A well logging curve missing intelligent reconstruction device based on model optimization and multi-dimensional evaluation model
By constructing an intelligent reconstruction device with a multi-dimensional evaluation model and a multi-algorithm library, the problem of missing logging curves was solved, and high-precision, continuous and geologically significant logging curve reconstruction was achieved, overcoming the limitations of existing technologies and the problem of high computational costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF PETROLEUM (EAST CHINA)
- Filing Date
- 2026-05-06
- Publication Date
- 2026-06-16
Smart Images
Figure FT_1 
Figure SMS_8 
Figure SMS_17
Abstract
Description
Technical Field
[0001] This invention belongs to the field of oil and gas exploration and development and artificial intelligence logging data processing technology, specifically involving an intelligent reconstruction device for missing logging curves based on model optimization and multi-dimensional evaluation models. Background Technology
[0002] In oil exploration and development, the continuity and integrity of logging curves are crucial for accurate geological interpretation, lithology identification, fluid identification, and reservoir evaluation. During downhole operations, due to uncontrollable factors such as complex downhole environments (e.g., wellbore collapse, enlargement, narrowing, mud intrusion), instrument malfunctions, obstructed logging cable transmission, or human error, the actual logging data often exhibits severe gaps or abnormal distortions in certain depth segments. This not only disrupts the analytical coherence of individual logging curves (e.g., natural gamma ray curve GR, neutron porosity curve CNL), but also significantly limits the effective depth range for joint interpretation of multiple logging curves (e.g., porosity-permeability cross-plot analysis), resulting in substantial exploration costs and a waste of valuable geological information.
[0003] To address the issue of missing logging curves, traditional data processing methods often employ mean filling, spline interpolation, or basic regression based on regional empirical formulas. However, in complex subsurface sedimentary environments, the response relationships between different logging curves are often highly nonlinear and dynamically change with depth and lithofacies. Traditional linear interpolation methods fail to effectively capture the deep intrinsic mapping relationships between multidimensional logging data, resulting in extremely low reconstruction accuracy.
[0004] With the development of machine learning technology, algorithms such as Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Support Vector Machine (SVM) have been introduced into this field, transforming curve reconstruction into a supervised learning regression task. However, existing single-algorithm applications in well logging data processing have significant drawbacks:
[0005] First, well logging data is characterized by high sampling rates (typically 0.125 meters or less), extremely large sample sizes, and complex feature dimensions. Faced with massive amounts of well logging data, traditional tree models (such as XGBoost and RF) need to traverse all feature values of all samples when searching for the optimal split point, resulting in extremely high computational costs and huge memory consumption.
[0006] Second, existing algorithms are largely unable to natively handle the objectively existing random missing values (NaN) in well logging data. Before inputting the data into the model, manual interpolation or mean filling must be performed. This forced preprocessing can easily alter the true distribution of the original data and introduce preprocessing bias.
[0007] Third, some distance-based algorithms (such as K-nearest neighbors) or models without morphological constraints are prone to producing geologically insignificant "step-like" artifacts or severe "spiculation" jitter when predicting continuous logging curves.
[0008] Fourth, and more seriously, the evaluation systems of existing reconstruction algorithms often only focus on absolute numerical errors (such as mean square error (MSE) and root mean square error (RMSE), lacking systematic morphological constraints based on the gradual changes in formation properties and macroscopic geological trends. Real well logging curves, in most cases, exhibit good smoothness and contain low-frequency geological cycle trends. Without these multi-dimensional evaluations and constraints, the model is highly susceptible to local overfitting, and even if the generated curve has a small MSE, it deviates from the overall geological context and cannot be used for actual well logging interpretation.
[0009] Therefore, how to provide a device that can integrate multiple advanced machine learning algorithms, automatically perform hyperparameter optimization, natively handle missing values, and intelligently select the most suitable well condition model based on a multi-dimensional evaluation model (covering smoothness, correlation, trend consistency, etc.) is a core technical problem that urgently needs to be solved in the field of well logging artificial intelligence. Summary of the Invention
[0010] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide an intelligent reconstruction device for missing logging curves based on model optimization and multi-dimensional evaluation model.
