Fluid identification method and visualization method based on logging data

Abstract

The invention discloses a fluid identification method based on logging data. The method comprises the following steps: making a logging data matrix as training data based on logging data and data processed by a gas logging method; inputting the training set into each base learning device to perform first-layer training, combining predicted values obtained by each base learning device based on the training set to serve as features of next-layer training, and processing predicted values obtained by each base learning device based on the test set to serve as test data of the next-layer training; and inputting the features into the meta-learner to train the meta-learner, and inputting the test data into the trained meta-learner to evaluate the performance of the meta-learner. The invention further provides a method for visually explaining the recognition result of the method. According to the recognition method, the advantages of each base model are combined, so that the prediction precision is obviously higher than that of each single model, the stability and the accuracy are obviously better than those of a classical single model, and the recognition accuracy is higher in the aspect of gas reservoir recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of oil development, and more specifically to a fluid identification method and a visualization method based on well logging data. Background Art

[0002] At present, the reservoir fluid identification method is mainly based on the difference between logging curves, empirical formulas and field core sampling, and the conventional chart method is used for qualitative identification. However, these methods are difficult to fully utilize the fluid characteristics contained in the logging curves, and the identification results are too dependent on regional experience judgments, and the nonlinear relationship between the logging curves and the gas content of the reservoir is not considered. Especially in high-temperature, high-pressure reservoirs, and environments with complex reservoir electrical parameters, it is often difficult to accurately find the electrical boundary between gas and water, resulting in low accuracy or failure of conventional methods in identifying reservoir fluid properties.

[0003] A large number of scholars have conducted relevant research based on the response characteristics of logging curves and various physical and chemical parameters derived from them, including the intersection plot method, triangular plate, curve overlap method, mathematical statistics method, and fluid identification methods that use special logging (electrical imaging logging, array acoustic logging, nuclear magnetic logging, etc.) to calculate various rock physical parameters and fluid characterization parameters. However, due to cost constraints, special logging is difficult to measure in most wells, and is not collected in most old wells. Therefore, fluid identification based on conventional logging curves and gas logging curves is still a research hotspot and faces great challenges.

[0004] For reservoirs with complex fluid distribution, more and more research tends to identify them based on machine learning methods. In recent years, many classification algorithms have been produced, such as clustering algorithms, logistic regression, support vector machines, fuzzy logic models, decision tree models, naive Bayes, neural networks and their derivative methods. In theory and practical applications, each machine learning algorithm has its own advantages and disadvantages. Therefore, it is difficult to fully mine the logging response information of high-temperature and high-pressure reservoirs using a single model. Therefore, it is urgent to study a new technology that can analyze the response relationship between gas layer and logging data, and can also ensure that this relationship has good discrimination ability for logging response data points with high similarity but different fluid properties. Summary of the invention

[0005] The object of the present invention is to provide a fluid identification method based on well logging data and a related visualization method to solve at least one of the above problems existing in the prior art.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0007] According to one aspect of the present invention, a fluid identification method based on well logging data is provided, comprising the following steps:

[0008] Step S1: Based on the well logging data and the data processed by the gas logging method, a logging data matrix is ​​prepared as the original training data, part of the data is extracted from the original training data, and the extracted data is divided into a training set and a test set;

[0009] Step S2: input the training set into each base learner among the multiple base learners for first-layer training, the target label variables of the training are the dry layer, the air layer, the wet layer and the water layer, and obtain the multiple trained base learners, combine the prediction values ​​obtained by each base learner based on the training set as the features for the next layer of training, and process the prediction values ​​obtained by each base learner based on the test set as the test data for the next layer of training;

[0010] Step S3: Input the features obtained in step S2 into the meta-learner to train the meta-learner, and input the test data into the trained meta-learner to evaluate the performance of the meta-learner.

[0011] According to one embodiment of the present invention, in step S1, the logging data includes natural potential and porosity curves.

[0012] According to one embodiment of the present invention, in step S1, the data processed by the gas measurement method includes at least one of the following data: data processed by the cross plot method, data processed by the light and heavy hydrocarbon ratio method, data processed by the triangle plate method, data processed by the 3H method, data representing the properties of hydrocarbons, and data representing the relative content of hydrocarbons.

