Intelligent stratum lithology identification method based on logging data

Abstract

The invention relates to an intelligent stratum lithology identification method based on logging data. The method comprises the following steps: collecting the logging data; preprocessing the data; extracting features; an LSTM-CRF model is adopted, training set data serve as input of the LSTM model, the CRF model is applied to output of the LSTM model, and stratum lithology is output; and evaluating the model. The method combines the long-term dependence of the recurrent neural network and a deep learning model of the sequence labeling capability of the CRF, and deduces the formation lithology of clay, sandstone, glutenite, limestone, mudstone and dolomite through logging data of natural gamma, sound wave, resistivity and the like. According to the method, the features are automatically extracted from the logging data through a machine learning algorithm, and the lithology identification is carried out by using the features, so that the identification accuracy can be effectively improved; the method does not depend on geological conditions of a specific region, is suitable for logging data of different regions, and has wide applicability.

Description

Technical Field

[0001] The invention relates to the technical field of geotechnical engineering, and in particular to an intelligent identification method of stratum lithology based on well logging data. Background Art

[0002] In the field of geotechnical engineering technology, the monitoring function of the drilling rig itself is combined with the sensors installed on the drill pipe short section to collect the logging data of the drilling rig when drilling in different strata. By preprocessing and cleaning the logging data data set, the main characteristics of different stratum lithology and geological structure are determined, and the stratigraphic model under different strata is established. The reliability of the stratigraphic model is verified by comparing the stratigraphic model with the existing geological exploration data. In geological exploration, accurate identification of stratigraphic lithology is crucial to understanding the underground structure and evaluating the potential of mineral resources. Traditional methods rely on manual interpretation by geologists, which is inefficient and highly subjective. In recent years, with the development of big data and cloud computing technology, it has become possible to use cloud computing platforms for large-scale logging data processing and analysis, especially the advantages of deep learning in processing sequence data, which provides more powerful computing power and storage capacity for the intelligent identification method of stratigraphic lithology based on logging data. Therefore, it is necessary to further study the intelligent identification method of stratigraphic lithology. Summary of the invention

[0003] In order to overcome the shortcomings of existing formation identification methods, such as being time-consuming, labor-intensive and having large errors, the present invention provides a formation lithology intelligent identification method based on well logging data, which can infer formation lithology through well logging data.

[0004] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0005] The present invention provides a method for intelligently identifying formation lithology based on well logging data, comprising the following steps:

[0006] (1) Collecting logging data: Collecting logging data of the formation to be identified;

[0007] (2) Data preprocessing: preprocessing the collected logging data;

[0008] (3) Feature extraction: Extract features from the preprocessed logging data and create training and test sets;

[0009] (4) Model training: The LSTM-CRF model is used, the training set data is used as the input of the LSTM model, and the CRF model is applied to the output of the LSTM model to output the lithology of the formation;

[0010] (5) Model evaluation: The performance of the model is evaluated through evaluation indicators: accuracy and recall.

[0011] Preferably, in step (1), the logging data includes any one or more of wellbore diameter, natural gamma ray spectrum, natural potential, density, neutron porosity, acoustic wave time difference, deep lateral resistivity, and shallow lateral resistivity.

[0012] More preferably, in step (1), the logging data include natural gamma ray spectrum, natural potential, density, acoustic time difference and deep lateral resistivity.

[0013] Among them, borehole diameter: measuring the diameter of the borehole, used to determine the size and shape of the borehole, is very important for designing and implementing drilling operations. Natural gamma ray spectrum: by measuring the intensity of gamma rays naturally emitted by rocks, we can understand the composition of rocks, especially the mud content and clay type, which is important for stratigraphic comparison and source rock identification. Natural potential: measuring the electrochemical properties of rocks can reflect the permeability of rocks and the properties of formation water, which is very helpful for evaluating the water content of reservoirs. Density: by measuring the density of rocks, we can understand the physical properties and porosity of rocks, which is important for determining rock types and reservoir properties. Neutron porosity: neutron logging can infer the porosity and fluid type of rocks by measuring the characteristics of the interaction between neutrons and rock nuclei. Acoustic time difference: measuring the propagation speed of sound waves in rocks, which can be used to calculate the elastic modulus of rocks and judge the degree of compaction of rocks. Deep lateral resistivity and shallow lateral resistivity: resistivity logging can reflect the conductivity of rocks, thereby inferring the water content and oil and gas content of rocks. Deep lateral and shallow lateral resistivity can provide conductivity information at different depths, which is very important for identifying oil and gas layers and evaluating their productivity.

