Well logging lithology identification method and system based on wavelet de-drifting and deep learning

Abstract

The present application relates to the oil and gas exploration technical field, specifically relates to the well logging lithology identification method and system based on wavelet de-drifting and deep learning, including the following steps: obtaining the well logging data for identifying well logging lithology;Using wavelet transform to eliminate the data drift of well logging data caused by burial depth;Based on the data set formed by the well logging data and well logging lithology, using convolution-self attention deep learning model for training, obtaining the well logging lithology identification model expressing the mapping relationship between well logging data and well logging lithology.The present application uses convolution-self attention deep learning model for training, obtains the well logging lithology identification model expressing the mapping relationship between well logging data and well logging lithology, realizes the well logging lithology according to well logging data, uses feature enhancement in feature processing to combine attention mechanism, completes the simulation brain signal processing mechanism, improves the feature extraction capability of lithology identification model, and feature utilization rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas exploration, and particularly relates to a well logging lithology identification method and system based on wavelet de-drifting and deep learning. BACKGROUND

[0002] In the process of oil and gas exploration and development, cores provide a large amount of detailed and important information. The distribution of underground lithology is a prerequisite for extending core information to an oilfield or an entire basin. Considering the large accumulation of traditional logging curve data and the scarcity of core data, it is crucial to have a robust method to extract the logging curve features corresponding to the core lithology. These methods can generally be divided into traditional methods (cross plot, curve morphology or empirical formula), traditional machine learning and deep learning methods.

[0003] Traditional methods require a lot of labor to extract and apply the rock geophysical response in the logging curve, and therefore are limited by their inefficiency and subjectivity. In addition, due to the influence of factors such as mineral composition, diagenesis, borehole environment and formation fluid, the rock geophysical response in the curve shows complex nonlinear characteristics, further limiting the application scenarios of traditional methods.

[0004] Machine learning and deep learning methods can automatically extract features from data and build complex mappings between these features, making the related applications of machine learning and deep learning in logging curve lithology identification increasingly growing. Although machine learning also shows good performance, its application is limited to a certain extent by its high data complexity and the need for manual feature engineering, resulting in problems such as poor feature extraction ability, low feature utilization rate and poor model performance.

[0005] Therefore, the existing technology for identifying logging lithology has the problems of poor feature extraction ability and low feature utilization rate of deep learning models. SUMMARY

[0006] The present application aims to provide a well logging lithology identification method and system based on wavelet de-drifting and deep learning to solve the technical problems of poor feature extraction ability and low feature utilization rate of deep learning models in the prior art for identifying logging lithology.

[0007] To solve the above technical problems, the present application specifically provides the following technical solutions:

[0008] A well logging lithology identification method based on wavelet de-drifting and deep learning, comprising the following steps:

[0009] Obtaining logging data for identifying logging lithology;

[0010] Using wavelet transform to eliminate data drift caused by burial depth of the logging data;

[0011] Based on the logging data and a data set composed of logging lithology, a convolution-self-attention deep learning model is trained to obtain a logging lithology identification model expressing the mapping relationship between logging data and logging lithology.

[0012] Logging lithology is obtained according to logging data by using the logging lithology identification model.

[0013] As a preferred scheme of the present application, the logging data includes density logging, natural gamma logging, deep-shallow dual lateral resistivity logging, acoustic logging and compensated neutron logging.

[0014] As a preferred scheme of the present application, the logging lithology categories include conglomerate, sandy conglomerate, coarse sandstone, medium sandstone, fine sandstone, siltstone and mudstone.

[0015] As a preferred scheme of the present application, the model structure of the convolution-self-attention deep learning model is as follows:

[0016] A convolution network module is configured to extract features related to logging lithology identification from the logging data to obtain logging data features.

[0017] A feature enhancement module is configured to enhance key features related to logging lithology identification in the logging data features to obtain logging key enhanced features.

[0018] An attention mechanism module is configured to assign feature weights to the logging key enhanced features, extract key features related to logging lithology identification, and obtain logging key features.

[0019] A fully connected network module is configured to perform probability calculation on each category of logging lithology by using the logging key features to obtain the probability of each category of logging lithology.

