CN117610429A

CN117610429A - Well logging curve completion model training method and device based on deep learning

Info

Publication number: CN117610429A
Application number: CN202311694040.6A
Authority: CN
Inventors: 孙晓明; 伍新明
Original assignee: University of Science and Technology of China USTC
Current assignee: University of Science and Technology of China USTC
Priority date: 2023-12-08
Filing date: 2023-12-08
Publication date: 2024-02-27

Abstract

The application discloses a logging curve completion model training method and device based on deep learning, and belongs to the technical field of artificial intelligence. The method comprises the following steps: inputting the logging curves in the sample curve set into a preset initial model to obtain normalized characteristics output by the heterogeneous graph neural network and statistical information of sampling points on the missing logging curves output by the fully connected neural network; performing inverse normalization operation on the normalization features according to the statistical information to convert the normalization features into corresponding missing logging curves; training the initial model based on the missing log curve to update parameters of the heterograph neural network and the fully connected neural network to obtain a log curve complement model. According to the method and the system, the missing logging curve prediction is realized through the graph neural network and the full-connection neural network, the known logging curve prediction missing logging curve can be applied, the prediction logging curve type is not required to be limited, the missing logging curve prediction is more flexible, more accurate and stable, and the efficiency is higher.

Description

Logging curve completion model training method and device based on deep learning

Technical Field

The application belongs to the technical field of artificial intelligence, and particularly relates to a logging curve completion model training method and device based on deep learning.

Background

Geophysical logging is a method of measuring geophysical parameters using the electrochemical, conductive, acoustic, radioactive, etc., geophysical properties of a formation; logging data obtained through logging is basic data of geological modeling and fine characterization of a reservoir, can be used for geological exploration, oil and gas field development and production management, and helps engineers and geologist know the properties of underground rock and stratum, so that drilling, completion and production operations are guided.

In particular use, geologists and engineers can build accurate geologic models based on well log data and design exploration and development strategies. The acquisition of logging data mainly goes into the well through logging tools, and physical properties of stratum are measured by utilizing sensors of different principles, and common logging tools comprise logging instruments, logging probes, logging cables and the like. However, the acquisition of log curves is often expensive and time consuming, and in actual measurements, due to various objective reasons, problems with missing log data often occur, and some of the entire log curves may be abandoned for cost reasons. Thus, completion and generation of well logs is an academic and engineering study.

However, because the stratum condition is complex and has anisotropy, the mapping relationship between different logging curves is extremely complex, and whether the mapping relationship is a traditional physical model or an empirical model, the relationship between the logging curves is difficult to accurately describe, and the missing logging curves cannot be complemented.

Disclosure of Invention

The present application aims to solve at least one of the technical problems existing in the prior art. Therefore, the application provides a training method and device for a completion model of a logging curve based on deep learning, so as to improve the completion efficiency of a missing logging curve.

In a first aspect, a method for training a completion model of a log based on deep learning, includes:

inputting the logging curves in the sample curve set into a preset initial model to obtain the normalization characteristics corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model and the statistical information of sampling points on the missing logging curves output by the full-connection neural network of the initial model;

performing inverse normalization operation on the normalization feature according to the statistical information so as to convert the normalization feature into a corresponding missing log curve;

training the initial model based on the missing log curves to update parameters of the heterogeneous graph neural network and the fully connected neural network, and obtaining a log curve completion model.

According to the training method of the deep learning-based logging curve complement model, the logging curves in the sample curve set are input into a preset initial model, so that the normalization characteristics corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model and the statistical information of sampling points on the missing logging curves output by the full-connection neural network of the initial model are obtained; performing inverse normalization operation on the normalization feature according to the statistical information so as to convert the normalization feature into a corresponding missing log curve; training the initial model based on the missing log curves to update parameters of the heterogeneous graph neural network and the fully connected neural network, and obtaining a log curve completion model. According to the embodiment of the application, the heterogeneous graph neural network is used for deducing the interrelationship among the multi-well logging curves so as to predict the normalized logging curves, the full-connection neural network is used for inversely normalizing the predicted normalized logging curves, the model is trained in a hybrid deep learning mode combining the heterogeneous graph neural network and the full-connection neural network, the consistency of network output is ensured, the completion of different logging curves is adaptively realized, and the completion efficiency of the missing logging curves is adaptively improved.

According to one embodiment of the present application, the method further comprises:

acquiring a group of logging curves corresponding to a target well; the set of well logs includes a plurality of well logs; the logging curve is correspondingly provided with a label for representing the shape of the logging curve;

carrying out random masking on at least one tag to obtain a set of masked log curves;

combining the masked set of log curves and the corresponding labels to form a pair of training sample pairs;

and constructing the sample curve set according to training sample pairs corresponding to the target drilling.

In this embodiment, the sample curve set is constructed by performing a random mask on the log, so that in the model training process, the loss function is calculated by predicting the log of the random mask and further comparing the predicted value with the true value corresponding to the label. Further, since the number of complete and undesireable log curves is relatively small, and in order to obtain a sufficient amount of training data, in this embodiment, by randomly masking the log curves, a plurality of random masking can be performed on the same set of log curves, so as to generate a plurality of pairs of training sample pairs, thereby increasing the amount of training data.

