Defect determination device, defect determination method, program, learning device, trained model, and model generation method

Abstract

To provide a technology that contributes to efficiently and accurately detecting defects. [Solution] The defect determination device has an image acquisition unit 102 that acquires a flaw detection image that shows the echo height of the reflected signal for each position of a reflection source inside the object, generated from inspection data of an ultrasonic flaw detection inspection performed on the object, a partial image extraction unit 103 that extracts a partial image from the flaw detection image that includes pixels whose pixel values ​​satisfy predetermined conditions, and a defect determination unit 104 that determines whether the partial image is a partial image that represents a defect in the object based on the output from a pre-trained machine learning model when the partial image is input into the machine learning model.

Description

[Technical Field]

[0001] The present invention relates to a defect determination device, a defect determination method, a program, a learning device, a trained model, and a model generation method. [Background technology]

[0002] Ultrasonic testing (UT) is known for detecting defects in products. Various techniques have been proposed for automatically identifying defects detected by UT. For example, Patent Document 1 discloses an apparatus for determining defects in an object to be inspected using a model trained by machine learning using test condition information for ultrasonic testing. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-004738 Summary of the Invention [Problem to be solved by the invention]

[0004] As described above, in recent years, attempts have been made to develop technology that applies machine learning to ultrasonic flaw detection, and new technology that can efficiently and accurately detect defects is required. Other issues and new features will become apparent from the description of this specification and the accompanying drawings. [Means for solving the problem]

[0005] A defect determination device according to one aspect of the present disclosure includes an image acquisition unit that acquires a flaw detection image that indicates the echo height of a reflected signal for each position of a reflection source inside the object, the flaw detection image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object; a partial image extraction unit that extracts a partial image from the flaw detection image that includes pixels whose pixel values ​​satisfy a predetermined condition; and a defect determination unit that determines whether the partial image is a partial image that represents a defect in the object based on the output from a pre-trained machine learning model when the partial image is input into the machine learning model.

[0006] A defect determination method according to one aspect of the present disclosure acquires a flaw detection image generated from inspection data of an ultrasonic flaw detection inspection performed on an object, showing the echo height of a reflected signal for each position of a reflection source inside the object, extracts a partial image from the flaw detection image, the partial image including pixels whose pixel values ​​satisfy predetermined conditions, and determines whether the partial image represents a defect in the object based on the output from a pre-trained machine learning model when the partial image is input into the machine learning model.

[0007] In addition, a program according to one aspect of the present disclosure causes a computer to execute an image acquisition step of acquiring a flaw detection image, which is generated from inspection data of an ultrasonic flaw detection inspection performed on an object and shows the echo height of a reflected signal for each position of a reflection source inside the object; a partial image extraction step of extracting a partial image from the flaw detection image, which includes pixels whose pixel values ​​satisfy a predetermined condition; and a defect determination step of determining whether the partial image is a partial image representing a defect in the object based on the output from a pre-trained machine learning model when the partial image is input into the machine learning model.

[0008] In addition, a learning device according to one aspect of the present disclosure includes a learning data acquisition unit that acquires learning data including a first type of partial image, a second type of partial image, and a third type of partial image, and a model learning unit that performs machine learning using the learning data to generate a machine learning model that takes as input a partial image cut out from a flaw detection image that shows the echo height of a reflected signal for each position of a reflection source inside the object, generated from inspection data of an ultrasonic flaw detection inspection performed on the object, and outputs information indicating whether the partial image is a partial image that represents a defect in the object, wherein the first type of partial image is a partial image of a flaw detection image for an area that includes a defect portion and represents a defect echo that is a reflected signal caused by the defect, the second type of partial image is a partial image of a flaw detection image for an area that does not include a defect portion and represents a pseudo echo that is a reflected signal caused by something other than the defect, and the third type of partial image is a partial image of a flaw detection image for an area that does not include a defect portion and does not represent the pseudo echo.

[0009] Furthermore, a trained model according to one aspect of the present disclosure is a trained model for causing a computer to function to determine whether a partial image cut out from a flaw detection image, which is generated from inspection data of an ultrasonic flaw detection inspection performed on an object and shows the echo height of a reflected signal for each position of a reflection source inside the object, is a partial image representing a defect in the object, and the trained model is a machine learning model generated by the above-mentioned learning device.

[0010] In addition, a model generation method according to one aspect of the present disclosure acquires learning data including a first type of partial image, a second type of partial image, and a third type of partial image, and uses as input a partial image cut out from a flaw detection image that shows the echo height of a reflected signal for each position of a reflection source inside the object, generated from inspection data of an ultrasonic flaw detection inspection performed on the object, to generate a machine learning model by performing machine learning using the learning data, and outputting information indicating whether the partial image is a partial image that represents a defect in the object, wherein the first type of partial image is a partial image of a flaw detection image for an area that includes a defect portion and represents a defect echo that is a reflected signal caused by the defect, the second type of partial image is a partial image of a flaw detection image for an area that does not include a defect portion and represents a pseudo echo that is a reflected signal caused by something other than the defect, and the third type of partial image is a partial image of a flaw detection image for an area that does not include a defect portion and does not represent the pseudo echo. [Effects of the Invention]

