Methods and apparatus for estimating s-wave velocity by learning from well logs

Abstract

This disclosure relates to a method and apparatus for estimating S-wave velocities by learning from well logging records, wherein the method includes: a model formation step of forming an S-wave estimation model for outputting S-wave velocities corresponding to the measurement depth when well logging records are input based on a training dataset including training data and labeled data having S-wave velocities corresponding to the measurement depth as a result, the training data having values ​​of a plurality of factors included in the well logging records, the values ​​being set to correspond to the measurement depth; and an S-wave velocity estimation step of inputting invisible data into the S-wave estimation model to estimate the S-wave velocity corresponding to the measurement depth, the invisible data having values ​​of a plurality of factors included in the well logging records obtained from the well from which the S-wave velocity is to be estimated, the values ​​being set to correspond to the measurement depth.

Description

[0001] Cross-reference to related applications

[0002] This application claims priority to Korean Patent Application No. 10-2020-0067930, filed on June 4, 2020, the entire contents of which are incorporated herein by reference for all purposes. Technical Field

[0003] This invention relates to a method and apparatus for estimating S-wave velocity by learning well logging records. Background Technology

[0004] Various resources such as coal, oil, natural gas, and minerals exist underground. To explore the possibility of these underground natural resources, drilling processes are performed in the formation to directly examine the formation. When drilling is performed, well logs are obtained, which are records of rock properties acquired during the drilling process in the formation. Shear wave (S-wave) velocity is one of the key elements considered in well logs when estimating subsurface physical properties. However, it is possible that S-wave velocity may not be measurable during the process of obtaining well logs.

[0005] In the absence of S-wave velocity in well logging records, other factors in the logging record can be analyzed to estimate the S-wave velocity in the formation. Traditionally, a method has been used where a small number of petrologists analyze well logging records and estimate S-wave velocity based on their experience. Manual velocity estimation by domain experts requires painstaking analysis of large amounts of diverse data, incurring high costs and time. Furthermore, high accuracy cannot be guaranteed, and results may even differ depending on the person performing the analysis.

[0006] [Related Technical Documents]

[0007] [Patent Documents]

[0008] (Patent Document 1) CN 103424772 A Summary of the Invention

[0009] The purpose of this invention is to provide a method and apparatus for estimating S-wave velocity using an artificial intelligence model that has been learned from well logging records.

[0010] According to one aspect of the present invention, the above and other objectives can be achieved by providing a method for estimating S-wave velocities by learning well logging records, the method comprising:

[0011] The model formation step involves forming an S-wave estimation model to output the S-wave velocity corresponding to the measurement depth when a training dataset, including training data and labeled data with S-wave velocities corresponding to the measurement depth, is input to the well log. This training data contains values ​​for multiple factors included in the well log, set to correspond to the measurement depth.

[0012] The S-wave velocity estimation step involves inputting unseen data into the S-wave estimation model to estimate the S-wave velocity corresponding to the measurement depth. This unseen data contains values ​​of multiple factors included in the logging records obtained from the well from which the S-wave velocity is to be estimated, and these values ​​are set to correspond to the measurement depth.

[0013] The model formation steps may include:

[0014] The training dataset generation step generates a training dataset including training data and, as a result, labeled data of S-wave velocities at the target measurement depth. This training data contains measurements of multiple factors included in the well logging record, based on the target measurement depth, measurement depths shallower than the target measurement depth, and measurement depths deeper than the target measurement depth. These measurements are configured as a two-dimensional matrix structure.

[0015] The model training step involves training an S-wave estimation model using a multi-point convolutional model structure. This multi-point convolutional model structure uses the training dataset to output the S-wave velocity at the target measurement depth, the S-wave velocity at a measurement depth shallower than the target measurement depth, and the S-wave velocity at a measurement depth deeper than the target measurement depth for each measurement depth.

[0016] The S-wave velocity estimation steps may include:

[0017] The invisible data generation step generates invisible data, which includes measurement values ​​of multiple factors from well logs based on the target measurement depth, measurement depths shallower than the target measurement depth, and measurement depths deeper than the target measurement depth. These measurement values ​​are set as a two-dimensional matrix structure based on well logs obtained from the well from which the S-wave velocity to be estimated.

[0018] The model uses the following steps: outputting the S-wave velocity at each measurement depth corresponding to the measurement depth included in the training data of the training dataset, as the output of the result of inputting invisible data into the S-wave estimation model, and determining the S-wave velocity at the measurement depth corresponding to the target measurement depth as the estimated S-wave velocity.

[0019] The model formation steps may include:

[0020] The training dataset generation step generates a training dataset that includes training data and, as a result, labeled data with S-wave velocities corresponding to the measurement depth. This training dataset contains values ​​for multiple factors included in the well logging record, which are set to correspond to the measurement depth. Different sampling techniques can be applied to the training dataset, thus making the generated training datasets distinct from each other.

[0021] The model training steps involve training an S-wave estimation model to output the S-wave velocity corresponding to the measured depth when input well logging records. This can be achieved by using multiple training datasets, each with different structures, to train multiple S-wave estimation models that differ from the training datasets in at least one structure.

[0022] The model selection step involves evaluating the performance of multiple S-wave estimation models that differ from the training dataset in at least one structure and selecting the S-wave estimation model with the highest performance.

[0023] The steps for generating the training dataset may include:

[0024] Generate multiple training datasets, each consisting of multiple well logging records, by performing at least one of the following:

[0025] Optimal ratio sampling is used to generate multiple training datasets at various ratios to determine the optimal ratio of data in well logging records to be used as training data and data to be used as test data.

[0026] Consistent lithofacies sampling is used to select data that ensures consistent lithofacies rates across well log records included in the training dataset.

[0027] Random repeated sampling is used to randomly extract data from one or more well logs, wherein it can be determined whether each lithofacies included in the final extracted data is present at a rate greater than a predetermined rate, and if the included predetermined lithofacies are less than a predetermined rate, the data can be extracted repeatedly;

[0028] Similarity pattern sampling is used to extract logging records on a well-by-well basis to generate a training dataset whose patterns resemble the patterns of values ​​for specific factors obtained from logging records of wells from which the S-wave velocity to be estimated.

[0029] Class sampling is used to select logging records obtained from wells to generate a training dataset, where the well belongs to a class of wells predicted to have formations similar to those of the well whose S-wave velocity is to be estimated; or

[0030] Depth factor sampling is used to selectively choose the range of measurement depths and the number and types of factors included in the training dataset configured as a two-dimensional matrix structure.

[0031] The model usage steps may further include: an integration process that determines a final S-wave velocity by synthesizing an S-wave velocity corresponding to a measurement depth corresponding to a target measurement depth, the final S-wave velocity being presented in each of a plurality of predicted labels output as a result of inputting all invisible data into the S-wave estimation model, all invisible data including the measurement depth corresponding to the target measurement depth.

[0032] According to another aspect of the present invention, an apparatus for estimating S-wave velocity by learning well logging records is provided, the apparatus comprising:

[0033] The logging database (DB) is configured to store logging records and S-wave velocities corresponding to the measurement depth, which are data obtained through measurement and analysis after drilling in the formation;

[0034] The training dataset generation unit is configured to generate a training dataset that includes training data and, as a result, labeled data with S-wave velocities corresponding to the measurement depth. The training data has values ​​for multiple factors in the logging records stored in the logging record DB, which are set to correspond to the measurement depth.

