A Fluid Identification Method for Igneous Rocks Based on CXDF Algorithm
By combining the CXDF algorithm with the XGBoost and DF algorithms, the data imbalance problem in the identification of fluids in complex igneous reservoirs was solved, and the accurate identification of fluids in complex igneous reservoirs was achieved, thus improving the identification effect.
Patent Information
- Application Number
- CN202411153892.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-08-21
AI Technical Summary
Existing traditional methods are limited by lithological inconsistencies and formation water salinity similarities in fluid identification in complex igneous reservoirs, making fluid identification difficult, and machine learning methods have not been widely applied to such reservoirs.
The CXDF algorithm based on the adaptive multi-objective group cross-optimization algorithm (AMSCO) is adopted, combined with the ensemble learning method of extreme gradient boosting tree (XGBoost) and deep forest (DF) algorithm. Through data preprocessing and model optimization, the imbalance problem of the initial dataset is solved, and fluid identification is realized.
The accuracy and recall of fluid identification in complex igneous rock reservoirs were improved, the dataset was optimized, the generalization ability of the model was enhanced, and effective identification of fluids in complex igneous rock reservoirs was achieved.
Smart Images

Figure CN119132460B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geophysics, specifically to a method for identifying fluids in igneous rocks based on the CXDF algorithm. Background Technology
[0002] Igneous reservoirs are widely distributed in major oil-bearing basins across the globe. With advancements in oil and gas exploration and development technologies, the industry is increasingly expanding into igneous reservoirs with complex lithologies, opening new avenues for increasing oil and gas production and reserves. Accurate identification of reservoir fluids is crucial for the effective development of such reservoirs. Traditional methods for reservoir fluid identification include standard water layer comparison, minimum reservoir resistivity, radial resistivity, and adjacent well curve comparison. However, these methods are often limited by reservoir conditions, such as lithological inconsistencies and similar formation water salinity, leading to difficulties in fluid identification. Although various machine learning methods (such as support vector machines, extreme gradient boosting, and random forests) have been used for reservoir evaluation, there is currently no research on applying machine learning methods to fluid identification in complex igneous reservoirs.
[0003] Ensemble learning methods combine multiple weak learners into a single strong learner, preserving the characteristics of weak learners while improving the overall model performance. Typical ensemble learning methods include Random Forest, AdaBoost, GBRT, and XGBoost. These methods have been widely applied in lithology identification, thin-layer identification, and reservoir property inversion. Ensemble learning methods offer potential advantages for the development of complex igneous reservoirs. However, the imbalance of conventional well logging datasets poses a challenge to fluid identification, as classification models based on imbalanced data may be biased towards the majority class, necessitating dataset rebalancing. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method for identifying igneous rock fluids based on the CXDF algorithm. This invention solves the problem of imbalanced initial datasets by using an adaptive multi-objective swarm optimization algorithm and employs an ensemble learning method combining extreme gradient boosting tree and deep forest algorithms, enabling effective identification of fluids in complex igneous rock reservoirs.
[0005] This invention provides a method for identifying igneous rock fluids based on the CXDF algorithm, which specifically includes the following modules:
[0006] S1 Data Preprocessing: Obtain core and well logging data from the study area and organize the data into an XLSX format table file; take the logarithm of the specified well logging data; normalize the well logging data; 78% of the preprocessed data is the training dataset and 22% is the validation dataset; divide the labels of the training dataset into two parts, namely high-hydrocarbon and low-hydrocarbon.
[0007] S2 Model Establishment: Based on the CXDF algorithm and other artificial intelligence algorithms, a fluid identification model is constructed; a simulated well is constructed based on the validation dataset; and the simulated well is identified according to the fluid identification model.
[0008] S3 Model Optimization: Model evaluation and comparison are performed based on tables, radar charts, and bar charts; the optimal fluid recognition model is determined.
[0009] Preferably, the fluid types in the study area described in S1 include dry layers, low-yield oil layers, water layers, oil layers, low-yield water layers, oil-bearing water layers, and gas layers.
[0010] Preferably, the lithology of the study area described in S1 includes basalt, trachyte, diabase, andesite, and tuff.
