System for detecting borehole breakouts on basis of machine learning model

Abstract

The present invention relates to a system for detecting borehole breakouts on the basis of a machine learning model and, more particularly, to a system for detecting borehole breakouts on the basis of a machine learning model, the system receiving learning borehole ultrasonic image logging data and inspection borehole ultrasonic image logging data so as to generate a plurality of pieces of learning image data and inspection image data, pre-processing the learning image data and the inspection image data so as to generate gray-scale learning image data and inspection image data that have a set resolution, augmenting the generated gray-scale learning image data, generating a machine learning model trained by the augmented gray-scale learning image data and a data set of coordinate and size information of a bounding box that encompasses borehole breakouts of the learning image data, inputting the inspection image data shifted in parallel by a rolling unit to the trained machine learning model so as to acquire the coordinate and size information (x, y, w, h) of the bounding box that encompasses the borehole breakouts, and providing borehole breakout attribute information (center depth, azimuth angle, opening angle and length).

Description

Compression fracture detection system based on a machine learning model

[0001] The present invention relates to a compression fracture zone detection system based on a machine learning model, and more particularly, to a compression fracture zone detection system based on a machine learning model, which receives training borehole ultrasonic image logging data and inspection borehole ultrasonic image logging data, generates a plurality of training image data and inspection image data, preprocesses the data to generate grayscale training image data and inspection image data having a set resolution, augments the generated grayscale training image data, and generates a machine learning model learned by a data set of coordinate and size information of a bounding box surrounding a compression fracture zone of the augmented grayscale training image data and the compression fracture zone of the training image data, and inputs inspection image data translated in parallel by a rolling unit into the learned machine learning model to obtain coordinate and size information (x, y, w, h) of a bounding box surrounding a compression fracture zone and provides compression fracture zone attribute information (center depth, azimuth, opening angle, and length).

[0002]

[0003] The characteristics of stresses acting on the Earth's crust (magnitude, direction, anisotropy, etc.) are crucial for understanding the stability of underground space development and the mechanisms of earthquakes and fault activity. Recently, shallow crust less than 1 km deep has been developed for various purposes, such as carbon dioxide geological storage and high-level nuclear waste disposal facilities, and thus, the identification of stresses in underground spaces is directly related to stability. Stress fields in shallow crust are measured using various geotechnical techniques, such as hydraulic fracturing, overcoring, fault line inversion, and borehole stress indicators. Among these, borehole stress indicators are widely used because they can obtain stress field information relatively quickly and continuously through compressive or tensile fracture zones observed in the borehole wall. Stress indicators appearing in the borehole wall include drilling-induced tensile fractures and compressive fracture zones (borehole breakouts). These fracture zones occur during the drilling process where stresses around the borehole are relaxed or concentrated, and can provide information on the direction and magnitude of in-situ stresses. Among these, compressive fracture zones occur when stress is concentrated in a direction orthogonal to the direction of maximum horizontal principal stress, and the stress magnitude at that point is greater than the strength of the rock. The occurrence of compressive fracture zones can be determined by analyzing amplitude and travel time data acquired through ultrasonic imaging logging. Due to the nature of principal stress, compressive fracture zones occur at 180° intervals, and at this time, the amplitude is small and the travel time is measured large. Therefore, the direction of maximum horizontal principal stress at each depth can be inferred from the log data.

[0004] Typically, compression fracture zones are manually identified by experts from image logs. This process can result in varying analysis results depending on the expert's skill level, making it difficult to eliminate subjectivity. Furthermore, analysis accuracy can vary depending on the available time and cost. Furthermore, analysis of high-density or deep-depth drilling data, with its large volume, can be time-consuming.

[0005] To address these issues, Dias et al. (2020) developed a model using the deep learning-based Fast-RCNN (fast-region convolution neural network) technique to improve the efficiency of manual image log analysis. This model automatically detects compression fractures and fractures in image logs.

[0006] However, models using the deep learning-based Fast-RCNN technique had the problem that their application was limited when available data was limited.