[0013] According to one embodiment of the present invention, the parameters involved in the intersection diagram method include C2 / C3, nC4 / iC4, C2 / C1, C3 / C1, C3+iC4, C1+C2, C1+C2+; the parameters involved in the light and heavy hydrocarbon ratio method include C2+%, C1 / C2+; the parameters involved in the triangle plate method include X, Y, Q; the parameters involved in the 3H method include WH%, BH%, CH%; the parameters involved in the properties of the hydrocarbons include TG%, heavy hydrocarbons%, light hydrocarbons%, non-hydrocarbons; the parameters involved in the relative content of hydrocarbons include C1%, C2%, C3%, iC4%, nC4%, iC5%, nC5.

[0014] According to an embodiment of the present invention, step S1 further includes normalizing and performing SMOTE oversampling processing on the numerical features in the data.

[0015] According to one embodiment of the present invention, the multiple base learners adopt a K nearest neighbor model, a support vector machine SVM model, an XGBoost extreme gradient boosting tree model and a deep learning model.

[0016] According to one embodiment of the present invention, the parameter that needs to be adjusted during training of the K nearest neighbor model is distance.

[0017] According to an embodiment of the present invention, the support vector machine (SVM) model needs to adjust the penalty coefficient C and the kernel function for the error term during training.

[0018] According to one embodiment of the present invention, the XGBoost extreme gradient boosting tree model needs to adjust the learning rate, minimum weight, tree depth, loss function and regularization parameters during training.

[0019] According to one embodiment of the present invention, the deep learning network model needs to adjust the number of hidden layers and the number of neurons during training.

[0020] According to one embodiment of the present invention, the training termination condition of each base learner is that when the difference between the accuracy loss value of each round of training and the previous round is less than 5% for 5 times, the parameters are fixed and the training of the base learner is terminated.

[0021] According to one embodiment of the present invention, the first layer of training is performed using a cross-validation method, including: dividing the training set and the test set into k parts, using the k data sets of the training set as validation sets in turn, and using the remaining k-1 data sets as training data sets; after the same base learner completes training on the training set, it predicts the validation set and the test set; finally, the prediction results on the k validation sets are merged as features for the next layer of training, and the prediction value of the final test set is obtained by averaging the k prediction results on the test set.

[0022] According to an embodiment of the present invention, the cross-validation method is a 5-fold cross-validation method.

[0023] According to one embodiment of the present invention, the meta-learner adopts a logistic regression model.

[0024] According to one embodiment of the present invention, the meta-learner "scores" the results based on the comparison between the predicted value and the true value, and suppresses the base learner expressions with low accuracy.

[0025] According to one embodiment of the present invention, the meta-learner adjusts the parameters of each base learner by using a stochastic gradient descent method.

[0026] According to one embodiment of the present invention, the method further includes inputting the well logging data of the well to be identified into the trained base learner and meta learner to predict the fluid of each reservoir of the well to be identified.

[0027] Another aspect of the present invention provides a method for visually interpreting the recognition results of the fluid recognition method based on logging data as described above, including calculating the SHAP value of the feature to quantify the contribution of each feature marker to the model recognition result to visually interpret the recognition result of the model.

[0028] According to one embodiment of the present invention, the SHAP value obeys the following equation:

[0029] y i =y base +f(x i1 )+f(x i2 )+…+f(x ik )

[0030] Among them, y i Represents the recognition value of the model for the i-th sample, y base represents the baseline of the model, x i represents the i-th sample, x ik represents the kth feature of the i-th sample, f(x ik ) is the final recognition result y of the kth feature tag pair of the i-th sample i contribution value.