[0011] Preferably, the logging data described in S1 includes neutron (CNL), density (DEN), natural gamma (GR), deep lateral resistivity (RLLD), and micro lateral resistivity (RMLL). Specifically, the logging data specified in S1 are deep lateral resistivity (RLLD) and micro lateral resistivity (RMLL).
[0012] Preferably, the multi-hydrocarbon layer mentioned in S1 includes an oil layer and a gas layer; the low-hydrocarbon layer mentioned in S1 includes a dry layer, a low-yield oil layer, a water layer, a low-yield water layer, and an oil-bearing water layer.
[0013] Preferably, the CXDF algorithm described in S2 is an optimized dataset based on the adaptive multi-objective group cross-optimization (AMSCO) algorithm, combined with an ensemble learning model of the XGBoost algorithm and the deep forest (DF) algorithm.
[0014] Preferably, the AMSCO algorithm proposes an adaptively optimized data sampling model to address the data imbalance problem in classification scenarios. To leverage the model's adaptive characteristics, a composite evaluation criterion of reliable accuracy (RA) is proposed.
[0015] Preferably, the rebalancing process of the AMSCO algorithm specifically includes: reading the preprocessed training dataset as the current training dataset; evaluating the reliable accuracy based on the current training dataset; dividing the current training dataset that does not meet the criteria into two subsets, a minority class and a majority class; optimizing the minority class subset using the oversampling technique (OSMOTE) and optimizing the majority class subset using the group instance selection technique (SIS(o_major)); coupling the two optimizations together as a unified iterator, that is, gradually enhancing the data mixing effect of the two optimizations by iterating and crossing the results of the two optimizations until a high-quality dataset is generated.
[0016] Preferably, the radar image described in module S3 is based on F betaThe three evaluation metrics—precision, recall, and accuracy—demonstrate the fluid identification model's performance in identifying seven fluids in a simulated well. Among them, the aforementioned F... beta It is a statistical metric used to evaluate the accuracy of binary classification models, with different beta values corresponding to different weights of recall and precision.
[0017] Furthermore, based on the combined recognition results of tables, radar charts, and bar charts, the optimal igneous fluid recognition model is determined to achieve accurate recognition of igneous fluids in complex lithologies. Attached Figure Description
[0018] Figure 1 The flowchart illustrates a method for identifying igneous rock fluids based on the CXDF algorithm provided by this invention.
[0019] Figure 2 This is a flowchart illustrating the rebalancing training dataset for the AMSCO algorithm provided in an embodiment of the present invention.
[0020] Figure 3 This is a cross-plot of the SVM classification boundaries of the preprocessed and rebalanced training datasets provided in this embodiment of the invention.
[0021] Figure 4 The diagram shows the structure of the XDF algorithm provided in this embodiment of the invention.
[0022] Figure 5 A radar image of the fluid identification model for model well fluid identification provided in an embodiment of the present invention.
[0023] Figure 6 A bar chart illustrating the fluid identification model for model wells provided in this embodiment of the invention. Detailed Implementation
[0024] To enable those skilled in the art to more fully understand the technical solutions of this specification, the embodiments of this specification will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments in this specification, and not all of them. All other embodiments obtained by those skilled in the art based on one or more embodiments in this specification without creative effort should fall within the protection scope of the embodiments in this specification.
[0025] This invention provides a method for identifying igneous rock fluids based on the CXDF algorithm, which specifically includes the following modules:
[0026] S1: Data preprocessing;
[0027] S2: Model building;
[0028] S3: Model optimization.
[0029] Specifically, the S1 module includes the following steps:
[0030] (1) Obtain core and well logging data of the study area and organize the data into a table file in xlsx format;
[0031] (2) Take the logarithm of the specified well logging data;
[0032] (3) Normalize the logging data;
[0033] (4) 78% of the preprocessed data is the training dataset and 22% is the validation dataset;
[0034] (5) Divide the labels of the training dataset into two parts, namely, high hydrocarbon and low hydrocarbon.