[0035] The model training unit is configured to train an S-wave estimation model to output an S-wave velocity corresponding to the measured depth when inputting well log records using a training dataset; and

[0036] The S-wave velocity estimation unit is configured to input invisible data into an S-wave estimation model trained by a model training unit to estimate the S-wave velocity corresponding to the measurement depth. The invisible data has values ​​of multiple factors included in the logging records obtained from the well from which the S-wave velocity is to be estimated, and these values ​​are set to correspond to the measurement depth.

[0037] The training dataset and the unseen data can be measurements of multiple factors included in the well log, based on the target measurement depth, measurement depths shallower than the target measurement depth, and measurement depths deeper than the target measurement depth. These measurements are set as a two-dimensional matrix structure, based on the well log obtained from the well where the S-wave velocity to be estimated. The S-wave estimation model can have a multi-point convolutional model structure configured to output the S-wave velocity at the target measurement depth, the S-wave velocity at the measurement depth shallower than the target measurement depth, and the S-wave velocity at the measurement depth deeper than the target measurement depth for each measurement depth using the training dataset.

[0038] The training dataset generation unit can generate a training dataset including training data and, as a result, labeled data with S-wave velocities corresponding to the measurement depth. This training dataset has values ​​for multiple factors included in the well logging record, set to correspond to the measurement depth. Different sampling techniques can be applied to the training dataset, making the generated training datasets distinct from each other. The model training unit can train an S-wave estimation model to output S-wave velocities corresponding to the measurement depth when input to the well logging record. This can be achieved by employing different model structures or by using partially different training datasets to uniquely train each of the multiple S-wave estimation models. The device may further include a model selection unit configured to evaluate the performance of multiple S-wave estimation models that differ from the training dataset in at least one structure and select the S-wave estimation model with the highest performance.

[0039] The training dataset generation unit can generate multiple training datasets, each containing multiple well log records, that are at least partially different from each other by performing at least one of the following:

[0040] Optimal rate sampling is used to generate multiple training datasets at various rates to determine the optimal ratio of data in well logging records to be used as training data and data to be used as test data.

[0041] Uniform lithofacies sampling is used to select data that ensures consistent lithofacies facies among the well log records included in the training dataset.

[0042] Random repetitive sampling is used to randomly extract data from one or more well logs, where it can be determined whether each lithofacies included in the final extracted data is present at a rate greater than a predetermined rate, and data can be extracted repeatedly if the included predetermined lithofacies are less than a predetermined rate.

[0043] Similar pattern sampling is used to extract well logs on a well-by-well basis to generate a training dataset whose patterns resemble the patterns of values ​​for specific factors obtained from well logs of the well from which the S-wave velocity is to be estimated.

[0044] Cluster sampling is used to select logging records from wells to generate a training dataset, where the well belongs to a class of wells predicted to have formations similar to those of the well whose S-wave velocity is to be estimated; or

[0045] Depth factor sampling is used to selectively choose the range of measurement depths and the number and types of factors included in the training dataset configured as a two-dimensional matrix structure.

[0046] The S-wave velocity estimation unit can further perform an integrated process to determine the final S-wave velocity:

[0047] By inputting invisible data into an S-wave estimation model trained by a model training unit, the S-wave velocity corresponding to the measurement depth is estimated. This invisible data contains values ​​of multiple factors included in the well logging records obtained from the well from which the S-wave velocity is to be estimated, and these values ​​are set to correspond to the measurement depth.

[0048] By synthesizing the S-wave velocity corresponding to the measurement depth corresponding to the target measurement depth in each of the multiple prediction labels output as a result of inputting all invisible data into the S-wave estimation model, all invisible data includes the measurement depth corresponding to the target measurement depth.

[0049] The features and advantages of the present invention will become clearer from the following detailed description taken in conjunction with the accompanying drawings.

[0050] It should be understood that the terms used in the specification and appended claims should not be construed as limited to their general or dictionary meanings, but should be interpreted in accordance with the meaning and concept of the invention, based on the principle that the inventors are allowed to define appropriate terms for the best interpretation. Attached Figure Description

[0051] The above and other objects, features, and other advantages of the present invention will become clearer from the following detailed description taken in conjunction with the accompanying drawings, wherein:

[0052] Figure 1 This is a block diagram illustrating an apparatus for estimating S-wave velocity by learning well logging records according to an embodiment of the present invention;

[0053] Figure 2 This is a view illustrating example data stored in a well logging database according to an embodiment of the present invention;

[0054] Figure 3 This is a flowchart illustrating a method for estimating S-wave velocity by learning well logging records according to an embodiment of the present invention;

[0055] Figure 4 This is a view showing an example training dataset according to an embodiment of the present invention;

[0056] Figure 5 This is a view illustrating an S-wave estimation model with a multi-point convolutional model structure according to an embodiment of the present invention;

[0057] Figure 6 This is a view showing example unseen data and predicted labels based on an S-wave estimation model with a multi-point convolutional model structure according to an embodiment of the present invention;

[0058] Figure 7 This is a flowchart illustrating a model formation step that further includes a model selection step according to an embodiment of the present invention;

[0059] Figure 8 This is a view illustrating the integration process of an S-wave estimation model with a multi-point convolutional model structure according to an embodiment of the present invention; and

[0060] Figure 9 This is a visualization showing the input and output of an S-wave estimation model according to an embodiment of the present invention. Detailed Implementation

[0061] The objects, advantages, and features of the present invention will become apparent from the detailed description of the embodiments with reference to the accompanying drawings. It should be noted that when assigning reference numerals to elements in the drawings, the same reference numerals are assigned to the same elements even if the same elements are shown in different drawings. Furthermore, the terms "first," "second," etc., are used to describe various elements without regard to order and / or importance, and to distinguish one element from another; the elements are not limited by the terminology. When reference numerals are used to denote elements having the terms "first," "second," etc., "-1," "-2," etc., may be added to the reference numerals. In the following description of embodiments of the present invention, detailed descriptions of known technologies incorporated herein will be omitted where such descriptions might obscure the subject matter of the embodiments of the present invention.

[0062] In the following, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

[0063] Figure 1 This is a block diagram illustrating an apparatus 100 for estimating S-wave velocity by learning well logging records according to an embodiment of the present invention.

[0064] Reference Figure 1 According to an embodiment of the present invention, an apparatus 100 for estimating S-wave velocity by learning well logging records may include a well logging record database (DB) 110, a training dataset generation unit 120, a model training unit 130, a model selection unit 14, an error correction unit 150, an S-wave velocity estimation unit 160, an input and output unit 170, and a storage unit 180.

[0065] Figure 2 This is a view showing example data stored in the well logging database 110 according to an embodiment of the present invention.

[0066] Reference Figure 2 The well logging database stores well logging records and S-wave velocities corresponding to the measurement depth. These records are obtained through measurement and analysis after drilling in the formation. The well logging database 110 can store well logging records and information about the well, including the measurement depth, the type of well logging factor, and the value corresponding to the measurement depth of the factor, as well as the lithofacies corresponding to the measurement depth. Each well has its own records. Depending on the drilling company or drilling method, well logging records can have various types of factors and can be provided in various forms of data.

[0067] The factors recorded in well logging can be obtained through direct measurement during drilling in the formation, or through calculation or analysis. Well logging records include the measured values ​​of each factor corresponding to the measurement depth.

