An improved decision tree lithology identification method based on logging data
The improved decision tree algorithm solves the problems of insufficient and imbalanced samples in well logging lithology identification, improves the identification rate, and provides interpretable classification rules applicable to well logging lithology classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHANGZHOU UNIV
- Filing Date
- 2022-12-13
- Publication Date
- 2026-07-24
Smart Images
Figure CN115809411B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of well logging lithology identification technology, and in particular to an improved decision tree lithology identification method based on well logging data. Background Technology
[0002] Traditional rock identification methods often involve using coring tools to extract large rock samples (cores) from underground during drilling to study basic data such as the lithology, physical properties, and thickness of oil and gas reservoirs, thereby inferring oil and gas conditions and reservoir characteristics. However, this method is difficult to obtain cores, costly, and often misses valuable unlabeled data.
[0003] In recent years, well logging data has been used to measure various parameters in different formations by using instruments that measure physical properties such as electricity, sound, heat, and radioactivity in boreholes. These parameters are then used to identify the properties of underground rocks and fluids. With the rapid development of machine learning, it has also been increasingly applied to well logging lithology classification and identification, achieving significant breakthroughs in this application.
[0004] Patent CN114991749A discloses a method for identifying fine-grained sedimentary lithology. This method uses SPSS software to perform multiple linear regression on well logging curves and measured mineral content to predict lithology. In multiple regression analysis, the selection of factors and their expressions are only speculative, and the goodness of fit is mostly between 0.6 and 0.8, which may lead to inaccurate lithology identification. Patent CN114462471A discloses a machine learning-based lithology identification method. This method uses the SSA algorithm to optimize the SVR (Support Vector Regression) model to predict lithology. Although the relevant parameters of the SSA algorithm have been optimized, support vector machines are mainly used to solve binary classification problems, and their performance is not ideal for multi-class classification problems.
[0005] With the popularity of neural networks and deep learning in recent years, invention patents such as CN111914478A use the PSO algorithm to optimize the BP neural network and improve the recognition accuracy; invention patents such as CN111783825A use convolutional neural networks to identify logging lithology; and invention patents such as CN109736790B use deep belief network methods to predict logging lithology types. Although there are many such methods and the recognition rate is high, these methods often require a large amount of sample data to ensure the performance of the model, and the prediction results cannot be analyzed and interpreted.
[0006] Therefore, while ensuring the recognition rate, the decision tree algorithm is simple to understand, highly interpretable, and easy to convert into well logging lithology classification rules. It also performs well in terms of recognition rate, especially when faced with imbalanced rock sample categories. Summary of the Invention
[0007] To address the shortcomings of existing algorithms, this invention not only effectively solves problems such as insufficient training samples, imbalanced sample categories, and insufficient recognition rate, but also establishes a tree model that is visualized and easily converted into well logging lithology classification rules, providing new ideas for subsequent well logging lithology classification and identification research and engineering applications.
[0008] The technical solution adopted in this invention is: an improved decision tree lithology identification method based on well logging data, comprising the following steps:
[0009] Step 1: Collect well logging data on lithological types and corresponding physical property parameters;
[0010] Furthermore, the rock types include: sandstone, mudstone, breccia, basalt, limestone, and dolomite;
[0011] Step 2: Preprocess the relevant data and then divide the dataset;
[0012] Further preprocessing includes: dimensionality reduction, imputation of missing values, handling of outliers, and One-Hot encoding of lithology categories;
[0013] Furthermore, the dimensionality reduction method employed the Pearson correlation coefficient method;
[0014] Furthermore, the reduced physical properties include: well depth, thorium, uranium, potassium, density, neutrons, and photoelectric factor.
[0015] Furthermore, the missing values are imputed using a chain equation multiple imputation method.
[0016] Step 3: Use the well logging lithology data training set to establish the corresponding decision tree classification and recognition model, and use the Monte Carlo algorithm to optimize the depth of the tree in the decision tree model;
[0017] Furthermore, the decision tree classification and recognition model divides the decision tree using information gain, specifically including:
[0018] (1) The entropy of the event is calculated before partitioning as entropy(before).
[0019] (2) After dividing the events according to their attributes, the entropy of the events is calculated again as entropy(after).
[0020] (3) After the information gain is calculated, Gain (information gain) = entroy (before) - entroy (after). After selecting the feature a, the well logging lithology sample set D is divided into n classes, which will generate v branch nodes.
[0021] Furthermore, the Monte Carlo algorithm optimizes the depth of the tree in the decision tree model, including:
[0022] The objective functions are set to maximize the correct identification rate of lithology types and minimize the depth of the tree, with constraints that the correct identification rate of lithology types is greater than 80% and the depth of the tree is 3 to 20.