[0011] According to the present disclosure, it is possible to provide a technique that contributes to efficiently and accurately detecting defects. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a block diagram showing an example of a configuration of a defect determination system according to an embodiment; [Figure 2] FIG. 1 is a schematic diagram showing a configuration example of an automatic UT device. [Figure 3] 1 is a graph showing an example of an echo waveform obtained from inspection data output by an automatic UT device. [Figure 4] FIG. 10 is a diagram showing an example of a B-Scan image. [Figure 5] 10A and 10B are schematic diagrams illustrating a peak pixel detection process performed by a partial image cutting unit. [Figure 6] FIG. 2 is a diagram showing an example of a defect candidate area; [Figure 7] 10 is a flowchart showing an example of an operation flow for defect determination in the information processing device according to the embodiment; [Figure 8]FIG. 10 shows a partial image on which labeling is performed. [Figure 9A] FIG. 10 is a diagram showing a specific example of a partial image used as learning data. [Figure 9B] FIG. 10 is a diagram showing a specific example of a partial image used as learning data. [Figure 9C] FIG. 10 is a diagram showing a specific example of a partial image used as learning data. [Figure 10] 10 is a table showing learning conditions used in model learning processing. [Figure 11A] FIG. 1 is a schematic diagram showing the results of evaluating the classification accuracy of a machine learning model using a confusion matrix. [Figure 11B] FIG. 1 is a schematic diagram showing the results of evaluating the classification accuracy of a machine learning model using a confusion matrix. [Figure 12] 10 is a table showing the defect detection accuracy by the information processing device for the four rotor members used in model learning. [Figure 13] 10 is a table showing the total number of images and the number of defect images for each of the 20 rotor components used in the experiment. [Figure 14] 14 is a graph showing the accuracy of defect detection by an information processing device for the rotor member shown in FIG. 13. [Figure 15A] 10 is a graph showing a comparison result between the positions of defects identified by an inspector and the positions of defects identified by an information processing device. [Figure 15B] 10 is a graph showing a comparison result between the positions of defects identified by an inspector and the positions of defects identified by an information processing device. [Figure 15C] 10 is a graph showing a comparison result between the positions of defects identified by an inspector and the positions of defects identified by an information processing device. [Figure 16] FIG. 2 is a block diagram illustrating an example of a configuration of a computer that realizes processing by the information processing device according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments will be described with reference to the drawings. For clarity of explanation, the following description and drawings have been omitted and simplified as appropriate. In addition, the same elements in each drawing are given the same reference numerals, and duplicate explanations are omitted as necessary. Furthermore, not all of the features or steps shown in any one drawing are necessarily required to explain an exemplary embodiment, and some features or steps may be omitted. Furthermore, the order of steps shown in any one drawing may be changed as appropriate.

[0014] FIG. 1 is a block diagram showing an example of the configuration of a defect determination system 10 according to an embodiment. As shown in FIG. 1, the defect determination system 10 includes an automatic UT device 200 and an information processing device 100. In this embodiment, the object inspected by the defect determination system 10 is, as an example, a turbine rotor shaft material (rotor) used in a power plant. However, the object for which the defect determination system 10 (information processing device 100) determines the presence or absence of a defect may be any object. The object for which the presence or absence of a defect is to be determined may be, in particular, any metal product such as a steel material. Hereinafter, the object for which the presence or absence of a defect is to be determined is also referred to as an inspection object. The defect determination system 10 is a system for determining whether a defect exists inside the inspection object.

[0015] The automated UT device 200 is a device that automatically performs ultrasonic flaw detection (UT), a type of non-destructive testing. FIG. 2 is a schematic diagram showing an example of the configuration of the automated UT device 200. The automated UT device 200 includes a turning roller 201 that rotates a rotor shaft material 300, which is an object to be inspected; a probe 202 that outputs ultrasonic waves toward the rotor shaft material 300 and receives reflected signals from the rotor shaft material 300; and a manipulator 203 that controls the position of the probe 202. The automated UT device 200 rotates the rotor shaft material 300 using the turning roller 201 while keeping the probe 202 in contact with the surface of the rotor shaft material 300, while fine-tuning the position of the probe 202 with the manipulator 203. This causes the probe 202 to scan in the circumferential direction of the rotor shaft material 300. The manipulator 203 can also scan the rotor shaft material 300 in the axial direction, and the automated UT device 200 continuously acquires inspection data. The automated UT device 200 particularly acquires the intensity (echo height) of the reflected signal from a reflection source inside the object to be inspected as inspection data. In this embodiment, the defect determination system 10 inspects the surface layer of the object to be inspected (for example, a range from the surface to a depth of about 120 mm), but the inspection range is not necessarily limited to the surface layer.

[0016] FIG. 3 is a graph showing an example of an echo waveform obtained from the inspection data output by the automated UT device 200. The horizontal axis of the graph in FIG. 3 represents the depth from the surface of the object being inspected, and the vertical axis represents the echo height. The echo height represents the strength of the reflected signal (reflection intensity) from the ultrasonic waves output from the probe 202. In other words, the graph shown in FIG. 3 shows the strength of the reflected signal for each depth of the reflection source inside the object being inspected. This type of data is also called A-Scan data. If a defect occurs inside the object being inspected, a reflected signal from the defect location is observed. Hereinafter, a reflected signal caused by a defect is referred to as a defect echo. The intensity (i.e., echo height) of this defect echo depends on the shape of the defect, but generally, the larger the defect, the greater the size. A reflected signal can occur even when a defect does not exist. Hereinafter, a reflected signal caused by something other than a defect is referred to as a false echo. In the example shown in FIG. 3, a false echo specific to the surface layer is observed near the defect echo. It is known that false echoes occur on the surface layer of the object being inspected. In addition, false echoes may occur at a certain depth as probe-specific false echoes, or may occur at various positions due to electrical noise. The occurrence of false echoes makes it difficult to determine whether or not a defect exists.

[0017] FIG. 4 is a diagram showing an example of an image that represents data obtained by linking A-scan data in the scanning direction of the probe 202 (i.e., the circumferential direction of the rotor shaft material 300), and that represents echo heights using pixel values. Such an image is also called a B-scan image or a flaw detection image. A B-scan image (flaw detection image) is an image that shows the echo height of a reflected signal for each position of a reflection source inside an object to be inspected. In a B-scan image (flaw detection image), the position of a pixel corresponds to the position (coordinate) of the reflection source of the reflected signal, and the pixel value indicates the echo height of the reflected signal. More specifically, the position of a pixel corresponds to the position (coordinate) of the reflection source, which is defined by the position in the scanning direction of the ultrasonic flaw detection inspection (i.e., the circumferential direction of the rotor shaft material 300, which is the scanning direction of the probe 202) and the position in the depth direction inside the object to be inspected (i.e., the radial direction of the rotor shaft material 300).

[0018] In FIG. 4, an enlarged view of a portion of the B-Scan image showing a defect is shown. When an inspector inspects an object to be inspected, the inspector visually checks the B-Scan image and A-Scan data to determine whether or not a defect exists. Since one B-Scan image is generated for each axial position of the rotor shaft material 300, the longer the rotor shaft material 300, the more B-Scan images are generated. Although this number varies depending on the inspection conditions and the axial length of the rotor shaft material 300, it is possible for the number of B-Scan images to exceed 1,000. Therefore, when an inspector inspects an object to be inspected, the inspection requires a long time. Therefore, in this embodiment, the information processing device 100 efficiently and accurately detects defects occurring inside the object to be inspected. In this embodiment, the inspection data output by the automatic UT device 200 is input to the information processing device 100.