[0068] Factors recorded in well logging can include measurement depth, wellbore diameter, gamma rays, resistivity, bulk density, neutron porosity, photoelectric factor, compression acoustic waves, shear acoustic waves, clay volume, calcite volume, quartz volume, tuff volume, effective porosity, water saturation, bulk modulus, P-wave velocity, and S-wave velocity.

[0069] The well logging DB 110 stores S-wave velocity, which can be directly measured or obtained through analysis of combined measurement depths. Lithofacies indicates the rock type at each measurement depth.

[0070] Lithofacies can include shale, sandstone, coal, calcareous shale, and limestone.

[0071] Well logging records can be stored together with information about the well (from which the logging records have already been obtained). Information about the well can include details such as serial number, name, location, and drilling date. Latitude and longitude can be used to indicate the well's location.

[0072] Re-reference Figure 1 .

[0073] The training dataset generation unit 120 can generate a training dataset TS including training data TD and, as a result, label data LD with S-wave velocities corresponding to the measurement depth. The training data TD has values ​​for multiple factors from well logs stored in the well log database 110, which are set to correspond to the measurement depth. The training dataset generation unit 120 uses data stored in the well log database 110 to generate the training dataset TS necessary for training the S-wave estimation model. Specifically, the training dataset generation unit 120 can sample well logs with known S-wave velocities in various ways to generate the training dataset TS including training data TD and, as a result, label data LD with S-wave velocities corresponding to the measurement depth. The training data TD has values ​​for multiple factors included in the well logs, which are set to correspond to the measurement depth. The training dataset generation unit 120 can sample some data from the data stored in the well log database 110 and set some data from the data stored in the well log database 110 into a structure set based on the type of S-wave estimation model to generate the training dataset TS.

[0074] In the case where the S-wave estimation model is a convolutional neural network structure, the convolutional neural network structure uses the training dataset TS to output the S-wave velocity at the target measurement depth, the measurement depth shallower than the target measurement depth, and the measurement depth deeper than the target measurement depth for each measurement depth. The training dataset TS and the invisible data UD generated by the training dataset generation unit 120 can be the measurement values ​​of multiple factors included in the well logs based on the target measurement depth, the measurement depth shallower than the target measurement depth, and the measurement depth deeper than the target measurement depth. Based on the well logs obtained from the well where the S-wave velocity is to be estimated, the measurement values ​​are set as a two-dimensional matrix structure.

[0075] The training dataset generation unit 120 can generate a training dataset TS that includes training data TD and, as a result, labeled data LD with S-wave velocities corresponding to the measurement depth. The training data TD has values ​​for multiple factors from the well logging record, which are set to correspond to the measurement depth. Different sampling techniques can be applied to the training dataset TS, resulting in training datasets TS that are distinct from each other. The training dataset generation unit 120, which samples data from the well logging record to generate the training dataset TS, will be described in detail below.

[0076] The training dataset generation unit 120 can generate the test data necessary to evaluate the performance of the S-wave estimation model. The training dataset generation unit 120 can generate test data using well logging records not included in the training dataset TS. The test data includes the same input and label data as the training dataset TS, and is not used to train the S-wave estimation model, but rather for the performance evaluation of the S-wave estimation model.

[0077] The training dataset generation unit 120 can generate invisible data UD, which has values ​​for multiple factors included in the logging records obtained from the well where the S-wave velocity to be estimated, and these values ​​are set to correspond to the measurement depth. The invisible data UD can be generated with the same structure as the training data TD of the training dataset TS, which is used to train the selected S-wave estimation model.

[0078] Model training unit 130 uses training dataset TS to train the S-wave estimation model. When well logs are input using the training dataset TS generated by training dataset generation unit 120, model training unit 130 trains the S-wave estimation model to output the S-wave velocity corresponding to the measurement depth. Model training unit 130 can input the label data LD of training dataset TS and compare the predicted label PL output from the S-wave estimation model with the label data LD of training dataset TS to repeatedly train the S-wave estimation model.

[0079] When well logging records are input, model training unit 130 trains an S-wave estimation model to output an S-wave velocity corresponding to the measured depth. Model training unit 130 can train multiple S-wave estimation models, either by employing different model structures or by uniquely training each of the multiple S-wave estimation models using partially different training datasets TS. S-wave estimation models may include linear regression, multinomial regression, single-point dense model (SPDM), single-point convolutional model (SPCM), multi-point dense model (MPDM), multi-point convolutional model (MPCM), and other artificial intelligence models. Model training unit 130 can train an S-wave estimation model for each of the training datasets TS generated by training dataset generation unit 120 to train different S-wave estimation models.

[0080] Model selection unit 140 evaluates the performance of the S-wave estimation model trained by model training unit 130 and selects the model with the highest performance. Model selection unit 140 can evaluate the performance of S-wave estimation models that differ from the training dataset in at least one structure and can select the S-wave estimation model with the highest performance. Model selection unit 140 can use evaluation methods such as MAE or MAPE to evaluate the performance of the S-wave estimation model. Model selection unit 140 can support the evaluation of model performance by visualizing the performance of the S-wave estimation model using graphs, thereby comparing the estimated S-wave velocity with the true S-wave velocity to determine how close the estimated velocity is to the result.

[0081] Error correction unit 150 can perform an error correction step on the estimated S-wave velocity derived from the S-wave estimation model. In the error correction step, error correction unit 150 can correct the offset between the estimated S-wave velocity and the true S-wave velocity, or it can perform statistical correction if the following conditions (Mathematical Formula 1 below) are not met.

[0082] [Mathematical Equation 1]

[0083]

[0084] (MD: range of measured depth values, d: measured depth) The estimated S-wave velocity value at the measurement depth. The true S-wave velocity at the measured depth, and θ: the critical value)

[0085] If, within the range of the measured depth values ​​(i.e., from the minimum to the maximum depth of the well), the difference between the estimated S-wave velocity value and the true S-wave velocity value exceeds a critical value even once, the error correction unit 150 can correct the error in the S-wave velocity estimated by the S-wave estimation model based on statistical data.

[0086] The S-wave velocity estimation unit 160 inputs invisible data UD into an S-wave estimation model trained by the model training unit 130 to estimate the S-wave velocity corresponding to the measurement depth. The invisible data UD has values ​​for multiple factors included in the well logging records obtained from the well where the S-wave velocity is to be estimated, and these values ​​are set to correspond to the measurement depth. The S-wave velocity estimation unit 160 can generate a visualization of the estimated S-wave velocity corresponding to the measurement depth. For example, the S-wave velocity estimation unit 160 can generate the visualization by displaying both the estimated S-wave velocity and the actual S-wave velocity corresponding to the measurement depth as two line graphs.

[0087] S-wave velocity estimation unit 160 can further perform an integrated process to determine the final S-wave velocity: by inputting invisible data UD into an S-wave estimation model trained by model training unit 130 to estimate the S-wave velocity corresponding to the measurement depth, the invisible data UD having values ​​of multiple factors included in the logging records obtained from the well from which the S-wave velocity is to be estimated, the values ​​being set to correspond to the measurement depth; and by synthesizing the S-wave velocity corresponding to the measurement depth corresponding to the target measurement depth in each of a plurality of predicted labels output as a result of inputting all invisible data into the S-wave estimation model, all invisible data including the measurement depth corresponding to the target measurement depth.

