Inematic rock lithology identification method based on engineering optimization cascade forest

Abstract

The invention discloses an igneous rock lithology identification method based on an engineering optimization cascade forest, and relates to the field of lithology identification, and the igneous rock lithology identification method based on the engineering optimization cascade forest comprises the following steps: 1, randomly selecting logging curves of different igneous rocks from a to-be-identified region, and carrying out the preprocessing of the logging curves; the CXDF method provided by the invention only depends on conventional logging data, can not depend on expensive data such as stratum element logging, and is low in cost, high in accuracy and higher in universality; the CXDF method combines a depth forest and an extreme gradient lifting method, so that the problem of overfitting is avoided, and excellent classification efficiency can be ensured when a back-to-undersampling data set is faced; in addition, the CXDF method also performs engineering optimization on the combined depth forest and extreme gradient lifting method, and column block parallel learning is adopted for extreme gradient lifting, so that the calculation amount is remarkably reduced, and the calculation cost is also greatly reduced.

Description

Technical Field

[0001] The invention relates to the field of rocks, and in particular to an igneous rock lithology identification method based on engineering optimized cascade forests. Background Art

[0002] As one of the main reservoir rocks for oil and gas, igneous rock is widely distributed in the main oil and gas basins on the earth, just like sedimentary rock oil and gas reservoirs, and has become an important field in oil and gas exploration and development. Accurate identification of lithology is the prerequisite for effective evaluation of igneous rocks.

[0003] Traditional lithology identification methods, such as the intersection diagram method and formation element logging technology, each have their own limitations. When facing igneous rocks, the intersection diagram method is difficult to effectively distinguish lithology due to the similar composition of various igneous rocks, resulting in slight differences in their physical property parameters. Although the formation element logging method can provide detailed information on the chemical composition of rocks, its high cost and lengthy analysis time greatly hinder the widespread application of this method. In order to seek a more cost-effective solution, researchers began to explore the application of artificial intelligence technology to identify the lithology of igneous rocks in an automated and intelligent manner. However, although these emerging methods have shown certain potential, in practical applications, their key performance indicators such as accuracy and recall rate are still difficult to reach a satisfactory level and still need to be further optimized and improved.

[0004] In summary, the existing methods for identifying igneous rocks do not have low-cost and high-accuracy methods and need to be improved. Summary of the invention

[0005] The object of the present invention is to provide an igneous rock lithology identification method based on engineering optimized cascade forests to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for identifying igneous rock lithology based on engineering optimized cascade forests, comprising the following steps: Step 1, randomly selecting well logging curves of different igneous rocks from the area to be identified, and pre-processing the well logging curves to obtain a training set to be processed; Step 2: Combine the deep forest (DF) and extreme gradient boosting (XGBoost) methods and obtain the lithology classification scheme CXDF after engineering optimization; Step 3: Use the CXDF method to train the dataset to be processed to obtain a model.

[0007] As a further solution of the present invention: Step 1 specifically comprises: Step 11, randomly selecting well logging curves of different igneous rocks from the area to be identified, and establishing an igneous rock training set data containing well logging curve data; Step 12: Use Z-score to preprocess the igneous rock training set data to obtain a training set to be processed.

[0008] As a further solution of the present invention: in step 11, the characteristic curves of the igneous rock training set data include acoustic wave (AC), natural gamma ray (GR), density (DEN), borehole diameter (CAL), compensated neutron (CNL), resistivity (RMLL), deep resistivity (RLLD) and shallow resistivity (RLLS).

[0009] As a further solution of the present invention: in step 11, different lithologies of igneous rocks are classified according to the distribution of oil and gas reservoirs, and the stratigraphic data in the data sets after different lithology classifications include dense layers, dry layers, water layers, low-yield layers, and oil layers.

[0010] As a further solution of the present invention: in step 12, a balanced data set consisting of samples of each igneous rock type to be identified is selected as a training set; in order to make the data distribution more balanced, the logarithm of the RLLD and RLLS curves is taken, and then all training set data are Z-score (a standardized method for evaluating the distance from a sample point to the overall mean) standardized based on the training set, and then a training-backtest-validation-prediction process is adopted to obtain the training set to be processed.

[0011] As a further solution of the present invention: Step 2 specifically includes: Step 21, combining the deep forest (DF) and extreme gradient boosting (XGBoost) methods, and performing the first engineering optimization to obtain the first optimized method; Step 22, performing a second engineering optimization on the method after the first optimization to obtain a lithology classification scheme CXDF.