According to one embodiment of the present application, the set of well logs further includes missing well logs;

the step of performing random masking on at least one tag to obtain a set of masked log curves includes:

and performing original masking on the labels corresponding to the missing log curves, and performing random masking on the labels corresponding to at least one of the missed log curves to obtain a set of log curves after masking.

In this embodiment, the missing log curves are further masked in the original mask manner, and the log curves which are not missing are randomly masked, so that in the model training process, the log curves with the random mask can still be predicted, and further, the predicted value and the true value corresponding to the label are compared to calculate the loss function. The number of missing logs is large, so that the amount of training data can be further increased by adaptively constructing training sample pairs in the manner described above.

According to one embodiment of the present application, the combining the masked set of log curves and the corresponding labels to form a pair of training sample pairs includes:

performing point multiplication on the logging curve after the random masking and the corresponding label to obtain a sampling point value at the masking;

Sampling the masked set of logging curves with different multiplying powers based on the sampling point values to obtain a set of logging curves with different sampling rates;

combining the set of log curves at different sampling rates with corresponding labels to form a pair of training sample pairs.

In this embodiment, by multi-scale sampling of the log, information of different logs can be captured with greater accuracy, thereby facilitating improved accuracy in extracting log features.

According to one embodiment of the application, the heterogeneous graph neural network outputs normalized features corresponding to missing log curves according to the following manner:

extracting first node features corresponding to different types of logging curves based on different encoders and different decoders;

calculating the characteristic relation of logging curves among different categories according to the first node characteristics;

predicting a second node characteristic corresponding to the missing logging curve according to the characteristic relation;

and outputting the second node characteristic, wherein the second node characteristic represents a normalization characteristic corresponding to the missing logging curve.

In this embodiment, if the uniqueness of different well profile features is greatly weakened by using a conventional single encoder to fix the well profile input extraction features, the embodiment can avoid interference between different well profile features by using different encoders and different decoders to perform feature extraction on different types of well profiles.

According to one embodiment of the application, the calculating the characteristic relation of the logging curves between different categories according to the first node characteristic includes:

calculating characteristic relations of logging curves among different categories based on an aggregation module, wherein the aggregation module is represented by the following formula:

wherein, the right side of the formula represents the characteristic relation of the logging curves among different categories, W _c Learning weights representing convolution operations, N (i) representing a set of adjacent nodes to node i, x _j First node characteristics, alpha, representing neighboring nodes _i,j Representing the attention coefficient between node i and node j, x' _i Representing the new feature of node i, namely the second node feature corresponding to the missing log.

wherein W is _i,j A learnable matrix is represented for enhancing attention.

According to an embodiment of the present application, the training the initial model based on the missing log to update parameters of the heterogeneous graph neural network and the fully connected neural network, to obtain a log complement model, includes:

Calculating the overall loss function of the heterograph neural network and the fully connected neural network according to the difference between the predicted missing log curve and the true value corresponding to the label value;

and updating parameters of the heterogeneous graph neural network and the fully connected neural network based on the total loss function to obtain a log curve completion model.

According to one embodiment of the present application, the overall loss function is calculated by the following formula:

L＝w _G L _G +w _F L _F

wherein L represents the overall loss function, L _F Representing the loss function, w, of a fully connected neural network _F Represents L _F Weights, L _G Representing the loss function of a heterogeneous graph neural network, w _G Represents L _G Weights, L _MSE And L _SSIM Respectively represent the mean square error and the structural similarity index, w ₁ And w ₂ As the weight coefficient, m _i ⁰ And m _i ^d Representing the original mask and the random mask (i=0, 1,2,3 …), y _i (i=0, 1,2,3 …) andrepresenting the label and predicted log values, respectively.

In a second aspect, the present application provides a method for completion of a log based on deep learning, comprising:

acquiring a well logging curve known by a target well;

selecting the type of the missing logging curve, inputting the known logging curve into a preset logging curve complement model, and obtaining the missing logging curve output by the logging curve complement model;

Wherein the log completion model is trained using the method as described in the first aspect.

According to the logging curve complement method based on deep learning, the logging curve complement model used in the method is used for deducing the interrelationship among the logging curves of multiple wells through the heterogeneous graph neural network so as to predict the normalized logging curve, the full-connection neural network is used for carrying out inverse normalization on the predicted normalized logging curve, the model is trained in a mixed deep learning mode combining the heterogeneous graph neural network and the full-connection neural network, consistency of network output is ensured, the complement of different logging curves is realized in a self-adaptive mode, and the complement efficiency of the missing logging curve is improved.

In a third aspect, the present application provides a logging curve completion model training device based on deep learning, the device comprising:

the input module is used for inputting the logging curves in the sample curve set into a preset initial model to obtain the normalization characteristics corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model and the statistical information of sampling points on the missing logging curves output by the full-connection neural network of the initial model;

The conversion module is used for carrying out inverse normalization operation on the normalization feature according to the statistical information so as to convert the normalization feature into a corresponding missing logging curve;

and the training module is used for training the initial model based on the missing log curve so as to update parameters of the heterogeneous graph neural network and the fully connected neural network and obtain a log curve complement model.

According to the deep learning-based well logging curve completion model training device, the correlation among the multi-well logging curves is deduced through the heterogeneous graph neural network so as to predict the normalized well logging curves, the full-connection neural network is used for reversely normalizing the predicted normalized well logging curves, the model is trained in a mixed deep learning mode combining the heterogeneous graph neural network and the full-connection neural network, consistency of network output is ensured, completion of different well logging curves is realized in a self-adaptive mode, and completion efficiency of missing well logging curves is improved.

