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

Abstract

The application discloses a kind of based on deep learning's well logging curve completion model training method and device, belong to artificial intelligence technical field.The method includes: the well logging curve in sample curve set is input into the initial model of preestablished, obtains the normalized feature output by heterogeneous graph neural network, and the statistical information of the up-sampling point on missing well logging curve output by fully connected neural network;According to statistical information, the normalized feature is carried out anti-normalization operation, to convert into corresponding missing well logging curve;Based on missing well logging curve, the initial model is trained, to update the parameters of heterogeneous graph neural network and fully connected neural network, obtains well logging curve completion model.The application realizes missing well logging curve prediction by graph neural network and fully connected neural network, can predict missing well logging curve by known well logging curve, does not need to limit the type of predicted well logging curve, is more flexible, more accurate and stable in missing well logging curve prediction, and is more efficient.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a training method and apparatus for a well logging curve completion model based on deep learning. Background Technology

[0002] Geophysical logging is a method of measuring geophysical parameters by utilizing the geophysical properties of rock formations, such as electrochemical, electrical, acoustic, and radioactive properties. The logging data obtained is the basic data for geological modeling and fine reservoir characterization. It can be used for geological exploration, oil and gas field development and production management, helping engineers and geologists understand the properties of underground rocks and formations, thereby guiding drilling, completion and production operations.

[0003] In practical applications, geologists and engineers can build accurate geological models and design exploration and development strategies based on well logging data. Well logging data acquisition primarily involves running logging tools into the well and using sensors based on different principles to measure the physical properties of the formation. Common logging tools include logging instruments, logging probes, and logging cables. However, acquiring well logging curves is often expensive and time-consuming. In actual measurements, due to various objective reasons, the problem of missing well logging data frequently occurs, and cost considerations may lead to the abandonment of measuring certain entire well logging curves. Therefore, the completion and generation of well logging curves is a research area with both academic and engineering value.

[0004] However, due to the complexity and anisotropy of the formation, the mapping relationship between different logging curves is extremely complex. Neither traditional physical models nor empirical models can accurately describe the relationship between logging curves and cannot complete missing logging curves. Summary of the Invention

[0005] This application aims to address at least one of the technical problems existing in the prior art. To this end, this application proposes a deep learning-based well logging curve completion model training method and apparatus to improve the efficiency of completing missing well logging curves.

[0006] Firstly, a deep learning-based method for training a well logging curve completion model includes:

[0007] The well logging curves in the sample curve set are input into a preset initial model to obtain the normalized features corresponding to the missing well logging curves output by the heterogeneous graph neural network of the initial model, and the statistical information of the sampling points on the missing well logging curves output by the fully connected neural network of the initial model.

[0008] The normalized features are denormalized based on the statistical information to convert them into corresponding missing logging curves.

[0009] The initial model is trained based on the missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, thereby obtaining a logging curve completion model.

[0010] According to the deep learning-based well logging curve completion model training method of this application, well logging curves from a sample curve set are input into a preset initial model to obtain normalized features corresponding to missing well logging curves output by the heterogeneous graph neural network of the initial model, and statistical information of sampling points on the missing well logging curves output by the fully connected neural network of the initial model. The normalized features are then denormalized based on the statistical information to convert them into corresponding missing well logging curves. The initial model is trained based on the missing well logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, resulting in a well logging curve completion model. This application embodiment uses a heterogeneous graph neural network to infer the interrelationships between well logging curves from multiple wells and predict normalized well logging curves, and a fully connected neural network to denormalize the predicted normalized well logging curves. The model is trained using a hybrid deep learning approach combining heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network outputs, adaptively complete different well logging curves, and adaptively improve the efficiency of missing well logging curve completion.

[0011] According to one embodiment of this application, the method further includes:

[0012] Obtain a set of logging curves corresponding to the target well; the set of logging curves includes multiple logging curves; each logging curve has a label indicating its shape;

[0013] At least one label is randomly masked to obtain a set of well logging curves after masking;

[0014] The masked set of logging curves and the corresponding labels are combined to form a training sample pair;

[0015] The sample curve set is constructed based on training sample pairs corresponding to multiple target wells.

[0016] In this embodiment, a sample curve set is constructed by randomly masking the well logging curves. During model training, the loss function is calculated by predicting the randomly masked well logging curves and comparing the predicted values ​​with the true values ​​corresponding to the labels. Furthermore, since the number of complete and unmissing well logging curves is relatively small, this embodiment uses random masking of the well logging curves to generate multiple pairs of training samples for the same set of well logging curves, thereby increasing the amount of training data.

[0017] According to one embodiment of this application, the set of logging curves also includes missing logging curves;

[0018] The step of randomly masking at least one label to obtain a set of masked logging curves includes:

[0019] The labels corresponding to the missing logging curves are masked in their original form, and the labels corresponding to at least one non-missing logging curve are masked randomly, resulting in a set of masked logging curves.

[0020] In this embodiment, missing logging curves are masked using the original masking method, while non-missing logging curves are randomly masked. This allows the model to still predict the randomly masked logging curves during training, and then compare the predicted values ​​with the true values ​​corresponding to the labels to calculate the loss function. Since the number of missing logging curves is large, adaptively constructing training sample pairs using the above method can further increase the amount of training data.

[0021] According to one embodiment of this application, the step of combining a masked set of well logging curves and corresponding labels to form a training sample pair includes:

[0022] The sampling point value at the mask is obtained by multiplying the randomly masked logging curve with the corresponding label;

[0023] Based on the sampling point values, a set of well logging curves after masking is sampled at different magnifications to obtain a set of well logging curves with different sampling rates;

[0024] The set of logging curves with different sampling rates and their corresponding labels are combined to form a training sample pair.

[0025] In this embodiment, by sampling the logging curves at multiple scales, information from different logging curves can be captured with higher precision, thereby improving the accuracy of feature extraction from the logging curves.