[0088] The input and output unit 170 may allow well logging records to be input from an external source, or may output estimation results or learning results to an external source. The input and output unit 170 may include a display capable of visually displaying data, and may further include a communication module for sending and receiving data, a port for sending and receiving data, a touch panel configured to receive user input, and input and output devices such as a keyboard or mouse.

[0089] Storage unit 180 may store program code, the structure of the S-wave estimation model, the trained S-wave estimation model, the error correction algorithm, the error correction result, visualizations and other information, wherein the program code is necessary to execute the method of learning logging records according to an embodiment of the present invention to estimate the S-wave velocity.

[0090] The training dataset generation unit 120, model training unit 130, model selection unit 140, error correction unit 150, and S-wave velocity estimation unit 160 according to embodiments of the present invention can be implemented as program code driven by an information processing device such as a processor, central processing unit (CPU), graphics processing unit (GPU), or neuromorphic chip.

[0091] Figure 3 This is a flowchart illustrating a method for estimating S-wave velocity by learning well logging records according to an embodiment of the present invention.

[0092] Reference Figure 3According to an embodiment of the present invention, a method for estimating S-wave velocity by learning well logging records may include: a model forming step (S10) of forming an S-wave estimation model to output an S-wave velocity corresponding to a measurement depth when a well logging record is input based on a training dataset TS including training data TD and labeled data having S-wave velocities corresponding to the measurement depth as a result, the training data having values ​​of a plurality of factors included in the well logging record, the values ​​being set to correspond to the measurement depth; and an S-wave velocity estimation step (S20) of inputting invisible data US into the S-wave estimation model to estimate the S-wave velocity corresponding to the measurement depth, the invisible data having values ​​of a plurality of factors included in the well logging record obtained from the well from which the S-wave velocity is to be estimated, the values ​​being set to correspond to the measurement depth.

[0093] The model formation step (S10) may include: a training dataset generation step (S11), generating a training dataset TS including training data TD and label data LD with S-wave velocities at the target measurement depth as a result. The training data TD has measurements of multiple factors included in well logs based on the target measurement depth, measurement depths shallower than the target measurement depth, and measurement depths deeper than the target measurement depth, and these measurements are set as a two-dimensional matrix structure; and a model training step (S12), training an S-wave estimation model using a multi-point convolutional model structure (MPCM). This MPCM uses the training dataset TS to output the S-wave velocity at the target measurement depth, the S-wave velocity at the measurement depth shallower than the target measurement depth, and the S-wave velocity at the measurement depth deeper than the target measurement depth for each measurement depth. The estimated S-wave velocity is the result value estimated by the S-wave estimation model as the S-wave velocity at the target measurement depth.

[0094] The training dataset generation unit 120 can perform the training dataset generation step (S11). In the training dataset generation step (S11), the training dataset generation unit 120 uses some of the well logging records stored in the well logging record DB 110 to form a training dataset. In the training dataset generation step (S11), the structure of the training dataset TS can be changed according to the structure of the S-wave estimation model to be trained.

[0095] Figure 4 This is a view showing an example training dataset TS according to an embodiment of the present invention. Figure 4 The sample training dataset structure used to train single-point dense model structures, single-point convolutional model structures, multi-point dense model structures, and multi-point convolutional model structures is shown.

[0096] Reference Figure 4The training dataset SPDM-TS for the single-point dense model (SPDM) can include training data SPDM-TD for the single-point dense model containing the first to third factors at a target measurement depth of 557m, and label data SPDM-LD for the single-point dense model containing S-wave velocity at a target measurement depth of 557m. The single-point dense model can generate a training dataset TS for each target measurement depth.

[0097] The training dataset SPCM-TS for the single-point convolutional model (SPCM) may include: training data SPCM-TD for single-point convolutional models containing the measurement depth at the target measurement depth of 559m and the first factor, the first to third factors, or the third to fifth factors, and label data SPCM-LD for single-point convolutional models containing the S-wave velocity at the target measurement depth of 559m.

[0098] As described above, according to the S-wave estimation model, the training dataset generated in the training dataset generation step (S11) can have different shapes.

[0099] Reference Figure 4 The training dataset for the Multipoint Dense / Convolutional Model (MPDM / MPCM) can include training data for the MPDM-TD / MPCM-TD model with some well logs arranged in a two-dimensional matrix, and label data for the MPDM-LD / MPCM-LD model with S-wave velocities corresponding to the target measurement depth. For example, the training data for the MPDM-TD / MPCM-TD model can have a matrix structure with rows (or columns) including the target measurement depth, measurement depths shallower than the target measurement depth, and measurement depths deeper than the target measurement depth, with the measurement depths located in the first column and the first to fifth factors located in the second to sixth columns, thus giving the well logs a number of rows (or columns) corresponding to the number of factors included in the training dataset TS. Rows and columns can be interchanged, and the positions of the factors can be changed. Preferably, the measurement depths are arranged sequentially. Figure 4 The sample training data for the multi-point dense / convolutional model, MPDM-TD / MPCM-TD, is shown. It has a 5×6 matrix structure and arbitrarily specifies the value of the depth-based factor.

[0100] The measurement depths included in the training data MPDM-TD / MPCM-TD for multi-point dense / convolutional models can include the target measurement depth, measurement depths shallower than the target measurement depth, and measurement depths deeper than the target measurement depth. Three measurement depths (target measurement depth, shallow measurement depth, and deep measurement depth), five measurement depths (one target measurement depth, two shallow measurement depths, and two deep measurement depths), or seven measurement depths (one target measurement depth, three shallow measurement depths, and three deep measurement depths) can be selected. For example, with a target measurement depth of 558m, the number of measurement depths included in the training data MPDM-TD / MPCM-TD for multi-point dense / convolutional models can be five, such as 558m (target measurement depth), 556m and 557m (shallower than the target measurement depth), and 559m and 560m (deeper than the target measurement depth).

[0101] The factors included in the training data MPDM-TD / MPCM-TD for multi-point dense / convolutional models may include measurement depth and other factors. With measurement depth and first to fifth factors selected, six factors are provided, and the values ​​of the first to fifth factors corresponding to the measurement depth are included in the training data MPDM-TD / MPCM-TD for multi-point dense / convolutional models.

[0102] Figure 5 This is a view illustrating an S-wave estimation model with a multi-point convolutional model structure according to an embodiment of the present invention. In this specification and the accompanying drawings, "multi-point convolutional model" may be simply referred to as "multi-point CNN".

[0103] like Figure 5 As shown, the S-wave estimation model according to an embodiment of the present invention can have a multi-point convolutional model structure. In the multi-point CNN structure, the filter (see...) Figure 4 It can have different sizes (1, 3, 5 and 7) and can provide four hidden layers.

[0104] An S-wave estimation model with a multi-point CNN architecture can output the S-wave velocity at a measurement depth corresponding to the measurement depth included in the training data TD of the training dataset TS. For example, in the training data TD of the training dataset TS (see... Figure 4 If the measurement depth is 556m to 560m and the target measurement depth is 558m, the S-wave velocity at the target measurement depth of 558m, the S-wave velocity at the measurement depths of 556m and 557m (shallower than the target measurement depth), and the S-wave velocity at the measurement depths of 559m and 560m (deeper than the target measurement depth) will all be output.

[0105] An S-wave estimation model with a multi-point CNN structure can determine the estimated S-wave velocity at the measurement depth corresponding to the target measurement depth in the predicted label PL. For example... Figure 5 As shown, the S-wave velocity corresponding to the target measurement depth of 558m in the prediction marker PL of the S-wave estimation model can be determined as the estimated S-wave velocity.