In a fourth aspect, the present application provides a logging curve completion device based on deep learning, comprising:

the acquisition module is used for acquiring a well logging curve known by the target well drilling;

the prediction module is used for selecting the type of the missing logging curve, inputting the known logging curve into a preset logging curve complement model and obtaining the missing logging curve predicted by the logging curve complement model;

According to the logging curve complement device based on deep learning, the logging curve complement model used in the method is used for deducing the interrelationship among the logging curves of multiple wells through the heterogeneous graph neural network so as to predict the normalized logging curve, the full-connection neural network is used for carrying out inverse normalization on the predicted normalized logging curve, the model is trained in a mixed deep learning mode combining the heterogeneous graph neural network and the full-connection neural network, consistency of network output is ensured, the complement of different logging curves is realized in a self-adaptive mode, and the complement efficiency of the missing logging curve is improved.

In a fifth aspect, the present application provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method according to the first or second aspect described above when executing the computer program.

In a sixth aspect, the present application provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements a method as described in the first or second aspect above.

In a seventh aspect, the present application provides a chip comprising a processor and a communication interface, the communication interface and the processor being coupled, the processor being configured to execute a program or instructions to implement a method as described in the first or second aspect.

In an eighth aspect, the present application provides a computer program product comprising a computer program which, when executed by a processor, implements a method as described in the first or second aspect above.

The above technical solutions in the embodiments of the present application have at least one of the following technical effects:

according to the training method of the deep learning-based logging curve complement model, the logging curves in the sample curve set are input into a preset initial model, so that the normalization characteristics corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model and the statistical information of sampling points on the missing logging curves output by the full-connection neural network of the initial model are obtained; performing inverse normalization operation on the normalization feature according to the statistical information so as to convert the normalization feature into a corresponding missing log curve; training the initial model based on the missing log curves to update parameters of the heterogeneous graph neural network and the fully connected neural network, and obtaining a log curve completion model. According to the embodiment of the application, the heterogeneous graph neural network is used for deducing the interrelationship among the multi-well logging curves so as to predict the normalized logging curves, the full-connection neural network is used for inversely normalizing the predicted normalized logging curves, the model is trained in a hybrid deep learning mode combining the heterogeneous graph neural network and the full-connection neural network, the consistency of network output is ensured, the completion of different logging curves is adaptively realized, and the completion efficiency of the missing logging curves is improved.

Further, in some embodiments, the sample curve set is constructed by randomly masking the log, so that in the model training process, the loss function is calculated by predicting the log of the random mask and then comparing the predicted value with the true value corresponding to the label. Further, since the number of complete and undesireable log curves is relatively small, and in order to obtain a sufficient amount of training data, in this embodiment, by randomly masking the log curves, a plurality of random masking can be performed on the same set of log curves, so as to generate a plurality of pairs of training sample pairs, thereby increasing the amount of training data.

Furthermore, in some embodiments, the missing log curves are further masked by an original mask method, and the random mask is performed on the log curves without missing, so that in the model training process, the log curves with random mask can still be predicted, and further, the predicted value and the true value corresponding to the label are compared to calculate the loss function. The number of missing logs is large, so that the amount of training data can be further increased by adaptively constructing training sample pairs in the manner described above.

Further, in some embodiments, by performing multi-scale sampling of the log, information of different logs can be captured with greater accuracy, thereby facilitating improved accuracy in extracting log features.

Still further, in some embodiments, where the uniqueness of different well profile features is substantially impaired by using a conventional single encoder to fix the well profile input extraction features, the embodiments may avoid interference between different well profile features by using different encoders and different decoders to extract features from different classes of well profiles.

Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the application.

Drawings

The foregoing and/or additional aspects and advantages of the present application will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

FIG. 1 is a schematic flow chart of a training method for a completion model of a logging curve based on deep learning according to an embodiment of the present application;

FIG. 2 is a schematic diagram of a network structure of an initial model according to an embodiment of the present application;

FIG. 3 is a schematic diagram of a training sample pair construction method employed in an embodiment of the present application;

FIG. 4 is a schematic diagram of the architecture of a neural network according to an embodiment of the present application;

FIG. 5 is a schematic diagram of a blind well (A3) verifying acoustic time Difference (DTC) prediction accuracy according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a blind well (A3) verifying acoustic time Difference (DTC) prediction accuracy according to an embodiment of the present application;

FIG. 7 is a schematic diagram of a blind well (A2) validating Neutron Porosity (NPHI) prediction accuracy in accordance with an embodiment of the present application;

FIG. 8 is a flow chart of a method for deep learning based log completion provided in an embodiment of the present application;

FIG. 9 is a schematic structural diagram of a training device for a deep learning-based log completion model according to an embodiment of the present application;

FIG. 10 is a schematic structural diagram of a depth-learning based log completion device provided in an embodiment of the present application;

fig. 11 is a schematic structural diagram of an electronic device according to an embodiment of the present application.

Detailed Description

Technical solutions in the embodiments of the present application will be clearly described below with reference to the drawings in the embodiments of the present application, and it is apparent that the described embodiments are some embodiments of the present application, but not all embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments in the present application are within the scope of the protection of the present application.