[0012] As a further solution of the present invention: in step 21, the data vector is input into the cascade forest as the input feature vector of the deep forest, and each layer of the cascade forest is composed of two or more different random forests, namely the initial random forest and the complete random forest; the complete random forest contains 1000 decision trees, and each decision tree randomly selects a feature as a bifurcation node; the initial random forest also has 1000 decision trees, and the number of features randomly selected by each tree is the square root of the number of sample features, and then the bifurcation node is determined by the Gini coefficient; each decision tree in each layer of the forest will output the feature vector of the data set, which will be further aggregated into the output feature vector of the layer through extreme gradient boosting. After the training of each layer of the cascade forest is completed, the sample will be used for K-fold back judgment; The extreme gradient boosting algorithm selects features in parallel when building each weak learner (such as the process of building a decision tree), and determines the subtree segmentation features and feature values. It determines the processing method of missing values ​​by enumerating whether all missing values ​​enter the left subtree or the right subtree of the current node. In addition, extreme gradient boosting itself adds L1 and L2 regularization terms, which helps prevent overfitting and has stronger generalization ability. The objective function of extreme gradient boosting is for: (1); in, is the loss function, is the true value, is the number of samples, is the number of iterations, express The predicted value at the time of the round iteration, For the sample, For the The predicted value of the iteration, is the regularization term, for The first derivative of for The second derivative of Represents the decision tree Node value, For leaf nodes The sample set at is the number of leaf nodes, and is the pruning parameter used to control the complexity of the tree.

[0013] As a further solution of the present invention: in step 22, the method after the first optimization is subjected to a second engineering optimization, and the extreme gradient boosting is designed to adopt column-block parallel learning to obtain the lithology classification scheme CXDF.

[0014] As a further solution of the present invention: in step 3, the CXDF method is used to train the processing data set to obtain a model. The core idea of ​​lithology identification is to transform the multi-classification problem into multiple binary classification problems, that is, all lithologies and other lithologies are classified into binary classifications respectively, and the lithologies with the best current classification effect are eliminated in turn; when the classification effect still fails to meet the established requirements, according to the results of feature importance analysis, the data set is reduced in dimension starting from the feature with the lowest importance until the established requirements are met; if the requirements are still not met, the lithology data with the best effect is output until all lithologies are distinguished.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: the CXDF method proposed in the present invention relies only on conventional logging data and does not rely on expensive data such as formation element logging, so it has low cost, high accuracy and stronger universality; The CXDF method combines the deep forest and extreme gradient boosting methods, which not only avoids the overfitting problem, but also ensures excellent classification efficiency when facing undersampled data sets; The CXDF method also performs engineering optimization on the combined deep forest and extreme gradient boosting methods, allowing extreme gradient boosting to adopt column-block parallel learning, which not only significantly reduces the amount of computation but also greatly reduces the computational cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the intersection diagram of the training set data.

[0017] Figure 2 Flowchart of the model classification scheme.

[0018] Figure 3 Simulated well performance graph for different machine learning methods.

[0019] Figure 4 This is the actual igneous rock lithology identification result diagram. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0021] A method for identifying igneous rock lithology based on engineering optimized cascade forests, comprising the following steps: Step 1, randomly selecting well logging curves of different igneous rocks from the area to be identified, and pre-processing the well logging curves to obtain a training set to be processed; Step 2: Combine the deep forest (DF) and extreme gradient boosting (XGBoost) methods and obtain the lithology classification scheme CXDF after engineering optimization; Step 3: Use the CXDF method to train the dataset to be processed to obtain a model.

[0022] In this embodiment: Step 1 specifically includes: Step 11, randomly selecting well logging curves of different igneous rocks from the area to be identified, and establishing an igneous rock training set data containing well logging curve data; Step 12: Use Z-score to preprocess the igneous rock training set data to obtain a training set to be processed.

[0023] In this embodiment: in step 11, the characteristic curves of the igneous rock training set data include acoustic wave (AC), natural gamma ray (GR), density (DEN), borehole diameter (CAL), compensated neutron (CNL), resistivity (RMLL), deep resistivity (RLLD) and shallow resistivity (RLLS).

[0024] In this embodiment: in step 11, different lithologies of igneous rocks are classified according to the distribution of oil and gas reservoirs, and the stratigraphic data in the data sets after different lithology classifications include dense layers, dry layers, water layers, low-yield layers, and oil layers.