[0020] As a preferred scheme of the present application, the enhancement method of the key features of the feature enhancement module includes:

[0021] A feature coordinate with the highest feature value in the logging data features is determined as an enhancement center.

[0022] A matrix with a size of r*r is expanded outward from the enhancement center in the logging data features as an enhancement region.

[0023] The enhancement region is multiplied by a Gaussian distribution matrix with a size of r*r to obtain the logging key enhanced features.

[0024] The enhancement expression of the logging key enhanced features is as follows:

[0025]

[0026] In the formula, gijk d,enhance is the feature value of the matrix position (i,j,k) of the key enhanced feature of the well logging at the well depth d, g ijk d is the feature value of the matrix position (i,j,k) of the well logging data feature at the well depth d, r*r is the size of the enhanced area, (i max max max is the matrix position of the enhancement center, (i,j,k) is the matrix position of the well logging data feature, s.t.1 is the first constraint condition, s.t.2 is the second constraint condition.

[0027] As a preferred scheme of the present application, the key feature extraction method of the attention mechanism module comprises:

[0028] inputting the key enhanced feature of the well logging into the attention mechanism and normalizing through a softmax function to obtain an attention weight;

[0029] weighting the key enhanced feature of the well logging by the attention weight to obtain the key feature of the well logging;

[0030] the weighting expression of the key feature of the well logging is:

[0031] S ijk d,enhance =W ijk d,enhance *g ijk d,enhance ;

[0032]

[0033] Vs d =Ks d ={g ijk d,enhance |i∈[1,X],j∈[1,Y],k∈[1,Z]};

[0034] Qs d =middle({g ijk d,enhance |i∈[1,X],j∈[1,Y],k∈[1,Z]});

[0035] In the formula, S ijk d,enhance is the feature value of the matrix position (i,j,k) of the key feature of the well logging at the well depth d, W ijk d ,enhance is the attention weight of g ijk d,enhance , and g ijk ​​d,enhance Qs d , Vs d and Ks d are Q, V and K values of the attention mechanism respectively, d Ksd is the dimension size of Ks d , X is the total number of matrix position horizontal coordinates, Y is the total number of matrix position vertical coordinates, Z is the total number of matrix position axis coordinates, softmax is a softmax function, and middle is a function of taking the middle term.

[0036] As a preferred scheme of the present application, the probability calculation method of the full connection network module comprises:

[0037] inputting the logging key features into the full connection network for processing, and then obtaining the probability distribution of the logging lithology category through the sigmoid activation function processing;

[0038] the calculation expression of the logging lithology category is:

[0039]

[0040] Y d ={Y L d |L∈[1,M]};

[0041] Y={Y d |d∈[1,D]};

[0042] In the formula, Y d is the probability distribution of the lithology category at the well depth d, Y L d is the probability of the lithology category L at the well depth d, Y is the probability distribution of the logging lithology, D is the total depth of the well depth, FC is the operation function of the full connection layer, M is the total number of the logging lithology categories, S ijk d,enhance is the feature value of the matrix position (i, j, k) of the logging key features at the well depth d, X is the total number of matrix position horizontal coordinates, Y is the total number of matrix position vertical coordinates, and Z is the total number of matrix position axis coordinates.

[0043] As a preferred scheme of the present application, the training loss of the convolution-self-attention deep learning model is:

[0044]

[0045] wherein,

[0046] In the formula, Loss is the loss function value, Y1 Y is a probability distribution of lithology class at logging depth 1 1 real Y is a true distribution of lithology class at logging depth 1 D Y is a probability distribution of lithology class at logging depth D D real Y is a true distribution of lithology class at logging depth D d Y is a probability distribution of lithology class at logging depth d d real Y is a true distribution of lithology class at logging depth d d-1 real Y is a true distribution of lithology class at logging depth d-1 d+1 real Y is a true distribution of lithology class at logging depth d+1 sim d Y is a stability of lithology class at logging depth d

[0047] are all L2 norm formulas.

[0048] As a preferred scheme of the present application, the method for constructing the logging lithology identification model comprises:

[0049] dividing a data set composed of logging data and logging lithology into a training set and a test set;

[0050] training the convolution-self-attention deep learning model based on the training set by using a loss function to obtain the logging lithology identification model;

[0051] performing classification performance evaluation of the logging lithology identification model based on the test set by using a mean square error.