[0026] According to one embodiment of this application, the heterogeneous graph neural network outputs the normalized features corresponding to the missing logging curves in the following manner:

[0027] Extracting first-node features corresponding to different categories of logging curves based on different encoders and decoders;

[0028] Calculate the characteristic relationship between logging curves of different categories based on the characteristics of the first node;

[0029] Predict the second node features corresponding to the missing logging curves based on the aforementioned feature relationships;

[0030] Output the second node feature, which represents the normalized feature corresponding to the missing logging curve.

[0031] In this embodiment, if the conventional method of extracting features by fixing the logging curve input with a single encoder is used, the uniqueness of the features of different well curves will be greatly weakened. This embodiment avoids the interference between the features of different logging curves by using different encoders and different decoders to extract features from different categories of logging curves.

[0032] According to one embodiment of this application, the step of calculating the characteristic relationship of logging curves between different categories based on the first node characteristics includes:

[0033] The aggregation module calculates the characteristic relationships between logging curves of different categories, and the aggregation module is represented by the following formula:

[0034]

[0035] The right side of the formula represents the characteristic relationship between well logging curves of different categories. Represents the learned weights of the convolution operation. Represents a node i The set of neighboring nodes, This represents the first node characteristic of the neighboring nodes. Represents a node i and nodes j Attention coefficient between them Represents a node i The new feature is the second node feature corresponding to the missing logging curve.

[0036] According to one embodiment of this application, the step of calculating the characteristic relationship of logging curves between different categories based on the first node characteristics includes:

[0037] The aggregation module calculates the characteristic relationships between logging curves of different categories, and the aggregation module is represented by the following formula:

[0038]

[0039] in, This represents a learnable matrix, used to enhance attention.

[0040] According to one embodiment of this application, training the initial model based on the missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network to obtain a logging curve completion model includes:

[0041] The overall loss function of the heterogeneous graph neural network and the fully connected neural network is calculated based on the difference between the predicted missing logging curve and the actual value corresponding to the label value.

[0042] The parameters of the heterogeneous graph neural network and the fully connected neural network are updated based on the overall loss function to obtain the well logging curve completion model.

[0043] According to one embodiment of this application, the overall loss function is calculated using the following formula:

[0044]

[0045]

[0046] in, L Represents the overall loss function. L F This represents the loss function of a fully connected neural network. w F express L F The weight, L G This represents the loss function of a heterogeneous graphical neural network. w G express L G The weight, L MSE and L SSIM These represent mean squared error and structural similarity index, respectively. w 1 and w 2 represents the weighting coefficient. m i 0 and m i d Represents the original mask and the random mask ( i =0, 1, 2, 3…), y i ( i =0, 1, 2, 3…) and i ( i =0, 1, 2, 3…) represent the label and the predicted logging curve value, respectively.

[0047] Secondly, this application provides a deep learning-based well logging curve completion method, including:

[0048] Obtain the known logging curves of the target well;

[0049] Select the type of missing logging curve, and input the known logging curve into the preset logging curve completion model to obtain the missing logging curve output by the logging curve completion model;

[0050] The well logging curve completion model is trained using the method described in the first aspect.

[0051] According to the deep learning-based well logging curve completion method of this application, the well logging curve completion model used in this application uses a heterogeneous graph neural network to infer the relationship between well logging curves of multiple wells and then predict the normalized well logging curve. A fully connected neural network is used to denormalize the predicted normalized well logging curve. The model is trained in a hybrid deep learning manner that combines heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network output, adaptively realize the completion of different well logging curves, and improve the completion efficiency of missing well logging curves.

[0052] Thirdly, this application provides a training device for a well logging curve completion model based on deep learning, the device comprising:

[0053] The input module is used to input the logging curves from the sample curve set into a preset initial model to obtain the normalized features corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model, and the statistical information of the sampling points on the missing logging curves output by the fully connected neural network of the initial model.

[0054] The conversion module is used to perform an inverse normalization operation on the normalized features based on the statistical information, so as to convert them into the corresponding missing logging curves;

[0055] The training module is used to train the initial model based on the missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, thereby obtaining a logging curve completion model.

[0056] According to the deep learning-based well logging curve completion model training device of this application, this application uses a heterogeneous graph neural network to infer the relationship between well logging curves of multiple wells and then predict the normalized well logging curve. A fully connected neural network is used to denormalize the predicted normalized well logging curve. The model is trained by a hybrid deep learning method that combines heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network output, adaptively realize the completion of different well logging curves, and improve the completion efficiency of missing well logging curves.

[0057] Fourthly, this application provides a deep learning-based well logging curve completion device, comprising:

[0058] The acquisition module is used to acquire known logging curves of the target well.

[0059] The prediction module is used to select the type of missing logging curve and input the known logging curve into a preset logging curve completion model to obtain the missing logging curve predicted by the logging curve completion model.

[0060] The well logging curve completion model is trained using the method described in the first aspect.

[0061] According to the deep learning-based well logging curve completion device of this application, the well logging curve completion model used in this application uses a heterogeneous graph neural network to infer the relationship between well logging curves of multiple wells and then predict the normalized well logging curve. A fully connected neural network is used to denormalize the predicted normalized well logging curve. The model is trained by a hybrid deep learning method that combines heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network output, adaptively realize the completion of different well logging curves, and improve the completion efficiency of missing well logging curves.

[0062] Fifthly, this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first or second aspect above.

[0063] In a sixth aspect, this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first or second aspect above.

[0064] In a seventh aspect, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the methods described in the first or second aspect above.

[0065] Eighthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first or second aspect above.

[0066] The above-described one or more technical solutions in the embodiments of this application have at least one of the following technical effects:

[0067] According to the deep learning-based well logging curve completion model training method of this application, well logging curves from a sample curve set are input into a preset initial model to obtain normalized features corresponding to missing well logging curves output by the heterogeneous graph neural network of the initial model, and statistical information of sampling points on the missing well logging curves output by the fully connected neural network of the initial model. The normalized features are then denormalized based on the statistical information to convert them into corresponding missing well logging curves. The initial model is trained based on the missing well logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, resulting in a well logging curve completion model. This embodiment uses a heterogeneous graph neural network to infer the interrelationships between well logging curves from multiple wells and predict normalized well logging curves, and a fully connected neural network to denormalize the predicted normalized well logging curves. The model is trained using a hybrid deep learning approach combining heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network outputs, adaptively complete different well logging curves, and improve the efficiency of missing well logging curve completion.