[0025] In this example: See Figure 1 In step 12, each igneous rock to be identified is selected (e.g. Figure 2 The samples of andesite, trachyte, tuff, granite, etc. shown in the figure constitute a balanced data set as the training set; in order to make the data distribution more balanced, the RLLD and RLLS curves are logarithmic ( Figure 1 The sixth and seventh from top to bottom), and then based on the training set, all training set data are standardized by Z-score (a standardized method for evaluating the distance from a sample point to the overall mean), and then the training-back-judgment-validation-prediction process is adopted to obtain the training set to be processed.

[0026] By adopting the training-review-validation-prediction process, the model is not guided by the prediction results of a single data set, which can avoid overfitting.

[0027] In this embodiment: Step 2 specifically includes: Step 21, combining the deep forest (DF) and extreme gradient boosting (XGBoost) methods, and performing the first engineering optimization to obtain the first optimized method; Step 22, performing a second engineering optimization on the method after the first optimization to obtain a lithology classification scheme CXDF.

[0028] In this embodiment: In step 21, the deep forest is a multi-layer ensemble method based on decision trees, and each layer consists of multiple forests. When sampling data, the deep forest adopts a multi-granularity scanning method, which enriches the sample data set of the algorithm. In the native deep forest, the average of the feature vectors output by each forest in each layer is the output vector of the layer. The number of layers of the deep forest is not a custom hyperparameter, but can be automatically adjusted according to the complexity of the data set. At the same time, the deep forest has good robustness, and good results can be obtained using the default parameters. This means that there is no need to perform a result-oriented parameter optimization process when using the deep forest method.

[0029] For continuous data, there is a strong spatial correlation between adjacent data. The multi-granularity scanning method can be used to fill the data samples to enhance the classifier effect. Taking the data of length 600 as an example, the length is set to 100 and the step size is 1, and the data vector of length 3006 can be obtained in the end. The data vector is input into the cascade forest as the input feature vector of the deep forest. Each layer of the cascade forest consists of two different random forests, namely the initial random forest and the complete random forest; the complete random forest contains 1000 decision trees, and each decision tree randomly selects a feature as a bifurcation node; the initial random forest also has 1000 decision trees, and the number of features randomly selected by each tree is the square root of the number of sample features, and then the bifurcation node is determined by the Gini coefficient; each decision tree in each layer of the forest will output the feature vector of the data set, which will be further aggregated into the output feature vector of the layer through extreme gradient boosting. After the training of each layer of the cascade forest is completed, the samples will be used for K-fold back-judgment; this has two purposes. First, it can avoid overfitting, and secondly, when the back-judgment effect reaches the expected effect of the algorithm, the output vector no longer enters the next layer but enters the final result output process. This means that even for undersampled data sets, Deep Forest can still guarantee excellent classification efficiency. After layer-by-layer output, the final prediction result is finally obtained through the extreme gradient boosting weak classifier.

[0030] The extreme gradient boosting algorithm selects features in parallel when building each weak learner (such as the process of building a decision tree), and determines the subtree segmentation features and feature values. It determines the processing method of missing values ​​by enumerating whether all missing values ​​enter the left subtree or the right subtree of the current node. In addition, extreme gradient boosting itself adds L1 and L2 regularization terms, which helps prevent overfitting and has stronger generalization ability. The objective function of extreme gradient boosting is for: (1); in, is the loss function, is the true value, is the number of samples, is the number of iterations, express The predicted value at the time of the round iteration, For the sample, For the The predicted value of the iteration, is the regularization term, for The first derivative of for The second derivative of Represents the decision tree Node value, For leaf nodes The sample set at is the number of leaf nodes, and is the pruning parameter used to control the complexity of the tree.

[0031] In this embodiment: in step 22, the method after the first optimization is subjected to a second engineering optimization, and the extreme gradient boosting is designed to adopt column-block parallel learning to obtain the lithology classification scheme CXDF.

[0032] Engineering optimization is to design extreme gradient boosting and adopt column-block parallel learning at the engineering implementation level. At this time, the pre-training data will be sorted according to the features and saved in the block structure. Each block structure is stored in a compressed sparse column format, and the block structure will be reused in subsequent training, which can significantly reduce the amount of calculation, thereby greatly reducing the calculation cost.

