A lithology identification method based on empirical mode decomposition and heterogeneous online learning

Abstract

The application aims to provide a lithology identification method based on empirical mode decomposition and heterogeneous online learning, comprising the following steps: firstly, data preprocessing is performed on the obtained drilling parameters, including feature selection, missing value processing, filtering illegal data, and feature data target data coding; then, the feature parameters are decomposed based on an empirical mode decomposition model; m offline models and m online models are trained for the decomposed feature parameters; finally, the heterogeneous models are fused by combining adaptive weights, and are used for identifying the formation lithology. The application applies empirical mode decomposition and machine learning algorithm to lithology identification, proposes a mode decomposition method for drilling parameters and a machine learning training architecture, and applies joint domain distribution (JDA) to machine learning model training, so that the identification accuracy and robustness of the original offline model can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the fields of machine learning technology and oil exploration and development technology, and particularly relates to a lithology identification method based on empirical mode decomposition and heterogeneous online learning. BACKGROUND

[0002] The EMD signal decomposition method can be decomposed by the time scale related characteristics of the data itself without manually setting any form of basis function in advance, and it has no specific requirements for the type of signal to be decomposed, has a very obvious advantage in processing non-stationary and nonlinear data, and has a very high signal-to-noise ratio. Machine learning refers to learning the general rules from limited observation data, then establishing a learning model, and using the established model to predict unknown data. The original data is extracted as a set of features through artificial experience or feature conversion method, and then the features are input into the learning model and the prediction result is output. Machine learning is based on analytical mathematics and statistics, but it is different from them. Machine learning is to learn the relationship between data by algorithm and to establish a mathematical model. Whether the learning criterion is met is judged, and the model is optimized by gradient descent in the iteration process. Common machine learning algorithms include perceptron, linear model, decision tree, support vector machine, Bayesian method, clustering method, graph method and ensemble learning. The initial learning method of machine learning is to learn some static data with the same part. There are some bottlenecks and limitations in the learning of non-stationary stream data. Online learning shows good results in processing non-stationary stream data. The basis of intelligent analysis and optimization of drilling parameters is the timely judgment of downhole rock properties. In recent years, with the revival of machine learning, the continuous optimization of artificial intelligence algorithms, the collection of high-quality data sets and the improvement of computer computing power, emerging information technology has achieved remarkable results in various industries. Machine learning algorithms also have powerful high-dimensional nonlinear mapping capabilities and big data processing capabilities, and play a huge role in different industries. Using classification models can accurately analyze related problems. Domestic and foreign research on drilling parameter analysis mainly focuses on establishing mathematical equations or relying on traditional machine learning methods using offline training and online identification mode. The present application uses empirical mode decomposition to decompose complex non-stationary logging data into high-frequency to low-frequency component information, considers using the decomposed parameter information for training of the machine learning model, the proposed algorithm can effectively use the adjacent well data, reduce the data distribution difference, combine the offline and online models heterogeneously, and under the adaptive weight setting, the algorithm is more suitable for stream data with concept drift, the new sample data has better promotion while maintaining the excellent characteristics of the algorithm for most samples. According to the judgment of the identification model, the drilling parameters are adjusted in time and effectively, which is of great significance to improve the drilling speed, reduce the drill bit wear and increase the downhole tool life. SUMMARY

[0003] The present application aims to lay the foundation for intelligent drilling, and designs a lithology identification method based on empirical mode decomposition and heterogeneous online learning.

[0004] A lithology identification method based on empirical mode decomposition and heterogeneous online learning, comprising the following steps:

[0005] S1: Empirical mode decomposition of drilling parameter signals;

[0006] S2: JDA mapping processing of new well data and adjacent well data;

[0007] S3: Based on the mapped data, m offline models are established for m adjacent wells respectively;

[0008] S4: m online models are established for new well data;

[0009] S5: Weighted combination to obtain a strong learner.