[0031] Due to the adoption of the above technical solution, the fluid identification method and visualization method based on logging data provided by the present invention has at least one of the following beneficial effects compared with the prior art:

[0032] (1) Using machine learning algorithms and combining well logging and mud logging data, we can achieve automatic and intelligent reservoir fluid identification;

[0033] (2) In the field of fluid identification based on gas logging data, different from other methods that only use a single machine learning machine for prediction, the Stacking meta-learner is applied to the overall fluid prediction for the first time, so that the model acquires an automatic parameter adjustment capability, allowing it to quickly learn new tasks based on the existing logging interpretation knowledge;

[0034] (3) Different from the previous method of using only lithology curves or gas logging curves, this method can evaluate the contribution of each parameter and compare the parameters of different methods together. The meta-learner enables the model to acquire comprehensive discrimination capabilities.

[0035] (4) Use multivariate learning methods to alleviate the overfitting problem caused by small well sections, little gas logging interpretation data, and unbalanced data results;

[0036] (5) Visualization based on the SHAP algorithm can be used to interpret and visualize the parameters of logging interpretation data for the first time, judge the rationality of the influence of hyperparameters calculated by traditional algorithms on oil and gas classification results, and compare parameters across different algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the present invention but do not constitute a limitation of the present invention. In the accompanying drawings:

[0038] Figure 1 An exemplary architecture diagram of a model for implementing the fluid identification method based on mud logging data according to the present invention;

[0039] Figure 2 It is a scatter plot of training data with C1+C2 and C3 / C1 as variables;

[0040] Figure 3 It is an overall flow chart of the fluid identification method based on well logging data according to the present invention;

[0041] Figure 4 A schematic diagram of interpretation accuracy in practical application of the fluid identification method based on logging data according to the present invention;

[0042] Figure 5 A diagram showing visualization using a waterfall chart in practical application of the visualization method according to the present invention;

[0043] Figure 6 A diagram of visualization in the form of a force diagram in practical application of the visualization method according to the present invention;

[0044] Figure 7 A diagram showing the visualization of multiple samples in the form of a force graph in practical application of the visualization method of the present invention;

[0045] Figure 8 is a feature dependency graph in practical application of the visualization method according to the present invention;

[0046] Fig. 9 It is an interactive SHAP summary diagram in the practical application of the visualization method according to the present invention. DETAILED DESCRIPTION

[0047] The specific implementation of the present disclosure is described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the present disclosure, and is not used to limit the present disclosure.

[0048] In addition, reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0049] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those skilled in the art to which this application belongs. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and will not be interpreted with an idealized or overly formal meaning unless specifically defined as in the embodiments of this application.

[0050] The fluid identification method based on logging data of the present invention generally adopts the Stacking fusion model method. This method integrates machine learning algorithms with good fluid identification effects (K nearest neighbor, support vector machine SVM, XGBoost extreme gradient boosting tree) and deep learning DNN network to train the parameters of each group of fluid interpretation and then predict the target reservoir properties.

[0051] Figure 1 The exemplary architecture diagram of the model for implementing the fluid identification method based on logging data according to the present invention is shown. The Stacking fusion model method can be divided into two layers according to the structure: the first layer is a plurality of base learners, and the original data set needs to be input into each base learner, and the second layer is a secondary learner. The output value obtained in the first layer will be input as an input value into the meta learner of the second layer to train the model and obtain the final result.

[0052] Figure 2 The scatter plot of the training data with C1+C2 and C3 / C1 as variables is shown. As can be seen from the figure, there is a lot of overlap in each explanation label. It is difficult to accurately find a good dividing line by traditional plate method alone. The data will be disturbed by various factors. Therefore, it is necessary to use multiple algorithms to make multi-faceted judgments and distinguish the internal relationship of data points as much as possible.

[0053] Figure 3 The overall flow chart of the fluid identification method based on logging data according to the present invention is shown in the figure. As shown in the figure, the fluid identification method based on logging data according to the present invention generally includes the following steps:

[0054] Step S1: Based on the well logging data and the data processed by the gas logging method, a logging data matrix is ​​prepared as the original training data, part of the data is extracted from the original training data, and the extracted data is divided into a training set and a test set;

[0055] Step S2: input the training set into each base learner among the multiple base learners for first-layer training, the target label variables of the training are the dry layer, the air layer, the wet layer and the water layer, and obtain the multiple trained base learners, combine the prediction values ​​obtained by each base learner based on the training set as the features for the next layer of training, and process the prediction values ​​obtained by each base learner based on the test set as the test data for the next layer of training;

[0056] Step S3: Input the features obtained in step S2 into the meta-learner to train the meta-learner, and input the test data into the trained meta-learner to evaluate the performance of the meta-learner.