[0068] Furthermore, in some embodiments, a sample curve set is constructed by randomly masking the logging curves. During model training, the loss function is calculated by predicting the randomly masked logging curves and comparing the predicted values ​​with the true values ​​corresponding to the labels. Furthermore, since the number of complete, unmissing logging curves is relatively small, this embodiment uses random masking of the logging curves to generate multiple pairs of training samples for the same set of logging curves, thereby increasing the amount of training data.

[0069] Furthermore, in some embodiments, missing logging curves are masked using the original masking method, while non-missing logging curves are randomly masked. This allows the model to still predict the randomly masked logging curves during training, and then compare the predicted values ​​with the true values ​​corresponding to the labels to calculate the loss function. Since the number of missing logging curves is large, adaptively constructing training sample pairs using the above method can further increase the amount of training data.

[0070] Furthermore, in some embodiments, by performing multi-scale sampling on the logging curves, information from different logging curves can be captured with higher precision, thereby improving the accuracy of feature extraction from the logging curves.

[0071] Furthermore, in some embodiments, if the conventional method of extracting features by using a single encoder to fix the logging curve input would significantly weaken the uniqueness of different logging curve features, this embodiment can avoid interference between different logging curve features by using different encoders and different decoders to extract features from different categories of logging curves.

[0072] Additional aspects and advantages of this 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 this application. Attached Figure Description

[0073] The above and / or additional aspects and advantages of this application will become apparent and readily understood in conjunction with the following description of the embodiments in conjunction with the accompanying drawings, wherein:

[0074] Figure 1 This is a flowchart illustrating the training method for a deep learning-based well logging curve completion model provided in an embodiment of this application.

[0075] Figure 2 This is a schematic diagram of the network structure of the initial model in the embodiments of this application;

[0076] Figure 3 This is a schematic diagram of the training sample pair construction method used in the embodiments of this application;

[0077] Figure 4 This is a schematic diagram of the neural network architecture in an embodiment of this application;

[0078] Figure 5 This is a schematic diagram of the accuracy of the prediction of the acoustic transit time (DTC) in a blind well (A3) according to an embodiment of this application;

[0079] Figure 6 This is a schematic diagram of the accuracy of the prediction of the acoustic transit time (DTC) in a blind well (A3) according to an embodiment of this application;

[0080] Figure 7 This is a schematic diagram illustrating the verification of neutron porosity (NPHI) prediction accuracy using a blind well (A2) according to an embodiment of this application.

[0081] Figure 8 This is a flowchart illustrating the deep learning-based well logging curve completion method provided in this application embodiment;

[0082] Figure 9 This is a schematic diagram of the structure of the deep learning-based well logging curve completion model training device provided in the embodiments of this application;

[0083] Figure 10 This is a schematic diagram of the structure of the deep learning-based well logging curve completion device provided in the embodiments of this application;

[0084] Figure 11 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0085] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0086] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0087] For some wells where logging curves are missing due to acquisition costs or equipment limitations, current solutions mainly include replacing missing sections with curves from adjacent wells or wells at other depths, using rock physics models, and employing deep learning-based logging curve completion. Curve replacement requires extremely high levels of geological structural information. Rock physics model completion requires expert knowledge to identify lithology and select appropriate models because different rocks have different physical models. Deep learning-based logging curve completion currently has relatively fixed network inputs and outputs, thus preventing the simultaneous completion of different logging curves and limiting its application in real-world data.

[0088] As can be seen from the existing technology, both traditional physical models and empirical models have relatively low efficiency in completing well logging curves. This application considers that a hybrid deep learning approach combining heterogeneous graph neural networks and fully connected neural networks can be used to extract the mapping relationship between different well logging curves, thereby ensuring the consistency of network output and improving the efficiency of completing missing well logging curves.

[0089] The following description, in conjunction with the accompanying drawings, details the deep learning-based well logging curve completion model training method and apparatus provided in this application through specific embodiments and application scenarios.

[0090] Among them, the deep learning-based well logging curve completion model training method can be applied to the terminal, specifically executed by the hardware or software in the terminal.

[0091] The terminal includes, but is not limited to, portable communication devices such as mobile phones or tablets with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads). It should also be understood that, in some embodiments, the terminal may not be a portable communication device, but rather a desktop computer with touch-sensitive surfaces (e.g., touchscreen displays and / or touchpads).

[0092] The following embodiments describe a terminal including a display and a touch-sensitive surface. 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.

[0093] The deep learning-based well logging curve completion model training method provided in this application embodiment can be executed by an electronic device or a functional module or entity in an electronic device that can implement the deep learning-based well logging curve completion model training method. The electronic devices mentioned in this application embodiment include, but are not limited to, mobile phones, tablets, computers, cameras, and wearable devices. The following uses an electronic device as the execution subject to illustrate the deep learning-based well logging curve completion model training method provided in this application embodiment.

[0094] like Figure 1 As shown, the training method for the deep learning-based well logging curve completion model includes steps 110, 120, and 130.

[0095] Step 110: Input the logging curves from the sample curve set into the preset initial model to obtain the normalized features corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model, and the statistical information of the sampling points on the missing logging curves output by the fully connected neural network of the initial model.

[0096] In this embodiment, well logging data refers to the physical property data of underground rocks and formations obtained through well logging tools, including information such as formation pressure, permeability, porosity, density, and resistivity. Well logging curves are graphs plotted based on well logging data, used to reflect changes in different physical properties of the formation. Common well logging curves include: resistivity curves, natural gamma curves, sonic transit time curves, density curves, and neutron porosity curves. Well logging curves are closely related to well logging data; by analyzing well logging curves, information on the physical properties of the formation can be obtained, guiding oil and gas exploration and production.

[0097] In this embodiment of the application, a sample curve set can be constructed by acquiring real logging data. For example, real logging curves of multiple wells in a certain work area can be acquired, where each well can include multiple different types of logging curves. The logging curves of the same well are grouped together to construct a sample curve set consisting of multiple groups of logging curves corresponding to multiple wells. The logging curves can also be labeled to indicate the shape of the logging curves.