[0011] The technical solution for achieving the purpose of this invention is: an intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation models, characterized in that it includes:
[0012] Step 1: Data preprocessing module, used to establish well logging data preprocessing and feature engineering models, and to divide the training set and prediction set;
[0013] Step 2: Construct a model library module to build a well logging curve prediction model library based on supervised learning. The well logging curve prediction model library includes multiple algorithms such as histogram gradient boosting (HGB), XGBoost, CatBoost, random forest (RF), support vector machine (SVM) and K-nearest neighbors (KNN).
[0014] Step 3, Model Optimization Module, is used to iteratively train the selected model based on the training set, optimize the model parameters by minimizing the objective function, and establish a nonlinear mapping relationship between the feature curve and the target curve;
[0015] Step 4: Construct a multidimensional evaluation module to establish a multidimensional evaluation model that includes smoothness, curve correlation, trend consistency, and prediction uncertainty, and to quantitatively evaluate and compare the prediction results of each candidate model under different parameters.
[0016] Step 5, Parameter Optimization Module, is used to intelligently select the model with the best overall performance and the optimal parameter combination based on the multi-dimensional evaluation results;
[0017] Step 6: Missing segment reconstruction module, which is used to input the feature data of the depth segment to be reconstructed into the optimal model to perform numerical inference of the missing segment, and to perform a seamless splicing operation of the original data and the predicted data, and output the final complete logging curve reconstruction result.
[0018] Furthermore, the data preprocessing module is specifically used for: processing the target logging curve Utilize well-developed deep segments to construct training sets Construct a prediction set using the missing depth segments. ;in This is a set of characteristic variables, containing the response values of other associated logging curves at the same depth point, excluding the target logging curve. The target variable is used; the feature data is cleaned and normalized to ensure the consistency of the feature distribution.
[0019] Furthermore, in the model library construction module and the model selection module, when the Histogram Gradient Boosting (HGB) algorithm is used as the prediction model, the specific steps are as follows: A histogram-based feature discretization strategy is adopted to discretize continuous feature values and place them into a fixed number of feature buckets; when constructing the decision tree, the feature buckets are traversed instead of the sample points, and the first-order gradient of the samples in each bucket is calculated. and second gradient Construct a gradient histogram; the formula for calculating the split gain at the optimal split point is:
[0020]
[0021] in, , These represent the sum of the gradients of the left and right child nodes, respectively; , These represent the sum of the Hessian values of the left and right child nodes, respectively. For regularization parameters; The minimum loss reduction required for node splitting; for missing values, the algorithm assigns them to a dedicated feature bucket during feature bucketing and automatically learns the optimal splitting direction containing missing values to obtain the maximum information gain during node splitting.
[0022] Furthermore, in the model optimization module, when the XGBoost algorithm is used, its first... Objective function of round iteration Represented as:
[0023]
[0024]
[0025] in, To measure the predicted value Compared with the true value The loss function for the difference; This is a regularization term used to control the complexity of the decision tree to prevent overfitting. The number of leaf nodes. The weight of the leaf node. and The penalty coefficient is used; the objective function is approximated by a second-order Taylor expansion to optimize the solution.
[0026] Furthermore, the smoothness and curve correlation evaluation model established in the multidimensional evaluation module is as follows: The smoothness index is defined as the mean of the sum of squares of the second differences of all points within the prediction segment, calculated using the following formula:
[0027]
[0028] in, To predict the number of sample points in the segment, For depth point The predicted value at that location; A smaller value indicates a smoother local variation in the curve. Curve correlation is defined using the Pearson correlation coefficient. Measuring the linear consistency between the predicted curve and the actual curve:
[0029]
[0030] in, For the true value, For predicted values, and These are the corresponding means.
[0031] Furthermore, the trend consistency and prediction uncertainty assessment model established in the multidimensional assessment module is as follows:
[0032] Trend consistency is defined as the root mean square error (TCE) between the original curve and the predicted curve on the trend line after moving average smoothing.