[0014] Preferably, in step (2), wavelet transform is first used for denoising, and then Z-score normalization is used to perform a normalization operation on the basis of the denoised data set to improve data quality.

[0015] The specific steps of wavelet transform denoising are: first analyze the wavelet coefficients to identify the noise component and the signal component, then perform threshold processing on the noise component (i.e., the high-frequency detail part), such as hard threshold or soft threshold method, to reduce or eliminate the influence of noise; retain the signal component (i.e., the low-frequency approximate part), which is the main information of the data set. Finally, perform inverse wavelet transform on the processed high-frequency detail part and low-frequency approximate part to obtain the denoised logging data set.

[0016] Preferably, in step (2), the formula for normalizing the Z score is: Where μ is the mean of the original logging data, and σ is the standard deviation of the original logging data.

[0017] Preferably, in step (3), the ratio of the training set to the test set is 8:2; the training set is used for LSTM-CRF model training, and the test set is used for LSTM-CRF model evaluation.

[0018] Preferably, in step (3), feature extraction is to extract features for formation lithology identification from preprocessed logging data; and reduce feature dimensions using principal component analysis;

[0019] Among them, the features used for formation lithology identification are features that are helpful in distinguishing formation lithology. Using the principal component analysis (PCA) method to reduce feature dimensions is to simplify the data set by reducing the number of variables, extract the main characteristic components of the data, and thus obtain the principal components that have the greatest impact on the data, while retaining the main information of the data as much as possible.

[0020] Preferably, in step (4), the LSTM model includes the input x at time t t , cell state C t , candidate cell states Hidden state h t , forget gate f t , input gate i t and output gate O t ;

[0021] Among them, the LSTM (Long Short-Term Memory Neural Network) model is a type of RNN (Recurrent Neural Network) model, which is suitable for processing sequence data and can learn the rules in the data from the input sequence. By forgetting the information in the cell state and inputting (memorizing) new information, the information useful for subsequent moment calculations can be transmitted, while the useless information is discarded, and the hidden state h is output at each time step. t , where forgetting, input and output are determined by the hidden state h at the previous moment t-1 And the forget gate f calculated by the current input t , input gate i t , output gate O t To control (such as Figure 1 shown).

[0022] Preferably, the formula involved in step (4) is as follows:

[0023] Forget Gate:

[0024] f t =σ(W f ·[h t-1 , x t ]+b f )

[0025] In the formula, f t is the output of the forget gate, σ is the Sigmoid function, W f is the weight matrix of the forget gate, b f is the bias term of the forget gate, h t-1is the hidden state of the previous time step, x t is the input at the current time step (time t):

[0026] Input Gate:

[0027] i t =σ(W i ·[h t-1 , x t ]+b i )

[0028] Candidate cell states:

[0029]

[0030] In the formula, i t is the input gate output, is the candidate cell state: tanh is the hyperbolic tangent function: Wi is the weight matrix of the input gate, b i is the bias term of the input gate, W c is the weight matrix of the candidate cell state, b c is the bias term of the candidate cell state,

[0031] The cell state at time t is:

[0032]

[0033] In the formula, C t is the cell state; C t-1 is the cell state at the previous time step;

[0034] Output Gate:

[0035] O t =σ(W 0 ·[h t-1 , x t ]+b 0 )

[0036] The hidden state at time t:

[0037] h t =O t tanh(C t )

[0038] In the formula, O t is the output gate output, h t is the hidden state at the current time step (time t), W 0 is the weight matrix of the output gate, b 0 is the bias term of the output gate.

[0039] Preferably, in step (4), the sequence output by the LSTM model is labeled using a CRF model, and the labeled sequence is used to train the CRF model to output the optimal formation lithology;

[0040] Preferably, the LSTM model sets hyperparameters before starting training; the CRF model is trained using the maximum likelihood estimation method; and the global optimal solution is output through Viterbi algorithm decoding.