[0098] In this embodiment, the initial model is constructed based on heterogeneous graph neural networks and fully connected neural networks. The network structure used in the initial model is as follows: Figure 2 As shown, heterogeneous graph neural networks can be used to infer the relationships between logging curves from multiple wells and then predict normalized logging curves. Fully connected layers are used to estimate the statistical information of the predicted logging curves, such as the mean and variance, in order to denormalize the predicted normalized logging curves.

[0099] Specifically, the heterogeneous graph neural network (HNN) uses a graph neural network as its core network. First, it performs mean-variance normalization on each input well logging curve. This normalization process allows the network to capture spatial features while reducing the exclusivity between sampling points from different well logging curves. Then, it performs upsampling and downsampling on the normalized well logging curves to enhance their variation characteristics. Subsequently, multi-scale sampling points are fed into the graph neural network for training. Through these steps, the heterogeneous graph neural network head can adaptively predict missing normalized well logging curves using all known well logging curves; that is, the heterogeneous graph neural network can output the normalized features corresponding to missing well logging curves.

[0100] Since the output of the heterogeneous graph neural network head is a normalized logging curve, it is necessary to perform an inverse normalization operation on the predicted missing logging curves. For the sampling points of the missing logging curves, statistical information such as the mean and variance are unknowns. To address this issue, this application employs a fully connected neural network head to estimate the statistical information required for inverse normalization of the missing logging curves. Specifically, this can involve multi-scale sampling of the unnormalized sampling points input to the heterogeneous graph neural network head and inputting them into a four-layer fully connected network. This fully connected neural network then accurately predicts the mean and variance to obtain the statistical information of the sampling points on the missing logging curves.

[0101] Step 120: Perform denormalization on the normalized features based on the statistical information to convert them into the corresponding missing logging curves.

[0102] Specifically, denormalization is the process of restoring normalized data to its original form. The method used for denormalization depends on the method used for normalization. After obtaining the statistical information of the sampling points on the missing logging curves, the normalized features can be denormalized based on this statistical information to obtain the corresponding missing logging curves.

[0103] Step 130: Train the initial model based on the missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, and obtain the logging curve completion model.

[0104] Specifically, after predicting the missing logging curves, the initial model can be trained by combining the labels of the logging curves in the sample. The loss function is calculated by comparing the prediction results with the true values ​​corresponding to the labels, providing input for backpropagation, and iteratively updating the parameters of the heterogeneous graph neural network and the fully connected neural network. After training is completed, the logging curve completion model is obtained.

[0105] According to the deep learning-based well logging curve completion model training method of this application, well logging curves from a sample curve set are input into a preset initial model to obtain normalized features corresponding to the missing well logging curves output by the heterogeneous graph neural network of the initial model, and statistical information of the sampling points on the missing well logging curves output by the fully connected neural network of the initial model. The normalized features are then denormalized based on the statistical information to convert them into the corresponding missing well logging curves. The initial model is trained based on the missing well logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, resulting in a well logging curve completion model. This embodiment uses a heterogeneous graph neural network to infer the interrelationships between well logging curves from multiple wells and predict normalized well logging curves, while a fully connected neural network is used to denormalize the predicted normalized well logging curves. The model is trained using a hybrid deep learning approach combining heterogeneous graph neural networks and fully connected neural networks, ensuring the consistency of network outputs and adaptively completing different well logging curves, thus improving the efficiency of missing well logging curve completion.

[0106] In some embodiments, the sample curve set can be constructed in the following manner:

[0107] Obtain a set of logging curves corresponding to the target well; a set of logging curves includes multiple logging curves; each logging curve has a label indicating its shape;

[0108] At least one label is randomly masked to obtain a set of well logging curves after masking;

[0109] The masked set of logging curves and the corresponding labels are combined to form a training sample pair;

[0110] A sample curve set is constructed based on training sample pairs corresponding to multiple target wells.

[0111] Typically, multiple different logging curves can be obtained for each well, and the shape of the logging curve itself can serve as a label. For example... Figure 3 As shown, taking the construction of training sample pairs based on four types of well logging curves as an example, Figure 3 middle y 0, y 1, y 2, y 3 represents the natural gamma curve, density curve, neutron porosity curve, and acoustic transit time curve, respectively. At least one tag is randomly masked (M... d The masked logging curves are obtained, and then the set of logging curves and the corresponding labels are combined to form a training sample pair.

[0112] In this embodiment, a sample curve set is constructed by randomly masking the well logging curves. During model training, the loss function is calculated by predicting the randomly masked well logging curves and comparing the predicted values ​​with the true values ​​corresponding to the labels. Furthermore, since the number of complete and unmissing well logging curves is relatively small, this embodiment uses random masking of the well logging curves to generate multiple pairs of training samples for the same set of well logging curves, thereby increasing the amount of training data.

[0113] In practice, the number of complete and unmissing logging curves is relatively small, while the number of missing logging curves is large but cannot be used as training data. The above-mentioned random masking method can increase the amount of training data. In order to further increase the amount of training samples, this application constructs a sample curve set by adaptively constructing training sample pairs.

[0114] In some embodiments, a set of logging curves may also include missing logging curves;

[0115] At least one label is randomly masked to obtain a set of masked logging curves, including:

[0116] The labels corresponding to the missing logging curves are masked in their original form, and the labels corresponding to at least one non-missing logging curve are masked randomly to obtain a set of masked logging curves.

[0117] Because of the missing logging curve (M0), the logging curve of this well cannot be used as training data under normal circumstances. However, by processing the logging curve of this well according to the embodiments of this application, it can be used as a training dataset. Specifically, firstly, at least one label is randomly masked (M0). dThe masked samples are obtained, and the original masks are applied to the labels corresponding to the missing well logging curves. The values ​​of the random mask and the original mask are either 0 or 1. If the value is 0, it means that the data at the training data mask is missing.