[0007] [Prior Art Literature]

[0008] [Non-patent literature]

[0009] (Non-patent Document 1) Dias, LO, Bom, CR, Faria, EL, Valentin, MB, Correia, MD, de Albuquerque, MP, de Albuquerque, MP, and Coelho, JM, 2020. Automatic detection of fractures and breakouts patterns in acoustic borehole image logs using fast-region convolutional neural networks, Journal of Petroleum Science and Engineering, 191, 107099.

[0010]

[0011] Accordingly, the present invention has been made in consideration of the above circumstances, and the purpose of the present invention is to provide a compression fracture detection system based on a machine learning model that can create a machine learning model for compression fracture detection even when available learning data is limited, and can improve the detection accuracy of compression fracture even when compression fractures exist at both ends in inspection image data.

[0012]

[0013] In order to achieve the above object, a compression fracture zone detection system based on a machine learning model according to an embodiment of the present invention comprises: an input data generation unit configured to receive training borehole ultrasonic image logging data and inspection borehole ultrasonic image logging data, in which the X-axis represents an angle of 0 to 360 degrees, the Y-axis represents a depth, and color represents an amplitude, and generate a plurality of training image data having a set interval from the top and bottom of the compression fracture zone, and a plurality of inspection image data obtained by dividing the inspection borehole ultrasonic image logging data at a set depth interval; a preprocessing unit configured to preprocess the plurality of training image data and the inspection image data to generate training image data and inspection image data in grayscale having a set resolution; a data augmentation unit configured to receive training image data in grayscale having the set resolution and generate augmented training image data in grayscale; a rolling unit configured to receive the preprocessed inspection image data and to parallel-shift an angle column having a lowest average amplitude value to a first column, and then parallel-shift it to the right; It is characterized by including a compression fracture zone object unit detection model generation unit configured to generate a machine learning model learned by a data set of coordinate and size information of a bounding box surrounding a compression fracture zone of the grayscale training image data augmented by the data augmentation unit, and the augmented grayscale training image data; and a compression fracture zone object unit detection unit configured to input the inspection image data moved in parallel by the rolling unit into the machine learning model generated from the compression fracture zone object unit detection model generation unit, thereby obtaining coordinate and size information of a bounding box surrounding the compression fracture zone.

[0014] The compression fracture detection system based on a machine learning model according to the above embodiment may further include a post-processing unit configured to receive coordinate and size information (x, y, w, h) of a bounding box from the compression fracture object unit detection unit, and to receive parallel-translated inspection image data and parallel-translation information from the rolling unit, and to post-process them to generate compression fracture attribute information for each bounding box and compression fracture attribute information for each pixel.

[0015] The compression fracture detection system based on a machine learning model according to the above embodiment may further include a result display unit configured to display coordinate and size information of a bounding box obtained from the compression fracture object unit detection unit, or compression fracture attribute information for each bounding box generated by the post-processing unit and compression fracture attribute information for each pixel unit.

[0016] In the compression fracture detection system based on the machine learning model according to the above embodiment, the preprocessing unit may include a resolution adjustment unit that receives the plurality of learning image data and inspection image data and adjusts them to have a set resolution; and a grayscale conversion unit that converts the learning image data and inspection image data having the set resolution into a plurality of grayscale learning image data and inspection image data.

[0017] In the compression fracture detection system based on a machine learning model according to the above embodiment, the post-processing unit receives coordinate and size information (x, y, w, h) of a bounding box from the compression fracture object unit detection unit, receives parallel-shifted inspection image data and parallel shift information from the rolling unit, and calculates the center depth, azimuth, opening angle, and length of the compression fracture for each bounding box using the input coordinate and size information of the bounding box, the parallel-shifted inspection image data, and the parallel shift information; a clustering model that classifies the compression fracture image data classified based on the bounding box in the parallel-shifted inspection image data into two types by using a clustering algorithm to assign a cluster having a relatively small average value of a pixel in the image data to 1 and a relatively large cluster to 0; It may include a second compression fracture zone attribute information calculation unit that calculates the median of a compression fracture zone cluster from the image data clustered into the above two types, and calculates the central depth, azimuth, opening angle, and length of the compression fracture zone for each pixel from the clustered image data from which the median has been calculated.