[0052] As a preferred scheme of the present application, the present application provides a logging lithology identification system based on wavelet de-drifting and deep learning, which is applied to a method for logging lithology identification based on wavelet de-drifting and deep learning, and the system comprises:

[0053] a data acquisition unit configured to acquire logging data for identifying logging lithology;

[0054] a data preprocessing unit configured to eliminate data drifting of the logging data caused by burial depth by using wavelet transform;

[0055] a deep learning unit configured to train a convolution-self-attention deep learning model based on a data set composed of the logging data and logging lithology to obtain a logging lithology identification model expressing a mapping relationship between the logging data and the logging lithology;

[0056] An identification output unit is configured to obtain the well logging lithology according to the well logging data by using the well logging lithology identification model.

[0057] Compared with the prior art, the present application has the following beneficial effects:

[0058] The present application uses a convolution-self-attention deep learning model for training to obtain a well logging lithology identification model expressing the mapping relationship between the well logging data and the well logging lithology, realizes the well logging lithology according to the well logging data, uses feature enhancement combined with an attention mechanism in feature processing to complete the simulation of the human brain signal processing mechanism, improves the feature extraction capability of the lithology identification model, the feature utilization rate, and finally improves the performance of the lithology identification model. BRIEF DESCRIPTION OF DRAWINGS

[0059] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other drawings can be obtained from the provided drawings without creative labor.

[0060] Figure 1 A well logging lithology identification method flowchart based on wavelet de-drifting and deep learning provided for the embodiments of the present application;

[0061] Figure 2 A well logging lithology identification system block diagram based on wavelet de-drifting and deep learning provided for the embodiments of the present application;

[0062] Figure 3 A network structure of a convolution-self-attention deep learning model provided for the embodiments of the present application;

[0063] Figure 4 A result graph of wavelet transform eliminating data drift provided for the embodiments of the present application;

[0064] Figure 5 A model effect comparison graph provided for the embodiments of the present application. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0066] As Figure 1As shown, the present application provides a well logging lithology identification method based on wavelet de-drifting and deep learning, comprising the following steps:

[0067] Obtaining well logging data for identifying well logging lithology;

[0068] Eliminating data drift caused by burial depth of well logging data by using wavelet transform;

[0069] Based on the data set composed of well logging data and well logging lithology, a convolution-self-attention deep learning model is used for training to obtain a well logging lithology identification model expressing the mapping relationship between well logging data and well logging lithology;

[0070] Using the well logging lithology identification model, the well logging lithology is obtained according to the well logging data.

[0071] The present application uses a convolution-self-attention deep learning model to train the data set composed of well logging data and well logging lithology, constructs the mapping relationship between well logging data and well logging lithology, realizes the well logging lithology according to the well logging data, realizes the standard model of well logging lithology identification, and improves the efficiency of well logging lithology identification.

[0072] The convolution-self-attention deep learning model used in the present application includes five modules, which are:

[0073] 1. Convolution network module, used for extracting features related to well logging lithology identification in well logging data to obtain well logging data features, usually using CNN neural network as the convolution network module;

[0074] 2. Feature enhancement module, used for enhancing key features related to well logging lithology identification in well logging data features to obtain well logging key enhanced features, realizing information enhancement of key features in well logging data features for well logging lithology identification, so as to enlarge the difference between key features and non-key features, so that the subsequent neural network can accurately and quickly identify the key features, and the quick and accurate identification of key features helps to improve the feature extraction capability of the model, and the information enhancement of key features lays a foundation for improving the feature utilization rate;

[0075] 3. The attention mechanism module is used for assigning feature weights to the logging key reinforcement features, extracting key features related to logging lithology identification, and obtaining logging key features. The attention mechanism simulates the human brain signal processing mechanism, changes the information weighting method, measures the similarity of the identification target and the input information, allocates resources in proportion according to the correlation between the two, realizes the extraction of key features required for identifying logging lithology on the basis of information enhancement, and obtains a high proportion of resources for key features, thereby improving the time sequence feature extraction capability of the prediction model. The introduction of the mechanism imitates the method of identifying lithology by using logging curves, that is, considering the absolute value of the curve, the combination of different curves, and fully considering the difference between the upper and lower curves. The sequence input realizes the ability of the model to consider the difference between the upper and lower curves. The convolution along the dimension direction of the curve type realizes the combination of different curves, and can also reduce the probability of parameter drift and reduce network loss.