The terms first, second and the like in the description and in the claims, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used may be interchanged, as appropriate, such that embodiments of the present application may be implemented in sequences other than those illustrated or described herein, and that the objects identified by "first," "second," etc. are generally of a type and not limited to the number of objects, e.g., the first object may be one or more. Furthermore, in the description and claims, "and/or" means at least one of the connected objects, and the character "/", generally means that the associated object is an "or" relationship.

For the problem that logging curves of partial wells are missing due to acquisition cost or instrument and equipment, the method mainly comprises the steps of completing missing well sections through temporary wells or other depth well curve replacement, completing missing well sections through a petrophysical model and completing missing well sections through logging based on deep learning. The method of curve replacement is used for complementing the well logging curve, and the geological structure information is extremely high in requirement. In the rock physical model complement mode, since different rocks have different physical models, expert knowledge is required to identify lithology and select a proper model. The completion of the logging curves through deep learning is fixed at present, so that the completion of different logging curves cannot be realized at the same time, and the application of incomplete methods in actual data is restricted.

From the prior art, the completion efficiency of the conventional physical model or the empirical model for the logging curve is low. According to the method and the system, if the mapping relation between different logging curves is extracted through the hybrid deep learning mode of combining the heterograph neural network and the full-connection neural network, the consistency of network output can be ensured, and the complement efficiency of the missing logging curves is improved.

The method and the device for training the completion model of the logging curve based on the deep learning provided by the embodiment of the application are described in detail through specific embodiments and application scenes thereof with reference to the accompanying drawings.

The well logging curve completion model training method based on deep learning can be applied to a terminal, and can be specifically executed by hardware or software in the terminal.

The terminal includes, but is not limited to, a portable communication device such as a mobile phone or tablet having a touch sensitive surface (e.g., a touch screen display and/or a touch pad). It should also be appreciated that in some embodiments, the terminal may not be a portable communication device, but rather a desktop computer having a touch-sensitive surface (e.g., a touch screen display and/or a touch pad).

In the following various embodiments, a terminal including a display and a touch sensitive surface is described. However, it should be understood that the terminal may include one or more other physical user interface devices such as a physical keyboard, mouse, and joystick.

According to the deep learning-based well logging completion model training method, an execution main body of the deep learning-based well logging completion model training method can be an electronic device or a functional module or a functional entity capable of realizing the deep learning-based well logging completion model training method in the electronic device, the electronic device comprises a mobile phone, a tablet personal computer, a camera, a wearable device and the like, and the deep learning-based well logging completion model training method provided by the embodiment of the application is described below by taking the electronic device as an execution main body as an example.

As shown in fig. 1, the training method of the logging curve complement model based on deep learning comprises the following steps: step 110, step 120 and step 130.

And 110, inputting the logging curves in the sample curve set into a preset initial model to obtain the normalized characteristics corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model and the statistical information of sampling points on the missing logging curves output by the full-connection neural network of the initial model.

In the embodiment of the application, the logging data refers to physical property data of underground rock and stratum acquired by logging tools, including information such as stratum pressure, permeability, porosity, density, resistivity and the like. The log is a graph drawn from the log data to reflect changes in different physical properties of the formation. Common well logs include: resistivity curves, natural gamma curves, sonic moveout curves, density curves, neutron porosity curves, and the like. The logging curve is closely related to logging data, and physical property information of stratum can be obtained through analyzing the logging curve, so that oil and gas exploration and production are guided.

In embodiments of the present application, a sample curve set may be constructed by acquiring real log data. For example, a real well-logging curve of a plurality of wells in a work area is obtained, wherein each well-logging can comprise a plurality of different types of well-logging curves, the well-logging curves of the same well-logging are divided into a group, and a sample curve set consisting of a plurality of groups of well-logging curves corresponding to the plurality of wells is constructed, wherein the well-logging curves can be provided with labels for marking the shape of the well-logging curves.

In the embodiment of the application, the initial model is constructed according to a heterogeneous graph neural network and a fully-connected neural network, and the network structure adopted by the initial model is shown in fig. 2. The heterogeneous graph neural network can be used to infer correlations between multiple well logs to predict a normalized well log, and the fully connected layer can be used to estimate statistical information, such as mean and variance, of the predicted well log to denormalize the predicted normalized well log.

Specifically, the heterogeneous graph neural network uses the graph neural network as a core network, firstly, the mean-variance normalization processing can be carried out on each logging curve input, and the exclusivity among sampling points of different logging curves can be reduced while the spatial characteristics are acquired by the network through the normalization processing. Further, the normalized log is up-sampled and down-sampled to enhance the changing characteristics of the log. The multiscale sampling points are then fed into a graph neural network for training. Through the steps, the heterogeneous graph neural network head can adaptively predict the missing normalized well-logging curve through all known well-logging curves, namely, the normalized features corresponding to the missing well-logging curve which can be output by the heterogeneous graph neural network.

Since the heterographic neural network head outputs as normalized log curves, a denormalization operation of the predicted missing log curves is required. For the sampling points of the missing log, the statistical information such as the mean and variance is an unknown quantity. The present application employs a fully connected neural network head to estimate the statistical information required for missing log inverse normalization to solve this problem. The specific modes can include: and carrying out multi-scale sampling on the non-normalized sampling points input by the heterogeneous graph neural network head, inputting the non-normalized sampling points into a four-layer fully-connected network, and accurately predicting the mean value and the variance through the fully-connected neural network to obtain the statistical information of the sampling points on the missing logging curve.