[0019] As shown in FIG. 1 , the information processing device 100 includes an image generation unit 101, an image acquisition unit 102, a partial image clipping unit 103, a defect determination unit 104, a defect output unit 105, a learning data acquisition unit 106, a model learning unit 107, and a model storage unit 108. Among these, the image generation unit 101, the image acquisition unit 102, the partial image clipping unit 103, the defect determination unit 104, and the defect output unit 105 are components primarily for determining defects in an inspection object. The learning data acquisition unit 106 and the model learning unit 107 are components primarily for generating a machine learning model used for defect determination. The model storage unit 108 can be referred to as either a component for determining defects in an inspection object or a component for generating a machine learning model. In the example shown in FIG. 1 , the information processing device 100 includes both a component group for determining defects in an inspection object and a component group for generating a machine learning model, but may not include part or all of either of the component groups. Furthermore, an information processing device having components for determining defects in an inspection object may be referred to as a defect determination device. Similarly, an information processing device having components for generating a machine learning model may be referred to as a learning device.

[0020] The image generation unit 101 generates a B-Scan image for defect determination by the information processing device 100 from inspection data of an ultrasonic flaw detection inspection performed on an object to be inspected. The image generation unit 101 generates a B-Scan image, which represents data obtained by linking inspection data (A-scan data) indicating echo heights output from the automatic UT device 200 in the scanning direction of the ultrasonic flaw detection inspection (i.e., the circumferential direction of the rotor shaft material 300), and which represents echo heights using pixel values. The size of this image in a first direction (the vertical direction of the image) corresponds to the inspection range from the surface of the object to a predetermined depth, and the size of the image in a second direction (the horizontal direction of the image) corresponds to the scanning range of the ultrasonic flaw detection inspection. Note that the scanning range of the ultrasonic flaw detection inspection can also be said to be the inspection range in the scanning direction of the ultrasonic flaw detection inspection (i.e., the circumferential direction of the rotor shaft material 300). As an example, the image generation unit 101 generates a B-Scan image having an image size of 1280 pixels x 720 pixels. The image generating unit 101 generates a B-Scan image in which the echo height of the reflected signal is corrected to correspond to the defect size, regardless of the depth position of the reflection source. That is, the image generating unit 101 performs a known TCG (Time Corrected Gain) correction. In this embodiment, an echo height of 100% is an echo height that represents a defect of a predetermined size (specifically, a diameter of 0.8 millimeters). Therefore, the echo height can also be considered an index value of signal intensity that is determined based on the intensity of a reflected signal corresponding to a defect of a predetermined size.

[0021] In this embodiment, the image generation unit 101 generates a color B-Scan image, but if a grayscale image is used for training a machine learning model (described later), the image generation unit 101 may generate a grayscale B-Scan image. When generating a color B-Scan image, the image generation unit 101 determines an RGB value corresponding to the echo height for each pixel, for example, by referring to a lookup table that defines a predetermined correspondence between echo height values ​​and RGB values.

[0022] The image acquisition unit 102 acquires the image generated by the image generation unit 101. In this embodiment, the image acquisition unit 102 particularly acquires a B-Scan image in which the average echo height indicated by the pixel values ​​of the pixel group that constitutes the B-Scan image is equal to or greater than a predetermined threshold value (also referred to as an error threshold value). In this embodiment, as an example, if the average echo height indicated by the pixel values ​​of the pixel group that constitutes the B-Scan image (hereinafter referred to as the average echo height) satisfies the condition shown in the following equation 1, the image acquisition unit 102 excludes the B-Scan image as an error image from those to be acquired.

[0023] [Average echo height] < 0.8% (Equation 1)

[0024] A B-Scan image whose average echo height is less than the error threshold corresponds to inspection data from an improperly performed ultrasonic flaw detection inspection. As described above, in this embodiment, the ultrasonic flaw detection inspection is performed using the automated UT device 200. Therefore, it is possible that the ultrasonic flaw detection inspection is performed without the probe 202 properly contacting the surface of the object being inspected. In such cases, a B-Scan image whose average echo height is less than the error threshold is generated. In this embodiment, such error images are excluded, thereby avoiding the execution of meaningless defect determination processing on data from an improperly performed ultrasonic flaw detection inspection. Note that, although the threshold in Equation 1 is 0.8%, this is merely an example, and any value can be set as the threshold. Note that, although it is preferable to exclude error images as described above, the image acquisition unit 102 may acquire B-Scan images including error images.

[0025] Note that some or all of the processing by the image generating unit 101 described above may be performed by another device, in which case the image acquiring unit 102 may acquire the B-Scan image described above from the other device. In this case, the information processing device 100 does not need to include the image generating unit 101.

[0026] The partial image cropping unit 103 performs processing to crop a partial image to be input into a machine learning model (described later) from the B-scan image acquired by the image acquisition unit 102. The partial image cropping unit 103 crops a partial image including pixels whose pixel values ​​satisfy a predetermined condition from the B-scan image acquired by the image acquisition unit 102. Specifically, the partial image cropping unit 103 detects a pixel corresponding to a peak echo height (hereinafter referred to as a peak pixel) from a group of pixels constituting the B-scan image acquired by the image acquisition unit 102. The partial image cropping unit 103 then crops the partial image including the peak pixel from the B-scan image. More specifically, the partial image cropping unit 103 detects the peak pixel by referring to the average echo height indicated by the pixel values ​​of a predetermined number of consecutive pixels in the B-scan image in a direction corresponding to the scanning direction of the ultrasonic flaw detection inspection, and the echo height indicated by the pixel value of one of the predetermined number of pixels (see Equation 2 below). Here, the direction corresponding to the scanning direction of the ultrasonic flaw detection inspection is the direction corresponding to the circumferential direction of the rotor shaft material 300 in the B-Scan image, and is the above-mentioned second direction (the horizontal direction of the image) in the B-Scan image. In this embodiment, the partial image cropping unit 103 selects a predetermined number of consecutive pixels in the direction corresponding to the scanning direction of the ultrasonic flaw detection inspection as a pixel group of 60 pixels or a pixel group of 120 pixels. That is, the partial image cropping unit 103 performs both a peak pixel detection process targeting a pixel group of 60 pixels and a peak pixel detection process targeting a pixel group of 120 pixels. Note that the reason for performing the peak pixel detection process on two pixel groups in this manner is to prevent peak pixels from being missed. Figure 5 is a schematic diagram showing the peak pixel detection process performed by the partial image cropping unit 103. The partial image cropping unit 103 determines whether or not a peak pixel exists in a pixel group based on the average echo height indicated by the pixel values ​​of circumferentially consecutive pixels that are a pixel group indicated by a dashed line 103a in Fig. 5 and the echo height indicated by the pixel value of each pixel in the pixel group. Specifically, if the echo height indicated by the pixel value satisfies the condition expressed by the following equation 2, the partial image cropping unit 103 detects the pixel having the pixel value as a peak pixel.