[0041] Among them, setting hyperparameters includes: Hidden layer dimension: determines the dimension of the hidden state in the LSTM unit. Generally speaking, a larger hidden layer dimension can provide more model capacity, but it also increases the computational cost. The recommended range is usually between 50 and 500. Learning rate: controls the degree to which the model updates the weights in each iteration. A learning rate that is too high may cause the model to fail to converge, while a learning rate that is too low may cause slow training. The recommended range is usually between 0.001 and 0.1. Batch size: specifies the number of samples used in each training update. A larger batch size can speed up training, but may reduce the generalization performance of the model. The recommended range is usually between 16 and 128. Number of training iterations: specifies the number of times the model iterates on the entire dataset. A larger number of iterations can improve the performance of the model, but it will also increase the training time. It is recommended to choose an appropriate number of iterations based on the convergence of the model and the time limit.

[0042] Among them, the CRF (Conditional Random Field) model is a probabilistic model. CRF is applied to the output of the LSTM layer to label the sequence. The CRF layer defines the conditional probability distribution of the output sequence given a set of input feature sequences. The model is trained using the maximum likelihood estimation method. The core of maximum likelihood estimation is to determine a set of parameters under the premise of given observation data so that these parameters can "explain" or "generate" these observation data to the greatest extent. In sequence labeling models such as conditional random fields (CRF), maximum likelihood estimation is used to optimize model parameters to improve labeling accuracy. The CRF model can consider the dependencies between labels, improve the accuracy of labeling, and use the Viterbi algorithm to obtain the global optimal solution.

[0043] Preferably, in step (4), the CRF model is given an input sequence X=(x 1 , x 2 , ..., x n ), predict the output sequence Y = (y 1 ,y 2 , ..., y n ). The CRF model defines a conditional probability distribution P(Y|X), which is expressed by the following formula:

[0044]

[0045] Where Z(X) is the normalized analysis function, λ j is the jth characteristic function f j The associated weight parameter, f j is the characteristic function.

[0046] Preferably, in step (4), the cross entropy loss function is used to measure the difference between the label distribution predicted by the model and the true label distribution; and the Adam optimization algorithm is used to adjust the weight of the model.

[0047] The formula of the cross entropy loss function is as follows:

[0048]

[0049] Where: y is the label value; is the predicted value;

[0050] Preferably, in step (4), the lithology of the formation is clay, sandstone, conglomerate, limestone, mudstone or dolomite.

[0051] Among them, the output formation lithology is the formation lithology closest to the actual situation, that is, the global optimal solution;

[0052] Preferably, in step (5), the formula for accuracy is:

[0053] In the formula, TP represents the number of samples correctly classified as positive, TN represents the number of samples correctly classified as negative, FP represents the number of samples incorrectly classified as positive, and FN represents the number of samples incorrectly classified as negative.

[0054] The formula for recall is:

[0055] Among them, the evaluation index of model evaluation can verify whether the model can accurately learn from the input data and make accurate predictions. The recall rate focuses on the proportion of samples that are correctly predicted as positive by the model among all samples that are actually positive.

[0056] The technical solution of the present invention achieves the following beneficial technical effects:

[0057] Intelligent identification of stratum lithology based on well logging data is an important technology in the field of geotechnical engineering. It can help geological engineers accurately determine the rock type of the stratum and provide key information for geological exploration. The stratum lithology identification method based on the LSTM-CRF model of the present invention combines the long-term dependency of the recurrent neural network (RNN) and the sequence labeling ability of CRF into a deep learning model, which is suitable for processing sequence data such as well logging data.

[0058] The present invention is an intelligent identification method for formation lithology based on logging data. The model selects the LSTM (Long Short-Term Memory Network) model and the CRF (Conditional Random Field) model for classification and identification of formation lithology. First, the natural gamma, acoustic time difference, resistivity and other logging curve data are collected and sorted, and the data is subjected to data denoising and normalization. Then, features that are helpful for formation lithology identification are extracted from the preprocessed data set, and the data set is used to establish a training set and a test set in a ratio of 8:2. The training set is used for LSTM-CRF model training, and the test set is used for LSTM-CRF model evaluation. The cross entropy loss function is used to measure the difference between the model's prediction results and the true label. The Adam optimization algorithm is selected to adjust the model's weights. Finally, the model performance is evaluated using indicators such as accuracy and recall. This method has the following advantages:

[0059] 1. Improve recognition accuracy: Automatically extract features from logging data through machine learning algorithms and use these features for lithology identification, which can effectively improve recognition accuracy and reduce subjective errors in human interpretation.