[0076] 4. The full connection network module is used for calculating the probability of each category of logging lithology by using the logging key features, and obtaining the probability of each category of logging lithology.

[0077] In the training of the convolution-self-attention deep learning model, a dynamic loss function is adopted. When ensuring the accuracy of lithology identification at all logging depths, the stability of the rock mass at the logging depth is considered, that is, the stability of the rock mass at each logging depth is measured. If the stability of the rock mass at the logging depth d is high, the difference between the rock mass categories at adjacent depths is small, and the learning results of the model at adjacent depths can be transferred, then the convolution-self-attention deep learning model is easier to learn at the logging depth d, and it is easier to obtain accurate results. If the stability of the rock mass at the logging depth d is low, the difference between the rock mass categories at adjacent depths is large, and it is difficult to transfer the learning results of the model at adjacent depths, then the convolution-self-attention deep learning model is more difficult to learn at the logging depth d, and it is more difficult to obtain accurate results.

[0078] Therefore, the attention degree will be allocated from the logging depth with high rock mass stability to the logging depth with low rock mass stability during model training. In order to realize the reasonable allocation of the attention degree during the model training process, the low weight is given to the logging depth with high rock mass stability and the high weight is given to the logging depth with low rock mass stability in the loss function, so that the model pays more attention to the logging depth with difficult learning during the training process, and the model can obtain better accuracy at each logging depth, thereby realizing the equalization of the model performance.

[0079] The logging data includes density logging (DEN), natural gamma logging (GR), deep and shallow dual lateral resistivity logging (RD and RS), acoustic logging (AC) and compensated neutron logging (CNL).

[0080] The well logging lithology categories include gravel (gs), sand gravel (sgs), coarse sand (css), medium sand (mss), fine sand (fss), siltstone (sss) and mudstone (mds).

[0081] Figure 3 The model structure of the convolution-self-attention deep learning model is shown in the figure:

[0082] The convolution network module (CNN) is used to extract features related to well logging lithology identification from the well logging data, to obtain well logging data features.

[0083] The feature enhancement module is used to enhance key features related to well logging lithology identification in the well logging data features, to obtain well logging key enhanced features.

[0084] The attention mechanism module (ATT) is used to assign feature weights to the well logging key enhanced features, to extract key features related to well logging lithology identification, to obtain well logging key features.

[0085] The fully connected network module (FC) is used to calculate the probabilities of each category of well logging lithology using the well logging key features, to obtain the probabilities of each category of well logging lithology.

[0086] The enhancement method of the key features of the feature enhancement module includes:

[0087] The feature coordinate with the highest feature value in the well logging data features is determined as the enhancement center.

[0088] A matrix of size r*r is expanded outward from the enhancement center in the well logging data features as the enhancement region.

[0089] The enhancement region is multiplied by a Gaussian distribution matrix of size r*r to obtain the well logging key enhanced features.

[0090] The enhancement expression of the well logging key enhanced features is:

[0091]

[0092] In the formula, g ijk d,enhance is the feature value at the matrix position (i,j,k) of the well logging key enhanced features at well depth d, g ijk d is the feature value at the matrix position (i,j,k) of the well logging data features at well depth d, r*r is the size of the enhancement region, (i max ,j max ,k max ) is the matrix position of the enhancement center, (i,j,k) is the matrix position of the well logging data features, s.t.1 is the first constraint condition, and s.t.2 is the second constraint condition.

[0093] First, the feature coordinate with the highest eigenvalue is located, that is, the center of the key feature is obtained, and then the features in the neighborhood thereof are enhanced, that is, the region where the key feature is located is enhanced, the key feature in the logging data feature for logging lithology identification is information enhanced, so as to be able to enlarge the difference between the key feature and the non-key feature, so that the subsequent neural network can accurately and quickly identify the key feature, and the fast and accurate identification of the key feature helps to improve the feature extraction capability of the model, and the information enhancement of the key feature lays a foundation for improving the feature utilization rate.