[0057] Each operation step is described in detail below.

[0058] In step S1, a logging data matrix is ​​prepared based on the well logging data and the data processed by the gas logging method as the training original data X total . Extract part of the data from the original training data, and divide the extracted data into a training set and a test set. Among them, the logging data includes natural potential and porosity curves. The data processed by the gas logging method include at least one of the following data: data obtained by the intersection diagram method (C2 / C3, nC4 / iC4, C2 / C1, C3 / C1, C3+iC4, C1+C2, C1+C2+); data obtained by the light and heavy hydrocarbon ratio method (C2+%, C1 / C2+); data obtained by the triangle plate method (X, Y, Q); data processed by the 3H method (WH%, BH%, CH%); data representing the properties of hydrocarbons (TG%, heavy hydrocarbon%, light hydrocarbon%, non-hydrocarbon%); data representing the relative content of hydrocarbons (C1%, C2%, C3%, iC4%, nC4%, iC5%, nC5%). The training sample set should be a logging data matrix. Optionally, since each type of reservoir data may be unbalanced, the data may be preprocessed before training, for example, the data may be normalized and balanced using the SMOTE (Synthetic Minority Oversampling Technique) oversampling algorithm. Normalization processing may be performed using, for example, standardization. SMOTE can avoid overfitting. Its basic idea is to take a subset of the data of the type with a small proportion and use it as a sample, use this sample to create similar samples, and then put these newly created samples into the original data set, thus forming a new data set, which is used to train the model.

[0059] In step S2, the training set is input into each base learner of the multiple base learners for the first layer of training, and the target label variables of the training are the dry layer, the air layer, the wet layer and the water layer. The multiple base learners after training are obtained, and the predicted values ​​obtained by each base learner based on the training set are combined as the features of the next layer of training. The predicted values ​​obtained by each base learner based on the test set are processed and used as the test data for the next layer of training.

[0060] In some embodiments, multiple base learners adopt four different machine learning algorithm models: K nearest neighbor model, support vector machine SVM model, XGBoost extreme gradient boosting tree model and deep learning model.

[0061] The K nearest neighbor model is a representative of cluster analysis, which mainly judges the pattern based on the distance between the predicted data point and the learning data point. Since there is no need to train the learning data, the K nearest neighbor model has a high computational efficiency, and because the predicted data points will only be clustered into the pattern of the nearest learning data point, even if the learning sample contains a small number of erroneous samples, the model can still make effective judgments, reflecting its strong fault tolerance. The method of the present invention adopts the traditional K nearest neighbor model.

[0062] The basis of the support vector machine (SVM) model is the perceptron. The purpose of the perceptron is to find a separating hyperplane to divide samples into different categories. However, one shortcoming of the perceptron is that the separating hyperplane it finds is not unique, and many hyperplanes that meet the conditions can be found. In contrast, the support vector machine finds the separating hyperplane with the largest interval based on the perceptron. Therefore, the ultimate goal of the support vector machine is to maximize the interval between points that are close to the hyperplane. The support vector machine uses the kernel function to solve the objective function using the idea of ​​nonlinear transformation. Commonly used kernel functions include linear kernel functions, polynomial kernel functions, Gaussian kernel functions, etc.

[0063] XGBoost continuously fits the approximate residual of the data until the target condition is reached, and then comprehensively calculates the value of each leaf node to obtain the final prediction result. The submodel of the XGBoost tree model is the CART tree. The results of multiple submodels are integrated through Gradient TreeBoosting to obtain the final result. The model training process of XGBoost is: decide whether to continue splitting based on the difference between the gain values ​​before and after the node split; after a tree is fully built, pruning is required to avoid overfitting; each round of trees built compares the "residual" size between the actual value of the sample and the predicted value of the previous round, so that the prediction result gradually approaches the actual value. When splitting leaf nodes, all features and the objective function of the final node splitting should be calculated, and the feature with the largest objective function of the final node splitting and its split point are used as the best feature and the best split point for splitting. After a new tree is fully built, XGBoost will also prune it until an optimal decision tree is obtained.