And 120, performing inverse normalization operation on the normalization feature according to the statistical information so as to convert the normalization feature into a corresponding missing log curve.

In particular, inverse normalization is a process of restoring normalized data to original data, and the method employed for inverse normalization depends on the method employed for normalization. After the statistical information of the sampling points on the missing log is obtained, the normalization feature can be subjected to inverse normalization operation based on the statistical information, so that the corresponding missing log is obtained through conversion.

And 130, training the initial model based on the missing log curve to update parameters of the heterograph neural network and the fully-connected neural network, so as to obtain a log curve complement model.

Specifically, after the missing log curves are predicted, the labels of the log curves in the samples can be combined, the initial model is trained, the real values corresponding to the labels are compared based on the prediction results, a loss function is calculated, an input basis is provided for back propagation, parameters of the heterogeneous graph neural network and parameters of the fully connected neural network are iteratively updated, and training is completed, so that a log curve completion model is obtained.

According to the training method of the deep learning-based logging curve complement model, the logging curves in the sample curve set are input into a preset initial model, so that the normalization characteristics corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model and the statistical information of sampling points on the missing logging curves output by the full-connection neural network of the initial model are obtained; performing inverse normalization operation on the normalization features according to the statistical information to convert the normalization features into corresponding missing logging curves; training the initial model based on the missing log curve to update parameters of the heterograph neural network and the fully connected neural network to obtain a log curve complement model. According to the embodiment of the application, the heterogeneous graph neural network is used for deducing the interrelationship among the multi-well logging curves so as to predict the normalized logging curves, the full-connection neural network is used for inversely normalizing the predicted normalized logging curves, the model is trained in a hybrid deep learning mode combining the heterogeneous graph neural network and the full-connection neural network, the consistency of network output is ensured, the completion of different logging curves is adaptively realized, and the completion efficiency of the missing logging curves is improved.

In some embodiments, the sample curve set may be constructed according to the following:

acquiring a group of logging curves corresponding to a target well; the set of well logs includes a plurality of well logs; the logging curve corresponds to a label representing the shape of the logging curve;

and constructing a sample curve set according to the training sample pairs corresponding to the plurality of target wells.

Typically, a plurality of different logs are available for each well, the shape of the logs themselves being used as a tag. As shown in FIG. 3, taking the example of constructing training sample pairs based on four log curves, y in FIG. 3 ₀ ，y ₁ ，y ₂ ，y ₃ Respectively representing a natural gamma curve, a density curve, a neutron porosity curve and an acoustic time difference curve. At least one tag is randomly masked (M _d ) The masked log is obtained and the set of logs and corresponding labels are then combined to form a pair of training sample pairs.

In practical situations, since the number of complete and undelayed log curves is relatively small, and the number of missed log curves is large but cannot be used as training data, the random masking mode can increase the amount of the training data, and in order to further increase the amount of training samples, the sample curve set is constructed by adaptively constructing a training sample pair.

In some embodiments, the missing log is also included in the set of logs;

randomly masking at least one tag to obtain a masked set of log curves, comprising:

and performing original masking on the labels corresponding to the missing log curves, and performing random masking on the labels corresponding to at least one of the missed log curves to obtain a set of masked log curves.

Due to missing part of the log (M ₀ ) Normally, the log of the well cannot be used as training data. The logging curve of the well can be used as a training data set after being processed through the method and the device. Specifically, first, at least one tag is subjected to random masking (M _d ) And obtaining a masked sample, and carrying out original masking on the label corresponding to the missing logging curve. The values of the random mask and the original mask are taken to be 0 or 1, if a value of 0 represents missing data at the training data mask.

In some embodiments, combining the masked set of log curves and corresponding labels to form a pair of training sample pairs includes:

a set of log curves of different sampling rates and corresponding labels are combined to form a pair of training sample pairs.

Specifically, as shown in FIG. 3, the reference number m may be ₀ ⁰ ，m ₁ ⁰ ，m ₂ ⁰ ，m ₃ ⁰ (corresponding to the original log y respectively) ₀ ，y ₁ ，y ₂ ，y ₃ ) Downsampling by 4, 8 and 16 times to obtain x _i ⁰ ，x _i ⁴ ，x _i ⁸ ，x _i ¹⁶ (i=0, 1,2, 3). Finally, the log curves of different sampling rates of the multi-channel input and the corresponding labels are combined to form a pair of training sample pairs.

In some embodiments, the heterogeneous graph neural network outputs normalized features corresponding to missing log curves according to the following:

calculating characteristic relations of logging curves among different categories according to the characteristics of the first node;

predicting the second node characteristics corresponding to the missing logging curve according to the characteristic relation;

and outputting a second node characteristic, wherein the second node characteristic represents a normalization characteristic corresponding to the missing logging curve.

Specifically, in this embodiment, as shown in fig. 4, the core network of the heterogeneous graph neural network head is a graph neural network, and the basic structure thereof is a U-Net with the improved aggregation module of the present application. The network may adaptively construct a separate graph structure corresponding to each sample point of the missing log, with the graph network defining each type of log as a separate node. In this way, the graph network can employ an aggregation module to aggregate the characteristics of all known log nodes, thereby predicting missing node characteristics (logs).