[0094] The key feature extraction method of the attention mechanism module comprises:

[0095] The logging key enhanced feature is input into the attention mechanism and normalized by a softmax function to obtain an attention weight;

[0096] The logging key feature is obtained by weighting the logging key enhanced feature by the attention weight;

[0097] The weighted expression of the logging key feature is:

[0098] S ijk d,enhance =W ijk d,enhance *g ijk d,enhance ;

[0099]

[0100] Vs d =Ks d ={g ijk d,enhance |i∈[1,X],j∈[1,Y],k∈[1,Z]};

[0101] Qs d =middle({g ijk d,enhance |i∈[1,X],j∈[1,Y],k∈[1,Z]});

[0102] In the formula, S ijk d,enhance is an eigenvalue at a matrix position (i,j,k) of the logging key feature at a well depth d, W ijk d ,enhance is an attention weight of g ijk d,enhance , g ijk d,enhance is an eigenvalue at a matrix position (i,j,k) of the logging key enhanced feature at the well depth d, and Qsd , Vs d , and Ks d are Q, V and K values of the attention mechanism respectively, is a dimension size of Ks d , X is a total number of matrix position horizontal coordinates, Y is a total number of matrix position vertical coordinates, Z is a total number of matrix position axis coordinates, softmax is a softmax function, middle is a function of taking the middle term, and T is a transpose operator.

[0103] The attention mechanism simulates the human brain signal processing mechanism, changes the information weighting method, measures the similarity of the recognition target and the input information, allocates resources in proportion according to the correlation between the two, realizes the extraction of the key features required for identifying the well logging lithology on the basis of information enhancement, and improves the time sequence feature extraction capability of the prediction model. The introduction of the mechanism imitates the method of identifying the lithology by using the well logging curve, that is, considering the absolute value of the curve, the combination of different curves, and fully considering the difference between the upper and lower curves. The sequence input realizes the ability of the model to consider the difference between the upper and lower curves, the convolution in the dimension direction of the curve type realizes the combination of different curves, and the probability of parameter drift can be reduced, and the network loss can be reduced.

[0104] The probability calculation method of the full connection network module includes:

[0105] The well logging key features are input into the full connection network for processing, and then the probability distribution of the well logging lithology category is obtained through the sigmoid activation function processing;

[0106] The calculation expression of the well logging lithology category is:

[0107]

[0108] Y d ={Y L d |L∈[1,M]};

[0109] Y={Y d |d∈[1,D]};

[0110] In the formula, Y d is the probability distribution of the lithology category at the well depth d, Y L d is the probability of the lithology category L at the well depth d, Y is the probability distribution of the well logging lithology, D is the total depth of the well depth, FC is the operation function of the full connection layer, M is the total number of the well logging lithology categories, S ijk d,enhanceis the eigenvalue of the key feature at the matrix position (i,j,k) of the well depth d, X is the total number of matrix position horizontal coordinates, Y is the total number of matrix position vertical coordinates, and Z is the total number of matrix position axis coordinates.

[0111] wherein the FC function is wx+b, x is {S ijk d,enhance , w is the weight of the full connection network, and b is the bias of the full connection network.

[0112] The training loss of the convolution-self-attention deep learning model:

[0113]

[0114] wherein,

[0115] In the formula, Loss is the loss function value, Y 1 is the probability distribution of the lithology category at the well depth 1, Y 1 real is the real distribution of the lithology category at the well depth 1, Y D is the probability distribution of the lithology category at the well depth D, Y D real is the real distribution of the lithology category at the well depth D, Y d is the probability distribution of the lithology category at the well depth d, Y d real is the real distribution of the lithology category at the well depth d, Y d-1 real is the real distribution of the lithology category at the well depth d-1, Y d+1 real is the real distribution of the lithology category at the well depth d+1, W sim d is the stability of the lithology category at the well depth d;

[0116] are all L2 norm formulas.