[0018] In the compression fracture detection system based on a machine learning model according to the above embodiment, the center depth, azimuth, opening angle, and length of the compression fracture for each bounding box produced by the first compression fracture attribute information calculation unit can be calculated by the following [Mathematical Formula 1].

[0019]

[0020] [Mathematical Formula 1]

[0021] Center depth (depth, m) = Start depth (start depth) + (y × tm)

[0022] Azimuth (°) = x × 360° + roll

[0023] Opening angle (°) = ω × 360°

[0024] length (m) = h × tm

[0025] [Here, the starting depth represents the top depth of the image data output from the rolling unit, t represents the vertical length of the image data output from the rolling unit, x and y represent the center coordinates of the bounding box, w and h represent the horizontal and vertical lengths of the bounding box, and roll represents information performed by the rolling unit, which represents information that translates the column with the lowest average amplitude value among the columns from 0 degrees to 360 degrees in the image data output from the rolling unit to the first column and then to the right again].

[0026]

[0027] In the compression fracture detection system based on a machine learning model according to the above embodiment, the center depth, azimuth, opening angle, and length of the compression fracture per pixel produced by the second compression fracture attribute information calculation unit can be calculated by the following [Mathematical Formula 2].

[0028]

[0029] [Equation 2]

[0030] Center depth (depth, m) = (depth min +depth max ) / 2

[0031] Azimuth (°) = average of the median values ​​by depth of the compression fracture zone cluster

[0032] Opening angle (°) = Average opening angle by depth of the compression fracture zone cluster

[0033] length (m) = depth max ― depth min

[0034] [Here, depth max and depth min [Indicates the highest and lowest depths of the compression fracture zone per pixel]

[0035]

[0036] According to a compression fracture zone detection system based on a machine learning model according to an embodiment of the present invention, a plurality of training image data and inspection image data are inputted, in which the X-axis represents an angle of 0 to 360 degrees, the Y-axis represents a depth, and color represents an amplitude, and grayscale training image data and inspection image data are generated and preprocessed to generate training image data and inspection image data having a set resolution, and the generated grayscale training image data are augmented, and a machine learning model learned by a data set of coordinate and size information of a bounding box surrounding a compression fracture zone of the augmented grayscale training image data and the compression fracture zone of the training image data is generated, and the inspection image data translated in parallel by a rolling unit is inputted into the learned machine learning model to acquire coordinate and size information (x, y, w, h) of a bounding box surrounding a compression fracture zone, thereby having an excellent effect of being able to generate a machine learning model for compression fracture zone detection even when the available training data is small.

[0037] Moreover, according to the compression fracture detection system based on a machine learning model according to an embodiment of the present invention, by receiving grayscale inspection image data and moving the column with the lowest average amplitude value among the columns from 0 degrees to 360 degrees in the image data in parallel to the first column, and then moving in parallel to the right again, there is another excellent effect that the detection accuracy of the compression fracture can be improved even if the compression fracture exists at both ends of the inspection image data.

[0038]

[0039] Figure 1 is a block diagram of a compression fracture detection system based on a machine learning model according to an embodiment of the present invention.

[0040] Figure 2 is a detailed block diagram of the preprocessing unit of Figure 1.

[0041] Figure 3 is a detailed block diagram of the post-processing unit of Figure 1.

[0042] Figure 4 is a drawing showing the learning and inspection borehole ultrasonic image logging data input to the input data generation unit of Figure 1.

[0043] Figure 5 is a diagram showing image data for learning and testing generated in the input data generation unit of Figure 1.

[0044] Fig. 6 is a diagram showing grayscale learning and inspection image data converted by the grayscale conversion unit of Fig. 2.

[0045] Figure 7 is a diagram showing the original image data, which is grayscale learning image data input to the data augmentation unit of Figure 2, parallel translation image data augmented by the data augmentation unit, left-right inversion image data, and 180-degree rotation image data.

[0046] Figure 8 is a flowchart showing the process performed by the rolling unit of Figure 3.