[0010] Further, a lithology identification method based on empirical mode decomposition and heterogeneous online learning, the step S1 comprises the following sub-steps:

[0011] S11: Obtain all local extreme points of the signal to be decomposed;

[0012] S12: According to the extreme points, a cubic spline interpolation method is used to construct the upper and lower envelope lines of the signal, respectively denoted as QUOTE and ;

[0013] S13: Calculate the mean value of the upper and lower envelope lines at each point, denoted as , and the calculation formula is:

[0014]

[0015] Wherein represents the mean value of the upper and lower envelope lines at each point, represents the upper envelope line, ;

[0016] S14: Screen out the envelope mean value from the original signal QUOTE , and the remaining is denoted as , and the calculation formula is:

[0017]

[0018] Wherein ​​all local extreme points of the original signal;

[0019] If satisfy two conditions of intrinsic mode function and stop criterion, QUOTE QUOTE QUOTE As the first intrinsic mode function, the current residual term is taken as the initial signal for the next iteration screening ;

[0020] If not, the is taken as the initial signal for the next iteration QUOTE The above four processes are repeated until the stop criterion is met, The value range of is generally between 0.2-0.3, and the calculation formula is:

[0021]

[0022] Wherein , represents the relative deviation;

[0023] The above steps are repeated until the residual term is a monotonic function, and finally the decomposition mode functions (IMFs) of the original signal are obtained.

[0024] Further, a lithology identification method based on empirical mode decomposition and heterogeneous online learning, the step S2 includes the following sub-steps:

[0025] S21: Joint domain distribution uses maximum mean difference, the processing process is to adjust the edge distribution and conditional distribution between samples by matrix operation, construct an effective and stable new feature representation method to reduce the sample distribution difference.

[0026] S22: Use the neighbor well data and part of the unknown label new well data to initialize the mapping matrix ,

[0027] In the offline process, each neighbor well data is directly mapped, and the calculation formula is:

[0028]

[0029] Wherein ;

[0030] In the online process, the same transformation matrix is used for processing, and with the continuous increase of new data, the difference between the edge distribution and the conditional distribution is adjusted using the new well data. In the online process, the is iteratively updated, and the update expression is:

[0031]

[0032] wherein represents the sample class, represents the average value of the characteristic attribute of the sample class in the adjacent well m, represents the average value of the characteristic attribute of the sample class in the new well, is a regularization coefficient.

[0033] Further, a lithology identification method based on empirical mode decomposition and heterogeneous online learning, the step S3 is for the training set Tr processed by steps S1 and S2 {( ),···( ),···, ( )}, iterates T rounds in extreme gradient boosting, including the following sub-steps:

[0034] S31: the first-order derivative of each sample ( ) in the current round loss function L is based on and the second-order derivative ; ;

[0035] The sum of the first-order derivatives of all samples is

[0036]

[0037] wherein represents the sum of the first-order derivatives, represents a certain sample,

[0038] The sum of the second-order derivatives of all samples is:

[0039]

[0040] wherein represents the sum of the second-order derivatives, represents a certain sample,

[0041] After XGBoost optimization, the following regular term is added to the loss function of GBDT:

[0042]

[0043] represents the number of leaf nodes in the decision tree, represents the optimal value of the th leaf node, is a weight coefficient;

[0044] S32: Attempt to split the decision tree according to the current node, set the default score =0, calculate the sum of the first derivative and the second derivative of the current node that needs to be split;

[0045] For feature sequence number k=1,2...K:

[0046] S321: Enumerate the first k samples in turn, and put the current sample into the left subtree. The sum of the first derivative and the second derivative of the left and right subtrees after the current sample is put into the left subtree, and the calculation formula is:

[0047] S3211:

[0048] = 0, = 0

[0049] Wherein represents the first derivative sum of the left subtree, represents the second derivative sum of the left subtree;

[0050] S3212:

[0051] ,

[0052] ,

[0053] Wherein represents the first derivative sum of the right subtree, represents the second derivative sum of the right subtree;

[0054] S322: Enumerate the first k samples in turn, and put the current sample into the right subtree. The sum of the first derivative and the second derivative of the left and right subtrees after the current sample is put into the right subtree, and the calculation formula is:

[0055] S3221:

[0056] = 0, = 0

[0057] S3222:

[0058] ,

[0059] ,

[0060] S323: Attempt to update score, and the update formula is:

[0061]

[0062] ​​wherein the score of the decision tree at the time of splitting;

[0063] S33: based on the maximum the leaf node of the sub-tree is split according to the corresponding division feature and the feature value;

[0064] S34:

[0065] If the maximum is 0, the current node is split, and the of other leaf regions is calculated to obtain the base learner , the strong learner is updated, and the next round of weak learner iteration is entered;

[0066] If the maximum is not 0, go to step S32 to continue to try to split the decision tree.