[0027] [Echo height] > Max([average echo height] × 1.5, 10%) (Equation 2)

[0028] That is, if the echo height indicated by the pixel to be determined is greater than a threshold (1.5 times the average value) based on the average value of the echo heights indicated by the pixel group and greater than a predetermined lower limit value (10%), the partial image cropping unit 103 detects the pixel to be determined as a peak pixel. In this embodiment, a predetermined lower limit value (10%) is set as a peak pixel detection condition in order to detect only pixels that exhibit a certain level of echo height as peak pixels. However, such a predetermined lower limit value does not necessarily have to be used as a detection condition. The above-described lower limit value is merely an example, and the lower limit value can be set to any value as appropriate. Similarly, a value 1.5 times the average value is used as a threshold based on the average value of the echo height. However, this is merely an example, and any value can be set as the threshold as appropriate. The partial image cropping unit 103 also performs similar detection processing on pixel groups adjacent in the direction indicated by the arrow 103b in FIG. 5. In this way, the partial image cropping unit 103 identifies an area consisting of peak pixels. The direction indicated by arrow 103b in Figure 5 corresponds to the radial direction of the rotor shaft material 300 in the B-Scan image, and is the first direction (vertical direction of the image) mentioned above in the B-Scan image. A region consisting of peak pixels is a region where a defect may exist, and therefore will be referred to as a defect candidate region hereinafter. Figure 6 is a diagram showing an example of a defect candidate region 90 identified by the above-mentioned process. As shown in Figure 6, a defect candidate region can be more specifically described as a region composed of a pixel group consisting of consecutive peak pixels.

[0029] In this embodiment, as described above, the partial image cropping unit 103 detects peak pixels by focusing on a predetermined number of consecutive pixels in a direction corresponding to the scanning direction of the ultrasonic flaw detection inspection. This allows peak pixels corresponding to defect echoes to be properly detected. This is for the following reason: Spurious echoes often occur throughout the entire surface layer of the object under inspection or throughout a predetermined depth. Therefore, pixel rows corresponding to depths at which spurious echoes occur, i.e., pixel rows aligned in the scanning direction of the ultrasonic flaw detection inspection, all exhibit large echo height values. In other words, pixel rows corresponding to spurious echoes do not exhibit peaks when viewed in the scanning direction of the ultrasonic flaw detection inspection. In contrast, defect echoes do not occur throughout the entire predetermined depth. Therefore, pixels corresponding to defect echoes exhibit peaks when viewed in the scanning direction of the ultrasonic flaw detection inspection. Therefore, by detecting peak pixels by focusing on pixel rows aligned in the scanning direction of the ultrasonic flaw detection inspection, it is possible to detect peak pixels corresponding to defect echoes while avoiding detection of peak pixels corresponding to spurious echoes as much as possible.

[0030] Once the partial image cropping unit 103 identifies a defect candidate area, it crops a partial image including the defect candidate area from the B-Scan image acquired by the image acquisition unit 102. However, in this embodiment, the partial image cropping unit 103 only crops a partial image including a defect candidate area where the maximum value of the echo height indicated by the pixel values ​​of the pixel group in the defect candidate area is equal to or greater than a predetermined threshold. Specifically, the partial image cropping unit 103 crops a partial image including a defect candidate area that satisfies the condition expressed by the following equation 3 from the B-Scan image.

[0031] [Maximum echo height within defect candidate area] ≥ 40% (Equation 3)

[0032] In Equation 3, the "maximum echo height within the defect candidate area" refers to the maximum echo height indicated by the pixel values ​​of the pixel group in the defect candidate area. In Equation 3, the threshold is 40%, but this is merely an example and can be set appropriately depending on the allowable defect size. This allows only defect candidate areas having echo heights equal to or greater than the threshold corresponding to the allowable defect size to be input to the machine learning model described below. In other words, in this embodiment, partial images including defect candidate areas having echo heights less than the threshold are not input to the machine learning model. This prevents an increase in processing load due to the detection of defects of an allowable size. Thus, in this embodiment, the partial image cropping unit 103 crops a partial image so as to include a defect candidate area in which the maximum value of the echo heights indicated by the pixel values ​​of the pixel group constituting the defect candidate area is equal to or greater than the threshold. In this embodiment, the partial image cropping unit 103 uses the maximum echo height condition as shown in Equation 3, but the partial image cropping unit 103 does not necessarily use such a condition. Furthermore, the partial image cropping unit 103 may crop a partial image including a defect candidate area of ​​a predetermined size or larger. That is, the partial image cutting unit 103 may cut out a partial image including a defect candidate area made up of peak pixels whose number is equal to or greater than a predetermined threshold.

[0033] The partial image cropping unit 103 crops a partial image of a predetermined size from the B-Scan image so that the defect candidate area is positioned at the center of the partial image. Specifically, this predetermined size is the same as the size of the image used for training the machine learning model described below.

[0034] The defect determination unit 104 determines whether or not a partial image cut out by the partial image cutout unit 103 is a partial image that represents a defect in the object to be inspected, based on the output from the machine learning model when the partial image is input into the machine learning model. That is, by inputting the partial image into the machine learning model, the defect determination unit 104 determines whether or not a defect exists at a position in the object to be inspected that corresponds to the partial image. The defect determination unit 104 uses a machine learning model that has been trained in advance by the model training unit 107. Details of model training by the model training unit 107 will be described later.