[0106] Once the estimated S-wave velocity is determined, the model training unit 130 compares the estimated S-wave velocity with the S-wave velocity of the label data LD. If the estimated S-wave velocity differs from the S-wave velocity of the label data LD, the model training unit 130 can repeat the training of the S-wave estimation model. If the difference between the estimated S-wave velocity and the S-wave velocity of the label data LD is within a predetermined range, the model training unit 130 can determine that training is complete and can stop training.

[0107] Figure 6 This is an exemplary view of the unseen data UD and predicted label PL based on an S-wave estimation model with a multi-point convolutional model structure according to an embodiment of the present invention.

[0108] The S-wave velocity estimation step (S20) can be performed by the S-wave velocity estimation unit 160. The S-wave velocity estimation step (S20) can be performed using an S-wave estimation model trained by the model training unit 130 using a training dataset TS generated by the training dataset generation unit 120. The S-wave velocity estimation step (S20) may include: an invisible data generation step, generating invisible data having measurements of multiple factors included in well logs based on a target measurement depth, measurement depths shallower than the target measurement depth, and measurement depths deeper than the target measurement depth, these measurements being set as a two-dimensional matrix structure based on well logs obtained from the well from which the S-wave velocity is to be estimated; and a model usage step, outputting the S-wave velocity at each measurement depth corresponding to the measurement depth included in the training data TD of the training dataset TS as the output of the result of inputting the invisible data UD into the S-wave estimation model, and determining the S-wave velocity at the measurement depth corresponding to the target measurement depth as the estimated S-wave velocity.

[0109] In the S-wave velocity estimation step (S20), the S-wave velocity can be estimated for each measurement depth to estimate the S-wave velocity at some or all depths of the total well depth, which is included in the logging records obtained from the well where the S-wave velocity estimation is required. In the S-wave velocity estimation step (S20), the invisible data generation step and the model usage step can be repeated for each target measurement depth to estimate the S-wave velocity at some or all measurement depths of the well where the S-wave velocity estimation is necessary.

[0110] For example, when the number of target measurement depths is 5, i.e., the target measurement depth is 558 to 562, such as Figure 6 As shown, invisible data UD with a 5×4 two-dimensional matrix structure, including five measurement depths and one measurement depth plus first and third factors, can be generated in the invisible data generation step. These five measurement depths include a target measurement depth, two measurement depths shallower than the target measurement depth, and two measurement depths deeper than the target measurement depth. Invisible data UD can be generated for each of the five target measurement depths, thus generating the first to fifth invisible data UD-1, UD-2, UD-3, UD-4, and UD-5. Here, the S-wave estimation model for this invisible data UD input is trained using a training dataset TS containing training data TD, which has a 5×4 two-dimensional matrix structure in the same manner as the invisible data UD.

[0111] When the first invisible data UD-1 is input into an S-wave estimation model with a multi-point CNN structure, a predicted label PL can be obtained for the S-wave velocity of each output in the measurement depths included in the first invisible data UD-1, such as... Figure 5 As shown, the S-wave velocity at the measurement depth corresponding to the target measurement depth can be determined as the estimated S-wave velocity. When the first invisible data UD-1 to the fifth invisible data UD-5 can be input into the S-wave estimation model to obtain the first prediction label PL-1 to the fifth prediction label PL-5, and the S-wave velocity at the measurement depth corresponding to the target measurement depth is determined as the estimated S-wave velocity, the following can be obtained: Figure 6 The estimated S-wave velocity results.

[0112] In embodiments of the present invention, training data TD and invisible data UD of the training dataset TS are generated using a two-dimensional matrix structure. These two data are then input into an S-wave estimation model with a multi-point CNN structure. This allows the S-wave estimation model to learn information about the formation at the target measurement depth, as well as information about formations shallower and deeper than the target measurement depth. Therefore, the S-wave estimation model can more accurately estimate the S-wave velocity at the formation corresponding to the target measurement depth.

[0113] Figure 7 This is a flowchart illustrating a model formation step (S10) that further includes a model selection step (S13) according to an embodiment of the present invention.

[0114] The model formation step (S10) may include: a training dataset generation step (S11), generating a training dataset TS comprising training data TD and label data LD with S-wave velocities corresponding to the measurement depth, wherein the training data TD has values ​​of multiple factors included in the well logging record, which are set to correspond to the measurement depth. Different sampling techniques may be applied to the training dataset TS, thereby making the generated training datasets TS distinct from each other; a model training step (S12), training an S-wave estimation model to output S-wave velocities corresponding to the measurement depth when inputting well logging records, wherein multiple training datasets, at least a portion of which are distinct from each other, are used to train S-wave estimation models with various structures, thereby training multiple S-wave estimation models that differ from the training dataset TS in at least one structure; and a model selection step (S13), evaluating the performance of the multiple S-wave estimation models that differ from the training dataset TS in at least one structure and selecting the S-wave estimation model with the highest performance.

[0115] Well logs stored in logging record DB 110 are obtained from wells formed in various formations. These wells are different from each other in terms of factors such as lithofacies type, lithofacies ratio and depth, types of factors measured, patterns of factor values, and well location. In order to accurately estimate the S-wave velocity at any well using logging records obtained from multiple wells with different properties, it is important that the data of the selected logging records are included in the training dataset TS.

[0116] In the training dataset generation step (S11), at least one of the following can be performed: optimal ratio sampling, consistent lithofacies sampling, random repeated sampling, similar pattern sampling, class set sampling, or depth factor sampling, to generate the training dataset TS. Two or more types of sampling can be performed simultaneously to generate the training dataset TS.

[0117] Optimal ratio sampling requires generating multiple training datasets (TS) at various ratios to determine the optimal ratio of data to be used as training datasets (TS) and data to be used as test data in the well logging records. For example, when using well logging records from wells 1 through 4 to generate training datasets (TS), 80% of the data from the well logging records from wells 1 through 4 can be used as training datasets (TS), and 20% of the data can be used as test data. When sampling training datasets (TS) at ratios of 80%, 70%, and 60%, three training datasets (TS) are generated. These three training datasets (TS) can be used to train three S-wave estimation models, and the performance of these three S-wave estimation models can be evaluated to determine the ratio of training datasets (TS) that produces the highest performance.

[0118] Consistent lithofacies sampling requires selecting data that ensures a consistent lithofacies ratio across the logging records included in the training dataset TS. This involves varying the range of S-wave velocities based on the lithofacies type, and also varying the range of factors included in the logging records based on the lithofacies type. In cases of hard and dense lithofacies, the range of S-wave velocities is larger. In cases of soft and porous lithofacies, the range of S-wave velocities is smaller. If a logging record includes five lithofacies A through E, consistent lithofacies sampling can be used to select data such that lithofacies A accounts for 20%, lithofacies B for 20%, lithofacies C for 20%, lithofacies D for 20%, and lithofacies E for 20%. Because specific lithofacies may be abundant while others may be scarce, the lithofacies distribution in logging records obtained from a single well can be inconsistent, depending on the formation and location of the well. When using well logs with inconsistent lithofacies distributions to generate a training dataset TS without any modifications, the estimation accuracy for the S-wave velocity range occurring in lithofacies with high distribution ratios may be high, but the estimation accuracy for the S-wave velocity range occurring in lithofacies with low distribution ratios may be low. However, when the S-wave estimation model is trained using the training dataset TS generated by performing consistent lithofacies sampling, the S-wave estimation model can consistently learn well logs corresponding to various ranges of S-wave velocities. Therefore, the S-wave estimation model can exhibit consistent estimation accuracy across various ranges of S-wave velocities.