[0067] Further, a lithology identification method based on empirical mode decomposition and heterogeneous online learning, the step S4 model establishment process is that the online extreme gradient boosting is improved on the basis of the overall structure of the algorithm for the flow data, and the basic principle and the offline model are the same.

[0068] Further, a lithology identification method based on empirical mode decomposition and heterogeneous online learning, the step S4 includes the following sub-steps:

[0069] S41: For the processing of data stream, the online extreme gradient boosting algorithm stores the sequentially arrived data in a buffer, and once the buffer reaches the upper limit, all data in the buffer is sent to the offline extreme gradient boosting in a small batch manner to construct a base learner based on the current data, iteratively initializes the buffer-buffer-training process, and trains the next base learner f i+1 ;

[0070] S42: The establishment of the first base learner is based on the newly obtained buffered data and the loss value generated by the previous base learner, and the calculation of the loss value is only based on the number ( T) of base learners contained in the current strong learner, and the calculation formula is:

[0071] wherein represents the number of leaf nodes of each base boosting tree, represents the optimal value of the th leaf node of the current boosting tree, , ; and

[0072] S43: In solving the problem of self-adaptation, the size of the cache window is defined, and the calculation formula is:

[0073]

[0074] The size of the cache window when the i-th batch of data is used to build the model;

[0075] S44: Before the number of base learners included in the online model reaches the maximum value, the new learner will be directly added to the strong learner;

[0076] S45: After reaching the maximum number , the update strategy of the online model is divided into push and replace two ways,

[0077] push uses the way of queue to delete the oldest base model and add the latest base model;

[0078] replace directly replaces the oldest base model;

[0079] S46: In terms of weight setting, let the weight will make the relatively old model in the push way have an impact on the prediction result, and in the replace way, the new model will have an impact on the prediction result.

[0080] Further, a lithology identification method based on empirical mode decomposition and heterogeneous online learning, the step S5 includes the following sub-steps:

[0081] S51: Add model weight to each model, and the weights of the m-th offline and online models are defined as , , and the weight is initialized as ;

[0082] It is hoped that the model weight can be self-adaptively corrected and updated, and the update formula of the weight is defined as follows:

[0083]

[0084]

[0085] wherein is a hyperparameter, represents the prediction result of the offline model m on the historical well m, represents the prediction result of the online model m on the new well m, represents the true label result represents the weight of the m-th offline model, represents the weight of the m-th online model;

[0086] S52: Based on the offline and online heterogeneous model and the model weight updating method, the adaptive weight combination strategy is constructed, and the calculation formula is as follows:

[0087]

[0088] Wherein the output represents a dimension vector, wherein the number of classes of samples.

[0089] The present application has the beneficial effects: the empirical mode decomposition and machine learning algorithm are used for lithology identification, the mode decomposition method of drilling parameters and the training architecture of machine learning are proposed, and the joint domain distribution (JDA) is applied to the training of machine learning model, which can improve the identification accuracy and robustness of the existing offline model. BRIEF DESCRIPTION OF DRAWINGS

[0090] Figure 1 is the algorithm flow structure diagram of the present application.

[0091] Figure 2 is the empirical mode decomposition flow structure diagram of the present application.

[0092] Figure 3 is the online machine learning schematic diagram of the present application. DETAILED DESCRIPTION

[0093] The present application is further described, but the protection scope of the present application is not limited to the following. A lithology identification method based on empirical mode decomposition and heterogeneous online learning, comprising the following steps:

[0094] S1: Empirical mode decomposition of drilling parameter signal;

[0095] S2: JDA mapping processing is performed on new well data and adjacent well data;

[0096] S3: Based on the mapped data, m offline models are established for m adjacent wells respectively;

[0097] S4: m online models are established for new well data;

[0098] S5: Weighted combination to obtain a strong learner;

[0099] Further, step S1 includes the following sub-steps:

[0100] S11. Obtain all local extreme points of the signal to be decomposed

[0101] S12. According to the extreme point, a cubic spline interpolation method is used to construct the upper and lower envelope lines of the signal, respectively denoted as QUOTE and

[0102] S13. The mean values of the upper and lower envelope lines at each point are calculated, denoted as as shown in the following formula

[0103]

[0104] S14. The original signal is screened from the envelope mean QUOTE , and the remaining is denoted as as shown in the following formula. If satisfies two conditions of the intrinsic modal function and the stopping criterion, QUOTE QUOTE QUOTE as the first intrinsic modal function, the current residual term is taken as the initial signal for the next iteration screening . If not, the is taken as the initial signal for the next iteration QUOTE , and the above four processes are repeated until the stopping criterion is met, The value range of is generally between 0.2-0.3.