[0064] The deep learning DNN network model is a method of learning data representation in machine learning, and it is also the most practical machine learning method at present. DNN is structurally divided into three categories: input layer, hidden layer, and output layer. Each layer is fully connected, and neurons in the same layer are not connected and are independent of each other. Taking fluid identification as an example, 4 sets of feature vectors are input from the input layer, and the predicted values ​​of reservoir fluid properties are output through feature extraction in the hidden layer, selection of activation function, debugging of hyperparameters, definition of loss function, and selection of optimizer.

[0065] The target label variables for training each base learner are the vector encodings of the dry layer, air layer, wet air layer, and water layer.

[0066] In some embodiments, the first layer of training can be performed using a cross-validation method, for example, a 5-fold cross-validation method can be used. First, the pre-processed data is divided into a training set and a test set. For the training data, a 5-fold cross-validation must first be performed, and it must be divided into 5 parts. Each fold uses 4 parts of the data to train the model, and the other 1 part of the data is used to predict its label. In this way, after the 5-fold cross-validation, the prediction of the entire training data label can be obtained. Stacking these predicted values ​​in their respective orders forms a new feature of the entire training data, which is a prediction of the label. Because each sample has only one predicted value, the new feature matrix has only one column, and each base model only generates one column of predicted values ​​as a new feature. In the test data, each fold uses four parts of the training data to build a model, and then substitutes the entire test data into this model to obtain the predicted value of the test data label. Therefore, after the 5-fold, all test samples will get five predicted values, so these five predicted values ​​should be simply arithmetic averaged, so that each test sample can get a unique predicted value, which is used as the feature of the test data of the new feature matrix required for the second layer model.

[0067] Each base learner will generate such a column of data. In other words, the new feature matrix must have as many columns as there are base models, the number of rows is the same as the number of rows in the training data, and the number of rows in the test data is the same as the number of rows in the original test data. When the first layer of learners has been fully learned, a new feature matrix is ​​obtained, and data is prepared for the second layer model. The new feature matrix is ​​equivalent to the second layer's X, and the label Y remains unchanged, which is still the label corresponding to the original data, which is equivalent to using the predicted label as the feature.

[0068] In the method of the present invention, the original data X total Five samples were extracted with replacement, one of which was used as the final prediction result test set, and the remaining four were sent to K nearest neighbor, support vector machine SVM, XGBoost extreme gradient boosting tree and deep learning DNN network for training and prediction.

[0069] For each base learner, when the difference in accuracy loss between each round of training and the previous round of training is less than 5% for 5 times, the parameters are considered stable and fixed. When all base learner parameters are fixed, the fitting limit is considered to have been reached and the training of the first layer is terminated.

[0070] In step S3, the features obtained in step S2 are input into the meta-learner to train the meta-learner, and the test data is input into the trained meta-learner to evaluate the performance of the meta-learner. After the meta-learner receives the training sets and prediction results from different base learners, the meta-learner directly uses the new feature matrix obtained in step S2 as X and the true label as Y for training, and then substitutes X of the test data into the model to generate the predicted value of the sample of the test data to evaluate the performance of the meta-learner. In some embodiments, the meta-learner adopts a logistic regression model. During the training process of the meta-learner, the prediction results of the base learner are "scored" according to the comparison between the predicted value of the base learner and the true value, suppressing the expression of the base learner with low accuracy, and adjusting the parameters of each base learner by stochastic gradient descent to search for the optimal parameter solution.

[0071] For the parameter adjustment of the base learner, the K nearest neighbor algorithm needs to adjust the distance parameter; SVM needs to adjust the penalty coefficient C for the error term and the kernel function; XGBoost needs to adjust the learning rate, minimum weight, tree depth, loss function and regularization parameter; the deep learning network needs to adjust the number of hidden layers and the number of neurons.