[0119] Random repeated sampling involves randomly extracting data from one or more well logs, where it is determined whether each lithofacies included in the final extracted data exists at a predetermined ratio. If the included specific lithofacies are less than the predetermined ratio, data extraction is repeated. In random repeated sampling, the ratio of each lithofacies is a configurable value. In cases where a lithofacies exhibits low S-wave velocity estimation accuracy, the ratio of the lithofacies can be adjusted to be higher, allowing the training dataset TS to include a large number of well logs associated with the specific lithofacies, and enabling the S-wave estimation model to learn from a larger amount of data related to lithofacies that are difficult to estimate. In this case, further learning from well logs associated with a specific range of S-wave velocities can improve the estimation accuracy of that specific range of S-wave velocities.

[0120] Similar pattern sampling requires extracting logging records on a well-by-well basis to generate a training dataset whose patterns resemble the patterns TS of specific factor values ​​from logging records obtained from the well from which the S-wave velocity to be estimated. For example, when the specific factor values ​​from logging records obtained from the well from which the S-wave velocity to be estimated are in the range of 130 to 140, logging records with similar patterns in the range of 130 to 140 or close to this range can be selected on a well-by-well basis from the logging records stored in logging record DB 110 from each well. Alternatively, only logging records with similar patterns can be selected to be included in the training dataset TS. Logging records with specific factor values ​​in the range of 50 to 60 can be excluded from the training dataset TS. In similar pattern sampling, logging records with value ranges similar to those obtained from the well from which the S-wave velocity to be estimated can be used to train the S-wave estimation model, thereby improving the estimation accuracy of the S-wave velocity.

[0121] Cluster sampling involves selecting logging records from wells belonging to a cluster to generate a training dataset TS, which is predicted to have formations similar to those of the wells whose S-wave velocities are to be estimated. In cluster sampling, logging records from a predetermined number of wells in order of proximity to the wells whose S-wave velocities are to be estimated can be selected to generate the training dataset TS. Alternatively, clustering algorithms can be used to classify logging records from the wells whose S-wave velocities are to be estimated, along with logging records stored in logging record DB 110, on a well-by-well basis, and logging records from wells classified into the same cluster can be selected to generate the training dataset TS. Well-known algorithms such as k-means can be used as clustering algorithms. Generally, wells that are close to each other are expected to have similar formation properties. Therefore, performing cluster sampling based on short distances can improve the accuracy of S-wave velocity estimation. However, even neighboring wells may have dissimilar formation properties due to factors such as dislocations between wells. Therefore, in cluster sampling, using clustering algorithms to select wells that typically have similar factor values ​​can improve the estimation accuracy of S-wave velocity.

[0122] Depth factor sampling requires selectively choosing the range of depth to be measured and the number and types of factors included in the training dataset TS, configured as a two-dimensional matrix structure. (See reference...) Figure 4Five measurement depths can be included in the training dataset TS, and a total of six factors, including measurement depths and the first through fifth factors, can be included in the training dataset TS. When performing depth factor sampling, multiple training datasets TS with various combinations can be generated, such as cases where the number of measurement depths is 3, 5, 7, 9, or more, and the number of factors is 2, 3, 4, 5, 6, 7, 9, or more. Additionally, multiple training datasets TS with the same number but different types of factors can be generated. Furthermore, multiple training datasets TS with the same number and types of factors but different sequences can be generated. An S-wave estimation model can be trained using each of the multiple training datasets TS, and the performance of the S-wave estimation model can then be evaluated to determine the number of measurement depths, the number of factors, the types of factors, and the factor sequences that yield the highest performance.

[0123] In the training dataset generation step (S11), at least one of the above-mentioned optimal ratio sampling, consistent lithofacies sampling, random repeated sampling, similar pattern sampling, class set sampling, or depth factor sampling can be performed to generate multiple training datasets TS, at least some of which include multiple additional well logging records. In the training dataset generation step (S11), when generating a training dataset TS, one or more sampling methods can be performed together.

[0124] In the model training step (S12), various training datasets TS generated in the training dataset generation step (S11) can be used to train S-wave estimation models with various structures. The S-wave estimation models that the model training unit 130 can use in the model training step (S12) can include linear regression, multinomial regression, single-point dense model (SPDM), single-point convolutional model (SPCM), multi-point dense / convolutional model (MPDM / MPCM), and other artificial intelligence models. The multiple training datasets TS generated by sampling in the training dataset generation step (S11) are generated such that at least some of the multiple training datasets TS are different from each other. Therefore, even when using the same S-wave estimation model, performance may vary due to the differences in the training datasets TS. Furthermore, even when using the same training dataset TS, performance may change depending on the structure of the S-wave estimation model. In the model training step (S12), the model training unit 130 can use various training datasets TS to train S-wave estimation models with various structures to generate multiple trained S-wave estimation models.

[0125] The model selection step (S13) can be performed by the model selection unit 140. The model selection unit 140 can input test data into multiple S-wave estimation models to evaluate their performance. Well-known evaluation methods such as MAE or MAPE can be used to evaluate performance in the model selection step (S13). In the model selection step (S13), the performance of multiple S-wave estimation models trained using the training dataset TS generated through sampling is evaluated, and the S-wave estimation model with the highest performance is selected. The selected S-wave estimation model can be used in the S-wave velocity estimation step (S20).

[0126] Table 1 below shows the accuracy evaluation results of the S-wave estimation model on a well basis.

[0127] [Table 1]

[0128] Regression analysis Multinomial Regression SPDM SPCM MPCM Well 1 72.9 73.2 95.2 95.8 96.4 Well 2 78.2 54.3 93.0 92.3 96.9 Well 3 65.2 68.5 95.2 97.1 98.3 Well 4 82.5 90.3 89.4 90.1 92.6 Well 5 71.9 62.8 97.4 95.9 96.8 Well 6 87.3 70.1 96.3 95.6 99.2 … … … … … …

[0129] As shown in Table 1, the accuracy ranges from 60 to 90% when the S-wave estimation model has a regression analysis or multinomial regression structure. When the S-wave estimation model has a single-point dense model (SPDM), a single-point convolutional model (SPCM), or a multi-point convolutional model (MPCM), the accuracy is 90% or higher. Therefore, it can be confirmed that the performance of AI-based S-wave estimation models is generally high. The structure with the highest accuracy is the multi-point convolutional model (MPCM), and the model selection unit 140 typically selects the multi-point convolutional model (MPCM).

[0130] In the following text, an S-wave estimation model with a multi-point convolutional model (MPCM) structure that has the highest performance as an S-wave estimation model according to an embodiment of the present invention will be described.

[0131] Figure 8 This is a view illustrating the integration process of an S-wave estimation model with a multi-point convolutional model (MPCM) structure according to an embodiment of the present invention.