In the embodiment of the present application, the graph network in the heterogeneous graph neural network extracts the spatial characteristics of each log by using a plurality of independent convolutional encoders, for example, encoder G extracts the natural gamma curve characteristics in fig. 4, encoder N extracts the neutron porosity curve characteristics, and encoder R and encoder D extract the other log characteristics.

In the up-sampling process of the decoder side, the same independent decoders are adopted to up-sample and decode each logging curve, and the first node characteristics corresponding to different types of logging curves are obtained. The jump connection in the graph network structure can integrate downsampling and upsampling information in different stages, so that logging curve information in different frequency ranges can be completely reserved.

After the encoder and the decoder extract the first node characteristics corresponding to the logging curves of different types, the characteristic relation of the logging curves among different types can be calculated according to the characteristics, so that the second node characteristics corresponding to the missing logging curve are predicted according to the characteristic relation, wherein the second node characteristics are normalized characteristics corresponding to the missing logging curve.

In some embodiments, calculating a characteristic relationship of the log between different categories from the first node characteristics includes:

In this embodiment, the improved aggregation module described above allows for flexible use of existing logs to predict other missing logs.

wherein W is _i,j A learnable matrix is represented for enhancing attention.

In this embodiment, the learnable matrix W is introduced in equation (1) _i,j The weight distribution of the correlation between different well curves can be further captured.

In some embodiments, training the initial model based on the missing log to update parameters of the heterograph neural network and the fully connected neural network to obtain a log complement model includes:

calculating the total loss function of the heterogeneous graph neural network and the fully-connected neural network according to the difference between the predicted missing logging curve and the true value corresponding to the label value;

and updating parameters of the heterograph neural network and the fully-connected neural network based on the total loss function to obtain a logging curve completion model.

In this embodiment, by training the heterogeneous graph neural network in combination with the fully connected nerves, consistency of the network output is ensured by the mixed overall loss function.

In some embodiments, the overall loss function is calculated by the following formula:

L＝w _G L _G +w _F L _F (3)

wherein L represents a totalVolume loss function, L _F Representing the loss function, w, of a fully connected neural network _F Represents L _F Weights, L _G Representing the loss function of a heterogeneous graph neural network, w _G Represents L _G Weights, L _MSE And L _SSIM Respectively represent the mean square error and the structural similarity index, w ₁ And w ₂ As the weight coefficient, m _i ⁰ And m _i ^d Representing the original mask and the random mask (i=0, 1,2,3 …), y _i (i=0, 1,2,3 …) andrepresenting the label and predicted log values, respectively.

In this embodiment, the overall loss function is composed of two parts, namely, the loss function of the fully connected neural network and the loss function of the heterogeneous graph neural network; the loss function of the heterogeneous graph neural network consists of a mean square error MSE and a structural similarity index SSIM; the loss function of the fully connected neural network may be composed of the mean absolute error MAE to calculate the global mean and variance of the missing log.

In order to verify the effect of the deep learning-based well logging curve complement model training method, the inventor performs effect verification by using well logging data disclosed in a certain work area. The work area contains 117 wells. The 117 wells contained log categories: natural gamma, sonic moveout, neutron porosity, density, resistivity log, etc. The example selects a natural gamma curve, a density curve, a neutron porosity curve, and an acoustic moveout curve for method testing. Of 117 wells, 102 wells were selected as training sets, 10 of the remaining 15 wells were used as validation sets, and 5 other wells were tested as blind wells. In the training process, the self-adaptive learning rate is adopted, and the network trains 1500 rounds. The validation set loss function reaches convergence on the 1000 th round of training.

The validity of the method of the present application was verified by comparing the predictive accuracy of the method of the present application (flexlog net) with other methods (bi-directional long short time memory (BiLSTM) and empirical formula (gamnder)) at 5 blind wells. The two-way long short term memory method requires a fixed network to input the type of log and can only predict a fixed missing log. Thus, natural Gamma (GR), density (RHOB), neutron Porosity (NPHI), and acoustic time Difference (DTC) predictive models were trained separately for comparison. For the method of the application, the network input is composed of a natural gamma curve and any two other logging curves, and the other one is selected as a prediction curve. In quantitative analysis of the predicted results, the model of the present application performed best on pearson correlation coefficient (maximum correlation coefficient) and minimized on root mean square error (tables 1 and 2) indicating that the present patent method was far more accurate than other methods in predicting missing well data.

Table 1 test well pearson correlation coefficient comparison

Table 2 comparison of root mean square errors for test wells

In the blind well test, the method has the pearson correlation coefficient reaching 0.926 and the root mean square error reaching 4.6614 on the acoustic wave time difference prediction. In contrast, the two indexes of the bidirectional long-short time memory method are respectively 0.916 and 5.0022, and the two indexes of the empirical formula method are respectively 0.755 and 21.4728. Taking the A3 well as an example, the method of the present application predicts that the result reaches 0.903 on the pearson correlation coefficient and 3.8774 on the root mean square error. In the whole well Duan Quxian result prediction, as shown in fig. 5, the prediction result of the method is closer to the real logging data, the deviation from the real logging curve is smaller, and the prediction accuracy on a thin layer (at the curve peak) is higher.