[0072] The Stacking model used in the present invention can combine the advantages of each base model and ignore the shortcomings of the base model itself, so the prediction accuracy is significantly higher than that of each single model. When the prediction effects of the base models are different and each has its own advantages and disadvantages, it is necessary to select a suitable meta-learner to optimize the final model stacking. After the model training is completed, the logging data of the well to be identified can be input into the trained base learner and meta-learner to predict the fluid of each reservoir of the well to be identified.

[0073] Since there are a large number of parameters to be trained, and the parameters of each interpretation method are random during training, the importance of each parameter in the learner needs to be explained. The SHAP algorithm (Shapley Additive ex Planations) is an algorithm for post-model interpretation, which can explain the recognition results of complex machine learning models. The fluid recognition model constructed by the Stacking fusion recognition algorithm based on logging interpretation performs well in terms of recognition accuracy, but the improvement in accuracy brings about a decrease in interpretability. The traditional chart method can draw the icon relationship between variables, while machine learning processes the parameters in a high-dimensional space, which is difficult to visualize. To this end, the present invention also provides a method for visually interpreting the recognition results of the fluid recognition method based on logging data as described above. The method includes calculating the SHAP value of the feature to quantify the contribution of each feature marker to the model recognition result to visualize the recognition result of the model.

[0074] The SHAP value obeys the following equation:

[0075] y i =y base +f(x i1 )+f(x i2 )+…+f(x ik )

[0076] Among them, y i Represents the recognition value of the model for the i-th sample, y base represents the baseline of the model, x i represents the i-th sample, x ik represents the kth feature of the i-th sample, f(x ik ) is the final recognition result y of the kth feature tag pair of the i-th sample i The contribution value of f(x ik )>0, it indicates that the feature marker improves the recognition value and has a positive effect on the recognition result; otherwise, it reduces the recognition value and has a negative effect on the recognition result.

[0077] Therefore, the contribution of specific gas logging processing parameters can be shown to guide logging and interpretation engineers in feature engineering and decision-making.

[0078] The following are relevant application examples.

[0079] Based on the logging data of single wells in Dagang Oilfield and Sulige Gasfield, the present invention uses four trained models (K nearest neighbor, support vector machine SVM, XGBoost extreme gradient boosting tree and deep learning network) to test the accuracy of the model stacking method on the validation set by cross-validation method. Figure 4 As shown in the figure, with the increase of the number of training batches, the explanation accuracy can reach 86.58%, which is about 3% higher than the traditional method.

[0080] Figure 5 The visualization method of the present invention is implemented based on the logging data of the single wells in Dagang Oilfield and Sulige Gasfield to visualize the waterfall chart. Figure 6 A diagram is provided to visualize the corresponding force diagram form. Figure 7 is the corresponding diagram for multi-sample visualization in the form of a force graph, Figure 8 is the corresponding feature dependency graph, Fig. 9 It is the corresponding interaction SHAP summary diagram.

[0081] Theoretically, the main basis for determining whether it is a gas layer is the content of related hydrocarbon substances, which is determined by Figure 5-9It can be seen that the hydrocarbon properties TG, intersection diagram method C3+iC4 and triangle diagram method X are the top three contributions, which intuitively and easily present the contribution of gas logging processing parameters. The present invention introduces a new visualization method into the study of gas logging interpretation parameters, visualizes the characteristic markers of dry layers, gas layers, wet gas layers and water layers, and deeply understands the important characteristic markers that determine the formation at the overall level.

[0082] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to the protection scope of the present invention.

[0083] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0084] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. A fluid identification method based on well logging data, characterized in that: The following steps are involved: Step S1: Based on the well logging data and the data processed by the gas logging method, a logging data matrix is ​​prepared as the original training data, part of the data is extracted from the original training data, and the extracted data is divided into a training set and a test set; Step S2: input the training set into each base learner among the multiple base learners for first-layer training, the target label variables of the training are the dry layer, the air layer, the wet layer and the water layer, and obtain the multiple trained base learners, combine the prediction values ​​obtained by each base learner based on the training set as the features for the next layer of training, and process the prediction values ​​obtained by each base learner based on the test set as the test data for the next layer of training; Step S3: Input the features obtained in step S2 into the meta-learner to train the meta-learner, and input the test data into the trained meta-learner to evaluate the performance of the meta-learner.