[0033] In this example: See Figure 2 In step 3, the CXDF method is used to train the processing data set to obtain the model. The core idea of ​​lithology identification is to transform the multi-classification problem into multiple binary classification problems, that is, all lithologies and the remaining lithologies are classified into binary classifications, and the lithologies with the best current classification effect are eliminated in turn; when the classification effect still fails to meet the established requirements, according to the results of feature importance analysis, the data set is reduced in dimension from the lowest important feature until the established requirements are met; if the requirements are still not met, the lithology data with the best effect is output until all lithologies are distinguished.

[0034] Theoretical verification, introducing the F-beta evaluation standard to compare the performance of the CXDF method with other machine learning methods, and using the CXDF method to classify simulated wells and actual formations to verify the classification effect includes the following steps: The performance of the CXDF method is compared with other machine learning methods, and the CXDF method is used to classify actual strata to verify the classification effect.

[0035] F-beta is introduced and a simulated well is established to evaluate the classifier effect. F-beta is a statistical indicator used to evaluate the accuracy of a binary classification model, which can be used to evaluate the classifier effect. First, you need to understand two concepts, namely Precision and Recall. Precision indicates the proportion of samples with correct current lithology prediction among the samples predicted to be the current lithology. Recall indicates the ratio of the number of samples with correct current lithology prediction to the number of all samples of the current lithology. F1, F2 and F0.5 are a comprehensive evaluation standard called F-beta, which can reflect the comprehensive performance of the classifier. Different beta values ​​mean different weights of Precision and Recall in the algorithm. The specific expressions are as follows: (2); Therefore, the performance of the CXDF method and other machine learning methods can be reasonably evaluated based on F-beta.

[0036] A simulated well was built and the CXDF method was verified in the actual formation. The validation set data was assembled into a simulated well to visualize the model effect. In addition to the CXDF method, in order to compare the classification effect, classifiers such as XDF (Cascade Forest), DF (Deep Forest), XGBoost (Extreme Gradient Boosting), and GBRT (Gradient Boosting of Regression Trees) were also used for classification.

[0037] See also Figure 3 The simulated well curve is composed of validation data of five lithologies, and the depths are andesite, trachyte, granite, tuff and basalt from low to high. The lithology profile data is the logging lithology profile constrained by the core data. Figure 3 The right side of the middle is the lithology profile predicted by the above five methods. It can be seen that the fluctuations of the lithology boundaries and conventional curves in the simulated wells are synchronized. It can be intuitively seen that the lithology prediction results of CXDF are most consistent with the lithology profile of the simulated wells established based on actual formation data, and the errors are far less than those of the other two methods. In other words, in the igneous rock simulation wells with five lithology compositions, the effect of the CXDF method is better than that of other methods, which shows the reliability and accuracy of the CXDF method.

[0038] See also Figure 4 In order to further verify the application effect of the CXDF model, actual igneous rock formations can also be selected for lithology identification, and the identification results can be compared with thin section data. The comparison results show that the igneous rock lithology identification results of the CXDF method are consistent with the actual lithology profile and rock thin sections, and the results are reliable.

[0039] Therefore, the present invention can cope with the problem of complex igneous rock lithology identification that is difficult to solve with conventional logging data; The CXDF method proposed in the present invention relies only on conventional logging data and does not rely on expensive data such as formation element logging, and has stronger universality; The CXDF method combines the deep forest and extreme gradient boosting methods, which not only avoids the overfitting problem, but also ensures excellent classification efficiency when facing undersampled data sets; The CXDF method also performs engineering optimization on the combined deep forest and extreme gradient boosting methods, allowing extreme gradient boosting to adopt column-block parallel learning, which not only significantly reduces the amount of computation but also greatly reduces the computational cost.

[0040] The typical classification effect evaluation index F-beta, the simulation well verification results and the actual formation igneous rock lithology identification results all show that the CXDF method has a relatively higher lithology identification accuracy and can effectively solve the lithology identification problem of igneous rock formations.

[0041] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered as exemplary and non-restrictive in all respects.

[0042] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for identifying igneous rock lithology based on engineering optimized cascade forests, characterized in that: The igneous rock lithology identification method based on engineering optimized cascade forests includes the following steps: Step 1, randomly selecting well logging curves of different igneous rocks from the area to be identified, and pre-processing the well logging curves to obtain a training set to be processed; Step 2: Combine the deep forest and extreme gradient boosting methods and obtain the lithology classification scheme CXDF after engineering optimization; Step 3: Use the CXDF method to train the dataset to be processed to obtain a model.