[0047] FIG. 9 is a drawing for explaining the coordinate and size information (x, y, w, h) of the bounding box input into the first compression fracture zone attribute information calculation unit of FIG. 3, the image data in which the compression fracture zone is detected, and the attribute information of the compression fracture zone for each bounding box calculated by the first compression fracture zone attribute information calculation unit.

[0048] FIG. 10 is a diagram for explaining the clustering process by the clustering model of FIG. 3, the 1, 0 value classification process of the clustered image data, and the median calculation process of the compressed crushing zone cluster.

[0049] FIG. 11 is a diagram for explaining the process of calculating the attribute information of the compression fracture zone on a pixel-by-pixel basis from the clustered image data for which the median value is calculated by the second compression fracture zone attribute information calculation unit of FIG. 3.

[0050]

[0051] In describing embodiments of the present invention, if a detailed description of a known technology related to the present invention is judged to unnecessarily obscure the gist of the present invention, the detailed description will be omitted. In addition, the terms described below are terms defined in consideration of their functions in the present invention, and this may vary depending on the intention or custom of the user or operator. Therefore, the definitions should be made based on the contents throughout this specification. The terms used in the detailed description are only for the purpose of describing embodiments of the present invention and should never be construed as limiting. Unless clearly used otherwise, the singular form includes the plural form. In this description, expressions such as "comprises" or "having" are intended to indicate certain features, numbers, steps, operations, elements, parts or combinations thereof, and should not be construed to exclude the presence or possibility of one or more other features, numbers, steps, operations, elements, parts or combinations thereof other than those described.

[0052] In each system depicted in the drawings, elements may in some cases have the same reference number or different reference numbers, indicating that the depicted elements may be different or similar. However, the elements may operate with any or all of the systems shown or described herein, with different implementations. Various elements depicted in the drawings may be the same or different. Which one is referred to as a first element and which one as a second element is arbitrary.

[0053] In this specification, the phrase “transmitting,” “transmitting,” or “providing” data or signals from one component to another component includes not only the direct transmission of data or signals from one component to another component, but also the transmission of data or signals from one component to another component via at least one other component.

[0054]

[0055] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0056] Figure 1 is a block diagram of a compression fracture detection system based on a machine learning model according to an embodiment of the present invention.

[0057]

[0058] The compression fracture detection system based on a machine learning model according to an embodiment of the present invention includes, as illustrated in FIG. 1, an input data generation unit (100), a preprocessing unit (200), a data augmentation unit (300), a rolling unit (350), a compression fracture object unit detection model generation unit (400), a compression fracture object unit detection unit (500), a postprocessing unit (600), and a result display unit (700). The input data generation unit (100), the preprocessing unit (200), the data augmentation unit (300), the rolling unit (350), the compression fracture object unit detection model generation unit (400), the compression fracture object unit detection unit (500), the postprocessing unit (600), and the result display unit (700) may be configured as a single terminal device (e.g., a laptop, a personal computer, a PDA, a PMP, a smartphone, etc.).

[0059]

[0060] The input data generation unit (100) receives training borehole ultrasonic image logging data and inspection borehole ultrasonic image logging data (the X-axis represents an angle of 0 to 360 degrees, the Y-axis represents depth, and color represents amplitude) and generates a plurality of training image data having a set interval from the top and bottom of the compression fracture zone and a constant size, and a plurality of inspection image data in which the inspection borehole ultrasonic image logging data is divided into set depth intervals.

[0061] Figure 4 is a diagram showing the learning borehole ultrasonic image logging data and the inspection borehole ultrasonic image logging data input to the input data generation unit of Figure 1, where the X-axis represents an angle from 0 to 360 degrees, the Y-axis represents depth, and the color represents the amplitude. Here, the darkest color represents the lowest amplitude and has a high probability of being a compression fracture zone.

[0062] FIG. 5 is a drawing showing learning image data and inspection image data generated in the input data generation unit of FIG. 1, wherein each of the plurality of learning image data has a set interval (e.g., 0.5 m) from the top and bottom of the compression crushing zone and has a constant size.

[0063]

[0064] The preprocessing unit (200) preprocesses the plurality of training image data and inspection image data generated by the input data generation unit (100) to generate training image data and inspection image data in grayscale with a set resolution. More specifically, as illustrated in FIG. 2, the preprocessing unit (200) includes a resolution adjustment unit (210) and a grayscale conversion unit (220).

