Ultrasonic flaw detection device and ultrasonic flaw detection method

Abstract

To provide an ultrasonic flaw detection device or the like capable of accurately determining a harmful degree of a flaw by determining a form such as a three-dimensional shape or the like of the flaw present inside an inspected material.SOLUTION: An ultrasonic flaw detection device 100 comprises: an ultrasonic probe 1 arranged to face an inspected material S; a three-dimensional voxel data generation section 31 for generating three-dimensional voxel data of the inspected material based on a flaw detection signal to be obtained from an echo received by the ultrasonic probe; a scope data generation section 32 for generating C-scop data or B-scop data for each one of a plurality of different cross sections of the inspected material based on the three-dimensional voxel data; and a determination section 33 including a machine learning model which defines data to be obtained based on a plurality of pieces of C-scop data or data to be obtained based on a plurality of pieces of B-scope data as an input and also defines a form of the flaw present inside the inspected material or a harmful degree based on the form of the flaw as an output, so as to determine quality of the flaw present inside the inspected material through the use of the machine learning model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to an ultrasonic flaw detection device and an ultrasonic flaw detection method for detecting flaws in a material to be detected, such as a steel material, using ultrasonic waves, and in particular to an ultrasonic flaw detection device and an ultrasonic flaw detection method that can accurately determine whether a defect is good or bad based on the morphology, such as the three-dimensional shape, of the defect present inside the material to be detected. [Background technology]

[0002] Conventionally, ultrasonic flaw detectors have been used to detect defects present inside test materials such as steel plates. In ultrasonic flaw detectors, the strength of the flaw detection signal obtained from the echo received by the ultrasonic probe is generally used as the criterion for determining whether a detected defect is defective. For example, since the strength of the flaw detection signal obtained from the echo reflected from a large defect is large, defects with a large intensity of the flaw detection signal are determined to be defective. However, the strength of the flaw detection signal varies greatly depending not only on the size of the defect but also on the defect's morphology, such as its three-dimensional shape, type, and location. Therefore, the size of the defect, and therefore whether the defect is good or bad, cannot be accurately determined based solely on the strength of the flaw detection signal.

[0003] On the other hand, conventionally, methods have been proposed in which an ultrasonic probe with a single vibrator is scanned two-dimensionally over the material to be inspected, or a one-dimensional array ultrasonic probe with vibrators arranged in a one-dimensional array is scanned one-dimensionally over the material to be inspected, or a two-dimensional array ultrasonic probe with vibrators arranged in a two-dimensional plane is used to obtain C-scope data or B-scope data of a single cross section of the material to be inspected, and the size of the defect is calculated using the -6 dB width of the intensity profile of the defect in this C-scope data, etc. (-6 dB echo drop method).However, this method cannot determine the shape of the defect other than its size, such as its three-dimensional shape.

[0004] That is, conventional ultrasonic flaw detectors have the problem that they cannot accurately determine the form of a defect, such as its three-dimensional shape, and therefore cannot accurately determine whether the defect is good or bad.

[0005] Patent Document 1 proposes a voxel data evaluation system that can improve the accuracy of detecting the position of defects (detection targets) by performing machine learning using voxel data obtained by an ultrasonic flaw detection device (Claim 1, paragraph 0012, etc. of Patent Document 1), but this does not solve the above problems. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Publication No. 2023-030258 Summary of the Invention [Problem to be solved by the invention]

[0007] The present invention has been made to solve the problems of the above-mentioned conventional technology, and its objective is to provide an ultrasonic flaw detection device and an ultrasonic flaw detection method that can accurately determine whether a defect is good or bad based on the morphology, such as the three-dimensional shape, of the defect present inside the material being detected. [Means for solving the problem]

[0008] In order to solve the above problems, the inventors focused on the fact that the diffusion pattern of an echo reflected from a defect differs depending on the form of the defect, such as its three-dimensional shape. They then discovered that the difference in the echo diffusion pattern due to the difference in the form of the defect becomes apparent in C-scope data and B-scope data for multiple cross sections of the material to be inspected that are separated by a certain distance or more, and that by using this, the form of the defect can be determined, and based on this, the pass / fail of the defect can be determined with high accuracy. The present invention has been completed based on the above findings of the present inventors.

[0009] That is, in order to solve the above-mentioned problems, the present invention provides an ultrasonic probe that is arranged opposite to a material to be detected, that transmits ultrasonic waves toward the inside of the material to be detected, and that receives echoes reflected inside the material to be detected; a three-dimensional voxel data generation unit that generates three-dimensional voxel data of the material to be detected, which is three-dimensional intensity data composed of the intensity of the flaw detection signal, based on the flaw detection signal obtained from the echo received by the ultrasonic probe; and a three-dimensional voxel data generation unit that generates C-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the material to be detected that are at least two voxels apart in the transmission direction of the ultrasonic waves, based on the three-dimensional voxel data, or The ultrasonic flaw detection device includes: a scope data generation unit that generates B-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the material to be flaw-detected; and a machine learning model that receives as input data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data and outputs the form of a defect present inside the material to be flaw-detected, or a machine learning model that receives as input data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data and outputs a harmfulness level based on the form of a defect present inside the material to be flaw-detected, and a determination unit that uses the machine learning model to determine whether a defect present inside the material to be flaw-detected is good or bad.