[0118] In this embodiment, missing logging curves are masked using the original masking method, while non-missing logging curves are randomly masked. This allows the model to still predict the randomly masked logging curves during training, and then compare the predicted values ​​with the true values ​​corresponding to the labels to calculate the loss function. Since the number of missing logging curves is large, adaptively constructing training sample pairs using the above method can further increase the amount of training data.

[0119] In some embodiments, a masked set of logging curves and corresponding labels are combined to form a training sample pair, including:

[0120] The sampling point value at the mask is obtained by multiplying the randomly masked logging curve with the corresponding label;

[0121] Based on the sampling point values, a set of well logging curves after masking is sampled at different magnifications to obtain a set of well logging curves with different sampling rates;

[0122] A set of well logging curves with different sampling rates and their corresponding labels are combined to form a training sample pair.

[0123] Specifically, such as Figure 3 As shown, it is possible to... m 0 0 , m 1 0 , m 2 0 , m 3 0 (corresponding to the original logging curves respectively) y 0, y 1, y 2, y 3) Perform downsampling at 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 logging curves with different sampling rates from the multi-channel input and the corresponding labels are combined to form a training sample pair.

[0124] In this embodiment, by sampling the logging curves at multiple scales, information from different logging curves can be captured with higher precision, thereby improving the accuracy of feature extraction from the logging curves.

[0125] In some embodiments, the heterogeneous graph neural network outputs the normalized features corresponding to the missing logging curves in the following manner:

[0126] Extracting first-node features corresponding to different categories of logging curves based on different encoders and decoders;

[0127] Calculate the characteristic relationships of logging curves between different categories based on the characteristics of the first node;

[0128] Predict the second node features corresponding to the missing logging curves based on feature relationships;

[0129] Output the second node feature, which represents the normalized feature corresponding to the missing logging curve.

[0130] Specifically, in this embodiment, such as Figure 4 As shown, the core network of the heterogeneous graph neural network head is a graph neural network, and its basic structure is a U-Net with the improved aggregation module of this application. This network can adaptively construct a separate graph structure corresponding to each sampling point of the missing logging curve, and the graph network defines each type of logging curve as a separate node. In this way, the graph network can apply the aggregation module to aggregate the features of all known logging curve nodes, and then predict the features of the missing node (logging curve).

[0131] In this embodiment of the application, the graph network in the heterogeneous graph neural network extracts the spatial features of each logging curve by employing multiple independent convolutional encoders, for example... Figure 4 Encoder G extracts natural gamma curve features, encoder N extracts neutron porosity curve features, and encoders R and D extract other logging curve features.

[0132] During the upsampling process at the decoder end, multiple independent decoders are used to upsample and decode each logging curve, obtaining the first node features corresponding to different categories of logging curves. The jump connections in the graph network structure can integrate downsampling and upsampling information from different stages, ensuring that logging curve information across different frequency bands is completely preserved.

[0133] After the encoder and decoder extract the first node features corresponding to different categories of logging curves, the feature relationship between different categories of logging curves can be calculated based on these features. Then, the second node features corresponding to the missing logging curves can be predicted based on the feature relationship. The second node features are the normalized features corresponding to the missing logging curves.

[0134] In this embodiment, if the conventional method of extracting features by fixing the logging curve input with a single encoder is used, the uniqueness of the features of different well curves will be greatly weakened. This embodiment avoids the interference between the features of different logging curves by using different encoders and different decoders to extract features from different categories of logging curves.

[0135] In some embodiments, calculating the characteristic relationship between logging curves of different categories based on the characteristics of the first node includes:

[0136] Based on the aggregation module, the characteristic relationships between logging curves of different categories are calculated. The aggregation module is represented by the following formula:

[0137]

[0138] The right side of the formula represents the characteristic relationship between well logging curves of different categories. Represents the learned weights of the convolution operation. Represents a node i The set of neighboring nodes, This represents the first node characteristic of the neighboring nodes. Represents a node i and nodes j Attention coefficient between them Represents a node i The new feature is the second node feature corresponding to the missing logging curve.

[0139] In this embodiment, the improved aggregation module described above can flexibly apply existing logging curves to predict other missing logging curves.

[0140] In some embodiments, calculating the characteristic relationship between logging curves of different categories based on the characteristics of the first node includes:

[0141] Based on the aggregation module, the characteristic relationships between logging curves of different categories are calculated. The aggregation module is represented by the following formula:

[0142]

[0143] in, This represents a learnable matrix, used to enhance attention.

[0144] In this embodiment, a learnable matrix is ​​introduced into formula (1). This allows for further capture of the weight distribution of the correlation between different well curves.

[0145] In some embodiments, an initial model is trained based on missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, resulting in a logging curve completion model, including:

[0146] The overall loss function of the heterogeneous graph neural network and the fully connected neural network is calculated based on the difference between the predicted missing logging curves and the actual values ​​corresponding to the label values.

[0147] The parameters of the heterogeneous graph neural network and the fully connected neural network are updated based on the overall loss function to obtain the well logging curve completion model.

[0148] In this embodiment, the network output consistency is ensured by training a combination of heterogeneous graph neural networks and fully connected neural networks, using a hybrid overall loss function.

[0149] In some embodiments, the overall loss function is calculated using the following formula:

[0150]

[0151] in, L Represents the overall loss function. L F This represents the loss function of a fully connected neural network. w F express L F The weight, L G This represents the loss function of a heterogeneous graphical neural network. w G express L G The weight, L MSE and L SSIM These represent mean squared error and structural similarity index, respectively. w 1 and w 2 represents the weighting coefficient. m i 0 and m i d Represents the original mask and the random mask ( i =0, 1, 2, 3…), y i ( i =0, 1, 2, 3…) and i ( i =0, 1, 2, 3…) represent the label and the predicted logging curve value, respectively.

[0152] In this embodiment, the overall loss function consists of two parts: 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 is composed of the mean squared error (MSE) and the structural similarity index (SSIM). The loss function of the fully connected neural network can be composed of the mean absolute error (MAE), which is used to calculate the global mean and variance of the missing logging curves.