[0065]

[0066] The resolution adjustment unit (210) receives a plurality of learning image data and inspection image data generated by the input data generation unit (100) and adjusts them to have a set resolution (e.g., 640×640).

[0067]

[0068] The grayscale conversion unit (220) converts the training image data and inspection image data having a set resolution adjusted by the resolution adjustment unit (210) into grayscale training image data and inspection image data. Fig. 6 is a diagram showing grayscale training image data and inspection image data converted by the grayscale conversion unit of Fig. 2.

[0069]

[0070] The data augmentation unit (300) uses the grayscale learning image data preprocessed by the preprocessing unit (200) as the original image data, and generates parallel displacement image data by moving the original image data parallel to a set angle, inverts the original image data left and right to generate horizontal inversion image data, inverts the original image data up and down to generate vertical inversion image data, and rotates the original image data 180 degrees to generate 180 degree rotation image data. That is, 20 image data can be augmented for one original image data. FIG. 7 is a diagram showing the original image data, which is grayscale learning image data input to the data augmentation unit of FIG. 1, the parallel displacement image data augmented by the data augmentation unit, the horizontal inversion image data, the vertical inversion image data, and the 180 degree rotation image data. In this way, by augmenting the original image data, which is the learning image learning data, with parallel translation image data, horizontally flipped image data, vertically flipped image data, and 180-degree rotated image data, a machine learning model for detecting a compression fracture zone can be created even when the available learning image data is limited.

[0071]

[0072] The rolling unit (350) receives grayscale inspection image data having a set resolution preprocessed by the preprocessing unit (200) as illustrated in Fig. 8 (S351), and functions to parallel-translate the angle column with the lowest average amplitude value to the first column (S352), and then parallel-translate it to the right again (S353). Furthermore, the rolling unit (350) can provide parallel-translation information to the postprocessing unit (600).

[0073]

[0074] The compression fracture zone object unit detection model generation unit (400) serves to generate a machine learning model learned by a data set of grayscale learning image data augmented by the data augmentation unit (300) and coordinate and size information (x, y, w, h) of a bounding box surrounding the compression fracture zone of the augmented grayscale learning image data.

[0075]

[0076] The compression crushing zone detection unit (500) inputs inspection image data moved in parallel by the rolling unit (350) into the machine learning model generated by the compression crushing zone object unit detection model generation unit (400), thereby acquiring coordinate and size information (x, y, w, h) of the bounding box surrounding the compression crushing zone.

[0077]

[0078] The post-processing unit (600) receives coordinates and size information (x, y, w, h) of the bounding box from the compression crushing zone object unit detection unit (500), receives parallel-moved inspection image data and parallel-movement information from the rolling unit (350), and performs post-processing to generate compression crushing zone attribute information for each bounding box and compression crushing zone attribute information for each pixel.

[0079] More specifically, the post-processing unit (600) includes a first compression fracture zone attribute information calculation unit (620), a clustering model (610), and a second compression fracture zone attribute information calculation unit (630), as illustrated in FIG. 3.

[0080]

[0081] The first compression fracture zone attribute information calculation unit (620) receives coordinate and size information (x, y, w, h) of a bounding box from the compression fracture zone object unit detection unit (500), receives parallel-shifted inspection image data and parallel shift information from the rolling unit (350), and calculates the center depth, azimuth, opening angle, and length of the compression fracture zone for each bounding box by using the input coordinate and size information of the bounding box, parallel-shifted inspection image data, and parallel shift information. The center depth, azimuth, opening angle, and length of the compression fracture zone for each bounding box can be calculated by the following [Mathematical Formula 1].