[0033]
[0034] in, and The true curve and the predicted curve are respectively passed through a window of size . The trend value after processing with a moving average filter; Prediction uncertainty is defined as the prediction variance of the base learners in the ensemble learning model:
[0035]
[0036] in, For the number of decision trees, For the first The predicted value of the trees, This represents the average prediction value of the ensemble model.
[0037] Furthermore, the parameter optimization module is specifically used for: recording the changes in multidimensional evaluation metrics of each algorithm under different learning rates and different training rounds within a set hyperparameter search space; analyzing the convergence state of each algorithm model; comparing the comprehensive quantitative scores of different algorithms under hyperparameter combinations where the model is fully converged and the smoothness index reaches a minimum value; and selecting the model that meets the preset accuracy threshold for curve correlation, has the highest trend consistency, and the lowest training memory consumption and time overhead as the best model, and outputting its corresponding optimal hyperparameter combination.
[0038] The present invention also claims protection for a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, performs the functions of the modules of the aforementioned device; and an electronic device comprising a memory and a processor, characterized in that the processor, when executing the program, performs the functions of the modules of the aforementioned device.
[0039] The significant advantages of this invention compared to the prior art are as follows:
[0040] (1) By introducing multi-dimensional evaluation models such as smoothness and consistency of macro-geological trends as the criteria for judgment, the "step-like" artifacts can be reduced, which is conducive to improving the consistency between the reconstructed curve and the actual geological laws.
[0041] (2) Integrate a multimodal algorithm library and establish an intelligent optimization mechanism to avoid the problem of poor generalization ability caused by blindly selecting a single algorithm;
[0042] (3) It solves the problem of missing data in a native and efficient way. Combined with histogram discretization technology, it reduces memory usage and improves training efficiency when dealing with large-scale well logging data. Attached Figure Description
[0043] Figure 1 This is a flowchart illustrating the operation of the intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation models in this invention. Detailed Implementation
[0044] The present invention will now be described in further detail with reference to the accompanying drawings.
[0045] This embodiment discloses an intelligent reconstruction device for missing well logging curves based on model optimization and multi-dimensional evaluation models. Combined with... Figure 1 The flowchart shown illustrates, through this embodiment, the theoretical basis of the invention and the entire working process of each module by addressing the data missing problem of a specific well (such as well K24) in an exploration area:
[0046] Step 1: Data Preprocessing Module
[0047] like Figure 1 As shown, the incomplete well logging dataset and the target curve to be reconstructed are first input. Due to the coupled physical characteristics of formation lithology, at the same depth point, the response value of one well logging curve will inevitably have a potential physical mapping relationship with the response values of other well logging curves. The device's bottom-level algorithm automatically scans and cleans the data of the entire well section, extracting the complete depth segment without missing data. The well logging curve to be reconstructed (taking the neutron porosity (CNL) curve as an example) is set as the target variable. Combine other high-quality correlation curves at the same depth coordinate (such as natural gamma GR, acoustic transit time AC, and deep and shallow lateral resistivity RT, etc.) into a multidimensional characteristic variable matrix. Through rigorous depth alignment, the device constructs a training set containing a massive amount of feature-object data pairs. And prediction sets with missing depth Subsequently, outlier removal and Z-score standardization were performed on the feature data to eliminate differences in physical dimensions among different logging instruments and ensure the consistency of the feature space distribution.
[0048] Step 2: Build the model library module
[0049] To overcome the limitations of a single algorithm, this invention constructs a multimodal prediction model library containing six cutting-edge machine learning algorithms in the device's underlying algorithm engine. Each algorithm is highly encapsulated and retains its complex internal derivation logic.
[0050] (1) Extreme Gradient Boosting (XGBoost): Integrates multiple regression trees through an additive model. To find the optimal split point for logging features, XGBoost introduces a second-order Taylor expansion into the objective function and adds a regularization term to suppress the complexity of the tree model. In its first The core objective of round iteration is to minimize: .
[0051] (2) Category Boosting (CatBoost): To address the gradient estimation bias (target leakage) problem in the traditional GBDT algorithm, CatBoost innovatively adopts an ordered boosting strategy. During training, it retains multiple randomly arranged sample sequences and uses the sequence preceding the current sample to calculate the unbiased gradient. Furthermore, it employs an oblivious tree structure, making the model less prone to overfitting to logging noise and resulting in extremely high execution efficiency in the prediction phase.