[0153] To verify the effectiveness of the deep learning-based well logging curve completion model training method proposed in this application, the inventors conducted a performance verification using publicly available well logging data from a specific drilling area. This area contains 117 wells. The well logging curves from these 117 wells include: natural gamma ray, sonic transit time, neutron porosity, density, and resistivity logging curves. This example selects natural gamma ray, density, neutron porosity, and sonic transit time curves for method testing. Of the 117 wells, 102 were selected as the training set, 10 of the remaining 15 were used as the validation set, and the other 5 were used as blind wells for testing. During training, an adaptive learning rate was used, and the network was trained for 1500 epochs. The validation set loss function converged on the 1000th training epoch.

[0154] The effectiveness of the proposed method (FlexLogNet) was verified by comparing its prediction accuracy with other methods (Bidirectional Long Short-Term Memory (BiLSTM) and Garnder) on five blind wells. BiLSTM requires a fixed type of logging curve as input to the network and can only predict fixed missing logging curves. Therefore, for comparison, prediction models for natural gamma (GR), density (RHOB), neutron porosity (NPHI), and sonic transit time (DTC) were trained separately. For the proposed method, the network input consists of the natural gamma curve and two other logging curves, with the third selected as the prediction curve. In the quantitative analysis of the prediction results, the proposed model showed the best performance in terms of Pearson correlation coefficient (highest correlation coefficient) and the smallest in terms of root mean square error (Tables 1 and 2), indicating that the proposed method significantly outperforms other methods in predicting missing well data.

[0155] Table 1 Comparison of Pearson correlation coefficients of test wells

[0156]

[0157] Table 2 Comparison of Root Mean Square Error of Test Wells

[0158]

[0159] In blind well testing, the method described in this application achieved a Pearson correlation coefficient of 0.926 and a root mean square error (RMSE) of 4.6614 for sonic transit time prediction. In comparison, the two-way long short-term memory (LSTM) method achieved 0.916 and 5.0022 respectively, while the empirical formula method achieved 0.755 and 21.4728. Taking well A3 as an example, the method described in this application achieved a Pearson correlation coefficient of 0.903 and a RMSE of 3.8774. Regarding the prediction of curve results across the entire well section, as shown... Figure 5 As shown, the prediction results of the method in this application are closer to the actual logging data, with smaller deviations from the actual logging curves, and higher prediction accuracy in thin layers (at the peaks of the curves).

[0160] The method described in this application achieves a Pearson correlation coefficient of 0.845 and a root mean square error of 0.0504 in density curve prediction. In comparison, the two-way long short-term memory method has indices of 0.726 and 0.0725, respectively, while the empirical formula method has indices of 0.821 and 0.1084. Taking well A0 as an example, the method described in this application achieves a Pearson correlation coefficient of 0.954 and a root mean square error of 0.0619. Regarding the prediction of the entire well section curve, as shown... Figure 6 As shown, the prediction results of the method in this application are closer to the actual logging data, and the deviation from the actual logging curve is smaller.

[0161] In predicting neutron porosity curves, the method in this application achieved a Pearson correlation coefficient of 0.885 and a root mean square error of 0.0267. In comparison, the two-way long short-term memory (LSTM) method achieved 0.864 and 0.0286 in these two indicators, respectively. Except for well A0, where the method performed slightly worse than the two-way LSTM method, the method outperformed the LSTM method in the other four blind wells. Taking well A2 as an example, although both the method in this patent and the LSTM method can capture the vertical variation characteristics of the curve, in thinner, more developed layers, such as... Figure 7 As indicated by the arrow, the method in this application can match real well logging data very well.

[0162] Therefore, as can be seen from the above comparison, this application performs excellently in predicting missing logging curves and is unaffected by the type of missing logging curve.

[0163] This application also provides a method for completing well logging curves based on deep learning, such as... Figure 8 As shown, the deep learning-based well logging curve completion method includes steps 810 and 820.

[0164] Step 810: Obtain the known logging curves of the target well;

[0165] Step 820: Select the type of missing logging curve and input the known logging curves into the preset logging curve completion model to obtain the missing logging curves output by the logging curve completion model;

[0166] The well logging curve completion model was trained using the deep learning-based well logging curve completion model training method described in this application.

[0167] According to the deep learning-based well logging curve completion method of this application, the well logging curve completion model used in the embodiments of this application uses a heterogeneous graph neural network to infer the relationship between well logging curves of multiple wells and then predict the normalized well logging curve. A fully connected neural network is used to denormalize the predicted normalized well logging curve. The model is trained in a hybrid deep learning manner that combines heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network output, adaptively realize the completion of different well logging curves, and improve the completion efficiency of missing well logging curves.

[0168] The deep learning-based well logging curve completion model training method provided in this application can be implemented by a deep learning-based well logging curve completion model training device. This application uses the deep learning-based well logging curve completion model training device executing the deep learning-based well logging curve completion model training method as an example to illustrate the deep learning-based well logging curve completion model training device provided in this application.

[0169] This application also provides a training device for a well logging curve completion model based on deep learning.

[0170] like Figure 9 As shown, the training device for the deep learning-based well logging curve completion model includes:

[0171] The input module 910 is used to input the logging curves in the sample curve set into the preset initial model to obtain the normalized features corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model, and the statistical information of the sampling points on the missing logging curves output by the fully connected neural network of the initial model.

[0172] The conversion module 920 is used to perform denormalization on the normalized features based on statistical information in order to convert them into the corresponding missing logging curves.

[0173] Training module 930 is used to train the initial model based on missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, thereby obtaining a logging curve completion model.

[0174] According to the deep learning-based well logging curve completion model training device of this application, this application uses a heterogeneous graph neural network to infer the relationship between well logging curves of multiple wells and then predict the normalized well logging curve. A fully connected neural network is used to denormalize the predicted normalized well logging curve. The model is trained by a hybrid deep learning method that combines heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network output, adaptively realize the completion of different well logging curves, and improve the completion efficiency of missing well logging curves.

[0175] In some embodiments, the input module 910 is further configured to:

[0176] Obtain a set of logging curves corresponding to the target well; a set of logging curves includes multiple logging curves; each logging curve has a label indicating its shape;

[0177] At least one label is randomly masked to obtain a set of well logging curves after masking;

[0178] The masked set of logging curves and the corresponding labels are combined to form a training sample pair;

[0179] A sample curve set is constructed based on training sample pairs corresponding to multiple target wells.