[0117] The application adopts a dynamic loss function in training the convolution-self-attention deep learning model, and considers the rock mass stability of the logging depth when ensuring the accuracy of lithology identification at all logging depths, that is, the rock mass stability of each logging depth is measured, if the rock mass stability of the logging depth d is high, the difference of the rock mass category is small compared with the adjacent depth, the result can be transferred by the model learning result of the adjacent depth, then the convolution-self-attention deep learning model is easier to learn at the logging depth d, and it is easier to obtain accurate results, if the rock mass stability of the logging depth d is low, the difference of the rock mass category is large compared with the adjacent depth, it is difficult to transfer the result by the model learning result of the adjacent depth, then the convolution-self-attention deep learning model is more difficult to learn at the logging depth d, and it is more difficult to obtain accurate results.

[0118] Therefore, the attention degree is allocated from the logging depth with high rock mass stability to the logging depth with low rock mass stability during model training, in order to realize the reasonable allocation of the attention degree during model training, the low weight is given to the logging depth with high rock mass stability, and the high weight is given to the logging depth with low rock mass stability in the loss function, so that the model pays more attention to the logging depth with difficult learning during the training process, and the model can obtain better accuracy at each logging depth, and the performance of the model is balanced.

[0119] The rock mass stability of the well depth d is measured by the difference between the real distribution of lithology at the well depth d-1 (the previous well depth of the well depth d) and the well depth d+1 (the next well depth of the well depth d), the greater the difference, the more the lithology mutation occurs between the well depth d-1 and the well depth d+1, and the well depth d is in the layer of lithology mutation, the smaller the difference, the less the lithology mutation occurs between the well depth d-1 and the well depth d+1, and the well depth d is in the layer of stable lithology.

[0120] The application utilizes the stability W of the lithology category sim d The weight of the logging depth is constructed as (1-W sim d )]log(W sim d , the low weight can be given to the logging depth with high rock mass stability, and the high weight can be given to the logging depth with low rock mass stability, so that the model pays more attention to the logging depth with difficult learning during the training process, and the model can obtain better accuracy at each logging depth, and the performance of the model is balanced.

[0121] The method for constructing the logging lithology identification model comprises:

[0122] The data set composed of the logging data and the logging lithology is divided into a training set and a test set;

[0123] Based on the training set, the convolutional-self-attention deep learning model is trained by using a loss function to obtain a well logging lithology identification model.

[0124] Based on the test set, the well logging lithology identification model is classified and performance evaluated by using a mean square error.

[0125] The present application considers the data drift caused by compaction, and the data drift phenomenon is obvious in thousands of kilometers of drilling. Figure 4 Taking the AC curve shown in the figure as an example, the value gradually decreases from shallow to deep, which is a typical phenomenon that the porosity decreases and the density increases with the increase of depth due to compaction, and further leads to faster sound wave transmission (smaller AC value), and the well logging curves CNL and DEN corresponding to the porosity and density also change naturally. Therefore, the present application adopts wavelet transform to process the well logging data, and eliminates the low-frequency signal in the original well logging curve, which can eliminate the data drift caused by compaction.

[0126] In order to imitate the way of humans to identify lithology according to the shape of well logging curve and well logging combination, the model structure and model data input mode of the well logging lithology identification model are designed in the present application. Figure 4 The well logging lithology identification model (i.e.

[0127] In fact, in order to verify the effectiveness of considering the shape change of well logging curve as lithology identification, the model is constructed to compare the model input of 3 depths, 5 depths, 7 depths, 9 depths and 11 depths (as shown in the figure). Figure 5 It is found that: a. With the increase of the number of input depths, the model obtains better accuracy, which confirms that the model considering the shape of well logging curve is effective; b. The model proposed in the present application obtains better results than ensemble learning when the input is 7 depths, which further confirms the advantage of the model. In addition, the combination of "wavelet drift removed data + the model proposed in the present application" obtains better accuracy than "data without wavelet drift removal + the model of the present application", and the combination of "wavelet drift removed data + ensemble learning" obtains worse accuracy than "data without wavelet drift removal + ensemble learning", and wavelet drift removal will bring certain information loss and thus make the accuracy of ensemble learning lower, while the accuracy of the model of the present application is increased, which further confirms the necessity of the model considering the shape of well logging curve. Among them, Figure 4 and Figure 5L3-L11 represent the number of depth sequences of the model input logging data proposed in the application, and ''-r'' represents that the logging data has been subjected to wavelet de-drifting.