[0052] (3) Support Vector Machine (SVM): Transforms nonlinear logging features into a linear regression problem in a high-dimensional feature space. This is achieved by solving a Lagrangian dual optimization problem and using the RBF (Radial Basis Function) kernel function. Under the condition of meeting the tolerance deviation Given the premise, find the weight vector that makes the weight vector Hyperplane with minimized norm.
[0053] (4) K-Nearest Neighbors (KNN): As an instance-based lazy learning algorithm, it does not pre-compute model parameters. When inferring the missing value of a certain depth point, the device calculates the Euclidean distance between its feature vector and all sample points in the training set. Select the closest For each stratigraphic sample, the target predicted value is calculated using an inverse distance weighting strategy.
[0054] (5) Random Forest (RF): Based on the Bagging ensemble concept. The device uses the Bootstrap method to sample with replacement from the well logging dataset and randomly selects a subset of features to split tree nodes, constructing thousands of independent decision trees. Finally, the arithmetic mean is used to calculate the result. The output results significantly reduced the overall variance of the model.
[0055] (6) Histogram Gradient Boosting (HGB): This is the core optimization algorithm for massive logging data. Traditional GBDT requires traversing... The complexity of HGB algorithm is as follows: During initial training, the continuous logging features are divided into intervals and discretized into... There are 10 feature buckets. When constructing the tree, the algorithm only traverses each feature bucket. Within each bucket, the first and second gradients are accumulated to form a histogram. This mechanism exponentially reduces computational complexity. In particular, when faced with missing values (NaN) that are objectively present in well logging, HGB assigns them to a dedicated "missing feature bucket" and automatically places them into the left or right subtree during splitting to maximize gain, achieving native intelligent missing value handling.
[0056] Step 3, Model Optimization Module
[0057] Within the hyperparameter search space defined by the device (e.g., setting the learning rate to a range of [0.01, 0.05, 0.1, 0.5, 0.9] and the number of training epochs to 500 to 5000), the device uses the training set... The six algorithms from the aforementioned library are executed in parallel for fitting. Each algorithm continuously updates the network weights and tree structure according to its own loss function, establishing a deep nonlinear mapping from known features to the target logging response value in the multidimensional logging feature space.
[0058] Step 4: Construct a multi-dimensional evaluation module
[0059] The uniqueness of well logging data lies in its representation of the evolution of the physical properties of continuously deposited strata. To overcome the "geological distortion" caused by traditional assessments based solely on MSE (Mean Sedimentary Sedimentation), this invention constructs a multi-dimensional evaluation mechanism within the device:
[0060] (A) Smoothness: defined as Natural sedimentary rock layers are in most cases a continuous and gradual change. This index uses a second-order difference effective penalty model to predict unreasonable "high-frequency jitter" and "spiculation".
[0061] (B) Curve correlation (Pearson R): measures the degree of linear consistency between the predicted logging sequence and the actual sequence in terms of macroscopic fluctuations. The closer the value is to 1, the better the model reproduces the fine features of the formation peaks and troughs.
[0062] (C) Trend Consistency (TCE): Using moving average filtering to extract low-frequency components (i.e. large-scale sedimentary cycles or compaction trends) in well logging sequences. The smaller the value, the more accurate the prediction curve is locally, and the less its macroscopic geological evolution background has shifted.
[0063] (D) Prediction Variance: In an ensemble model, the prediction variance of multiple trees for the same depth is calculated. A larger variance indicates that the data characteristics at that depth are unusual, the model confidence is low, and it can be used to guide subsequent manual intervention.