[0035] Specifically, the S2 module includes the following steps:
[0036] (6) Construct a fluid recognition model based on the CXDF algorithm and other artificial intelligence algorithms;
[0037] (7) Construct simulated wells based on the validation dataset;
[0038] (8) Identify the simulated well based on the fluid identification model.
[0039] Specifically, the S3 module includes the following steps:
[0040] (9) Model evaluation and comparison are based on tables, radar charts and bar charts;
[0041] (10) Determine the optimal fluid recognition model.
[0042] The fluid types in the study area include dry layers, low-yield oil layers, water layers, oil layers, low-yield water layers, oil-bearing water layers, and gas layers.
[0043] The lithology of the study area includes basalt, trachyte, diabase, andesite, and tuff.
[0044] The logging data includes neutron (CNL), density (DEN), natural gamma (GR), deep lateral resistivity (RLLD), and micro lateral resistivity (RMLL).
[0045] As shown in Table 1, the core and logging data of the study area were organized into an .xlsx format table file based on Excel to facilitate subsequent processing.
[0046] Table 1: Partial Initial Dataset
[0047]
[0048] The specified logging data are deep lateral resistivity (RLLD) and micro lateral resistivity (RMLL). The RLLD and RMLL are logarithmically processed using the "LOG" function in the Excel spreadsheet. For example, the initial value of RLLD is 6.720, and the processed value is 0.827; the initial value of RMLL is 5.310, and the processed value is 0.725.
[0049] Specifically, based on equation (1), the logging data is normalized as follows:
[0050]
[0051] In the formula, X represents the initial dataset. represents the mean of the dataset, and s represents the standard deviation of the dataset.
[0052] As shown in Table 2, in the preprocessed dataset, the number of oil layer samples is significantly higher than that of other fluid samples, while gas layer samples account for only 1.98% of the total sample size. This dataset exhibits a clear data imbalance characteristic.
[0053] Table 2: Distribution of reservoir types after pretreatment
[0054]
[0055] Multi-hydrocarbon reservoirs include oil and gas reservoirs; low-hydrocarbon reservoirs include dry reservoirs, low-yield oil reservoirs, water reservoirs, low-yield water reservoirs, and oil-bearing water reservoirs.
[0056] The CXDF algorithm and other artificial intelligence algorithms were used to train the training dataset separately, and corresponding fluid recognition models were established for each. The fluid recognition models built by other artificial intelligence algorithms served as a control group, and were compared with the fluid recognition model built by the CXDF algorithm to select the best model.
[0057] The CXDF algorithm is an optimization dataset based on the Adaptive Multi-Objective Group Cross-Optimization (AMSCO) algorithm, combined with an ensemble learning model that integrates the Extreme Gradient Boosting Tree (XGBoost) algorithm and the Deep Forest (DF) algorithm.
[0058] The core optimizer of the AMSCO algorithm is Particle Swarm Optimization (PSO). PSO is a typical optimization algorithm for simulating the foraging process of birds, and it has the advantages of simple implementation, fast convergence speed and few parameters.
[0059] The optimization process for PSO is as follows:
[0060] Read the initial dataset as the current dataset; verify the quality of the dataset through candidate classifiers; perform parallel optimization on the dataset twice to change the sample distribution until the performance of the candidate classifiers reaches the threshold.
[0061] Specifically, the two parallel optimizations generated two candidate solutions with different performances. By selecting the optimal dataset instance and performing cross-output merging, a comprehensive dataset with a size close to the initial dataset was finally formed. An iterative optimization strategy was employed during the optimization process. When the performance of a candidate classifier fell below a preset threshold, the dataset was re-split and two parallel population optimization processes were performed.
[0062] During the iterative optimization process, each particle iterates according to its own velocity and position. The mathematical process of particle iteration is as follows:
[0063]
[0064]
[0065] In the formula, ω is the inertia weight, d is the spatial dimension, i is the number of particles, j is the iteration number, c1 and c2 are the velocity weights, r1 and c2 are the random weights, and V is the velocity of the current particle. At the beginning of the iteration, the particle parameters are random, P is the fitness index, and X is the set of all particles.