[0060] 2. Improve work efficiency: This method has a high degree of automation and can quickly process a large amount of logging data, greatly improving the work efficiency of formation lithology identification.

[0061] 3. Enhance data utilization efficiency: Through data preprocessing and feature extraction, the information in the logging data can be better utilized, and even data with high noise can be effectively utilized.

[0062] 4. Strong versatility: This method does not depend on the geological conditions of a specific area, is applicable to logging data in different regions, and has wide applicability.

[0063] 5. Reduce costs: The automated identification process reduces manual intervention and reduces labor costs, while improving decision-making efficiency, helping to reduce overall project costs.

[0064] 6. Result visualization: The identification results are intuitively displayed in the form of charts, etc., which facilitates geological engineers to analyze and make decisions, and improves the efficiency and understanding of information transmission.

[0065] 7. Promote technological innovation: The application of this method has promoted the integration of well logging technology and artificial intelligence technology, and promoted technological innovation and development in related fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 The unit structure diagram of the LSTM model in the method of the present invention;

[0067] Figure 2Flowchart of LSTM-CRF model classification and recognition in the method of the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical scheme and technical effect of the invention of this application clearer, the technical scheme in the embodiment of this application will be clearly and completely described below in conjunction with the drawings in the embodiment of this application. Obviously, the described embodiment is a part of the embodiment of this application, not all the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0069] The following is a specific embodiment of the present invention, a formation lithology intelligent identification method based on well logging data (such as Figure 2 Detailed description is given below.

[0070] Example

[0071] This embodiment provides a method for intelligently identifying formation lithology based on well logging data, comprising the following steps:

[0072] (1) Collecting logging data: Collect logging data of the formation to be identified, including natural gamma ray spectrum, natural potential, density, acoustic time difference, and deep lateral resistivity;

[0073] (2) Data preprocessing: preprocessing the collected logging data;

[0074] Firstly, wavelet transform is used for denoising, and then Z-score normalization is used to normalize the denoised data set to improve data quality.

[0075] Among them, the formula for Z score normalization is: Where μ is the mean of the original logging data, and σ is the standard deviation of the original logging data.

[0076] (3) Feature extraction: From the preprocessed logging data, features that are helpful for distinguishing formation lithology are extracted, and the feature dimension is reduced using the principal component analysis method. A training set and a test set are established, with a ratio of 8:2. The training set is used for LSTM-CRF model training, and the test set is used for LSTM-CRF model evaluation.

[0077] (4) Model training: The LSTM-CRF model is used. The training set data is used as the input of the LSTM model. The CRF model is applied to the output of the LSTM model to output the stratum lithology that is closest to the actual situation: clay, sandstone, conglomerate, limestone, mudstone, or dolomite.

[0078] Among them, the CRF model is used to annotate the sequence output by the LSTM model, and the annotated sequence is used to train the CRF model to output the optimal stratum lithology; the LSTM model sets hyperparameters before starting training; the CRF model is trained using the maximum likelihood estimation method; and the global optimal solution is output through Viterbi algorithm decoding.

[0079] ①The LSTM model includes the input x at time t t , cell state C t , candidate cell states Hidden state h t , forget gate f t , input gate i t and output gate O t ;

[0080] The formulas involved in the LSTM model are as follows:

[0081] Forget gate: f t =σ(W f ·[h t-1 , x t ]+b f )

[0082] In the formula, f t is the output of the forget gate, σ is the Sigmoid function, W f is the weight matrix of the forget gate, b f is the bias term of the forget gate, h t-1 is the hidden state of the previous time step, x t is the input of the current time step (time t);

[0083] Input gate: i t =σ(W i ·[h t-1 , x t ]+b i )

[0084] Candidate cell states:

[0085] In the formula, i t is the input gate output, is the candidate cell state; tanh is the hyperbolic tangent function; W i is the weight matrix of the input gate, b i is the bias term of the input gate, W c is the weight matrix of the candidate cell state, b c is the bias term of the candidate cell state,

[0086] The cell state at time t is:

[0087] In the formula, C t is the cell state; C t-1 is the cell state at the previous time step;

[0088] Output gate: O t =σ(W 0 ·[h t-1 , x t ]+b 0 )

[0089] Hidden state at time t: h t =O t tanh(C t )

[0090] In the formula, O t is the output of the output gate, ht is the hidden state at the current time step (time t), and W 0 is the weight matrix of the output gate, b 0 is the bias term of the output gate.