[0132] According to embodiments of the present invention, an S-wave estimation model with a multi-point convolutional model (MPCM) structure is as follows: Figure 5 As shown, the model can further include an integration process to synthesize predicted labels (PLs) to further improve accuracy, and this integration process can be further performed in the model usage step. The model usage step can further include: an integration process that determines a final S-wave velocity by synthesizing the S-wave velocity corresponding to the measurement depth corresponding to the target measurement depth in each of a plurality of predicted labels presented as the output of the S-wave estimation model, which is the result of inputting all invisible data into the model. All invisible data includes the measurement depth corresponding to the target measurement depth.

[0133] Reference Figure 8 The ensemble process performed in the model usage step is a process of improving the estimation accuracy of S-wave velocity at the target measured depth. The ensemble process will be described as an example for a target measured depth of 744.4. Assume the unseen data UD has a 5×4 two-dimensional matrix structure, which includes five measured depths and a total of four factors, such as the measured depth and first and third factors. By generating unseen data UD at each measured depth where the S-wave velocity is to be estimated and inputting the unseen data UD into an S-wave estimation model with a multi-point CNN structure, predicted labels PL can be obtained. For example, in the case of a target measured depth of 744.2, unseen data UD is generated in the first part to include two measured depths shallower than the target measured depth and two measured depths deeper than the target measured depth. By inputting the unseen data UD from the first part into the S-wave estimation model, predicted labels PL with S-wave velocities estimated at the measured depths included in the first part are generated. By performing this process for each target depth measurement, invisible data UD can be generated in the first through fifth parts, thereby enabling the acquisition of the predicted label PL.

[0134] Each of the prediction labels PL corresponding to the first through fifth parts includes an S-wave velocity corresponding to the target measurement depth of 744.4. Even at the same measurement depth, the estimated S-wave velocities differ from each other if the portions are different. For example, the prediction label PL for the first part includes an S-wave velocity of 739.229 at the target measurement depth of 744.4, while the prediction label PL for the second part includes an S-wave velocity of 722.493 at the target measurement depth of 744.4. This is because the portions of the invisible data UD input to the S-wave estimation model differ from each other, resulting in different input well logs.

[0135] During the integration process, the S-wave velocity at the measurement depth corresponding to the target measurement depth can be averaged to determine the final S-wave velocity. When the number of measurement depths included in the invisible data UD is five, the number of S-wave velocities to be averaged to determine any target measurement depth in the integration process can also be five. During the integration process, well logs corresponding to a wider range of measurement depths can be synthesized compared to well logs included in the invisible data UD with any measurement depth as the target measurement depth. For example, the invisible data UD with a target measurement depth of 744.4 corresponds to the third part, therefore well logs with measurement depths of 744.0, 744.1, 744.7, and 744.8 are not considered. However, when performing the integration process, well logs with measurement depths of 744.0 to 744.8 corresponding to the first to fifth parts can also be considered.

[0136] As mentioned above, refer to Figure 8 In the case of performing the ensemble process based on a multi-point CNN structure according to an embodiment of the invention, S-wave velocities at a measurement depth corresponding to the target measurement depth can be synthesized to estimate the final S-wave velocity. When the ensemble process is further performed, the estimation accuracy of the S-wave velocity is higher than that of determining the S-wave velocity included in the prediction label PL of a single part.

[0137] Figure 9 This is a visualization showing the input and output of an S-wave estimation model according to an embodiment of the present invention. Figure 9 The left side shows the logging records corresponding to the training data TD or the invisible data UD. Figure 9 The right side shows the predicted label PL or estimated S-wave velocity of the S-wave estimation model and the label data LD compared with the predicted label PL or estimated S-wave velocity.

[0138] like Figure 9 As shown, a visualization of S-wave velocity can be prepared by the S-wave velocity estimation unit 160, and the visualization of S-wave velocity can be provided visually by the input and output unit 170. The S-wave velocity estimation unit 160 can prepare the estimated S-wave velocity or the final S-wave velocity output from the S-wave estimation model as a result of inputting invisible data UD into the S-wave estimation model in graphical form to generate a visualization. The visualization can show at least one of the training data TD, label data LD, test data, or predicted label PL corresponding to the measurement depth, or show at least one of the training data TD, label data LD, test data, or predicted label PL for each well. The model selection unit 140 can generate, for example... Figure 9 The visualization shown allows people to visually identify the accuracy of the S-wave estimation model.

[0139] As described above, in the method and apparatus for estimating S-wave velocity by learning well logging records according to embodiments of the present invention, in order to learn and estimate the S-wave velocity in the formation at the target measurement depth, not only the well logging information measured at the target measurement depth is learned, but also information about well logging records measured at measurement depths shallower than the target measurement depth and information about well logging records measured at measurement depths deeper than the target measurement depth is learned, thereby enabling more accurate measurement of the S-wave velocity at the target measurement depth. To learn information about the formation above / below the target measurement depth, as described above, a training dataset TS with a two-dimensional matrix structure is generated, and an ensemble process is further performed using an S-wave estimation model with a multi-point CNN structure, thereby constructing an optimal S-wave estimation model. This multi-point CNN structure is capable of effectively learning from the training dataset TS with a two-dimensional matrix structure.

[0140] Furthermore, in the method and apparatus for estimating S-wave velocity by learning well logging records according to embodiments of the present invention, multiple training datasets TS are generated using various sampling methods, at least some of which are different from each other. S-wave estimation models with various structures are trained, the performance of multiple S-wave estimation models that differ from the training datasets TS in at least one structure is evaluated, and the S-wave estimation model with the highest performance is selected. Therefore, well logging records obtained from wells from which the S-wave velocity to be estimated can be analyzed effectively, thereby enabling accurate estimation of the S-wave velocity.

[0141] It is evident from the above description that, according to embodiments of the present invention, an artificial intelligence model that has learned from well logging records can be used to accurately and quickly predict S-wave velocities.

[0142] Although the invention has been described in detail with reference to embodiments, and embodiments are provided to further illustrate the invention, the method and apparatus for estimating S-wave velocities by learning well logging records according to the invention are not limited thereto, and those skilled in the art will appreciate that various modifications, additions and substitutions may be made without departing from the scope and spirit of the invention as disclosed in the appended claims.

[0143] It will be understood that the scope and spirit of the invention include simple variations and modifications thereof, and the scope of protection of the invention will be defined by the appended claims.

Claims

1. A method of estimating s-wave velocity by learning a log, the method comprising: a model forming step of forming an s-wave estimation model for outputting s-wave velocity corresponding to a measurement depth when a log is input based on a training data set including training data and, as a result, label data having s-wave velocity at a target measurement depth, the training data having values of a plurality of factors included in the log set to correspond to measurement depths; and an s-wave velocity estimating step of inputting unseen data into the s-wave estimation model to estimate s-wave velocity corresponding to a measurement depth, the unseen data having values of a plurality of factors included in a log acquired from a well for which s-wave velocity is to be estimated, the values being set to correspond to measurement depths, wherein the model forming step includes: a training data set generating step of generating a training data set including training data and, as a result, label data having s-wave velocity at a target measurement depth, the training data having measured values of a plurality of factors included in the log based on a target measurement depth, a measurement depth shallower than the target measurement depth, and a measurement depth deeper than the target measurement depth, the measured values being set to a two-dimensional matrix structure, the two-dimensional matrix structure including: measurement depths in a first column; and one or more factors in subsequent columns, the measurement depths being arranged in odd-numbered rows, the target depth being placed in the center, and the measurement depths being arranged in order, the s-wave estimation model learning information about a formation at the target measurement depth and learning information about a formation shallower than the target measurement depth and a formation deeper than the target measurement depth.