[0066] In a preferred embodiment of the present invention, the AMSCO algorithm proposes an adaptively optimized data sampling model aimed at solving the data imbalance problem in classification scenarios. To address the adaptive characteristics of the model, a composite evaluation criterion for reliable accuracy (RA) is proposed, the expression of which is as follows:
[0067]
[0068]
[0069]
[0070]
[0071]
[0072] In the formula, Accuracy is the accuracy rate, Kappa is the confidence index, P represents positive, N represents negative, TP represents true positive, TN represents true negative, FP represents false positive, FN represents false negative, P0 represents the percentage of concordance, and P c It is a consistency trend indicator.
[0073] The rebalancing process of the AMSCO algorithm specifically includes: reading the preprocessed training dataset as the current training dataset; evaluating reliable accuracy based on the current training dataset; dividing the current training dataset that does not meet the criteria into two subsets: minority and majority classes; optimizing the minority subset using the oversampling technique (OSMOTE) and the majority subset using the group instance selection technique (SIS(o_major)); coupling the two optimizations together as a unified iterator, that is, gradually enhancing the data mixing effect of the two optimizations by iteratively crossing the results of the two optimizations until a high-quality dataset is generated, such as... Figure 2 As shown. In this process, there are three types of evaluation rules:
[0074]
[0075]
[0076]
[0077] In equations (4)-(6) above, each subgroup partitions and searches the space in parallel and shares the best information to find the optimal solution in the shortest time.
[0078] The OSMOTE technique inserts the generated samples into any one or a few nearest neighbor segments of the minority class within the data range and optimizes them using PSO; the SIS(o_major) technique uses the learner's classification evaluation metric to select the optimal length between instances.
[0079] The AMSCO algorithm has the following beneficial effects:
[0080] (1) The dataset is gradually optimized by rebalancing instances through swarm optimization until the optimal state is reached.
[0081] (2) Optimize the imbalanced dataset by using evaluation metrics to make it a balanced dataset.
[0082] (3) Balance the dataset without changing the size of the dataset.
[0083] Furthermore, a cross plot comparing the classification boundaries of the preprocessed and rebalanced well logging data was drawn based on the support vector machine (SVM) algorithm to intuitively evaluate the optimization effect of the rebalanced dataset.
[0084] Specifically, the Support Vector Machine (SVM) algorithm is based on the RBF kernel function, and the accuracy metric is used to compare the performance of the two-dimensional classification boundary. The well logging data mentioned above uses natural gamma (GR) and neutron (CNL) data. Accuracy represents the overall predictive ability of the model, as shown in Equation (5).
[0085] Figure 3 In the diagram, blue scatter dots represent the majority class of the dry layer data, and a blue background represents the classification range of the dry layer data; red scatter dots represent the minority class of the oil layer data, and a red background represents the classification range of the oil layer data. After rebalancing, the classification range of the oil layer expanded, and the classification boundary shifted towards the dry layer. Simultaneously, the classification accuracy improved from 76.596% to 82.879%. This indicates that after the AMSCO algorithm rebalanced the dataset, the classification range of the minority class was expanded, the classification boundary was compressed into the majority class region, and the classification accuracy was improved.
[0086] Other artificial intelligence algorithms in this invention include the SVM algorithm, the XGBoost algorithm, and the XDF algorithm. The specific learning process of each algorithm is as follows:
[0087] Support Vector Machine (SVM) is a machine learning algorithm designed to separate data points into different classes by finding the optimal hyperplane. SVM utilizes support vectors from the training dataset to establish the optimal decision boundary and can handle both linear and nonlinear classification problems as well as regression modeling. When dealing with nonlinear problems, SVM introduces a kernel function to map the input data to a high-dimensional feature space, thereby performing inner product operations in this higher-dimensional space to improve classification and regression capabilities.
[0088] The SVM algorithm in this embodiment of the invention is a machine learning algorithm based on the radial basis function (RBF kernel function), and the specific formula is as follows:
[0089]
[0090] In the formula, K(x) i ,x j ) is the inner product kernel, x i and x j These are two distinct samples; σ is a free parameter.