[0091] ②CRF model is given an input sequence X = (x 1 , x 2 , ..., x n ), predict the output sequence Y = (y 1 ,y 2 , ..., y n ). The CRF model defines a conditional probability distribution P(Y|X), which is expressed by the following formula:

[0092]

[0093] Where Z(X) is the normalized analysis, λ j is the jth characteristic function f j The associated weight parameter, f j is the characteristic function.

[0094] ③ Use the cross entropy loss function to measure the difference between the label distribution predicted by the model and the true label distribution; use the Adam optimization algorithm to adjust the weight of the model.

[0095] The formula of the cross entropy loss function is as follows:

[0096]

[0097] (5) Model evaluation: The performance of the model is evaluated through evaluation indicators: accuracy and recall.

[0098] The formula for accuracy is:

[0099] In the formula, TP represents the number of samples correctly classified as positive, TN represents the number of samples correctly classified as negative, FP represents the number of samples incorrectly classified as positive, and FN represents the number of samples incorrectly classified as negative.

[0100] The formula for recall is:

[0101] Among them, the evaluation index of model evaluation can verify whether the model can accurately learn from the input data and make accurate predictions. The recall rate focuses on the proportion of samples that are correctly predicted as positive by the model among all samples that are actually positive.

[0102] Application Examples

[0103] The formation lithology is identified by using the intelligent identification method of formation lithology based on well logging data of the present invention. Among them, in the case of too many lithology categories in the well logging data, the problem can be simplified by reducing the dimension of the labels. In the collected original data, sandy clay and gravel clay are classified as clay, coarse sandstone, medium sandstone, fine sandstone, and gray fine sandstone are classified as sandstone, conglomerate and muddy conglomerate are classified as conglomerate, limestone, muddy limestone, dolomite limestone, and sandy limestone are classified as limestone, mudstone, sandy mudstone, and silty mudstone are classified as mudstone, and dolomite and gray dolomite are classified as dolomite. Finally, the lithology labels are divided into six categories, namely clay, sandstone, conglomerate, limestone, mudstone, and dolomite.

[0104] The purpose of label dimensionality reduction is to reduce the number of categories, thereby reducing the complexity of the classification problem. This method usually includes eliminating lithology categories with extremely low proportions, selecting highly correlated parameters and lithology for dimensionality reduction, and interpolating missing values. These steps help improve the effect of network training and thus improve the accuracy of recognition.

[0105] In order to effectively extract features from logging data, the present invention adopts LSTM neural network. In the collected original logging data, the logging parameters include well diameter, natural gamma spectrum, natural potential, density, neutron, acoustic time difference, deep lateral resistivity, and shallow lateral resistivity. The LSTM network can extract features from multiple logging curve data, effectively explore the essential characteristics of the data, and reduce the time and cost of extracting features. Through layer-by-layer feature changes, the LSTM network can solve complex classification problems and predict lithology. Finally, five logging parameters are selected, namely natural gamma spectrum, natural potential, density, acoustic time difference, and deep lateral resistivity. Compared with the existing methods, this method has improved recognition accuracy and response speed, making lithology identification convenient, fast, objective and reliable.

[0106] The data set was preprocessed by first checking the missing data and processing it using the interpolation method. The well logging data samples were obtained as shown in Table 1, and the lithology labels were given as shown in Table 2.

[0107] Table 1 Well logging data

[0108]

[0109] Table 2 Lithology labels

[0110] Lithology clay sandstone Conglomerate Limestone Mudstone Dolomites Label Encoding 1 2 3 4 5 6

[0111] After completion, the dataset is normalized and then divided into training and test sets in a ratio of 8:2.