In the method, on the basis of density curve prediction, the pearson correlation coefficient reaches 0.845, and the root mean square error is 0.0504. In contrast, the two indexes of the bidirectional long-short time memory method are respectively 0.726 and 0.0725, and the two indexes of the empirical formula method are respectively 0.821 and 0.1084. Taking the A0 well as an example, the prediction result of the method reaches 0.954 on the pearson correlation coefficient and 0.0619 on the root mean square error. In the overall well Duan Quxian result prediction, as shown in fig. 6, the method of the present application predicts results that are closer to the true log data and deviate less from the true log curve.

In the method, on neutron porosity curve prediction, the pearson correlation coefficient reaches 0.885, and the root mean square error is 0.0267. In contrast, the two indexes of the bidirectional long-short time memory method are respectively 0.864 and 0.0286. The method of the application is superior to the long short-time memory method except that the A0 well is slightly lower than the two-way long short-time memory method in index, and other 4 blind wells are superior to the long short-time memory method. Taking an A2 well as an example, although the vertical change characteristics of the curve can be captured by the method and the long-short time memory method, the method can well match with real logging data at the position of an arrow shown in FIG. 7 of a thin layer comparatively developing layer section.

Thus, from the above comparison, the present application is excellent in predicting a missing log and can be free from the influence of the type of missing log.

The embodiment of the application also provides a logging curve completion method based on deep learning, as shown in fig. 8, the logging curve completion method based on deep learning comprises the following steps: step 810 and step 820.

Step 810, acquiring a well logging curve known by the target well;

step 820, selecting the type of the missing logging curve, and inputting the known logging curve into a preset logging curve complement model to obtain the missing logging curve output by the logging curve complement model;

the well logging curve completion model is trained by the well logging curve completion model training method based on deep learning.

According to the logging curve completion method based on deep learning, the logging curve completion model used in the embodiment of the application is used for deducing the interrelationship among the logging curves of multiple wells through the heterogeneous graph neural network so as to predict the normalized logging curve, the full-connection neural network is used for carrying out inverse normalization on the predicted normalized logging curve, the model is trained in a mixed deep learning mode combining the heterogeneous graph neural network and the full-connection neural network, the consistency of network output is ensured, the completion of different logging curves is realized in a self-adaptive mode, and the completion efficiency of the missing logging curve is improved.

According to the well logging curve completion model training method based on the deep learning, which is provided by the embodiment of the application, an execution subject can be a well logging curve completion model training device based on the deep learning. In the embodiment of the application, a method for performing a well-logging curve completion model training based on deep learning by using the well-logging curve completion model training device based on deep learning as an example is described.

The embodiment of the application also provides a logging curve completion model training device based on deep learning.

As shown in fig. 9, the training device for the completion model of the logging curve based on deep learning comprises:

the input module 910 is configured to input the log curves in the sample curve set into a preset initial model, to obtain normalized features corresponding to the missing log curve output by the heterogeneous graph neural network of the initial model, and statistical information of sampling points on the missing log curve output by the full-connected neural network of the initial model;

the conversion module 920 is configured to perform inverse normalization operation on the normalized feature according to the statistical information, so as to convert the normalized feature into a corresponding missing log;

the training module 930 is configured to train the initial model based on the missing log to update parameters of the heterograph neural network and the fully connected neural network to obtain a log complement model.

In some embodiments, the input module 910 is further configured to:

In this embodiment, the sample curve set is constructed by randomly masking the log, so that in the model training process, the loss function is calculated by predicting the log of the random mask and then comparing the predicted value with the true value corresponding to the label. Further, since the number of complete and undesireable log curves is relatively small, and in order to obtain a sufficient amount of training data, in this embodiment, by randomly masking the log curves, a plurality of random masking can be performed on the same set of log curves, so as to generate a plurality of pairs of training sample pairs, thereby increasing the amount of training data.

In some embodiments, the missing log is also included in the set of logs; the input module 810 is further configured to:

In some embodiments, the input module 910 is further configured to:

wherein W is _i,j A learnable matrix is represented for enhancing attention.

In some embodiments, training module 930 is further to:

the overall loss function is calculated by the following formula:

L＝w _G L _G +w _F L _F

According to the logging curve complement method based on the deep learning, which is provided by the embodiment of the application, the execution main body can be the logging curve complement device based on the deep learning. In the embodiment of the application, taking a logging curve completion method based on deep learning as an example, the logging curve completion device based on deep learning is executed, and the logging curve completion device based on deep learning provided in the embodiment of the application is described.

The embodiment of the application also provides a logging curve complement device based on deep learning.

As shown in fig. 10, the deep learning-based log completion apparatus includes:

an acquisition module 1010 for acquiring a well log known to the target well;

the prediction module 1020 is used for selecting the type of the missing logging curve, inputting the known logging curve into a preset logging curve complement model, and obtaining the missing logging curve predicted by the logging curve complement model;

The device for training the deep learning-based logging completion model or the device for training the deep learning-based logging completion model in the embodiment of the application may be an electronic device, or may be a component in the electronic device, such as an integrated circuit or a chip. The electronic device may be a terminal, or may be other devices than a terminal. By way of example, the electronic device may be a mobile phone, tablet computer, notebook computer, palm computer, vehicle-mounted electronic device, mobile internet appliance (Mobile Internet Device, MID), augmented reality (augmented reality, AR)/Virtual Reality (VR) device, robot, wearable device, ultra-mobile personal computer, UMPC, netbook or personal digital assistant (personal digital assistant, PDA), etc., but may also be a server, network attached storage (Network Attached Storage, NAS), personal computer (personal computer, PC), television (TV), teller machine or self-service machine, etc., and the embodiments of the present application are not limited in particular.