[0035] The defect output unit 105 performs processing to output information about defects in the inspection object based on the output results from the machine learning model. The defect output unit 105 may output the information about the defects on a display or may transmit it to another device. For example, the defect output unit 105 may identify the position of the defect based on the position of a partial image determined to represent a defect in the inspection object in the B-Scan image from which the partial image was cut out, and output information indicating the identified position. In this case, for example, the partial image cutout unit 103 manages the position from which the partial image was cut out in association with the cutout partial image.

[0036] Furthermore, the defect output unit 105 may output an index value of the size of the defect based on the echo height indicated by the pixel value of the partial image determined to be a partial image representing a defect in the object to be inspected. The index value of the size of the defect may be an estimated value of the size of the defect estimated from the echo height, or may be a grade of the defect (minor defect, medium defect, major defect, etc.) specified according to the size of the defect estimated from the echo height.

[0037] The information output by the defect output unit 105 is not limited to the information described above. For example, the defect output unit 105 may output a partial image determined to represent a defect in the object to be inspected, or may output an echo height corresponding to a defect in the object to be inspected. The defect output unit 105 may output any one type of information from the various types of information described above, or may output multiple types of information. The defect output unit 105 may also output information about defects in any format. For example, information about defects may be output in list format, or a defect that appears across multiple partial images or B-Scan images may be treated as a single defect.

[0038] Next, a flowchart of the operation of the above-described information processing device 100 is shown. Fig. 7 is a flowchart showing an example of the flow of the operation for defect determination in the information processing device 100. Hereinafter, the flow of the operation for defect determination will be described with reference to Fig. 7.

[0039] In step S100, the image generating unit 101 generates a B-Scan image for defect determination by the information processing device 100 from inspection data of an ultrasonic flaw detection inspection performed on an inspection object.

[0040] Next, in step S101, the image acquisition unit 102 acquires all the images generated in step S100 except for the error image.

[0041] Next, in step S102, the partial image cutting unit 103 cuts out a partial image to be input to the machine learning model from the B-Scan image acquired in step S101.

[0042] Next, in step S103, the defect determination unit 104 uses a machine learning model to determine whether or not a defect exists in the position within the inspection object corresponding to the partial image.

[0043] Next, in step S104, information regarding defects in the inspection object is output based on the determination result in step S103.

[0044] Next, a description will be given of a machine learning model used by the defect determination unit 104. In this embodiment, a deep learning model is used as an example of the machine learning model. More specifically, the machine learning model used by the defect determination unit 104 is a convolutional neural network (CNN).

[0045] The learning data acquisition unit 106 acquires learning data to be used in machine learning of the model used by the defect determination unit 104. The learning data acquisition unit 106 may acquire learning data input from another device, or may acquire the learning data by reading out learning data stored in a storage device such as the memory 502 (described later) of the information processing device 100. The learning data acquired by the learning data acquisition unit 106 is data consisting of pairs of partial images of a B-Scan image and correct labels for the partial images.

[0046] In this embodiment, each partial image used as training data is labeled based on the inspector's judgment. The correct label assigned to each partial image used as training data indicates whether the partial image is an image of an area containing a defect. In other words, the correct label indicates whether the partial image contains a defect echo. In this embodiment, as an example, labeling is performed based on the inspector's judgment results for partial images obtained from B-Scan images of four rotor members. These B-Scan images are B-Scan images generated by a process similar to the generation process of the image generation unit 101 described above. Therefore, the training data acquired by the training data acquisition unit 106 may be data obtained from B-Scan images generated by the image generation unit 101 based on inspection data from an ultrasonic flaw detection inspection performed on an arbitrary training object. Hereinafter, a label indicating that the image is an area containing a defect will also be referred to as a defect label, and a partial image to which the label is assigned will also be referred to as a defect image. Similarly, a label indicating that the image is not of an area including a defect portion will also be referred to as a healthy label, and a partial image to which this label is assigned will also be referred to as a healthy image.

[0047] Fig. 8 is a diagram showing a partial image to be labeled. As shown in Fig. 8, labeling is performed on a partial image 92 cut out from a B-Scan image 91. The size of the partial image 92 is, for example, 128 pixels x 128 pixels. A defect label is assigned to a partial image of a partial region of the image region of the B-Scan image that an inspector has determined to include a defect portion. Note that in this embodiment, for example, a defect label is assigned only to a partial image in which the maximum value of the echo height indicated by the pixel values ​​of the pixels constituting the partial image is 20% or more.

[0048] In this embodiment, two types of images are used as healthy images. Specifically, the training data includes two types of images as partial B-Scan images of areas that do not include defects. One is a partial image that displays a spurious echo. This healthy image is a partial image of an area of ​​a B-Scan image that does not include a defect, but displays a spurious echo, which is a reflected signal caused by something other than a defect. Note that a partial image of an area that does not include a defect can also be considered a partial image that an inspector has determined not to be a defective image. The other is a partial image that does not display a spurious echo. This healthy image is a partial image of an area of ​​a B-Scan image that does not include a defect, and also does not display a spurious echo. These two types of healthy images are assigned the healthy label described above. Specific examples of partial images used as training data are shown in FIGS. 9A to 9C. FIG. 9A shows an example of a defect image. FIG. 9B shows an example of a healthy image that does not display a spurious echo. 9C is a diagram showing an example of a healthy image in which a pseudo echo is displayed. The reason why the partial image in which the pseudo echo is displayed is also used as training data is to prevent the machine learning model from erroneously detecting the pseudo echo as a defect. In other words, by using the partial image in which the pseudo echo is displayed as training data, the judgment accuracy of the machine learning model can be improved.

[0049] As described above, in this embodiment, the training data acquisition unit 106 acquires training data including a first type of partial image, a second type of partial image, and a third type of partial image. The first type of partial image is the defect image described above. That is, the first type of partial image is a partial image of a B-Scan image of an area including a defect, and is an image that represents a defect echo, which is a reflection signal caused by the defect. The second type of partial image is the healthy image described above in which a spurious echo is displayed. That is, the second type of partial image is a partial image of a B-Scan image of an area not including a defect, and is an image that represents a spurious echo, which is a reflection signal caused by something other than the defect. The third type of partial image is the healthy image described above in which a spurious echo is not displayed. That is, the third type of partial image is a partial image of a B-Scan image of an area not including a defect, and is an image that does not represent a spurious echo.

