules (ablation experiment); The technical effectiveness of each module was verified through ablation experiments. All experiments were performed on the same dataset of 20 reference wells, with random forest regression as the base model (100 trees, minimum number of samples per leaf node of 30). The evaluation metrics were the coefficient of determination R² for leave-one-well cross-validation and the nominal 95% calibration prediction interval coverage.

[0047] Experiment A: Co-validation of steps S2 and S5. With random sample partitioning (without performing S2), the model's R² is 0.9016; with leave-one-well cross-validation (performing S2), the R² is 0.4005, indicating that random partitioning artificially inflates the validation metrics. Based on the prediction residuals generated by leave-one-well cross-validation, the uncalibrated nominal 95% prediction interval coverage is 59.8%. After monotonic calibration in step S5, the nominal 95% calibrated prediction interval coverage increases to 89.3%.

[0048] Experiment B: Verification of the necessity of deployable feature constraints in step S3. Based on the well-level isolated data in S2, if the features constructed in step S3 include features related to the true value of the target curve (such as the value calculated using the Gardner formula when calculating density using sonic transit time), the model can obtain a high coefficient of determination when trained on a reference well where the target curve is known. However, in deployment scenarios where the target curve is missing, this feature cannot be calculated, and the model's prediction performance drops sharply. The results show that deployable feature constraints are a necessary condition to ensure that the model can be practically deployed.

[0049] Experiment C: Functional verification of step S5. Under the leave-one-well cross-validation condition, the nominal 95% prediction interval was constructed directly using the standard deviation of the predicted values ​​of each tree in the random forest, with an actual coverage rate of 59.8%. After monotonic calibration, the coverage rate of the nominal 95% calibrated prediction interval increased to 89.3%. The results show that the original prediction uncertainty cannot be directly equated to the prediction interval with a preset coverage level, and needs to be corrected by monotonic calibration based on the well-level isolation verification residuals.

[0050] The above ablation experiments collectively demonstrate that: step S2 provides a validation benchmark without leakage in the same well; step S3 ensures the model can be deployed to wells with missing target curves; step S5 improves the nominal 95% prediction interval coverage from 59.8% to 89.3%; step S6a establishes a performance knowledge base; and steps S6b and S7 transform the validation performance and input distribution differences into pre-prediction safety gates and output controls. This closed-loop technology does not guarantee high-precision reconstruction of all target curves; when low predictability or out-of-distribution risks are triggered, prohibiting the output itself serves as a safety control effect.

[0051] Example 4: Uncertainty Monotonic Calibration Based on Well-Level Isolation Verification Benchmark; Under the leave-one-well cross-validation framework, 20 reference wells were used as validation wells, and the remaining reference wells were used as training wells. After the model generated predictions on the validation wells, 85,551 pairs of raw prediction uncertainties and prediction residual calibration data were collected. The uncalibrated nominal 95% prediction interval coverage was 59.8%. After monotonic calibration, the nominal 95% calibrated prediction interval coverage increased to 89.3%, an improvement of 29.5 percentage points. The interval half-width was determined by the calibrated uncertainty output by the monotonic mapping function and the empirical quantile of the well-level isolation validation residuals, and was used to balance coverage and interval width under a given operational risk level.

[0052] In a preferred embodiment, monotonic calibration is implemented using isotonic regression. Specifically, the calibration data pairs collected from well-level isolation verification are arranged in ascending order of their original prediction uncertainties. An isotonic regression solver (one implementation is the Pool Adjacent Violators Algorithm, PAVA) is called to scan from front to back. When an adjacent data pair's prediction residual violates the monotonic non-decreasing constraint, the violating adjacent data pairs are merged into an interval and their mean is taken. This process continues until the entire sequence satisfies the monotonic non-decreasing constraint, resulting in a monotonic mapping function. The PAVA algorithm has a time complexity of O(n), where n is the number of calibration data pairs, making it suitable for efficient processing of large-scale datasets. It should be understood that isotonic regression and PAVA are only preferred implementations of monotonic calibration and do not constitute a limitation on the monotonic calibration model; other regression methods with monotonic constraints can also be used to implement monotonic calibration in this step.