[0128] The table 1 obtained by the application proves that the model is better than the machine learning model, and the result of the ensemble learning is the best in the machine learning. Figure 4 and Figure 5 When comparison is made in Figure 4 analysis, it can be seen that the result obtained by ''wavelet de-drifting data + the model proposed in the application'' is more consistent with the geological fact. ''Non-wavelet de-drifting data + the model proposed in the application'', ''non-wavelet de-drifting data + ensemble learning'' and ''wavelet de-drifting data + ensemble learning'' all fail to effectively identify the shallow sand body. It is further proved that removing the logging data drift caused by compaction can enhance the representativeness and universality of the data set, and the model proposed in the application has the ability to consider the logging shape, so that the combination of ''wavelet de-drifting data + the model proposed in the application'' has the advantage in lithology identification.

[0129] The deep learning method has more flexible and variable input methods, and the customized model structure has more flexible input methods, which can fully adapt and extract data features. Through comparison between models, the model method proposed in the application has better performance (table 1).

[0130] Table 1 model effect comparison table

[0131]

[0132] As shown in Figure 2 , the application provides a logging lithology identification system based on wavelet de-drifting and deep learning, which is applied to a logging lithology identification method based on wavelet de-drifting and deep learning, and the system comprises:

[0133] A data acquisition unit is configured to acquire logging data for identifying logging lithology.

[0134] A data preprocessing unit is configured to eliminate data drift caused by burial depth by using wavelet transform.

[0135] A deep learning unit is configured to train a convolution-self-attention deep learning model based on a data set composed of logging data and logging lithology, so as to obtain a logging lithology identification model expressing the mapping relationship between logging data and logging lithology.

[0136] An identification output unit is configured to obtain logging lithology according to logging data by using the logging lithology identification model.

[0137] The present application utilizes wavelet transform to process well logging data, eliminates low-frequency signals in the original well logging curve, can eliminate data drift caused by compaction, and then adopts a convolution-self attention deep learning model for training to obtain a well logging lithology identification model expressing the mapping relationship between well logging data and well logging lithology, so that the well logging lithology is obtained according to the well logging data, the feature enhancement is utilized in the feature processing to combine the attention mechanism, the simulation of the brain signal processing mechanism is completed, the feature extraction capability of the lithology identification model is improved, the feature utilization rate is improved, and finally the performance of the lithology identification model is improved.

[0138] The above examples are only exemplary embodiments of the present application and are not used to limit the present application, and the protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements are also regarded as falling within the protection scope of the present application.

Claims

1. A well logging lithology identification method based on wavelet de-drifting and deep learning, characterized in that, The method comprises the following steps: obtaining logging data for identifying logging lithology; eliminating data drift caused by burial depth of the logging data by using wavelet transform; training a convolution-self-attention deep learning model based on a data set composed of the logging data and the logging lithology to obtain a logging lithology identification model expressing a mapping relationship between the logging data and the logging lithology; obtaining the logging lithology according to the logging data by using the logging lithology identification model; a model structure of the convolution-self-attention deep learning model: a convolution network module for extracting features related to logging lithology identification from the logging data to obtain logging data features; a feature enhancement module for enhancing key features related to logging lithology identification in the logging data features to obtain logging key enhanced features; an attention mechanism module for assigning feature weights to the logging key enhanced features to extract key features related to logging lithology identification to obtain logging key features; a fully connected network module for calculating probabilities of each category of logging lithology by using the logging key features to obtain probabilities of each category of logging lithology; an enhancement method of the key features of the feature enhancement module comprises: determining a feature coordinate with the highest feature value in the logging data features as an enhancement center; expanding a matrix with a size of r*r outward from the enhancement center in the logging data features as an enhancement region; multiplying the enhancement region and a Gaussian distribution matrix with a size of r*r to obtain the logging key enhanced features; an enhancement expression of the logging key enhanced features is: ； ； ； wherein is the matrix position of the key enhanced feature of the well log at depth d is the eigenvalue at the feature position, is the matrix position of the well log data feature at depth d is the eigenvalue at the feature position, r*r is the size of the enhancement region, is the matrix position of the enhancement center, is the matrix position of the well log data feature, is the first constraint condition, is the second constraint condition; logging lithology identification based on wavelet de-drifting and deep learning.