[0050] The model learning unit 107 generates a machine learning model by performing machine learning using the learning data acquired by the learning data acquisition unit 106. This machine learning model is the model used by the defect determination unit 104 described above. That is, the machine learning model generated by the model learning unit 107 is a model that receives as input a partial image cut out from a B-Scan image generated from inspection data of an ultrasonic flaw detection inspection performed on an object, and outputs information indicating whether the partial image is a partial image that represents a defect in the object. In this embodiment, specifically, as described above, the model learning unit 107 generates a CNN model using the learning data acquired by the learning data acquisition unit 106. The trained model generated by the machine learning process of the model learning unit 107 is stored in the model storage unit 108. Then, the defect determination unit 104 uses the trained model stored in the model storage unit 108 to determine the presence or absence of a defect in the inspection object. That is, the trained model generated by the model learning unit 107 is used as a computer program module for causing a computer to function to determine whether a partial image cut out from a B-Scan image is a partial image that represents a defect in the inspection object. In other words, the trained model generated by the model training unit 107 is used as a computer program module to cause a computer to function to determine whether or not there is a defect in an object to be inspected.

[0051] FIG. 10 is a table showing the learning conditions used by the model learning unit 107 for model learning processing. As shown in FIG. 10, in this embodiment, a GoogLeNet model, a type of CNN model, was generated as an example. In this embodiment, the parameter values ​​of the first 10 layers were not changed, and fine tuning was performed, focusing on learning the final layer. The solver used was stochastic gradient descent with momentum (SGDM). In this embodiment, the images used as learning data were color images composed of pixels with RGB values. However, learning may also be performed using grayscale images. Half of the learning data was used for model training, and the other half was used only for model verification. Hereinafter, the former is referred to as model learning data, and the latter is referred to as model verification data. Under the conditions shown in FIG. 10, the number of defect images included in the learning data was 934. The number of healthy images was 6,999, of which at least 1,914 were healthy images containing pseudo echoes. Of the 6,999 healthy images, 5,085 were randomly extracted from B-Scan images. Most of these images were healthy images without any pseudo echoes, but some images contained some pseudo echoes.

[0052] 11A and 11B are schematic diagrams showing the results of evaluating the classification accuracy of a generated machine learning model using a confusion matrix. FIG. 11A shows the classification accuracy of the machine learning model for model training data, and FIG. 11B shows the classification accuracy of the machine learning model for model validation data. As can be seen from FIG. 11A, the recall and precision for the model training data are both 99% or higher, indicating sufficient learning. Furthermore, as can be seen from FIG. 11B, the recall and precision for the model validation data are 97% and 96%, respectively, indicating that the generated model has sufficiently high accuracy.

[0053] The inventors conducted an experiment in which the information processing device 100 executed the process described in FIG. 7 on the four rotor components used in model training. FIG. 12 is a table showing the defect detection accuracy by the information processing device 100 for the four rotor components used in model training. In the table shown in FIG. 12, the total number of images is the number of B-Scan images used in the experiment. Also, in the table shown in FIG. 12, the number of defect images is the number of B-Scan images that an inspector determined to contain a defect with an echo height of 40% or more. Also, in FIG. 12, the overdetection rate and the detection failure rate are defined by the following Equations 4 and 5, respectively. Note that the defect image referred to here does not refer to an image in units of partial images, but to an image in units of B-Scan images.

[0054] [Overdetection rate] = [Number of overdetected defect images] / ([Total number of images] - [Number of defect images]) × 100 (Equation 4)

[0055] [Detection failure rate] = [Number of missed defect images] / [Number of defect images] × 100 (Equation 5)

[0056] The overdetection rate was 4.0% for component No. 2, but was kept low overall. In addition, no detection errors occurred in any of the components.

[0057] The inventors also conducted an experiment in which the information processing device 100 performed the process described in FIG. 7 on 20 rotor components different from the four rotor components used in model training. FIG. 13 is a table showing the total number of images and the number of defect images for each of the 20 rotor components used in the experiment. The definitions of the total number of images and the number of defect images are the same as those in FIG. 12. FIG. 14 is a graph showing the defect detection accuracy of the information processing device 100 for the rotor components shown in FIG. 13. In FIG. 14, the overdetection rate and the detection miss rate are also defined as shown in Equation 4 and Equation 5. As shown in FIG. 14, for all components, defects detected by inspectors were detected by the information processing device 100, and no detection misses were observed. Since the average overdetection rate was 2.4%, defect detection by the information processing device 100 was slightly biased toward overdetection. However, considering that it is important to prevent defects from being overlooked when using a machine learning model, the defect classification accuracy of the information processing device 100 can be said to be good.

[0058] 15A to 15C are graphs showing the results of comparing the positions of defects identified by an inspector with the positions of defects identified by the information processing device 100. FIG. 15A shows the positions of defects in the axial direction, FIG. 15B shows the positions of defects in the circumferential direction, and FIG. 15C shows the positions of defects in the radial direction. Note that, here, defects detected consecutively in the axial direction are treated as one, and the position at which the echo height is maximum is shown on the graph as a representative coordinate. From FIGS. 15A to 15C, it can be seen that the positions of defects identified by the inspector and the positions of defects identified by the information processing device 100 generally coincide in all directions: axial, circumferential, and radial.

[0059] The above describes the embodiment. As described above, the information processing device 100 according to the present embodiment extracts a partial image from a B-Scan image, the partial image including pixels whose pixel values ​​satisfy a predetermined condition, and inputs the extracted partial image into a machine learning model to determine the presence or absence of a defect. Therefore, the information processing device 100 can efficiently and accurately detect defects. A supplementary explanation is provided regarding this. The portion of the B-Scan image corresponding to the defect (i.e., the portion showing the defect echo) is very small compared to the size of the B-Scan image. For this reason, it is difficult to achieve high determination accuracy in a determination process in which the B-Scan image is directly applied to a machine learning model. In contrast, in the present embodiment, as described above, a partial image satisfying a predetermined condition is extracted, and the extracted partial image is input into the machine learning model. Therefore, the above-described high determination accuracy can be achieved.

[0060] The above-described functions (processing) of the information processing device 100 may be realized by a computer 500 having the following configuration, for example.