2. The method according to claim 1, characterized in that In step S1, the logging data includes natural potential and porosity curves.

3. The method according to claim 1, characterized in that In step S1, the data processed by the gas measurement method includes at least one of the following data: data processed by the cross plot method, data processed by the light and heavy hydrocarbon ratio method, data processed by the triangle plate method, data processed by the 3H method, data representing the properties of hydrocarbons, and data representing the relative content of hydrocarbons.

4. The method according to claim 3, characterized in that The parameters involved in the cross-plot method include C2 / C3, nC4 / iC4, C2 / C1, C3 / C1, C3+iC4, C1+C2, C1+C2+; the parameters involved in the light and heavy hydrocarbon ratio method include C2+%, C1 / C2+; the parameters involved in the triangle plate method include X, Y, Q; the parameters involved in the 3H method include WH%, BH%, CH%; the parameters involved in the properties of the hydrocarbons include TG%, heavy hydrocarbons%, light hydrocarbons%, non-hydrocarbons; the parameters involved in the relative content of hydrocarbons include C1%, C2%, C3%, iC4%, nC4%, iC5%, nC5.

5. The method according to claim 1, characterized in that Step S1 also includes normalizing and SMOTE oversampling the numerical features in the data.

6. The method according to claim 1, characterized in that The multiple base learners adopt a K nearest neighbor model, a support vector machine SVM model, an XGBoost extreme gradient boosting tree model and a deep learning model.

7. The method according to claim 6, characterized in that The parameter that needs to be adjusted during training of the K nearest neighbor model is the distance.

8. The method according to claim 6, characterized in that The support vector machine (SVM) model needs to adjust the penalty coefficient C and kernel function for the error term during training.

9. The method according to claim 6, characterized in that The XGBoost extreme gradient boosting tree model requires adjusting the learning rate, minimum weight, tree depth, loss function and regularization parameters during training.

10. The method according to claim 6, characterized in that The deep learning network model needs to adjust the number of hidden layers and the number of neurons during training.

11. The method according to claim 6, characterized in that The training termination condition for each base learner is that when the difference between the accuracy loss value of each round of training and the previous round is less than 5% for 5 times, the parameters are fixed and the training of the base learner is terminated.

12. The method according to claim 1, characterized in that The first-level training is performed using a cross-validation method, including: dividing both the training set and the test set into k parts, using the k data sets of the training set as validation sets in turn, and using the remaining k-1 data sets as training data sets. After the same base learner completes training on the training set, it predicts the validation set and the test set. Finally, the prediction results on the k validation sets are merged as features for the next level of training, and the prediction value of the final test set is obtained by averaging the k prediction results on the test set.

13. The method according to claim 12, characterized in that The cross-validation method is a 5-fold cross-validation method.

14. The method according to claim 1, characterized in that The meta-learner adopts a logistic regression model.

15. The method according to claim 14, characterized in that The meta-learner "scores" the prediction results of the base learner according to the comparison between the predicted value and the true value, suppressing the base learner expression with low accuracy.

16. The method according to claim 15, characterized in that The meta-learner adjusts the parameters of each base learner by using stochastic gradient descent.

17. The method according to claim 1, characterized in that It also includes predicting the fluid of each reservoir of the well to be identified by inputting the logging data of the well to be identified into the trained base learner and meta learner.

18. A method for visually interpreting the identification result of the fluid identification method based on logging data according to any one of claims 1 to 17, characterized in that: It includes calculating the SHAP value of the feature to quantify the contribution of each feature marker to the model recognition result to visually explain the model's recognition results.

19. The method according to claim 18, characterized in that The SHAP value obeys the following equation: y i =y base +f(x i1 )+f(x i2 )+…+f(x ik ) Among them, y i Represents the recognition value of the model for the i-th sample, y base represents the baseline of the model, x i represents the i-th sample, x ik represents the kth feature of the i-th sample, f(x ik ) is the final recognition result y of the k-th feature label of the i-th sample i contribution value.

Images