[0010] According to the present invention, a three-dimensional voxel data generating unit generates three-dimensional voxel data of the material to be detected, which is three-dimensional intensity data composed of the intensity of the flaw detection signal, based on the flaw detection signal obtained from the echo received by the ultrasonic probe arranged opposite the material to be detected. When generating the three-dimensional voxel data, an ultrasonic probe having a single oscillator may be caused to scan the material to be detected two-dimensionally, a one-dimensional array ultrasonic probe in which oscillators are arranged one-dimensionally may be caused to scan the material to be detected one-dimensionally, or a two-dimensional array ultrasonic probe in which oscillators are arranged in a two-dimensional plane may be used.

[0011] Next, according to the present invention, the scope data generating unit generates C-scope data, which is two-dimensional intensity data composed of the intensity of the inspection signal, for each of a plurality of different cross sections of the material to be inspected that are at least two voxels apart in the direction of ultrasonic transmission (in other words, a plurality of discrete C-scope data are generated in the direction of ultrasonic transmission), or generates B-scope data, which is two-dimensional intensity data composed of the intensity of the inspection signal, for each of a plurality of different cross sections of the material to be inspected that are at least two voxels apart in a direction perpendicular to the direction of ultrasonic transmission (in other words, a plurality of discrete B-scope data are generated in the direction perpendicular to the direction of ultrasonic transmission). The discrete C-scope data or B-scope data generated by the scope data generating unit reveals differences in the echo diffusion patterns due to differences in the morphology of defects present inside the material to be inspected.

[0012] Finally, according to the present invention, the quality of defects present inside the material to be inspected is determined by a judgment unit equipped with one or more machine learning models described below. The machine learning model included in the judgment unit receives input data obtained based on a plurality of discrete C-scope data sets or a plurality of discrete B-scope data sets, and outputs the morphology of defects present inside the material to be inspected. Examples of the defect morphology output by the machine learning model include the three-dimensional shape, size, type, and location of the defect. The training data used in the machine learning model includes known C-scope data or B-scope data obtained by ultrasonically inspecting the actual material to be inspected, and information on the defect morphology contained in the known C-scope data or B-scope data. The C-scope data or B-scope data in the training data may be data obtained using known ultrasonic inspection simulation software. When the judgment unit includes the machine learning model, the judgment unit converts the defect morphology obtained from the machine learning model into a harmfulness level associated with the defect morphology using a correspondence table data prepared and stored in advance. The judgment unit then compares the harmfulness level value obtained by the conversion process with a predetermined, stored threshold value to determine whether the defect is acceptable or not. For example, if the obtained harmfulness value is equal to or greater than a threshold value, the defect is determined to be bad (harmful), and if the value is less than the threshold value, the defect is determined to be good (harmless). The other machine learning model included in the judgment unit receives input data obtained based on multiple discrete C-scope data or multiple discrete B-scope data, and outputs a harmfulness level based on the morphology of defects present inside the material to be inspected. Examples of defect morphology that serve as the criteria for setting the harmfulness level output by the other machine learning model include the three-dimensional shape, size, type, and location of the defect. The training data used in the other machine learning model includes known C-scope data or B-scope data obtained by ultrasonically inspecting an actual material to be inspected, and a harmfulness level set based on the morphology of the defect contained in the known C-scope data or B-scope data. The harmfulness level value can be a score assigned to the defect by an inspector or other person when inspecting the defect, corresponding to the defect's pass / fail judgment. The C-scope data or B-scope data in the training data may be data obtained using known ultrasonic inspection simulation software. When the judgment unit includes the other machine learning model, the judgment unit compares the obtained harmfulness level (the harmfulness level output from the machine learning model) with a predetermined and stored threshold value to judge whether the defect is pass / fail. For example, if the obtained harmfulness value is equal to or greater than a threshold value, the defect is determined to be bad (harmful), and if the value is less than the threshold value, the defect is determined to be good (harmless). Alternatively, the determination unit may include a plurality of machine learning models that output different defect morphologies (e.g., four types of machine learning models that output the three-dimensional shape, size, type, and position of the defect), or a plurality of other machine learning models that output different defect morphologies that serve as criteria for setting the harmfulness levels (e.g., four types of other machine learning models that use the three-dimensional shape, size, type, and position of the defect as criteria for setting the harmfulness levels), and the determination unit may store threshold values ​​for the harmfulness levels obtained from the morphologies output from the plurality of machine learning models or the harmfulness levels output from the plurality of other machine learning models. In this case, a possible configuration is one in which the defect is determined to be bad (harmful) if the harmfulness level value based on any of the machine learning models is equal to or greater than the corresponding threshold, and the defect is determined to be good (harmless) if the harmfulness level values ​​based on all of the machine learning models are less than the corresponding threshold.