[0091] Extreme Gradient Boosting (XGBoost) is an ensemble learning method, belonging to the additive model within the boosting algorithm family. XGBoost effectively suppresses overfitting by combining gradient boosting with L1 and L2 regularization, thereby enhancing its generalization ability. Its calculation formula is as follows:
[0092]
[0093] In the formula, This represents the prediction result for sample i after the t-th iteration of the model. f represents the prediction results of the first t-1 weak classifiers. t (X iLet F be the model of the t-th weak classifier, and F be the set of all possible weak classifiers.
[0094] Specifically, the objective function of the XGBoost algorithm consists of a loss function and a regularization term. The objective function is minimized through iteration, and the expression of the objective function after iteration is as follows:
[0095]
[0096]
[0097] In the formula, L is the loss function, n is the number of samples, and Obj (t) Let Ω be the objective function for the t-th iteration, and let Ω be the regularization term to suppress model complexity and prevent overfitting.
[0098] Specifically, the characteristic of the Boosting algorithm is that it integrates multiple decision trees (weak classifiers) together to build a strong classifier. The decision tree uses the CART regression tree model, and its implementation process includes the following steps:
[0099] The first step is to determine the optimal splitting variable j and splitting point s in order to obtain the (j,s) with the minimum error.
[0100]
[0101] The second step involves using the optimal (j,s) to divide the sample features into two nodes, and repeating the above steps until the termination condition is met. Finally, the input space is divided into M regions, thereby generating a decision tree.
[0102] c m =average(y i |x i ∈R m (j,s)) (17)
[0103]
[0104] The XDF algorithm is an ensemble learning model that combines the Extreme Gradient Boosting Tree (XGBoost) algorithm and the Deep Forest (DF) algorithm. The essential difference between the XDF algorithm and the CXDF algorithm is that the XDF algorithm does not use the AMSCO algorithm to rebalance the initial dataset.
[0105] Specifically, Deep Forest (DF) is a multi-layer ensemble learning method based on decision trees, where each layer consists of multiple forests. During the data sampling phase, Deep Forest employs a multi-granularity scanning method to enrich the sample dataset. In the basic Deep Forest model, the output feature vectors of each layer are averaged to synthesize the output vector of that layer. The number of layers in a Deep Forest is not a fixed hyperparameter but can be automatically adjusted according to the complexity of the dataset. Furthermore, Deep Forest exhibits strong robustness, typically achieving good performance using default parameters, indicating that no special parameter optimization is required when applying Deep Forest.
[0106] In this embodiment of the invention, for continuous data, there is usually a strong spatial correlation between adjacent data. To improve the performance of the classifier, a multi-granularity scanning method can be used to fill the data samples. Specifically, taking a data sequence of length 600 as an example, the scanning window length is set to 100, the step size is set to 1, and the final generated data length is 3006, as shown below. Figure 4 As shown.
[0107] Figure 4 In this process, data is input into a deep forest structure, which consists of two or more different types of random forests at each layer: a random forest and a full random forest. A full random forest contains 1000 decision trees, each randomly selecting a feature when choosing a branch node. A random forest also contains 1000 decision trees, but each tree selects the square root of the total number of features in the dataset, using the Gini coefficient to determine branch nodes. In each layer of the forest, each decision tree generates a feature vector, which are then aggregated using extreme gradient boosting to form the output feature vector for that layer. After training each layer, K-fold cross-validation is used to validate the sample data. The entire deep forest structure outputs layer by layer, and the final results are aggregated using the XGBoost algorithm as a weak learner.
[0108] Furthermore, based on 226 sets of data for 7 fluid types in the validation dataset, a reservoir model with a depth of 28.126 meters was constructed using the simulated well.
[0109] The tables, based on two evaluation metrics, precision and recall, show the effectiveness of the fluid identification model in identifying seven fluids in the simulated well, as shown in Tables 3, 4, 5, and 6.