[0053] Example 5: Determination of predictability grading threshold; The predictability grading threshold is determined based on the historical statistical distribution of the coefficient of determination (R²) of each target curve in well-level isolation verification. In one implementation, the reference well is divided into an inner optimization set and an outer verification set: on the inner optimization set, a first threshold and a second threshold are determined based on the statistical distribution characteristics of the R² of each target curve in historical well-level isolation verification, with the first threshold being greater than the second threshold; on the outer verification set, the stability of the grading results corresponding to the first and second thresholds is verified. This nested cross-validation framework is only one exemplary implementation of threshold determination; the specific threshold selection criteria and optimization strategies can be adjusted according to the characteristics of the work area, business risk preferences, and data distribution characteristics.

[0054] In one exemplary implementation, the first threshold is set to 0.50, and the second threshold is set to 0.05. Cross-validation performance above the first threshold is classified as high predictability, below the second threshold as low predictability, and in between as medium predictability. Specific thresholds should be determined based on the historical performance distribution within the training well range. The above values ​​are merely illustrative and do not constitute a limitation on the scope of protection.

[0055] Example 6: Practical Application; Example 1: For target wells with only GR, RT, and CAL, the system first constructs deployable features, then queries the performance knowledge base according to the target curve type and actual input combination, and completes pre-prediction grading and out-of-distribution determination. When the predictability is low or out of distribution, a blocking signal is generated, and no reconstructed value that can enter the downstream module is generated; when the predictability is medium or high and not out of distribution, the regression prediction model is called, and the well enters the risk indicator channel or the normal channel respectively.

[0056] Example 2: For blind logging wells that have not participated in model training, the system only uses the actual input curves available for the well to perform performance knowledge base queries and distribution difference calculations. It does not need to, and must not, use the true value of the target curve. When the proportion of sampling points exceeding the threshold or the length of consecutive well segments exceeding the threshold reaches a preset condition, the target well or corresponding well segment is determined to be out of distribution, triggering out-of-distribution determination and generating a blocking signal.

[0057] Example 7: Cross-structural unit generalization verification; During cross-tectonic unit validation, wells within a single tectonic unit are retained as independent test wells, while wells in other tectonic units are used for training and validation. Test wells are not involved in any pre-prediction statistical parameter calculations, feature transformations, hyperparameter selection, calibration parameter fitting, or grading threshold determination. This validation method is used to examine the adaptability of the model and risk control framework under conditions of geological distribution variation, providing a basis for setting the distribution difference threshold in step S6.

[0058] Example 8: Stepwise characteristic group ablation experiment with AC as the target curve; Example 8 uses AC as the target curve and gradually adds different feature groups under leave-one-well cross-validation to evaluate the contribution of increasing feature dimensions to AC reconstruction performance. The Base group uses basic and derived features related to GR, RT, and CAL, with an R² of 0.2529; after adding the Core Curves feature group, the R² is 0.4005. Example 2 predicts AC, CNL, and DEN using only GR, RT, and CAL derived features to compare the predictability of different target curves. Since the target settings, feature dimensions, and experimental purposes differ between the two examples, values ​​such as 0.1882 and 0.2529 are not directly equivalent and should not be used as substitutes for each other as the same experimental result.

[0059] It should be understood that the above embodiments are only used to illustrate the technical solutions and effects of the present invention, and not to limit its scope of protection. The technical effect of the present invention is not to guarantee high-precision reconstruction of all target curves, but rather to automatically generate blocking signals based on well-level isolation verification performance, calibration errors, and differences in target well input distribution when target curves are missing and input conditions are insufficient or out-of-distribution risks are high, thus preventing low-reliability reconstruction values ​​from entering the downstream process. The prohibition of output for low-predictability target curves is itself one of the technical effects. The number of wells, sampling scale, model parameters, grading thresholds, and curve combinations in the specific embodiments are all exemplary descriptions.