[0082]

[0083] [Mathematical Formula 1]

[0084] Center depth (depth, m) = Start depth (start depth) + (y × tm)

[0085] Azimuth (°) = x × 360° + roll

[0086] Opening angle (°) = ω × 360°

[0087] length (m) = h × tm

[0088] [Here, the starting depth represents the top depth of the image data output from the rolling unit, t represents the vertical length of the image data output from the rolling unit, x and y represent the center coordinates of the bounding box, w and h represent the horizontal and vertical lengths of the bounding box, and roll represents information performed in the rolling unit, which represents information that translates the column with the lowest average amplitude value among the columns from 0 degrees to 360 degrees in the image data output from the rolling unit into the first column and then translates it to the right again]

[0089]

[0090] FIG. 9 is a drawing for explaining the coordinate and size information (x, y, w, h) of the bounding box input into the first compression fracture zone attribute information calculation unit of FIG. 3, the image data in which the compression fracture zone is detected, and the attribute information of the compression fracture zone for each bounding box calculated by the first compression fracture zone attribute information calculation unit.

[0091]

[0092] The clustering model (610) uses a clustering algorithm to classify the compressed crushing zone image data classified based on the bounding box from the parallel-moved inspection image data acquired by the rolling unit (350) into two types by assigning a cluster with a relatively small average pixel value as 1 and a cluster with a relatively large average pixel value as 0.

[0093] Figure 10 illustrates the clustering process by the clustering model of Figure 3, the 1, 0 value classification process of the clustered image data, and the median calculation process of the compressed crushing zone cluster.

[0094] The clustering process by the clustering model (610) is explained.

[0095] First, the compression fracture zone obtained by the rolling unit (350) is classified into left and right images based on the bounding box for inspection image data located at the center of the image (classified into left and right images based on the compression fracture zone detection image data in the middle in FIG. 10).

[0096] Next, the compressed crushed zone image data classified based on the bounding box in the parallel-shifted inspection image data is classified into two types using a clustering algorithm, with the cluster having a relatively small pixel average value in the image data being 1 and the cluster having a relatively large pixel average value being 0 (clustering into white cluster and black cluster in Fig. 10).

[0097]

[0098] The second compression fracture zone attribute information calculation unit (630) calculates the median of the compression fracture zone cluster (median of the black cluster in the left and right images of Fig. 10) from the image data clustered into two types of 1 and 0 by the clustering model (610), and calculates the central depth, azimuth, opening angle, and length of the compression fracture zone for each pixel from the clustered image data for which the median has been calculated using the following [Mathematical Formula 2] (see Fig. 11).

[0099] [Equation 2]

[0100] Center depth (depth, m) = (depth min +depth max ) / 2

[0101] Azimuth (°) = average of the median values ​​by depth of the compression fracture zone cluster

[0102] Opening angle (°) = Average opening angle by depth of the compression fracture zone cluster

[0103] length (m) = depth max ― depth min

[0104] [Here, depth max and depth min [Indicates the highest and lowest depths of the compression fracture zone per pixel]

[0105]

[0106] The result display unit (600) serves to display coordinate and size information (x, y, w, h) of the bounding box obtained from the compression fracture zone object unit detection unit (500), or the compression fracture zone attribute information for each bounding box generated from the post-processing unit (600) (center depth, azimuth, opening angle, and length of the compression fracture zone for each bounding box) and the compression fracture zone attribute information for each pixel (center depth, azimuth, opening angle, and length of the compression fracture zone for each pixel).

[0107]

[0108] The operation of the compression fracture detection system based on a machine learning model according to an embodiment of the present invention configured as described above will be described.

[0109]

[0110] First, the input data generation unit (100) receives training borehole ultrasonic image logging data and inspection borehole ultrasonic image logging data, in which the X-axis represents an angle from 0 to 360 degrees, the Y-axis represents depth, and color represents amplitude, and generates a plurality of training image data having a set interval from the top and bottom of the compression fracture zone, and a plurality of inspection image data in which the inspection borehole ultrasonic image logging data is divided into a set depth interval.

[0111]

[0112] Next, the preprocessing unit (200) preprocesses the learning image data and inspection image data generated by the input data generation unit (100) to generate grayscale learning image data and inspection image data having a set resolution.

[0113]

[0114] Next, the data augmentation unit (300) receives grayscale learning image data having a set resolution generated by the preprocessing unit (200) and generates augmented grayscale learning image data.