[0061] 16 is a block diagram showing an example of the configuration of a computer 500 that realizes the processing of the information processing device 100. As shown in FIG.

[0062] The input / output interface 501 is an interface for connecting to other devices (for example, the automatic UT device 200, etc.).

[0063] The memory 502 is configured, for example, by a combination of a volatile memory and a non-volatile memory. The memory 502 is used to store software (computer programs) including one or more instructions executed by the processor 503, as well as data used for various processes. The model storage unit 108 is realized, for example, by the memory 502, but may also be realized by any storage device other than the memory 502.

[0064] The processor 503 reads and executes software (computer programs) from the memory 502 to perform the above-described processing of the information processing device 100. The processor 503 may be, for example, a microprocessor, an MPU (Micro Processor Unit), a CPU (Central Processing Unit), or a GPU (Graphics Processing Unit). The processor 503 may include multiple processors.

[0065] The program is included in a computer program product. Furthermore, the program includes a set of instructions (or software code) that, when loaded into a computer, causes the computer to perform one or more functions described in the embodiments. The program may be stored in a non-transitory computer-readable medium or a tangible storage medium. By way of example and not limitation, computer-readable media or tangible storage media include random-access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD) or other memory technologies, CD-ROM, digital versatile disc (DVD), Blu-ray (registered trademark) disc or other optical disk storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device. The program may also be transmitted on a transitory computer-readable medium or communication medium. By way of example and not limitation, transitory computer-readable media or communication media include electrical, optical, acoustic, or other forms of propagated signals.

[0066] The present invention is not limited to the above-described embodiment, and can be modified as appropriate within the scope of the invention.

[0067] Furthermore, some or all of the above-described embodiments can be described as, but are not limited to, the following supplementary notes. (Appendix 1) an image acquisition unit that acquires a flaw detection image that is generated from inspection data of an ultrasonic flaw detection inspection performed on an object and indicates the echo height of a reflected signal for each position of a reflection source inside the object; a partial image cutting unit that cuts out a partial image including pixels whose pixel values ​​satisfy a predetermined condition from the flaw detection image; a defect determination unit that determines whether the partial image represents a defect of the object based on an output from a machine learning model that has been trained in advance when the partial image is input to the machine learning model; A defect determination device having the above structure. (Appendix 2) The partial image cutting unit detects a peak pixel, which is a pixel corresponding to a peak of the echo height, from a group of pixels constituting the flaw detection image, and cuts out the partial image including the peak pixel from the flaw detection image. 2. The defect determination device according to claim 1. (Appendix 3) The partial image cutting unit detects the peak pixel by referring to an average of the echo heights indicated by pixel values ​​of a predetermined number of pixels in the flaw detection image that are continuous in a direction corresponding to the scanning direction of the ultrasonic flaw detection inspection, and the echo height indicated by a pixel value of one pixel out of the predetermined number of pixels. 3. A defect determination device according to claim 2. (Appendix 4) the partial image cropping unit crops out the partial image including a region configured by a pixel group consisting of consecutive peak pixels, where the maximum value of the echo height indicated by the pixel values ​​of the pixel group is equal to or greater than a predetermined first threshold; The defect determination unit determines whether the partial image represents a defect of the object based on an output from the machine learning model when the partial image including the region is input to the machine learning model. 4. The defect determination device according to claim 2 or 3. (Appendix 5) The image acquisition unit acquires the flaw detection image in which the average of the echo heights indicated by pixel values ​​of a pixel group constituting the flaw detection image is equal to or greater than a predetermined second threshold value. 5. A defect determination device according to any one of appendices 1 to 4. (Appendix 6) the machine learning model is a model that is machine-learned using learning data including a first type of partial image, a second type of partial image, and a third type of partial image; the first type of partial image is a partial image of a flaw detection image of an area including a defect portion, and represents a defect echo, which is a reflection signal caused by the defect; the second type partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and represents a spurious echo that is a reflected signal caused by something other than a defect; The third type of partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and does not represent the pseudo echo. 6. A defect determination device according to any one of appendices 1 to 5. (Appendix 7) The method further includes a defect output unit that identifies and outputs the position of the defect based on the position of the partial image determined to be a partial image representing a defect of the object in the original flaw detection image from which the partial image was extracted. 7. A defect determination device according to any one of appendices 1 to 6. (Appendix 8) The defect output unit further outputs an index value of the size of the defect based on the echo height indicated by a pixel value of the partial image determined to be a partial image representing a defect of the object. 8. The defect determination device according to claim 7. (Appendix 9) Acquiring a flaw detection image that indicates the echo height of a reflected signal for each position of a reflection source inside the object, the image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object; A partial image including pixels whose pixel values ​​satisfy a predetermined condition is extracted from the flaw detection image; When the partial image is input to a pre-trained machine learning model, it is determined whether the partial image represents a defect of the object based on an output from the machine learning model. Defect determination method. (Appendix 10) an image acquisition step of acquiring a flaw detection image that indicates the echo height of a reflected signal for each position of a reflection source inside the object, the image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object; a partial image cutting step of cutting out a partial image including pixels whose pixel values ​​satisfy a predetermined condition from the flaw detection image; a defect determination step of determining whether or not the partial image represents a defect of the object based on an output from a machine learning model that has been trained in advance when the partial image is input to the machine learning model; A program that causes a computer to execute the following. (Appendix 11) a learning data acquisition unit that acquires learning data including a first type of partial image, a second type of partial image, and a third type of partial image; a model learning unit that receives as input a partial image cut out from a flaw detection image that indicates the echo height of a reflected signal for each position of a reflection source inside the object, the partial image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object, and that outputs information indicating whether the partial image is a partial image that represents a defect in the object, by performing machine learning using the learning data; and and the first type of partial image is a partial image of a flaw detection image of an area including a defect portion, and represents a defect echo, which is a reflection signal caused by the defect; the second type partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and represents a spurious echo that is a reflected signal caused by something other than a defect; The third type of partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and does not represent the pseudo echo. Learning device. (Appendix 12) A trained model for causing a computer to function to determine whether a partial image cut out from a flaw detection image that shows the echo height of a reflected signal for each position of a reflection source inside an object, the partial image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object, is a partial image that represents a defect in the object, The trained model is a machine learning model generated by the learning device described in Supplementary Note 11. Trained model. (Appendix 13) acquiring learning data including a first type of partial image, a second type of partial image, and a third type of partial image; a machine learning model that receives as input a partial image cut out from a flaw detection image that shows the echo height of a reflected signal for each position of a reflection source inside the object, the partial image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object, and outputs information indicating whether the partial image is a partial image that represents a defect in the object, by performing machine learning using the learning data; the first type of partial image is a partial image of a flaw detection image of an area including a defect portion, and represents a defect echo, which is a reflection signal caused by the defect; the second type partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and represents a spurious echo that is a reflected signal caused by something other than a defect; The third type of partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and does not represent the pseudo echo. Model generation method. [Explanation of symbols]