[0023] Using a linear weighting method, with a weight of 0.7, and a weight of 0.3 for the depth of the objective function tree, the formula is as follows:
[0024] minf(x) = -0.7(x) + 0.3f²(x)
[0025]
[0026] Where f(x) is the new objective function after linear weighting, f1(x) is the correct identification rate of lithology types obtained by the decision tree, and f2(x) is the depth of the tree.
[0027] Step 4: Improve the decision tree using post-pruning methods and calculate the computational accuracy after pruning;
[0028] Furthermore, specifically including:
[0029] First, generate a complete decision tree from the training set, and then examine the non-leaf nodes from the bottom up. Calculate the accuracy of the validation set before partitioning based on the existing labels. If partitioning is assumed, delete the tree with this node as the root node to make it a leaf node. Compare the performance before and after deletion to see if there is an improvement. If there is, delete it; otherwise, keep it.
[0030] Furthermore, the formula for calculating the accuracy after pruning is as follows:
[0031]
[0032] Where ACC2 is the accuracy after deletion, and A is the number of training set samples. i Let α be the number of child nodes of the i-th node that have the same target value as the target value in the training set, and let α be an undetermined coefficient. This indicates the depth of the binary tree.
[0033] The beneficial effects of this invention are:
[0034] 1. It is suitable for handling imbalanced sample problems and can be easily transformed into well logging lithology classification rules, providing new ideas for subsequent well logging lithology classification and identification research;
[0035] 2. Optimize the depth of the decision tree using the Monte Carlo algorithm to find the optimal value of the tree depth;
[0036] 3. The improved post-pruning algorithm takes into account both the complexity and accuracy of the tree, which can be used to solve the overfitting problem of decision trees, and the recognition accuracy on the test set is significantly improved. Attached Figure Description
[0037] Figure 1 This is a flowchart of the improved decision tree lithology identification method based on well logging data of the present invention;
[0038] Figure 2 This is the frequency diagram of logging lithology types of the present invention;
[0039] Figure 3 This is a scatter plot of the physical property parameters of sandstone at different well depths according to the present invention;
[0040] Figure 4 This is a schematic diagram illustrating the recognition rate of the decision tree of the present invention at different binary tree depths;
[0041] Figure 5 This is a comparison chart of the results of automatic lithology prediction for new wells and core logging data from this invention. Detailed Implementation
[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.
[0043] like Figure 1 As shown, an improved decision tree-based lithology identification method based on well logging data includes the following steps:
[0044] Step 1: Collect well logging lithology-related data;
[0045] Taking well data from a block in Daqing Oilfield as an example, a tool containing various sensors is lowered into the borehole. The sensors then measure various physical properties of the surrounding rocks, including the electrical, acoustic, and nuclear characteristics of the surrounding medium. The data collected by the sensors is then transmitted to a computer. In this example, the characteristic values are various physical properties, and the target values are lithological types (sandstone, mudstone, breccia, basalt, limestone, and dolomite).
[0046] Step 2: Preprocess the acquired data and then divide the dataset;
[0047] First, the frequencies of lithological types were statistically analyzed; then, the acquired data underwent preprocessing operations (dimensionality reduction of physical property parameters, imputation of missing values, handling of outliers, and One-Hot encoding of lithological categories); finally, the dataset was partitioned, with the specific steps as follows:
[0048] 1. Frequency statistics of lithological types:
[0049] Statistical analysis of the frequency of lithological types, such as Figure 2 As shown, the rock sample categories are unbalanced, making it suitable for using decision tree algorithms to solve the well logging lithology identification problem.
[0050] 2. Dimensionality reduction of physical property parameters:
[0051] Raw well logging data may contain various physical property parameters. Taking well data from a block in Daqing Oilfield as an example, there are as many as 26 physical property parameters. Higher feature dimensions often lead to more complex learning models and are prone to overfitting. Therefore, for some highly correlated physical property parameter categories, dimensionality reduction can be used, while other unrelated categories are retained. This is called physical property parameter dimensionality reduction. The physical property parameters selected for dimensionality reduction include: well depth, thorium, uranium, potassium, density, neutrons, and photoelectric factor. The dimensionality reduction method used is the Pearson correlation coefficient method, and dimensionality reduction is performed on two or more parameter categories with a correlation coefficient higher than a preset value (correlation coefficient > 0.8) (selecting one of them). The formula for the Pearson correlation coefficient method is:
[0052]
[0053] Where x and y are physical property parameters, and r is the Pearson correlation coefficient.
[0054] 3. Imputation of missing values
[0055] In well logging, data loss often occurs due to machine malfunctions or human error. Therefore, it is necessary to perform missing value detection and select an appropriate method for interpolation. When the detected lithology category is missing, it can be directly assigned the "unknown" category.