[0115]

[0116] Next, the rolling unit (350) receives the inspection image data in grayscale with the set resolution generated by the preprocessing unit (200), moves the angle column with the lowest average amplitude value parallel to the first column, and then moves it parallel to the right again.

[0117]

[0118] The compression fracture zone object unit detection model generation unit (400) generates a machine learning model learned by a data set of grayscale learning image data augmented by the data augmentation unit (300) and coordinate and size information (x, y, w, h) of a bounding box surrounding the compression fracture zone of the augmented grayscale learning image data.

[0119]

[0120] Next, the compression crushing zone object unit detection unit (500) loads the machine learning model generated from the compression crushing zone object unit detection model generation unit (400), and inputs the inspection image data moved in parallel by the rolling unit (350) into the machine learning model to obtain coordinate and size information (x, y, w, h) of the bounding box surrounding the compression crushing zone.

[0121]

[0122] Next, the post-processing unit (600) receives coordinates and size information (x, y, w, h) of the bounding box from the compression crushing zone object unit detection unit (500), and receives parallel-moved inspection image data and parallel-movement information from the rolling unit (350) and performs post-processing to generate compression crushing zone attribute information for each bounding box and compression crushing zone attribute information for each pixel.

[0123]

[0124] Next, one or more of the coordinate and size information (x, y, w, h) of the bounding box obtained from the compression crushing zone object unit detection unit (500), the compression crushing zone attribute information for each bounding box generated from the post-processing unit (600), and the compression crushing zone attribute information for each pixel unit are displayed through the result display unit (700).

[0125]

[0126] According to a compression fracture zone detection system based on a machine learning model according to an embodiment of the present invention, a plurality of training image data and inspection image data are inputted, in which the X-axis represents an angle of 0 to 360 degrees, the Y-axis represents a depth, and color represents an amplitude, and grayscale training image data and inspection image data are generated and preprocessed to generate training image data and inspection image data having a set resolution, and the generated grayscale training image data are augmented, and a machine learning model learned by a data set of coordinate and size information of a bounding box surrounding a compression fracture zone of the augmented grayscale training image data and the compression fracture zone of the training image data is generated, and the inspection image data translated in parallel by a rolling unit is inputted into the learned machine learning model to acquire coordinate and size information (x, y, w, h) of a bounding box surrounding a compression fracture zone, thereby enabling the generation of a machine learning model for compression fracture zone detection even when the available training data is small.

[0127] Furthermore, according to the compression fracture detection system based on a machine learning model according to an embodiment of the present invention, by receiving grayscale inspection image data and moving the column with the lowest average amplitude value among the columns from 0 degrees to 360 degrees in the image data in parallel to the first column, and then moving parallel to the right again, even if compression fractures exist at both ends of the inspection image data, the detection accuracy of the compression fracture can be improved.

[0128]

[0129] The drawings and specification disclose optimal embodiments, and while specific terminology may be used, it is solely for the purpose of describing embodiments of the present invention and is not intended to limit the meaning or scope of the invention as defined in the claims. Therefore, those skilled in the art will appreciate that various modifications and equivalent embodiments are possible. Therefore, the true technical protection scope of the present invention should be defined by the technical spirit of the appended claims.

Claims

1. An input data generation unit (100) that receives learning borehole ultrasonic image logging data and inspection borehole ultrasonic image logging data, in which the X-axis represents an angle from 0 to 360 degrees, the Y-axis represents depth, and color represents amplitude, and generates a plurality of learning image data having a set interval from the top and bottom of a compression fracture zone, and a plurality of inspection image data in which the inspection borehole ultrasonic image logging data is divided into a set depth interval; A preprocessing unit (200) configured to preprocess the plurality of learning image data and inspection image data to generate grayscale learning image data and inspection image data having a set resolution; A data augmentation unit (300) configured to input grayscale learning image data having the above-described resolution and generate augmented grayscale learning image data; A rolling unit (350) that receives the preprocessed inspection image data, moves the angle column with the lowest average amplitude value to the first column, and then moves it to the right again; A compression fracture zone object unit detection model generation unit (400) configured to generate a machine learning model trained by a data set of coordinate and size information (x, y, w, h) of a bounding box surrounding a compression fracture zone of the grayscale training image data augmented by the data augmentation unit, and the augmented grayscale training image data; and A compression fracture detection system based on a machine learning model, comprising a compression fracture object unit detection unit (500) configured to input inspection image data moved in parallel by the rolling unit into a machine learning model generated from the compression fracture object unit detection model generation unit to obtain coordinate and size information (x, y, w, h) of a bounding box surrounding the compression fracture.