[0110] Precision represents the accuracy of a model's predictions of positive samples, and its expression is as follows:
[0111]
[0112] Recall represents the model's ability to identify positive samples, and its expression is as follows:
[0113]
[0114] Table 3: Reservoir Identification Results of the CXDF Model
[0115]
[0116] Table 4: Reservoir Identification Results of SVM Model
[0117]
[0118] Table 5: Reservoir Identification Results of XGboost Model
[0119]
[0120] Table 6: Reservoir Identification Results of XDF Model
[0121]
[0122] Tables 3-6 show that the fluid identification models constructed by the SVM, XGBoost, and XDF algorithms performed poorly in identifying water layer data, with most water layers being misidentified as low-yield oil reservoirs or oil reservoirs. The recall rates of these three methods were 0.104, 0.125, and 0.042, respectively. The CXDF algorithm showed improvement in this regard, correctly identifying all 48 water layer samples. Meanwhile, in the control group, gas layers were also misidentified as oil layers. Specifically, the SVM algorithm failed to identify gas layers; the XGBoost algorithm accurately identified only 13 out of 30 gas layer samples; and the XDF algorithm accurately identified 11 gas layer samples. In contrast, the CXDF algorithm accurately identified gas layer data, with both precision and recall of 1. In the CXDF algorithm's identification results, the precision for low-yield water reservoirs was 0.769, mainly due to the insufficient dataset for low-yield water reservoirs, which also affected the recall rate for oil reservoirs. Furthermore, there were errors in the identification between water layers and oil-bearing water layers. Four oil-bearing water layer samples were misidentified as water layers, which is related to the similarity of the physical properties of the two fluids, but the error is within an acceptable range. Overall, the CXDF method improves the model's precision and recall metrics by addressing data imbalance.
[0123] Radar image based on F beta The three evaluation metrics—precision, recall, and accuracy—demonstrate the fluid identification model's ability to identify seven fluids in a simulated well.
[0124] F betaBeta is a statistical metric used to evaluate the accuracy of binary classification models. It is a comprehensive evaluation standard designed to reflect the overall performance of the classifier. Different beta values correspond to different weights for recall and precision, thus adjusting the emphasis of model evaluation. Its expression is as follows:
[0125]
[0126] Figure 5 In the diagram, each radar chart represents a fluid category, composed of prediction rate, recall rate, and F... 0.5 It consists of evaluation indicators such as F1, F2, etc. Multiple F... beta Evaluation metrics can assess the fluid identification performance of the model from different perspectives. The red broken line in the figure represents the evaluation metrics of the CXDF algorithm. Although its identification performance of low-yield water layers is somewhat inferior to other methods, overall, the CXDF algorithm outperforms other methods in the control group in the radar chart.
[0127] Figure 6 The results show that the CXDF algorithm produced errors in identifying oil layers and the boundaries between water-bearing oil layers and water layers. According to the conventional curve, incorrectly identified layers exhibited corresponding fluctuations on the curve. However, the identification results for other fluids were quite accurate. Meanwhile, the identification results of the control group were inferior to the CXDF method in all cases.
[0128] In summary, the CXDF algorithm outperforms other artificial intelligence algorithms in terms of tables, radar charts, and bar charts. A fluid identification model based on the CXDF algorithm is constructed to achieve accurate identification of fluids in igneous rocks with complex lithology.