[0064] Step 5, Parameter Optimization Module
[0065] The device collects a large amount of validation experimental data in the background for horizontal and vertical comparisons:
[0066] In the longitudinal comparison of hyperparameter optimization, for the gradient boosting family (XGBoost, CatBoost, HGB), the data shows that as the learning rate decreases from 0.9 to 0.01, the model's "curve correlation" climbs from around 0.5 to its peak; simultaneously, the "smoothness" score also drops sharply and reaches a minimum (the curve tends to become naturally smooth). Therefore, the device automatically locks 0.01 as the optimal learning rate. Analysis of training epochs shows that HGB converges to its optimal performance across all metrics after 1000 epochs; while RF, due to its Bagging characteristics, requires an extremely large number of 5000 trees to stabilize the variance.
[0067] In the horizontal comparison of algorithm models, the device outputs a rigorous quantification report:
[0068] (1) Support Vector Machine (SVM) has the best smoothness performance (score of 0.4577), but its curve correlation is low (only 0.6206). Excessive smoothing causes it to lose a lot of detailed information about thin interlayers in the strata.
[0069] (2) Due to the simple inverse distance weighting, the K-Nearest Neighbor (KNN) algorithm causes the reconstructed curve to completely lose continuity and presents a "step-like" artifact, resulting in the reconstruction result being inconsistent with the actual continuous change characteristics of the strata, and its correlation is the lowest (0.5604).
[0070] (3) Random forest (RF) achieved a high curve correlation (0.6841) due to its strong ensemble nonlinear fitting ability. However, the large number of trees that need to be built leads to serious memory consumption of the device and a large training time cost.
[0071] (4) Ultimately, the Histogram Gradient Boosting (HGB) model demonstrated superior overall performance in the comprehensive comparison. Its curve correlation was as high as 0.6749, its trend consistency TCE was excellent (0.3274), and its smoothness was robust (1.1879). More significantly, when faced with the massive amount of logging data from well K24, thanks to the histogram discretization optimization of HGB, its training speed was several times faster than RF, and its memory usage was significantly reduced. Based on the principle of "achieving the required accuracy while minimizing engineering overhead," the device automatically selected HGB combined with the optimal parameters (learning rate 0.01, training epochs 1000) as the best prediction model for this task.
[0072] Step 6: Missing Reconstruction Module
[0073] like Figure 1 As shown in the final step, the core processing module of the device will predict the set. (i.e., the effective correlation feature curve where the target curve is within the missing depth range) is input into the trained and converged optimized HGB model. The model predicts the response numerical sequence at the missing depth. Finally, the device performs rigorous splicing and alignment of the complete retained portion of the original curve with the reconstructed portion generated by the model on the physical logging depth sequence. Thanks to the strong constraints of multi-dimensional geological indicators (especially smoothness and trend consistency), the numerical transition at the splicing junction is extremely smooth and natural, achieving seamless depth fusion. The device ultimately derives a complete, continuous, high-precision reconstructed logging curve with clear formation physical meaning, providing data support for downstream reservoir comprehensive evaluation and 3D geological modeling.
[0074] The above content describes the theoretical basis, comprehensive architecture, implementation process, and multi-dimensional evaluation advantages of this invention. Those skilled in the art should understand that, without departing from the core principles of multi-algorithm optimization and multi-dimensional geological assessment, the algorithm engine library of this invention can be further expanded, and the weights of the evaluation indicators can be fine-tuned. These variations and improvements to the isomorphic architecture all fall within the scope of this invention as claimed.
Claims
1. A smart reconstruction device for missing logging curves based on model optimization and multidimensional evaluation models, characterized in that, include: Step 1: Data preprocessing module, used to establish well logging data preprocessing and feature engineering models, and to divide the training set and prediction set; Step 2: Construct a model library module to build a well logging curve prediction model library based on supervised learning. The well logging curve prediction model library includes multiple algorithms such as histogram gradient boosting (HGB), XGBoost, CatBoost, random forest (RF), support vector machine (SVM) and K-nearest neighbors (KNN). Step 3, Model Optimization Module, is used to iteratively train the selected model based on the training set, optimize the model parameters by minimizing the objective function, and establish a nonlinear mapping relationship between the feature curve and the target curve; Step 4: Construct a multidimensional evaluation module to establish a multidimensional evaluation model that includes smoothness, curve correlation, trend consistency, and prediction uncertainty, and to quantitatively evaluate and compare the prediction results of each candidate model under different parameters. Step 5, Parameter Optimization Module, is used to intelligently select the model with the best overall performance and the optimal parameter combination based on the multi-dimensional evaluation results; Step 6: Missing segment reconstruction module, which is used to input the feature data of the depth segment to be reconstructed into the optimal model to perform numerical inference of the missing segment, and to perform a seamless splicing operation of the original data and the predicted data, and output the final complete logging curve reconstruction result.