2. The method according to claim 1, wherein the model forming step further includes: a model training step of training the s-wave estimation model with a multi-point convolution model structure that outputs s-wave velocity at the target measurement depth, s-wave velocity at a measurement depth shallower than the target measurement depth, and s-wave velocity at a measurement depth deeper than the target measurement depth for each measurement depth using the training data set.

3. The method according to claim 2, wherein the s-wave velocity estimating step includes: an unseen data generating step of generating unseen data having measured values of the plurality of factors included in the log based on a target measurement depth, a measurement depth shallower than the target measurement depth, and a measurement depth deeper than the target measurement depth, based on the log acquired from the well for which s-wave velocity is to be estimated, the measured values being set to a two-dimensional matrix structure; and a model using step of outputting s-wave velocity for each of the measurement depths corresponding to the measurement depths included in the training data of the training data set as a result of inputting the unseen data into the s-wave estimation model, and determining s-wave velocity at the measurement depth corresponding to the target measurement depth as an estimated s-wave velocity.

4. The method according to claim 1, wherein the model forming step includes: a training data set generation step of generating a training data set including training data and resulting label data having an s-wave velocity corresponding to a measured depth, the training data having values of a plurality of factors included in the log record set to correspond to the measured depth, wherein different sampling techniques can be applied to the training data set, thereby making the generated training data sets different from each other; a model training step of training an s-wave estimation model to output an s-wave velocity corresponding to a measured depth when the log record is input to the s-wave estimation model, wherein the s-wave estimation model having various structures is trained using at least a part of a plurality of training data sets different from each other to train a plurality of s-wave estimation models different in at least one structure from the training data set; and a model selection step of evaluating the performance of a plurality of s-wave estimation models different in at least one structure from the training data set and selecting an s-wave estimation model having the highest performance.

5. The method of claim 4, the training data set generation step comprising: a plurality of training data sets including a plurality of log records are generated by performing at least one of: optimal ratio sampling for generating a plurality of training data sets at various ratios to determine an optimal ratio of data to be used as the training data set and data to be used as test data in the log record; consistent facies sampling for selecting data so that a facies ratio of the log record included in the training data set is consistent, random repeated sampling for randomly extracting data from one or more of the log records, wherein it is determined whether each facies included in finally extracted data exists at more than a predetermined ratio, and in the case where a predetermined facies included is less than the predetermined ratio, data is repeatedly extracted; similar pattern sampling for extracting the log record in a well unit so as to generate the training data set, the pattern of the log record being similar to a pattern of values of a specific factor of the log record taken from the well to be estimated for s-wave velocity, class set sampling for selecting the log record taken from a well to generate the training data set, the well belonging to a class of wells predicted to have a formation similar to the formation of the well to be estimated for s-wave velocity; or depth factor sampling for differentially selecting a range of measured depth and a number and kind of factors included in the training data set configured to have a two-dimensional matrix structure.

6. The method of claim 3, wherein, The model using step further includes an integration process of determining a final s-wave velocity by synthesizing an s-wave velocity corresponding to a measured depth corresponding to the target measured depth present in each of a plurality of predicted labels output as a result of inputting all of the invisible data to the s-wave estimation model, all invisible data including a measured depth corresponding to the target measured depth.

7. An apparatus of estimating an s-wave velocity by learning a log record, the apparatus comprising: a log record database configured to store the log record and s-wave velocities corresponding to measured depths, the log record being data obtained by measurement and analysis after drilling a well in a formation; a training data set generation unit configured to generate a training data set including training data having values of a plurality of factors included in the log record stored in the log record database, the values being set to correspond to measured depths, and label data having s-wave velocities corresponding to measured depths as a result; a model training unit configured to train an s-wave estimation model to output s-wave velocities corresponding to measured depths when the log record is input using the training data set; and an s-wave velocity estimation unit configured to input unseen data having values of a plurality of factors included in a log record acquired from a well in which s-wave velocities are to be estimated, the values being set to correspond to measured depths, into the s-wave estimation model trained by the model training unit, to estimate s-wave velocities corresponding to measured depths, wherein the training data set and the unseen data are based on measured values of a plurality of factors included in log records of a target measured depth, measured depths shallower than the target measured depth, and measured depths deeper than the target measured depth, the measured values being set to a two-dimensional matrix structure based on a log record acquired from the well in which s-wave velocities are to be estimated, the two-dimensional matrix structure including: measured depths in a first column; and one or more factors in subsequent columns, measured depths are arranged in odd rows, a target depth is placed in the center, and measured depths are arranged in order, the s-wave estimation model learns information about a formation at the target measured depth and learns information about formations shallower than the target measured depth and formations deeper than the target measured depth.

8. The apparatus of claim 7, wherein, the s-wave estimation model has a multi-point convolution model structure configured to output, for each measured depth, s-wave velocities at the target measured depth, s-wave velocities at measured depths shallower than the target measured depth, and s-wave velocities at measured depths deeper than the target measured depth using the training data set.

9. The apparatus of claim 8, wherein, the training data set generation unit generates a training data set including training data having values of a plurality of factors included in the log record and label data having s-wave velocities corresponding to measured depths as a result, wherein different sampling techniques can be applied to the training data set, such that the generated training data sets are different from each other, The model training unit trains the s-wave estimation model to output s-wave velocities corresponding to measurement depths when the log records are input into the s-wave estimation model, wherein each of a plurality of s-wave estimation models is uniquely trained by employing different model structures or by using partially different training data sets, and The device further includes a model selection unit configured to evaluate the performance of a plurality of s-wave estimation models that are different in at least one structure from the training data set and select the s-wave estimation model having the highest performance.

10. The device of claim 9, the training data set generation unit generates a plurality of training data sets including a plurality of log records by performing at least one of: optimal ratio sampling for generating the plurality of training data sets at various ratios to determine an optimal ratio of data to be used as the training data set and data to be used as test data in the log records; uniform facies sampling for selecting data so that the facies ratio of the log records included in the training data sets is uniform, random repeated sampling for randomly extracting data from one or more of the log records, wherein it is determined whether each facies included in the finally extracted data exists at a rate greater than a predetermined ratio, and in the case where the included predetermined facies is less than the predetermined ratio, the data is repeatedly extracted; similar pattern sampling for extracting the log records in well units so as to generate the training data sets, the pattern of the log records being similar to the pattern of values of a specific factor of the log records taken from the well for which the s-wave velocity is to be estimated; class set sampling for selecting log records taken from wells to generate the training data sets, the wells belonging to a class of wells predicted to have a formation similar to the formation of the well for which the s-wave velocity is to be estimated; or depth factor sampling for differentially selecting a range of measurement depths and the number and kind of factors included in the training data set configured to have a two-dimensional matrix structure.

11. The device of claim 8, wherein the s-wave velocity estimation unit further performs an integration process of determining the final s-wave velocity: by inputting the invisible data into the s-wave estimation model trained by the model training unit to estimate s-wave velocities corresponding to measurement depths, the invisible data having values of a plurality of factors included in the log records taken from the well for which the s-wave velocity is to be estimated, the values being set to correspond to measurement depths; and by synthesizing the s-wave velocity corresponding to the measurement depth corresponding to the target measurement depth present in each of a plurality of predicted labels output as a result of inputting all invisible data into the s-wave estimation model, all invisible data including the measurement depth corresponding to the target measurement depth.

Images