[0105]

[0106]

[0107] The above steps are repeated until the residual term is a monotonic function, and finally the decomposition modal functions (IMFs) of the original signal are obtained.

[0108] Further, the procedure of step S2 is as follows:

[0109] The joint domain distribution uses the maximum mean difference to process the edge distribution and the conditional distribution between the samples by matrix operation to construct an effective and stable new feature representation method to reduce the sample distribution difference. In this step, the adjacent well data and part of the unknown label new well data are used to initialize the mapping matrix , and the offline process directly maps each adjacent well data , and the same transformation matrix is used in the online process, but as new data is continuously added, more new well data can be used to adjust the difference between the edge distribution and the conditional distribution, so the online process is used to The iterative update, update expression as follows.

[0110]

[0111] wherein represents the sample category, represents the average value of the feature attribute of the sample category in the adjacent well m, represents the average value of the feature attribute of the sample category in the new well.

[0112] Further, the step S3 off-line model establishment step as follows:

[0113] For the training set Tr= {( ),···( ),···, ( )} processed in step S1 and S2, in the extreme gradient lifting iteration T rounds, the steps of the algorithm as follows

[0114] S31. Each sample ( ) in the current round loss function L based on first derivative and second derivative ; the sum of the first derivative of all samples is , the sum of the second derivative is . It is worth noting that XGBoost optimization after the loss function in GBDT added as follows regular term:

[0115]

[0116] In the formula is the number of leaf nodes in the decision tree, is the optimal value of the th leaf node.

[0117] S32. According to the current node to try to split the decision tree, set the default score =0, calculate the sum of the first derivative and the second derivative of the node to be split.

[0118] For feature serial number k=1,2...K:

[0119] S321. In the form of enumeration, the th sample is taken out in turn, and the sum of the first and second derivatives of the left and right child trees after the current sample is put into the left child tree is calculated

[0120] S3211. = 0, = 0

[0121] S3212. ,

[0122] ,

[0123] S322. Enumerate the way to take out the first sample in turn, calculate the sum of the first and second derivatives of the left and right sub-trees after the current sample is put into the right sub-tree

[0124] S3221. = 0, = 0

[0125] S3222. ,

[0126] ,

[0127] S323. Try to update score

[0128]

[0129] S33. Based on the maximum corresponding split feature and feature value to split the leaf node of the sub-tree

[0130] S34. If the maximum is 0, the current node is split, and the of other leaf regions is calculated to obtain the base learner , the strong learner is updated , and the next round of weak learner iteration is entered. If the maximum is not 0, go to step S32 to continue trying to split the decision tree.

[0131] Further, the model establishment process of step S4 is as follows:

[0132] Online extreme gradient boosting is an improved algorithm for the overall structure of the algorithm based on offline gradient boosting, and its basic principle is the same as that of the offline model.

[0133] For the processing of data streams, the online extreme gradient boosting algorithm stores the sequentially arriving data in a buffer, and once the buffer reaches the upper limit, all data in the buffer is sent to the offline extreme gradient boosting in small batches to build a base learner based on the current data , and iteratively initializes the cache-cache-training process to train the next base learner . Unlike offline XGB, the establishment of the first The loss value generated by the base learner, and the calculation of the loss value is only based on the number of base learners contained in the current strong learner T, as shown in the following formula, is the number of each base boosting tree leaf node, is the optimal value of the current boosting tree th leaf node.

[0134]

[0135] In solving the adaptive problem, the size of the cache window is defined as follows, indicates the size of the cache window when the i-th batch of data is used to build the model.