[0060] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

[0061] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A pre-prediction safety control method for well logging curve reconstruction, characterized in that: It is divided into a training and verification phase and a deployment and control phase; The training and validation phase includes full-process well-level data isolation, unified construction of deployable features, regression model training, monotonic calibration of prediction uncertainty, and construction of performance knowledge base and risk judgment threshold. The performance knowledge base is built and associated with the well-level isolation verification results. The base stores historical well-level isolation verification performance indicators, monotonic calibration function parameters, input feature distribution parameters, distribution difference thresholds, predictability grading thresholds and supporting output control strategies. Based on the real cross-well generalization ability obtained from well isolation verification, it realizes pre-prediction risk management. The deployment control phase sets up a pre-prediction safety gate mechanism. Before calling the regression prediction model to perform reconstruction calculation, the performance knowledge base is first retrieved to complete the target well input distribution difference detection and predictability classification. Reconstruction prediction is only allowed to start when the target well input distribution is medium / high predictable and the input distribution matches and meets the preset safety judgment conditions. Based on the safety gate determination results, the output of the reconstruction curve is controlled in a graded manner: when the reconstructed value is low or outside the input distribution, the reconstruction value output is blocked; when the reconstructed value is medium / high and the distribution matches, a logging reconstruction curve with a calibration uncertainty interval is generated.

2. The pre-prediction safety control method for well logging curve reconstruction according to claim 1, characterized in that: The training and verification phase includes the following steps: Step S1, multi-well logging data acquisition and depth alignment, constructing a dataset carrying well identifiers, and distinguishing between reference wells with complete target curves and target wells with missing target curves; Step S2: Establish a full-process well-level data isolation mechanism. The dataset is divided using well identifiers as the smallest unit of division. Group cross-validation or leave-one-well cross-validation is used to divide the dataset. Various preprocessing and feature standardization statistical parameters are calculated independently only by training wells. After the parameters are frozen, they are applied to the verification / test wells separately to prevent cross-well information leakage. Step S3: Based on the available input curves of the target well, deployable features are uniformly constructed. The features are obtained only from the input curves and their depth window statistics and curve coupling transformation quantities, without containing the true value of the target curve to be reconstructed and its derived quantities. Training and deployment adopt completely consistent feature calculation rules. Step S4: Train the regression model using deployable features under well isolation constraints, obtain the original prediction uncertainty and prediction residuals based on the validation samples, and freeze the model and feature processing parameters after training and store them in the performance knowledge base. Step S5: Construct a calibration function with monotonic constraints using well isolation verification residuals, solve the post-calibration uncertainty based on the original prediction uncertainty, and generate a calibration prediction interval with a specified coverage level by combining the verification residual quantiles. Step S6a: During the training phase, a performance knowledge base is built, pre-storing target curve types, input curve combinations, well-level isolation verification indicators, monotonic calibration function parameters, input feature distribution parameters, distribution difference thresholds, predictability grading thresholds, and corresponding output control strategies.

3. The pre-prediction safety control method for well logging curve reconstruction according to claim 2, characterized in that: When constructing deployable features in step S3, at least one of the following is calculated within the depth window: mean, standard deviation, range, first / second difference, and local slope of the input curve; at least one of the following is calculated: ratio, difference, correlation, and mutual information among multiple curves; and a logarithmic transformation is performed on the RT curve as a derived feature. The normalized statistics within each cross-validation iteration are calculated only from the current training wells, without introducing validation set data, further eliminating information leakage between iterations.

4. The pre-prediction safety control method for well logging curve reconstruction according to claim 2, characterized in that: The monotonic calibration function in step S5 is solved by order-preserving regression combined with PAVA adjacent violation constraint algorithm. The constraint condition is that the greater the uncertainty of the original prediction, the upper limit of the residual after calibration is monotonically non-decreasing. The interval half-width adopts or , where q α and c α All are determined by the well-level isolation verification residuals during the training and verification phase.