[0056] Since the differences in the physical properties of lithology mainly depend on the type of lithology, multiple interpolation using chain equations is particularly useful. The steps of the multiple interpolation method using chain equations are as follows:
[0057] (1) During the imputation process, missing variables can be predicted from other variables, and the random forest algorithm is selected as the prediction model;
[0058] (2) Apply statistical models to the five datasets generated after iteration, evaluate the fitting effect of the five models, integrate the results of individual analysis into a set of results, and evaluate the merits of the interpolation model.
[0059] (3) Fill in the missing values.
[0060] 4. Handling outliers
[0061] Scatter plots of various physical properties of different lithologies at different depths were generated, and outliers deviating from the overall trend of the scatter plots were removed. Taking sandstone as an example, photoelectric factor parameters were measured at different well depths, such as... Figure 3 As shown, the photoelectric factor value fluctuates with increasing well depth and shows a continuous decreasing trend. No outliers were found that are particularly far from the overall trend of the scatter plot of well logging lithological and physical property parameters, so no outlier processing was performed.
[0062] 5. Lithology Category Code
[0063] The lithology category coding method adopts One-Hot coding, as detailed in Table 1:
[0064] Table 1 One-Hot coded lithological labels
[0065]
[0066]
[0067] 6. Dataset partitioning
[0068] The dataset is divided into two parts: one well logging lithology data is randomly selected as the test set, and the rest are used as the training set, in order to solve the problem of identifying different new wellhead lithologies.
[0069] Step 3: Use the well logging lithology data training set to establish the corresponding decision tree classification and recognition model, and use the Monte Carlo algorithm to optimize the depth of the tree in the decision tree model;
[0070] Decision trees are a supervised learning algorithm based on if-then-else rules. During prediction, a certain attribute value is used to make a judgment at the internal nodes of the tree, and the decision is made on which branch node to enter based on the judgment result, until the leaf node is reached to obtain the classification result.
[0071] In most cases, information gain is used to determine the basis for splitting a decision tree. Information gain is the difference in entropy before and after splitting a dataset using a certain feature. Entropy can represent the uncertainty of a sample set. The greater the entropy, the greater the uncertainty of the sample. Therefore, the difference in information entropy before and after splitting can be used to measure the effectiveness of using the current feature to split the sample set D.
[0072] The steps for splitting a decision tree using information gain are as follows:
[0073] (1) The entropy of the event is calculated before partitioning as entropy(before).
[0074] (2) After dividing the events according to their attributes, the entropy of the events is calculated again as entropy(after).
[0075] (3) After the partitioning, the information gain is calculated as Gain(information gain) = entroy(before) - entroy(after). After selecting the feature a, the well logging lithology sample set D is divided into n classes, which will generate v branch nodes. The specific steps of the decision tree partitioning are as follows:
[0076] 1. Calculate the category information entropy Ent(D):
[0077]
[0078] Among them, C k Let be the number of samples in the k-th category, and D be the total number of samples.
[0079] 2. Calculate the information entropy of attributes:
[0080]
[0081] Among them, D v C represents the number of samples contained in the v-th branch node of attribute a. kv This represents the number of samples in the k-th category among the samples contained in the v-th branch node of attribute a.
[0082] 3. Calculate the information gain Gain(D,a):
[0083]
[0084] Among them, C k Let D be the number of samples in the k-th category. v C represents the number of samples contained in the v-th branch node of attribute a. kv This represents the number of samples in the k-th category among the samples contained in the v-th branch node of attribute a.
[0085] 4. Build a decision tree:
[0086] Starting from the root node, calculate the information gain of all features for the node, select the feature with the largest information gain as the node feature, create child nodes according to the different values of the feature, and generate new child nodes in the same way for each child node until the information gain is very small or there are no features to choose from.
[0087] The basic idea of the Monte Carlo method is that when the problem to be solved is the probability of a certain random event or the expected value of a random variable, the probability of the random event can be estimated by means of some "experiment" based on the frequency of the event's occurrence, or certain numerical characteristics of the random variable can be obtained and used as the solution to the problem. Therefore, the idea of the Monte Carlo method is applied to the field of computer stochastic simulation to solve multi-objective programming problems (ensuring the correct identification rate while minimizing the tree depth). Its objective function is to maximize the correct identification rate of lithology types and minimize the tree depth, with constraints that the correct identification rate of lithology types is greater than 80% and the tree depth is 3 to 20. In summary, the model of this problem can be described as follows:
[0088] min{-f1(x),f2(x)}
[0089]
[0090] Where f1(x) is the correct identification rate of lithology types obtained from the decision tree, and f2(x) is the depth of the tree.