[0180] In this embodiment, a sample curve set is constructed by randomly masking the logging curves. During model training, the loss function is calculated by predicting the randomly masked logging curves and comparing the predicted values ​​with the true values ​​corresponding to the labels. Furthermore, since the number of complete and unmissing logging curves is relatively small, this embodiment uses random masking of the logging curves to generate multiple pairs of training samples for the same set of logging curves, thereby increasing the amount of training data.

[0181] In some embodiments, the set of logging curves may also include missing logging curves; the input module 810 is further configured to:

[0182] At least one label is randomly masked to obtain a set of masked logging curves, including:

[0183] The labels corresponding to the missing logging curves are masked in their original form, and the labels corresponding to at least one non-missing logging curve are masked randomly, resulting in a set of masked logging curves.

[0184] In some embodiments, the input module 910 is further configured to:

[0185] The sampling point value at the mask is obtained by multiplying the randomly masked logging curve with the corresponding label;

[0186] Based on the sampling point values, a set of well logging curves after masking is sampled at different magnifications to obtain a set of well logging curves with different sampling rates;

[0187] A set of well logging curves with different sampling rates and their corresponding labels are combined to form a training sample pair.

[0188] In some embodiments, the input module 910 is further configured to:

[0189] Extracting first-node features corresponding to different categories of logging curves based on different encoders and decoders;

[0190] Calculate the characteristic relationships of logging curves between different categories based on the characteristics of the first node;

[0191] Predict the second node features corresponding to the missing logging curves based on feature relationships;

[0192] Output the second node feature, which represents the normalized feature corresponding to the missing logging curve.

[0193] In some embodiments, the input module 910 is further configured to:

[0194] Based on the aggregation module, the characteristic relationships between logging curves of different categories are calculated. The aggregation module is represented by the following formula:

[0195]

[0196] The right side of the formula represents the characteristic relationship between well logging curves of different categories. Represents the learned weights of the convolution operation. Represents a node i The set of neighboring nodes, This represents the first node characteristic of the neighboring nodes. Represents a node i and nodes j Attention coefficient between them Represents a node i The new feature is the second node feature corresponding to the missing logging curve.

[0197] In some embodiments, the input module 910 is further configured to:

[0198] Based on the aggregation module, the characteristic relationships between logging curves of different categories are calculated. The aggregation module is represented by the following formula:

[0199]

[0200] in, This represents a learnable matrix, used to enhance attention.

[0201] In some embodiments, the training module 930 is further configured to:

[0202] The overall loss function of the heterogeneous graph neural network and the fully connected neural network is calculated based on the difference between the predicted missing logging curves and the actual values ​​corresponding to the label values.

[0203] The parameters of the heterogeneous graph neural network and the fully connected neural network are updated based on the overall loss function to obtain the well logging curve completion model.

[0204] In some embodiments, the training module 930 is further configured to:

[0205] The overall loss function is calculated using the following formula:

[0206]

[0207]

[0208] in, L Represents the overall loss function. L F This represents the loss function of a fully connected neural network. w F express L F The weight, L G This represents the loss function of a heterogeneous graphical neural network. w G express L G The weight, L MSE and L SSIM These represent mean squared error and structural similarity index, respectively. w 1 and w 2 represents the weighting coefficient. m i 0 and m i d Represents the original mask and the random mask ( i =0, 1, 2, 3…), y i ( i =0, 1, 2, 3…) and i ( i =0, 1, 2, 3…) represent the label and the predicted logging curve value, respectively.

[0209] The well logging curve completion method based on deep learning provided in this application can be executed by a well logging curve completion device based on deep learning. This application uses the execution of the well logging curve completion method based on deep learning by a well logging curve completion device as an example to illustrate the well logging curve completion device based on deep learning provided in this application.

[0210] This application also provides a deep learning-based well logging curve completion device.

[0211] like Figure 10 As shown, the deep learning-based well logging curve completion device includes:

[0212] Module 1010 is used to acquire known logging curves of the target well.

[0213] The prediction module 1020 is used to select the type of missing logging curve and input the known logging curves into the preset logging curve completion model to obtain the missing logging curves predicted by the logging curve completion model.

[0214] The well logging curve completion model was trained using the deep learning-based well logging curve completion model training method described in this application.

[0215] According to the deep learning-based well logging curve completion device of this application, the well logging curve completion model used in this application uses a heterogeneous graph neural network to infer the relationship between well logging curves of multiple wells and then predict the normalized well logging curve. A fully connected neural network is used to denormalize the predicted normalized well logging curve. The model is trained by a hybrid deep learning method that combines heterogeneous graph neural networks and fully connected neural networks to ensure the consistency of network output, adaptively realize the completion of different well logging curves, and improve the completion efficiency of missing well logging curves.

[0216] The deep learning-based well logging curve completion model training device or deep learning-based well logging curve completion device in the embodiments of this application can be an electronic device or a component in an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal or other devices besides a terminal. For example, the electronic device can be a mobile phone, tablet computer, laptop computer, handheld computer, in-vehicle electronic device, mobile internet device (MID), augmented reality (AR) / virtual reality (VR) device, robot, wearable device, ultra-mobile personal computer (UMPC), netbook or personal digital assistant (PDA), etc. It can also be a server, network attached storage (NAS), personal computer (PC), television (TV), ATM or self-service machine, etc. The embodiments of this application do not specifically limit it.

[0217] The deep learning-based well logging curve completion model training device or the deep learning-based well logging curve completion device in the embodiments of this application can be a device with an operating system. This operating system can be a Microsoft (Windows) operating system, an Android operating system, an iOS operating system, or other possible operating systems; this application embodiment does not specifically limit it.

[0218] In some embodiments, such as Figure 11 As shown, this application embodiment also provides an electronic device 1100, including a processor 1101, a memory 1102, and a computer program stored in the memory 1102 and executable on the processor 1101. When the program is executed by the processor 1101, it implements the various processes of the above-described deep learning-based well logging curve completion model training method or deep learning-based well logging curve completion method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0219] It should be noted that the electronic devices in the embodiments of this application include the aforementioned mobile electronic devices and non-mobile electronic devices.