5. The pre-prediction safety control method for well logging curve reconstruction according to claim 2, characterized in that: The predictability grading threshold is determined through nested cross-validation: Inner layer optimization set based on well isolation verification R 2 The thresholds are divided into high, medium, and low levels, and the outer validation set verifies the stability of the threshold levels. R 2 Values ​​above the high threshold are considered highly predictable, values ​​between the two thresholds are considered moderately predictable, and values ​​below the low threshold are considered lowly predictable.

6. The pre-prediction safety control method for well logging curve reconstruction according to claim 1, characterized in that: The deployment control phase includes the following steps: Step S6b, construct deployable features for the target well, query the performance knowledge base using the target curve type and input curve combination as joint retrieval keys, and obtain the reference feature distribution mean vector μ, covariance matrix Σ, regularization parameter λ, and distribution difference threshold τ. Calculate the regularized squared Mahalanobis distance between the target well characteristics and the reference distribution: ； in, ; If the proportion of sampling points exceeding the threshold or the length of consecutive well segments exceeding the threshold exceeds the preset limit, the corresponding well / well segment is determined to be an out-of-distribution sample. Step S7: Based on the predictable level and the out-of-distribution judgment result, output in stages: for low predictability or out-of-distribution scenarios, generate blocking signals, set the reconstructed values ​​to empty or invalid masks, and prohibit downstream output; When the prediction is medium to high and the distribution matches, the frozen regression model is invoked to generate reconstructed values ​​and the original uncertainty, and then the calibration prediction interval is output through the monotonic calibration function. Medium predictability outputs reconstructed data with risk labels, while high predictability outputs standard reconstructed data without risk labels.

7. The pre-prediction safety control method for well logging curve reconstruction according to claim 6, characterized in that: When the performance knowledge base contains multiple matching entries, the entry that best matches the geological range, sampling scale, feature definition, and deployment conditions of the training well is selected as the query result.

8. A pre-prediction safety control system for well logging curve reconstruction, characterized in that, The method for pre-prediction safety control according to any one of claims 1-7 includes: a data acquisition module, a well-level data isolation module, a feature construction module, a performance knowledge base module, a regression prediction module, an uncertainty calibration module, a predictability classification and safety gate determination module, a risk classification output module, a memory, and a processor. The processor schedules each module to run in two phases: training and verification, and deployment and control. The performance knowledge base module is used to store historical well-level isolation verification performance, monotonic calibration function parameters, input feature distribution parameters, distribution difference threshold, predictability grading threshold and supporting output control strategies constructed from well-level isolation verification results. During the deployment phase, the predictability classification and safety gate determination module is run first to complete the safety gate verification before prediction: This module retrieves data from the performance knowledge base module and sequentially completes the target well input distribution difference detection and predictability classification determination. Only after the verification is passed will the regression prediction module and uncertainty calibration module be started to calculate the reconstructed curve and the calibration prediction interval. The risk classification output module outputs data for normal channels, risk-marked channels, and blocked channels based on the safety door determination results. In scenarios with low predictability or outside the input distribution, the risk classification output module generates a blocking signal, sets the reconstructed value to empty or adds an invalid mask, and prohibits the output of the reconstructed result downstream.

9. A pre-prediction safety control system for well logging curve reconstruction according to claim 8, characterized in that: The system is applicable to multiple oil and gas exploration blocks and can safely reconstruct missing logging curves such as sonic transit time, neutron porosity, and bulk density. Based on the predictability classification and safety gate determination module and the performance knowledge base module, the system can complete the prediction risk identification and graded output control of target wells across structural units and geological blocks.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the pre-prediction safety control method for well logging curve reconstruction as described in any one of claims 1 to 7.

Images