2. The igneous rock lithology identification method based on engineering optimized cascade forest according to claim 1 is characterized in that: Step 1 specifically includes: Step 11, randomly selecting well logging curves of different igneous rocks from the area to be identified, and establishing an igneous rock training set data containing well logging curve data; Step 12: Use Z-score to preprocess the igneous rock training set data to obtain a training set to be processed.

3. The igneous rock lithology identification method based on engineering optimized cascade forest according to claim 2 is characterized in that: In step 11, the characteristic curves of the igneous rock training set data include acoustic waves, natural gamma rays, density, borehole diameter, compensated neutrons, resistivity, deep resistivity, and shallow resistivity.

4. The igneous rock lithology identification method based on engineering optimized cascade forest according to claim 2 is characterized in that: In step 11, different lithologies of igneous rocks are classified according to the distribution of oil and gas reservoirs. The stratigraphic data in the data sets after different lithology classifications include dense layers, dry layers, water layers, low-yield layers, and oil layers.

5. The method for identifying igneous rock lithology based on engineering optimized cascade forest according to any one of claims 2 to 4, characterized in that: In step 12, a balanced data set consisting of samples of each igneous rock type to be identified is selected as a training set; in order to make the data distribution more balanced, the logarithm of the RLLD and RLLS curves is taken, and then all training set data are Z-score standardized based on the training set as a benchmark, and then the training-back-judgment-validation-prediction process is used to obtain the training set to be processed.

6. The method for identifying igneous rock lithology based on engineering optimized cascade forest according to claim 1, characterized in that: Step 2 specifically includes: Step 21, combining the deep forest and extreme gradient boosting methods, and performing a first engineering optimization to obtain a first optimized method; Step 22, performing a second engineering optimization on the method after the first optimization to obtain a lithology classification scheme CXDF.

7. The method for identifying igneous rock lithology based on engineering optimized cascade forest according to claim 6, characterized in that: In step 21, the data vector is input into the cascade forest as the input feature vector of the deep forest. Each layer of the cascade forest consists of two different random forests, namely the initial random forest and the complete random forest. The complete random forest contains 1000 decision trees, and each decision tree randomly selects a feature as a bifurcation node. The initial random forest also has 1000 decision trees. The number of features randomly selected by each tree is the square root of the number of sample features, and then the bifurcation node is determined by the Gini coefficient. Each decision tree in each layer of the forest will output the feature vector of the data set, which will be further aggregated into the output feature vector of the layer through extreme gradient boosting. After the training of each layer of the cascade forest is completed, the samples will be used for K-fold back-judgment. The extreme gradient boosting algorithm selects features in parallel when constructing each weak learner, and determines the subtree segmentation features and feature values. It determines how to handle missing values ​​by enumerating whether all missing values ​​enter the left subtree or the right subtree of the current node. In addition, extreme gradient boosting itself adds L1 and L2 regularization terms, which helps prevent overfitting and has stronger generalization ability. The objective function of extreme gradient boosting is for: （1）； in, is the loss function, is the true value, is the number of samples, is the number of iterations, express The predicted value at the time of the round iteration, For the sample, For the The predicted value of the iteration, is the regularization term, for The first derivative of for The second derivative of Represents the decision tree Node value, For leaf nodes The sample set at is the number of leaf nodes, and is the pruning parameter used to control the complexity of the tree.

8. The method for identifying igneous rock lithology based on engineering optimized cascade forest according to claim 6, characterized in that: In step 22, the method after the first optimization is subjected to a second engineering optimization, and the extreme gradient boosting is designed to adopt column-block parallel learning to obtain the lithology classification scheme CXDF.

9. The method for identifying igneous rock lithology based on engineering optimized cascade forest according to claim 1, characterized in that: In step 3, the CXDF method is used to train the processing data set to obtain a model. The core idea of ​​lithology identification is to transform the multi-classification problem into multiple binary classification problems, that is, to classify all lithologies and the remaining lithologies into binary classifications, and then eliminate the lithologies with the best current classification effect in turn; When the classification effect still fails to meet the established requirements, the data set is reduced in dimension starting from the least important feature according to the results of feature importance analysis until the established requirements are met; if the requirements are still not met, the lithology data with the best effect is output until all lithologies are distinguished.

Images