The well logging completion model training device based on deep learning or the well logging completion device based on deep learning in the embodiment of the application may be a device with an operating system. The operating system may be a microsoft (Windows) operating system, an Android operating system, an IOS operating system, or other possible operating systems, which are not specifically limited in the embodiments of the present application.

In some embodiments, as shown in fig. 11, the embodiment of the present application further provides an electronic device 1100, including a processor 1101, a memory 1102, and a computer program stored in the memory 1102 and capable of running on the processor 1101, where the program when executed by the processor 1101 implements the above-mentioned method for training the completion model of the logging curve based on deep learning or the respective processes of the embodiment of the method for completing the logging curve based on deep learning, and can achieve the same technical effects, so that repetition is avoided and will not be repeated here.

The electronic device in the embodiment of the application includes the mobile electronic device and the non-mobile electronic device.

The embodiment of the application also provides a non-transitory computer readable storage medium, on which a computer program is stored, which when executed by a processor, implements each process of the above-mentioned embodiment of the well logging curve completion model training method based on deep learning or the well logging curve completion method based on deep learning, and can achieve the same technical effect, so that repetition is avoided and redundant description is omitted here.

The processor is a processor in the electronic device in the above embodiment. Readable storage media include computer readable storage media such as computer readable memory ROM, random access memory RAM, magnetic or optical disks, and the like.

The embodiment of the application also provides a computer program product, which comprises a computer program, wherein the computer program is executed by a processor to realize the well logging curve complement model training method based on deep learning or the well logging curve complement method based on deep learning.

The embodiment of the application further provides a chip, the chip includes a processor and a communication interface, the communication interface is coupled with the processor, the processor is used for running a program or instructions, the above-mentioned logging curve completion model training method based on deep learning or the processes of the logging curve completion method embodiment based on deep learning are realized, the same technical effects can be achieved, and in order to avoid repetition, the description is omitted here.

It should be understood that the chips referred to in the embodiments of the present application may also be referred to as system-on-chip chips, chip systems, or system-on-chip chips, etc.

It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element. Furthermore, it should be noted that the scope of the methods and apparatus in the embodiments of the present application is not limited to performing the functions in the order shown or discussed, but may also include performing the functions in a substantially simultaneous manner or in an opposite order depending on the functions involved, e.g., the described methods may be performed in an order different from that described, and various steps may also be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

From the above description of the embodiments, it will be clear to those skilled in the art that the above-described embodiment method may be implemented by means of software plus a necessary general hardware platform, but of course may also be implemented by means of hardware, but in many cases the former is a preferred embodiment. Based on such understanding, the technical solutions of the present application may be embodied essentially or in a part contributing to the prior art in the form of a computer software product stored in a storage medium (such as ROM/RAM, magnetic disk, optical disk), comprising several instructions for causing a terminal (which may be a mobile phone, a computer, a server, or a network device, etc.) to perform the methods described in the embodiments of the present application.

The embodiments of the present application have been described above with reference to the accompanying drawings, but the present application is not limited to the above-described embodiments, which are merely illustrative and not restrictive, and many forms may be made by those of ordinary skill in the art without departing from the spirit of the present application and the scope of the claims, which are also within the protection of the present application.

In the description of the present specification, reference to the terms "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present application. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

While embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that: many changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the application, the scope of which is defined by the claims and their equivalents.

Claims

1. A logging curve completion model training method based on deep learning is characterized by comprising the following steps:

2. The method according to claim 1, wherein the method further comprises:

3. The method of claim 2, wherein the set of well logs further comprises missing well logs;

4. The method of claim 2, wherein combining the masked set of logs and corresponding labels forms a pair of training sample pairs, comprising:

5. The method of claim 1, wherein the heterogeneous graph neural network outputs normalized features corresponding to missing log curves according to:

6. The method of claim 5, wherein calculating a characteristic relationship of the log between different categories from the first node characteristic comprises:

7. The method of claim 5, wherein calculating a characteristic relationship of the log between different categories from the first node characteristic comprises:

wherein W is _i,j A learnable matrix is represented for enhancing attention.

8. The method of claim 1, wherein training the initial model based on the missing log to update parameters of the heterogeneous graph neural network and the fully connected neural network results in a log complement model, comprising:

9. The method of claim 8, wherein the overall loss function is calculated by the formula:

L＝w _G L _G +w _F L _F

10. A logging completion method based on deep learning, comprising:

acquiring a well logging curve known by a target well;

Wherein the log completion model is trained using the method of any of claims 1-9.

11. Logging curve completion model trainer based on deep learning, characterized by comprising:

12. A logging completion device based on deep learning, comprising:

13. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any of claims 1-10 when the program is executed by the processor.

14. A non-transitory computer readable storage medium, having stored thereon a computer program, which when executed by a processor, implements the method according to any of claims 1-10.

15. A computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1-10.