2. The intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation model according to claim 1, characterized in that, The data preprocessing module is specifically used for: processing the target logging curve Utilize well-developed deep segments to construct training sets Construct a prediction set using the missing depth segments. ;in This is a set of characteristic variables, containing the response values of other associated logging curves at the same depth point, excluding the target logging curve. The target variable; The feature data is cleaned and normalized, and the processing logic is encapsulated in the core data processing component to ensure the consistency of feature distribution.
3. The intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation model according to claim 1, characterized in that, In the model library construction module and model selection module, when the Histogram Gradient Boosting (HGB) algorithm is used as the prediction model, the specific steps are as follows: A histogram-based feature discretization strategy is adopted to discretize continuous feature values and place them into a fixed number of bins; when constructing the decision tree, the bins are traversed instead of the sample points, and the first-order gradient of the samples within each bin is calculated. and second gradient Construct a gradient histogram; The formula for calculating the split gain at the optimal split point is: ； in, , These represent the sum of the gradients of the left and right child nodes, respectively; , These represent the sum of the Hessian values of the left and right child nodes, respectively. For regularization parameters; The minimum loss reduction required for node splitting; for missing values, the algorithm assigns them to a dedicated bucket during feature bucketing and automatically learns the optimal splitting direction for missing values during splitting.
4. The intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation model according to claim 1, characterized in that, In the model optimization module, when the XGBoost algorithm is used, its first... Objective function of round iteration Represented as: ； ； in, To measure the predicted value Compared with the true value The loss function for the difference; This is a regularization term used to control the complexity of the tree; The number of leaf nodes. The weight of the leaf node. and The penalty coefficient is used; the objective function is approximated by a second-order Taylor expansion to optimize the solution.
5. The intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation model according to claim 1, characterized in that, The smoothness and curve correlation index established in the multidimensional evaluation module is as follows: The smoothness index is defined as the mean of the sum of squares of the second differences of all points within the prediction segment, and the calculation formula is: ； in, To predict the number of sample points in the segment, For depth point The predicted value at that location; A smaller value indicates a smoother local change in the curve; the Pearson correlation coefficient is used to define curve correlation. Measuring the linear consistency between the predicted curve and the actual curve: ； in, For the true value, For predicted values, and These are the corresponding means.
6. The intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation model according to claim 1, characterized in that, The trend consistency and prediction uncertainty indices established in the multi-dimensional evaluation module are as follows: Trend consistency is defined as the root mean square error (TCE) between the original curve and the predicted curve on the trend line after moving average smoothing. ； in, and The true curve and the predicted curve are respectively passed through a window of size . The trend value after processing with a moving average filter; Prediction uncertainty is defined as the prediction variance of the base learners in the ensemble learning model: ； in, For the number of decision trees, For the first The predicted value of the trees, This represents the average prediction value of the ensemble model; The larger the value, the lower the model confidence level.
7. The intelligent reconstruction device for missing logging curves based on model optimization and multidimensional evaluation model according to claim 1, characterized in that, The parameter optimization module is specifically used for: Within the defined hyperparameter search space, the changes in multidimensional evaluation metrics for each algorithm under different learning rates and training epochs are recorded. The convergence state of each algorithm model is analyzed. Under the hyperparameter combination where the model is fully converged and the smoothness index reaches a minimum value, the comprehensive quantitative scores of different algorithms are compared. The model that satisfies the preset accuracy threshold for curve correlation, has the highest trend consistency, and the lowest training memory consumption and time overhead is selected as the best model, and its corresponding optimal hyperparameter combination is output.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program performs the functions of each module of the apparatus as described in any one of claims 1-7.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it performs the functions of each module of the device as described in any one of claims 1-7.