[0136]

[0137] The number of base learners contained in the online model reaches the maximum value before the new learner is directly added to the strong learner; after reaching the maximum number The update strategy of the online model is divided into push and replace two ways, push uses the way of queue to delete the oldest base model, and adds the latest base model; replace directly replaces the oldest base model. In addition, in the weight setting aspect, let the weight will make the relatively old model in the push way have a greater impact on the prediction result, and in the replace way, the new model will have a greater impact on the prediction result.

[0138] Further, step S5 details the steps:

[0139] The algorithm increases the model weight for each model, and the weight of the m-th offline and online model is defined as , , and the weight is initialized as . In the algorithm, the model weight is expected to be self-adaptively corrected and updated, and the update method of the weight is defined as shown in the following formula.

[0140]

[0141]

[0142] wherein is a hyperparameter, indicates the prediction result of the offline model m on the data history well m, indicates the prediction result of the online model m on the data new well m, indicates the true label result.

[0143] ​Based on the offline and online heterogeneous model and the model weight updating method, the adaptive weight combination strategy is constructed as follows: the output is a dimensional vector, wherein

[0144] Embodiment 1

[0145] The steps of the present application are described in detail below with the drilling data of an oilfield as an example.

[0146] A lithology identification method based on empirical mode decomposition and heterogeneous online learning:

[0147] Initialize the weight parameters

[0148] Input: pretreated drilling parameters as input;

[0149] Step 1: consider the sequence mode, and construct an empirical mode decomposition method;

[0150] Step 2: use joint domain distribution to reduce the distribution difference of interwell data for the decomposed data;

[0151] Step 3: train the offline model for the adjacent well data;

[0152] Step 4: construct the online model for the streaming data of the new well;

[0153] Step 5: update the offline model and online model weights according to the new drilling data;

[0154] The weight parameter is finally adjusted to through experiments;

[0155] Step 6: integrate the learner to output the identification result and draw the confusion matrix;

[0156] Output: lithology classification confusion matrix of the new well.

[0157] The present scheme uses empirical mode decomposition and machine learning algorithm for lithology identification through a lithology identification method based on empirical mode decomposition and heterogeneous online learning, proposes a mode decomposition method of drilling parameters and a training architecture of machine learning, and applies joint domain distribution (JDA) to the training of machine learning model, which can improve the recognition accuracy and robustness of the existing offline model.