[0068] 10 Defect Judgment System 90 Defect candidate areas 91 B-Scan images 92 partial images 100 Information processing device 101 Image generation unit 102 Image acquisition unit 103 Partial image extraction section 103a dashed line 103b Arrow 104 Defect Judgment Unit 105 Defect output unit 106 Learning data acquisition unit 107 Model Learning Department 108 Model Memory Unit 200 Automatic UT equipment 201 Turning Roller 202 Probe 203 Manipulator 300 rotor shaft material 500 computers 501 Input / Output Interface 502 memory 503 processor

Claims

1. an image acquisition unit that acquires a flaw detection image that is generated from inspection data of an ultrasonic flaw detection inspection performed on an object and indicates the echo height of a reflected signal for each position of a reflection source inside the object; a partial image cutting unit that cuts out a partial image including pixels whose pixel values ​​satisfy a predetermined condition from the flaw detection image; a defect determination unit that determines whether the partial image represents a defect of the object based on an output from a machine learning model that has been trained in advance when the partial image is input to the machine learning model; A defect determination device having the above structure.

2. The partial image cutting unit detects a peak pixel, which is a pixel corresponding to a peak of the echo height, from a group of pixels constituting the flaw detection image, and cuts out the partial image including the peak pixel from the flaw detection image. The defect determination device according to claim 1 .

3. The partial image cutting unit detects the peak pixel by referring to an average of the echo heights indicated by pixel values ​​of a predetermined number of pixels in the flaw detection image that are continuous in a direction corresponding to the scanning direction of the ultrasonic flaw detection inspection, and the echo height indicated by a pixel value of one pixel out of the predetermined number of pixels. The defect determination device according to claim 2 .

4. the partial image cropping unit crops out the partial image including a region configured by a pixel group consisting of consecutive peak pixels, where the maximum value of the echo height indicated by the pixel values ​​of the pixel group is equal to or greater than a predetermined first threshold; The defect determination unit determines whether the partial image represents a defect of the object based on an output from the machine learning model when the partial image including the region is input to the machine learning model.

4. The defect determination device according to claim 2 or 3.

5. The image acquisition unit acquires the flaw detection image in which the average of the echo heights indicated by pixel values ​​of a pixel group constituting the flaw detection image is equal to or greater than a predetermined second threshold value. The defect determination device according to claim 1 .

6. the machine learning model is a model that is machine-learned using learning data including a first type of partial image, a second type of partial image, and a third type of partial image; the first type of partial image is a partial image of a flaw detection image of an area including a defect portion, and represents a defect echo, which is a reflection signal caused by the defect; the second type partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and represents a spurious echo that is a reflected signal caused by something other than a defect; The third type of partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and does not represent the pseudo echo. The defect determination device according to claim 1 .

7. The method further includes a defect output unit that identifies and outputs the position of the defect based on the position of the partial image determined to be a partial image representing a defect of the object in the original flaw detection image from which the partial image was extracted. The defect determination device according to claim 1 .

8. The defect output unit further outputs an index value of the size of the defect based on the echo height indicated by a pixel value of the partial image determined to be a partial image representing a defect of the object. The defect determination device according to claim 7 .

9. Acquiring a flaw detection image that indicates the echo height of a reflected signal for each position of a reflection source inside the object, the image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object; A partial image including pixels whose pixel values ​​satisfy a predetermined condition is extracted from the flaw detection image; When the partial image is input to a pre-trained machine learning model, it is determined whether the partial image represents a defect of the object based on an output from the machine learning model. Defect determination method.

10. an image acquisition step of acquiring a flaw detection image that indicates the echo height of a reflected signal for each position of a reflection source inside the object, the image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object; a partial image cutting step of cutting out a partial image including pixels whose pixel values ​​satisfy a predetermined condition from the flaw detection image; a defect determination step of determining whether or not the partial image represents a defect of the object based on an output from a machine learning model that has been trained in advance when the partial image is input to the machine learning model; A program that causes a computer to execute the following.

11. a learning data acquisition unit that acquires learning data including a first type of partial image, a second type of partial image, and a third type of partial image; a model learning unit that receives as input a partial image cut out from a flaw detection image that indicates the echo height of a reflected signal for each position of a reflection source inside the object, the partial image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object, and that outputs information indicating whether the partial image is a partial image that represents a defect in the object, by performing machine learning using the learning data; and and the first type of partial image is a partial image of a flaw detection image of an area including a defect portion, and represents a defect echo, which is a reflection signal caused by the defect; the second type partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and represents a spurious echo that is a reflected signal caused by something other than a defect; The third type of partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and does not represent the pseudo echo. Learning device.

12. A trained model for causing a computer to function to determine whether a partial image cut out from a flaw detection image that shows the echo height of a reflected signal for each position of a reflection source inside an object, the partial image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object, is a partial image that represents a defect in the object, The trained model is a machine learning model generated by the learning device according to claim 11. Trained model.

13. acquiring learning data including a first type of partial image, a second type of partial image, and a third type of partial image; a machine learning model that receives as input a partial image cut out from a flaw detection image that shows the echo height of a reflected signal for each position of a reflection source inside the object, the partial image being generated from inspection data of an ultrasonic flaw detection inspection performed on the object, and outputs information indicating whether the partial image is a partial image that represents a defect in the object, by performing machine learning using the learning data; the first type of partial image is a partial image of a flaw detection image of an area including a defect portion, and represents a defect echo, which is a reflection signal caused by the defect; the second type partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and represents a spurious echo that is a reflected signal caused by something other than a defect; The third type of partial image is a partial image of a flaw detection image of an area that does not include a defect portion, and does not represent the pseudo echo. Model generation method.

Images