[0220] This application also provides a non-transitory computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the various processes of the above-described deep learning-based well logging curve completion model training method or deep learning-based well logging curve completion method embodiments, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0221] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0222] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described deep learning-based well logging curve completion model training method or deep learning-based well logging curve completion method.

[0223] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0224] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described deep learning-based well logging curve completion model training method or deep learning-based well logging curve completion method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0225] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0226] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0227] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0228] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

[0229] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0230] Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

Claims

1. A method for training a well logging curve completion model based on deep learning, characterized in that, include: The well logging curves in the sample curve set are input into a preset initial model to obtain the normalized features corresponding to the missing well logging curves output by the heterogeneous graph neural network of the initial model, and the statistical information of the sampling points on the missing well logging curves output by the fully connected neural network of the initial model. The normalized features are denormalized based on the statistical information to convert them into corresponding missing logging curves. The initial model is trained based on the missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, thereby obtaining a logging curve completion model. Obtain a set of logging curves corresponding to the target well; the set of logging curves includes multiple logging curves; each logging curve has a label indicating its shape; At least one label is randomly masked to obtain a set of well logging curves after masking; The masked set of logging curves and their corresponding labels are combined to form a training sample pair; the sample curve set is constructed based on the training sample pairs corresponding to multiple target wells.

2. The method according to claim 1, characterized in that, The set of logging curves also includes missing logging curves; The step of randomly masking at least one label to obtain a set of masked logging curves includes: The labels corresponding to the missing logging curves are masked in their original form, and the labels corresponding to at least one non-missing logging curve are masked randomly, resulting in a set of masked logging curves.

3. The method according to claim 1, characterized in that, The step of combining a masked set of logging curves and their corresponding labels to form a training sample pair includes: The sampling point value at the mask is obtained by multiplying the randomly masked logging curve with the corresponding label; Based on the sampling point values, a set of well logging curves after masking is sampled at different magnifications to obtain a set of well logging curves with different sampling rates; The set of logging curves with different sampling rates and their corresponding labels are combined to form a training sample pair.

4. The method according to claim 1, characterized in that, The heterogeneous graph neural network outputs the normalized features corresponding to the missing logging curves in the following manner: Extracting first-node features corresponding to different categories of logging curves based on different encoders and decoders; Calculate the characteristic relationship between logging curves of different categories based on the characteristics of the first node; Predict the second node features corresponding to the missing logging curves based on the aforementioned feature relationships; Output the second node feature, which represents the normalized feature corresponding to the missing logging curve.

5. The method according to claim 4, characterized in that, The step of calculating the characteristic relationship between logging curves of different categories based on the characteristics of the first node includes: The aggregation module calculates the characteristic relationships between logging curves of different categories, and the aggregation module is represented by the following formula: The right side of the formula represents the characteristic relationship between well logging curves of different categories. Let N(i) represent the learned weights of the convolution operation, and let N(i) represent the set of neighboring nodes of node i. This represents the first node characteristic of the neighboring nodes. This represents the attention coefficient between node i and node j. This represents the new feature of node i, namely the second node feature corresponding to the missing logging curve.

6. The method according to claim 5, characterized in that, The step of calculating the characteristic relationship between logging curves of different categories based on the characteristics of the first node includes: Based on the aggregation module, the characteristic relationships between logging curves of different categories are calculated. The aggregation module is represented by the following formula: in, This represents a learnable matrix, used to enhance attention.

7. The method according to claim 1, characterized in that, The step of training the initial model based on the missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, thereby obtaining a logging curve completion model, includes: The overall loss function of the heterogeneous graph neural network and the fully connected neural network is calculated based on the difference between the predicted missing logging curve and the actual value corresponding to the label value. The parameters of the heterogeneous graph neural network and the fully connected neural network are updated based on the overall loss function to obtain the well logging curve completion model.

8. The method according to claim 7, characterized in that, The overall loss function is calculated using the following formula: Where L represents the overall loss function, This represents the loss function of a fully connected neural network. express The weight, This represents the loss function of a heterogeneous graphical neural network. express The weight, and These represent mean squared error and structural similarity index, respectively. and These are the weighting coefficients. and Represents the original mask and the random mask. and These represent the label and the predicted logging curve value, respectively. .

9. A method for completing well logging curves based on deep learning, characterized in that, include: Obtain the known logging curves of the target well; Select the type of missing logging curve, and input the known logging curve into the preset logging curve completion model to obtain the missing logging curve output by the logging curve completion model; The well logging curve completion model is trained using the method described in any one of claims 1-8.

10. A training device for a well logging curve completion model based on deep learning, characterized in that, include: The input module is used to input the logging curves from the sample curve set into a preset initial model to obtain the normalized features corresponding to the missing logging curves output by the heterogeneous graph neural network of the initial model, and the statistical information of the sampling points on the missing logging curves output by the fully connected neural network of the initial model. The conversion module is used to perform an inverse normalization operation on the normalized features based on the statistical information, so as to convert them into the corresponding missing logging curves; The training module is used to train the initial model based on the missing logging curves to update the parameters of the heterogeneous graph neural network and the fully connected neural network, thereby obtaining a logging curve completion model. The input module is also used to obtain a set of logging curves corresponding to the target well; the set of logging curves includes multiple logging curves; the logging curves are labeled with labels representing the shape of the logging curves; at least one label is randomly masked to obtain a set of masked logging curves; the masked set of logging curves and the corresponding labels are combined to form a pair of training samples; the sample curve set is constructed based on the training sample pairs corresponding to multiple target wells.

11. A well logging curve completion device based on deep learning, characterized in that, include: The acquisition module is used to acquire known logging curves of the target well. The prediction module is used to select the type of missing logging curve and input the known logging curve into a preset logging curve completion model to obtain the missing logging curve predicted by the logging curve completion model. The well logging curve completion model is trained using the method described in any one of claims 1-8.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-9.

13. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-9.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-9.

Images