2. In paragraph 1, A compression fracture detection system based on a machine learning model, further comprising a post-processing unit (600) configured to receive coordinate and size information (x, y, w, h) of a bounding box from the compression fracture object unit detection unit (500), and to receive parallel-moved inspection image data and parallel-movement information from the rolling unit (350) and post-process them to generate compression fracture attribute information for each bounding box and compression fracture attribute information for each pixel.

3. In paragraph 1 or 2, A compression fracture detection system based on a machine learning model further comprising a result display unit (700) configured to display at least one of the coordinate and size information (x, y, w, h) of the bounding box obtained from the compression fracture object unit detection unit (500), the compression fracture attribute information for each bounding box generated from the post-processing unit (600), and the compression fracture attribute information for each pixel.

4. In paragraph 1, The above preprocessing unit (200) A resolution adjustment unit (210) that receives the plurality of learning image data and inspection image data and adjusts them to have a set resolution; and A compression fracture detection system based on a machine learning model, comprising a grayscale conversion unit (220) that converts the learning image data and inspection image data having a set resolution into a plurality of grayscale learning image data and inspection image data.

5. In paragraph 2, The above post-processing unit (600) A first compression fracture zone attribute information calculation unit (620) that receives coordinate and size information (x, y, w, h) of a bounding box from the compression fracture zone object unit detection unit (500), receives parallel-moved inspection image data and parallel-movement information from the rolling unit (350), and calculates the center depth, azimuth, opening angle, and length of the compression fracture zone for each bounding box by using the input coordinate and size information of the bounding box, parallel-moved inspection image data, and parallel-movement information; A clustering model (610) that classifies the compressed crushed zone image data classified based on the bounding box in the above parallel-shifted inspection image data into two types by using a clustering algorithm and assigning a cluster with a relatively small pixel average value as 1 and a cluster with a relatively large pixel average value as 0; A compression fracture detection system based on a machine learning model, comprising a second compression fracture attribute information calculation unit (630) that calculates the median value of a compression fracture cluster by depth from the image data clustered into the above two types, and calculates the central depth, azimuth, opening angle, and length of the compression fracture cluster by pixel from the clustered image data from which the median value has been calculated.

6. In paragraph 5, A compression fracture detection system based on a machine learning model in which the center depth, azimuth, opening angle and length of the compression fracture zone for each bounding box calculated by the first compression fracture zone attribute information calculation unit (620) are calculated by the following [Mathematical Formula 1]. [Mathematical Formula 1] Center depth (depth, m) = Start depth (start depth) + (y × tm) Azimuth (°) = x × 360° + roll Opening angle (°) = ω × 360° length (m) = h × tm [Here, the starting depth represents the top depth of the image data output from the rolling unit, t represents the vertical length of the image data output from the rolling unit, x and y represent the center coordinates of the bounding box, w and h represent the horizontal and vertical lengths of the bounding box, and roll represents information performed in the rolling unit, which represents information that translates the column with the lowest average amplitude value among the columns from 0 degrees to 360 degrees in the image data output from the rolling unit into the first column and then translates it to the right again] 7. In paragraph 5, A compression fracture detection system based on a machine learning model in which the center depth, azimuth, opening angle and length of the compression fracture zone per pixel, which are calculated by the second compression fracture zone attribute information calculation unit (630), are calculated by the following [Mathematical Formula 2]. [Equation 2] Center depth (depth, m) = (depth min +depth max ) / 2 Azimuth (°) = average of the median values ​​by depth of the compression fracture zone cluster Opening angle (°) = Average opening angle by depth of the compression fracture zone cluster length (m) = depth max ― depth min [Here, depth max and depth min [Indicates the highest and lowest depths of the compression fracture zone per pixel]

Images