[0112] The model training steps are as follows:

[0113] 1. Forward calculation:

[0114] At the beginning of each time step, the hidden state and cell state are initialized. The input gate is calculated to decide what information to discard from the cell state. The forget gate is calculated to decide what information in the cell state to keep. Based on the output of the input gate and the forget gate, the cell state is updated. The output gate is calculated to decide what information in the cell state is passed to the hidden state of the next time step. Based on the cell state and the output of the output gate, the hidden state for the current time step is calculated. The hidden state is often used as the output of the model or further processed.

[0115] 2. Back propagation calculation:

[0116] Calculate the loss using the hidden state and the actual label at the current time step. Calculate the gradient of the loss function with respect to the model parameters (weights and biases). Start at the output layer and traverse the model backwards, calculating the gradient of each parameter. This involves applying the chain rule to calculate the gradient of each parameter. Apply an optimization algorithm (such as gradient descent) based on the gradient to update the weights and biases.

[0117] The CRF layer is used to decode sequence data to obtain the optimal label sequence. The output of the LSTM network is usually a sequence, each element represents a hidden state of a time step, and these states contain the long-term dependency information of the input sequence. The CRF layer takes these hidden states as input and decodes the optimal label sequence by considering the dependency between labels. The CRF layer finally outputs the optimal label sequence, which reflects the label distribution of the input sequence.

[0118] The trained model is saved and tested with a test set to evaluate the accuracy of classification recognition. In multi-classification problems, model training usually involves minimizing the loss function to improve the accuracy of prediction. The change in the value of the loss function with the number of iterations can reflect the learning process of the model. The performance of the evaluation model of the present invention not only depends on the loss function, but also includes indicators such as accuracy and recall. Accuracy refers to the proportion of positive classes correctly predicted by the model, while recall refers to the proportion of all positive classes correctly predicted by the model. In the result evaluation, the accuracy of the model of the present invention is 86.61%, and the predicted sequence has a high degree of fit with the original sequence, which can be applied to actual engineering projects.

[0119] After the lithology recognition model is trained, the real-time logging data is input into the model for calculation. The model predicts the lithology based on the input data and outputs the results.

[0120] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the claims of this patent application.

Claims

1. A method for intelligent identification of formation lithology based on well logging data, characterized in that: The following steps are involved: (1) Collecting well logging data: Collecting well logging data of the formation to be identified; (2) Data preprocessing: preprocessing the collected logging data; (3) Feature extraction: Extract features from the preprocessed logging data and create training and test sets; (4) Model training: The LSTM-CRF model is used, the training set data is used as the input of the LSTM model, and the CRF model is applied to the output of the LSTM model to output the lithology of the formation; (5) Model evaluation: The performance of the model is evaluated through evaluation indicators: accuracy and recall.

2. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In the step (1), the logging data include natural gamma ray spectrum, natural potential, density, acoustic time difference and deep lateral resistivity.

3. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In the step (2), wavelet transform is first used for denoising, and then Z-score normalization is used to perform normalization processing on the basis of the denoised data set.

4. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In the step (3), the ratio of the training set to the test set is 8:2; the training set is used for LSTM-CRF model training, and the test set is used for LSTM-CRF model evaluation.

5. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In the step (3), feature extraction is to extract features for formation lithology identification from the preprocessed logging data, and reduce the feature dimension using principal component analysis.

6. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In step (4), the LSTM model includes the input x at time t t , cell state C t , candidate cell states Hidden state h t , forget gate f t , input gate i t and output gate O t .

7. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In the step (4), the sequence output by the LSTM model is labeled using the CRF model, and the CRF model is trained using the labeled sequence to output the optimal stratum lithology.

8. The method for intelligently identifying formation lithology based on well logging data according to claim 7, characterized in that: In the step (4), the LSTM model sets hyperparameters before starting training; the CRF model is trained using the maximum likelihood estimation method; and the global optimal solution is output through Viterbi algorithm decoding.

9. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In the step (4), the cross entropy loss function is used to measure the difference between the label distribution predicted by the model and the actual label distribution; and the Adam optimization algorithm is used to adjust the weight of the model.

10. The method for intelligently identifying formation lithology based on well logging data according to claim 1, characterized in that: In the step (4), the lithology of the formation is clay, sandstone, conglomerate, limestone, mudstone or dolomite.

Images