[0129] Those skilled in the art should understand that the above one or more embodiments are only used to describe the technical solutions of the present invention, and not to limit them. Various modifications and variations can be made to one or more embodiments of this specification without causing the essence of the corresponding technical solutions to depart from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for identifying igneous rock fluids based on the CXDF algorithm, characterized in that, Specifically, it includes the following modules: S1 Data Preprocessing: Core and well logging data of the study area were acquired and organized into XLSX format spreadsheet files. The well logging data included neutron (CNL), density (DEN), natural gamma (GR), deep lateral resistivity (RLLD), and micro lateral resistivity (RMLL). Logarithmic values were taken from the specified well logging data, which were RLLD and RMLL. The well logging data were then normalized. 78% of the preprocessed data was used as the training dataset, and 22% as the validation dataset. The training dataset labels were divided into two parts: high-hydrocarbon and low-hydrocarbon. High-hydrocarbon layers included oil and gas layers; low-hydrocarbon layers included dry layers, low-yield oil layers, water layers, low-yield water layers, and oil-bearing water layers. S2 Model Establishment: A fluid recognition model is constructed based on the CXDF algorithm; the CXDF algorithm is an optimized dataset based on the adaptive multi-objective group cross-optimization AMSCO algorithm, combined with an ensemble learning model of the XGBoost algorithm and the deep forest algorithm; The rebalancing process of the AMSCO algorithm specifically includes: reading the preprocessed training dataset as the current training dataset; evaluating the reliability accuracy based on the current training dataset; dividing the current training dataset that does not meet the criteria into two subsets, a minority class and a majority class; optimizing the minority class subset using the oversampling technique OSMOTE, and optimizing the majority class subset using the group instance selection technique SIS(o_major); coupling the two optimizations together as a unified iterator, that is, gradually enhancing the data mixing effect of the two optimizations by iterating and crossing the results of the two optimizations until a high-quality dataset is generated; constructing simulated wells based on the validation dataset; and identifying the simulated wells according to the fluid identification model. S3 Model Optimization: Model evaluation and comparison are performed based on tables, radar charts, and bar charts to determine the optimal fluid recognition model.
2. The method for identifying igneous rock fluids based on the CXDF algorithm according to claim 1, characterized in that: The fluid types in the study area described in S1 include dry layers, low-yield oil layers, water layers, oil layers, low-yield water layers, oil-bearing water layers, and gas layers.
3. The method for identifying igneous rock fluids based on the CXDF algorithm according to claim 1, characterized in that: The lithology of the study area described in S1 includes basalt, trachyte, diabase, andesite, and tuff.
4. The method for identifying igneous rock fluids based on the CXDF algorithm according to claim 1, characterized in that: The AMSCO algorithm proposes an adaptively optimized data sampling model to address the data imbalance problem in classification scenarios. Based on the model's adaptive characteristics, a composite evaluation criterion for reliable accuracy (RA) is proposed, expressed as follows: Wherein, Accuracy is the accuracy rate, Kappa is the reliability index, P is positive, N is negative, TP is true positive, TN is true negative, FP is false positive, FN is false negative, P0 is the consistency percentage, P c is the consistency trend index.
5. The method for identifying igneous rock fluids based on the CXDF algorithm according to claim 1, characterized in that: The radar image described in S3 is based on F beta The three evaluation metrics—precision, recall, and accuracy—demonstrate the fluid identification model's effectiveness in identifying seven fluids in a simulated well. Meanwhile, the F... beta Beta is a statistical metric used to evaluate the accuracy of binary classification models. Different beta values correspond to different weights for recall and precision, and its expression is as follows: Precision is the accuracy rate, and Recall is the recall rate.
6. The method for identifying igneous rock fluids based on the CXDF algorithm according to claim 1, characterized in that: Based on Equation 1, the logging data is normalized, and the expression is as follows: In the formula, X represents the initial dataset. Let S represent the mean of the dataset, and let S represent the standard deviation of the dataset.
7. The method for identifying igneous rock fluids based on the CXDF algorithm according to claim 1, characterized in that: The specified logging data are deep lateral resistivity (RLLD) and micro lateral resistivity (RMLL), where the logarithm of RLLD and RMLL is taken using the LOG function in the Excel spreadsheet.
8. The method for identifying igneous rock fluids based on the CXDF algorithm according to claim 1, characterized in that: The data is input into a deep forest structure, which consists of two or more different types of random forests at each layer: a random forest and a full random forest. A full random forest contains 1000 decision trees, each randomly selecting a feature when choosing a branch node. A random forest also contains 1000 decision trees, but each tree selects the square root of the total number of features in the dataset, using the Gini coefficient to determine branch nodes. In each layer of the forest, each decision tree generates a feature vector, which are then aggregated using extreme gradient boosting to form the output feature vector for that layer. After training at each layer, K-fold cross-validation is used to validate the sample data. The entire deep forest structure outputs layer by layer, and the final results are aggregated using the XGBoost algorithm as a weak learner.
Citation Information
Patent Citations
Logging lithology identification method based on deep belief network
CN109736790A
Logging reservoir identification and prediction method based on machine learning
CN113642772A