[0158] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A lithology identification method based on empirical mode decomposition and heterogeneous online learning, characterized in that, Comprise the following steps: S1: Empirical mode decomposition of drilling parameter signal; S2: JDA mapping processing of new well data and adjacent well data, comprising the following sub-steps: S21: Joint domain distribution uses maximum mean difference, the processing process is to adjust the edge distribution and conditional distribution between samples through matrix operation, to construct an effective and stable new feature representation method to reduce the sample distribution difference; S22: Use the adjacent well data and the new well data with partial unknown labels to initialize the mapping matrix ; In the offline process, the mapping processing is directly performed on each adjacent well data, and the calculation formula is: wherein ; The same transformation matrix will be used in the online process, and as new data is constantly added, the difference between the marginal distribution and the conditional distribution is adjusted using the new well data, and the iteration is updated in the online process The update expression is: wherein represents the sample class, represents the average of the characteristic attribute of the sample class in the neighboring well m, represents the average of the characteristic attribute of the sample class in the new well, a regularization coefficient; S3: Based on the mapped data, m offline models are established for m adjacent wells respectively; S4: m online models are established for new well data, comprising the following sub-steps: S41: For the processing of data stream, the online extreme gradient boosting algorithm stores the sequentially arrived data in a buffer, once the buffer reaches the upper limit, all data in the buffer is sent to the offline extreme gradient boosting in small batches to build the base learner based on the current data , iteratively initializes the buffer-buffer-training process to train the next base learner f i+1 ; S42: No. The base learner is built based on the newly acquired cached data and the previous data. The loss value generated by each base learner is calculated based solely on the number of base learners contained in the current strong learner. ( T), the calculation formula is: wherein represents the number of each base boosting tree leaf node, w T′j represents the optimal value of the current boosting tree th leaf node, , ; S43: In solving the adaptive problem, the size of the cache window is defined, and the calculation formula is: a cache window size when building a model representing the ith batch of data; S44: Before the number of base learners contained in the online model reaches the maximum value, the new learner will be directly added to the strong learner; S45: In reaching the maximum number After that, the update strategy of online model is divided into two ways: push and replace. Push uses the way of queue to delete the oldest base model and add the latest base model; Replace directly replaces the oldest base model; S46: In terms of weight settings, let the weight will make the relatively old model in the push way will have an impact on the prediction results, in replace mode, the new model will have an impact on the prediction results. S5: Weighted combination to obtain a strong learner, comprising the following sub-steps: S51: increase the model weight for each model, the weight of the mth offline and online model is defined as , respectively, and the weight is initialized as ; The model weight is expected to be self-adaptive to correct and update, and the weight update formula is defined as follows: wherein hyperparameters of the offline model, represents the prediction result of the offline model m on the data history well m, represents the prediction result of the online model m on the data new well m, represents the real label result represents the weight of the mth offline model, represents the weight of the mth online model; S52: Based on the heterogeneous model of offline and online and the model weight update method, the following adaptive weight combination strategy is constructed, and the calculation formula is: where the output represent dimensional vector, where represents the number of classes of samples. 2.The lithology identification method based on empirical mode decomposition and heterogeneous online learning according to claim 1, characterized in that, The step S1 comprises the following sub-steps: S11: obtaining a signal to be decomposed all local extreme points of the signal S12: according to the extreme point, a cubic spline interpolation method is used to construct the upper and lower envelope lines of the signal, respectively denoted as and ; S13: Calculate the mean of the upper and lower envelope lines at each point, denoted as , with the following formula: wherein represents the mean value of the upper and lower envelope at the respective point, represents the upper envelope, ; S14: original signal Mid-sifting envelope mean The remainder is noted as The calculation formula is: wherein represent all local extremum points of the original signal; If the two conditions of the intrinsic mode function are satisfied and the stopping criterion is met, the first intrinsic mode function is taken as the initial signal for the next iteration of the screening of the current residual term ; If not, then as the initial signal for the next iteration The above four procedures are repeated until the stopping criterion is satisfied, The value of is generally between 0.2 and 0.3, and the calculation formula is: wherein , represents the relative deviation; Repeat the above steps until the residual term is a monotonic function, and finally obtain each decomposition modal function (IMF) of the original signal. 3.The lithology identification method based on empirical mode decomposition and heterogeneous online learning according to claim 1, characterized in that, The step S3 is for the training set Tr= {( ),···( ),···, ( )} processed by steps S1 and S2, and is iterated T rounds in extreme gradient boosting, including the following sub-steps: S31: each sample ( ) at the current round loss function L based on first derivative and second derivative ; The sum of the first derivatives of all samples is wherein representing the first derivative of the sum, representing a certain sample, The sum of the second derivatives of all samples is: wherein representing the second derivative of the sum, representing a certain sample, XGBoost adds the following regular term to the loss function of GBDT after optimization: a number of leaf nodes in the decision tree, an optimal value of the i-th leaf node, an optimal value of the i-th leaf node, a weight coefficient; S32: try to split the decision tree according to the current node, set the default score = 0, calculate the sum of the first derivative and the second derivative of the current node that needs to be split; For feature serial number k=1,2...K: S321: enumerated in turn to take the first sample, the current sample into the left child tree after the sum of the first and second derivatives of the left and right child trees, the formula is: S3211： = 0， = 0 wherein represents the first order sum of the left sub-tree, represents the second order sum of the left sub-tree; S3212： , , wherein right child first order sum, representing the right child second order sum; S322: The enumerated way takes out the first sample in turn, and the sum of the first and second derivatives of the left and right sub-trees after the current sample is put into the right sub-tree, and the calculation formula is: S3221： = 0， = 0 S3222： , , S323: Try to update the score, and the update formula is: wherein Split score at decision tree split; S33: based on the maximum The corresponding division feature and feature value split the leaf nodes of the sub-tree; S34: If the maximum number of splits is 0, the current node is split completely, and the other leaf regions are calculated , and the base learners are obtained. The strong learner is updated, and the next round of weak learner iteration is entered. ​ If the maximum If not 0, go to step S32 to continue trying to split the decision tree.

4. The lithology identification method based on empirical mode decomposition and heterogeneous online learning according to claim 1, characterized in that, The model establishment process of the step S4 is that the online extreme gradient boosting improves the overall structure of the algorithm based on the offline gradient boosting, and the basic principle is the same as that of the offline model.

Images