2. The well logging lithology identification method based on wavelet de-drifting and deep learning according to claim 1, characterized in that: The logging data comprises density logging, natural gamma logging, deep and shallow dual-lateral resistivity logging, acoustic logging and compensated neutron logging.

3. The well logging lithology identification method based on wavelet de-drifting and deep learning according to claim 1, characterized in that: The logging lithology categories comprise conglomerate, sandy conglomerate, coarse sandstone, medium sandstone, fine sandstone, siltstone and mudstone.

4. The well logging lithology identification method based on wavelet de-drifting and deep learning according to claim 1, characterized in that: An extraction method of the key features of the attention mechanism module comprises: inputting the logging key enhanced features into the attention mechanism and normalizing the logging key enhanced features by using a softmax function to obtain attention weights; weighting the logging key enhanced features by using the attention weights to obtain the logging key features; a weighting expression of the logging key features is: * ； ； ； ； In the formula, is the matrix position of the key feature of the well depth d is the eigenvalue at , is the attention weight of is the matrix position of the key reinforcement feature of the well depth d is the eigenvalue at , and are Q, V and K values of the attention mechanism respectively, is the dimension size of , X is the total number of matrix position horizontal coordinates, Y is the total number of matrix position vertical coordinates, Z is the total number of matrix position axis coordinates, softmax is a softmax function, and middle is a function of taking the middle term.

5. The well logging lithology identification method based on wavelet de-drifting and deep learning according to claim 4, characterized in that: A probability calculation method of the fully connected network module comprises: inputting the logging key features into the fully connected network for processing, and then processing the logging key features by using a sigmoid activation function to obtain a probability distribution of the logging lithology categories; a calculation expression of the logging lithology categories is: ； ； ； wherein, is a probability distribution of lithology classes at well depth d, is a probability of lithology class L at well depth d, Y is a probability distribution of well logging lithology, D is a total depth of well depth, FC is an operation function of fully connected layer, M is a total number of well logging lithology classes, is a feature value at matrix position of well logging key features at well depth d X is a total number of matrix position horizontal coordinates, Y is a total number of matrix position vertical coordinates, Z is a total number of matrix position axial coordinates.

6. The well logging lithology identification method based on wavelet de-drifting and deep learning according to claim 5, characterized in that: a training loss of the convolution-self-attention deep learning model is: ； wherein ); wherein Loss is a loss function value, is a probability distribution of the lithology class at well depth 1, is a true distribution of the lithology class at well depth 1, is a probability distribution of the lithology class at well depth D, is a true distribution of the lithology class at well depth D, is a probability distribution of the lithology class at well depth d, is a true distribution of the lithology class at well depth d, is a true distribution of the lithology class at well depth d-1, is a true distribution of the lithology class at well depth d+1, is a stability of the lithology class at well depth d; , , , are L2-norm.

7. The well logging lithology identification method based on wavelet de-drifting and deep learning according to claim 6, characterized in that: A construction method of the logging lithology identification model comprises: dividing a data set composed of the logging data and the logging lithology into a training set and a test set; training the convolution-self-attention deep learning model by using a loss function based on the training set to obtain the logging lithology identification model; performing classification performance evaluation of the logging lithology identification model by using a mean square error based on the test set. 8.A well logging lithology identification system based on wavelet de-drifting and deep learning, characterized in that, The system is applied to the logging lithology identification method based on wavelet de-drifting and deep learning in any one of claims 1-7 and comprises: a data acquisition unit for obtaining logging data for identifying logging lithology; The data preprocessing unit is configured to eliminate data drift caused by burial depth of the well logging data by using wavelet transform. The deep learning unit is configured to train a convolutional-self-attention deep learning model based on the well logging data and a data set of well logging lithology, to obtain a well logging lithology identification model expressing a mapping relationship between the well logging data and the well logging lithology. The identification output unit is configured to obtain well logging lithology according to the well logging data by using the well logging lithology identification model.

Images