[0091] Using a linear weighted method, since the objective function requires a high accuracy rate in identifying lithology types, a weight of 0.7 is applied. The depth of the objective function tree is weighted at 0.3. This transforms the multi-objective programming problem into a linear programming problem, with the following model:
[0092] minf(x) = -0.7(x) + 0.3f²(x)
[0093]
[0094] Here, f(x) is the new objective function after linear weighting, and the Monte Carlo method is used to find f(x), the objective function recognition rate, and the tree depth.
[0095] Figure 4 The diagram illustrates the recognition accuracy at different binary tree depths. The calculated recognition rate is 97.30%, and the tree depth of 9 meets the requirements.
[0096] Step 4: Improve the decision tree using post-pruning methods and calculate the computational accuracy after pruning;
[0097] First, a complete decision tree is generated from the training set. Then, non-leaf nodes are examined from the bottom up. The accuracy of the validation set before splitting is calculated based on the existing labels. If a split is assumed, the tree rooted at that node is deleted, making it a leaf node. The performance (accuracy) before and after deletion is compared to see if there is an improvement. If there is, the node is deleted; otherwise, it is retained. The formula for calculating accuracy is:
[0098]
[0099] Where ACC1 is the accuracy before deletion, A is the number of training set samples, and A i This represents the number of child nodes of the i-th node that have the same target value as the training set.
[0100]
[0101] Where ACC2 is the accuracy after deletion, and A is the number of training set samples. i This represents the number of child nodes of the i-th node that have the same target value as the training set, where α is an undetermined coefficient used to penalize the number of nodes and guide the model to use fewer nodes. This indicates the depth of the binary tree.
[0102] The data from the well logging lithology test samples are used as input, and the output of the improved decision tree model is used to verify the actual results.
[0103] Step four yields the decision tree model for the optimal depth. Then, step four performs post-pruning on the model from step three to obtain an improved decision tree model. Data from well logging lithology test samples are used as input, and the output of the improved decision tree model is cross-validated with the actual results. Figure 5 The image shown is a comparison between the results of automatic lithology prediction for new wells and core logging data.
[0104] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.
Claims
1. An improved decision tree-based lithology identification method based on well logging data, characterized in that, Includes the following steps: Step 1: Collect well logging data on lithological types and corresponding physical property parameters; The rock types include: sandstone, mudstone, breccia, basalt, limestone, and dolomite; Step 2: Preprocess the relevant data and then divide the dataset; The preprocessed dimensionality-reduced physical properties include: well depth, thorium, uranium, potassium, density, neutrons, and photoelectric factor; The dataset is divided into two parts: one well logging lithology data is randomly selected as the test set, and the rest are used as the training set. Step 3: Use the well logging lithology data training set to establish the corresponding decision tree classification and recognition model, and use the Monte Carlo algorithm to optimize the depth of the tree in the decision tree model; The Monte Carlo algorithm optimizes the depth of trees in decision tree models, including: The objective functions are set to maximize the correct identification rate of lithology types and minimize the depth of the tree, with the constraints being that the correct identification rate of lithology types is greater than 80% and the depth of the tree is 3 to 20. Using a linear weighting method, with a weight of 0.7, and a weight of 0.3 for the depth of the objective function tree, the formula is as follows: ； in, The new objective function after linear weighting. To obtain the correct identification rate of lithology types from the decision tree, The depth of the tree; Step 4: Improve the decision tree using post-pruning methods and calculate the computational accuracy after pruning; Step four specifically includes: First, generate a complete decision tree from the training set, and then examine the non-leaf nodes from the bottom up. Calculate the accuracy of the validation set before partitioning based on the existing labels; Assuming a partition, delete the tree rooted at this node, making it a leaf node. Compare the performance before and after the deletion to see if there is an improvement. If there is, delete it; otherwise, keep it. The formula for calculating accuracy after pruning is: ； in, For the accuracy after deletion, The number of samples in the training set. The number of child nodes of the i-th node that have the same target value as the training set. For undetermined coefficients, This indicates the depth of the binary tree.
2. The improved decision tree lithology identification method based on well logging data according to claim 1, characterized in that, Preprocessing includes: Dimensionality reduction, imputation of missing values, handling of outliers, and One-Hot coding of lithology categories.
3. The improved decision tree lithology identification method based on well logging data according to claim 2, characterized in that: Dimensionality reduction was performed using the Pearson correlation coefficient method.
4. The improved decision tree lithology identification method based on well logging data according to claim 2, characterized in that: The imputation of missing values is performed using the chain equation multiple imputation method.
5. The improved decision tree lithology identification method based on well logging data according to claim 1, characterized in that, Decision tree classification and recognition models divide decision trees using information gain, specifically including: (1) The entropy of the event before partitioning is calculated as entroy(before). (2) The entropy of the event is calculated again after dividing it according to the attributes. (3) The information gain after partitioning is calculated as Gain(information gain) = ; Select good features Subsequently, the well logging lithology sample set was analyzed. Divided into A class will then produce Branch nodes.