[0013] As described above, according to the present invention, by inputting data obtained based on multiple discrete C-scope data of the material to be inspected, in which differences in the diffusion patterns of echoes reflected by defects are apparent, into a machine learning model provided in the judgment unit, it is possible to accurately judge whether a defect is good or bad, taking into account the shape of the defect.

[0014] According to the findings of the present inventors, the differences in the diffusion patterns of echoes reflected from defects are not necessarily limited to C-scope data and B-scope data, but also become apparent in two-dimensional amplitude data composed of amplitudes and two-dimensional phase data composed of phases, which are obtained by subjecting these data to a two-dimensional Fourier transform. Therefore, in the present invention, preferably, the data obtained based on the C-scope data input to the machine learning model is the C-scope data (the C-scope data itself), first two-dimensional amplitude data consisting of amplitudes obtained by two-dimensional Fourier transform of the C-scope data, or first two-dimensional phase data consisting of phases obtained by two-dimensional Fourier transform of the C-scope data, and the data obtained based on the B-scope data input to the machine learning model is the B-scope data (the B-scope data itself), second two-dimensional amplitude data consisting of amplitudes obtained by two-dimensional Fourier transform of the B-scope data, or second two-dimensional phase data consisting of phases obtained by two-dimensional Fourier transform of the B-scope data.

[0015] According to the findings of the inventors, by using the data obtained based on the C-scope data or the data obtained based on the B-scope data input into the machine learning model as differential data, it may be possible to make the differences in the diffusion patterns of echoes reflected from defects even more apparent. Therefore, in the present invention, preferably, the data obtained based on the C-scope data input to the machine learning model is difference data of the C-scope data adjacent to the transmission direction of the ultrasound, difference data of first two-dimensional amplitude data consisting of amplitudes obtained by two-dimensional Fourier transform of the C-scope data adjacent to the transmission direction of the ultrasound, or difference data of first two-dimensional phase data consisting of phases obtained by two-dimensional Fourier transform of the C-scope data adjacent to the transmission direction of the ultrasound, and the data obtained based on the B-scope data input to the machine learning model is difference data of the B-scope data adjacent to the direction perpendicular to the transmission direction of the ultrasound, difference data of second two-dimensional amplitude data consisting of amplitudes obtained by two-dimensional Fourier transform of the B-scope data adjacent to the direction perpendicular to the transmission direction of the ultrasound, or difference data of second two-dimensional phase data consisting of phases obtained by two-dimensional Fourier transform of the B-scope data adjacent to the direction perpendicular to the transmission direction of the ultrasound.

[0016] In the present invention, the defect morphology preferably includes at least the three-dimensional shape of the defect (including the orientation of the defect). However, the present invention is not limited to this, and the defect morphology may also include the size, type, and position.

[0017] Further, in order to solve the above-mentioned problems, the present invention provides an ultrasonic transmission / reception step of arranging an ultrasonic probe opposite a material to be detected, transmitting ultrasonic waves from the ultrasonic probe toward the inside of the material to be detected, and receiving echoes reflected inside the material to be detected by the ultrasonic probe; a three-dimensional voxel data generation step of generating, based on the flaw detection signal obtained from the echo received by the ultrasonic probe, three-dimensional voxel data of the material to be detected, which is three-dimensional intensity data composed of the intensity of the flaw detection signal; and a step of generating, based on the three-dimensional voxel data, C-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the material to be detected that are at least two voxels apart in the transmission direction of the ultrasonic waves, or The present invention is also provided as an ultrasonic flaw detection method, comprising: a scope data generation step of generating B-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the flaw detection material that are at least two voxels apart; and a determination step of determining whether a defect present inside the flaw detection material is good or bad using a machine learning model that takes as input data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data and outputs the form of a defect present inside the flaw detection material, or a machine learning model that takes as input data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data and outputs a harmfulness level based on the form of a defect present inside the flaw detection material. [Effects of the Invention]

[0018] According to the present invention, it is possible to accurately determine whether a defect is good or bad based on the form, such as the three-dimensional shape, of a defect present inside a material to be inspected. [Brief explanation of the drawings]

[0019] [Figure 1] 1 is a diagram illustrating a schematic configuration of an ultrasonic flaw detection device according to an embodiment of the present invention. [Figure 2] 2 is a diagram for schematically explaining C-scope data and B-scope data generated by the scope data generating unit 32 shown in FIG. 1. FIG. [Figure 3] FIG. 10 is a diagram for schematically explaining that differences in echo diffusion patterns due to differences in defect morphology become apparent in a plurality of discrete C-scope data. DETAILED DESCRIPTION OF THE INVENTION

[0020] An ultrasonic flaw detector according to one embodiment of the present invention will be described below with reference to the accompanying drawings as needed. Fig. 1 is a diagram showing a schematic configuration of an ultrasonic flaw detection device according to this embodiment. In Fig. 1, the Z direction means the transmission direction of ultrasonic waves, the X direction means the direction perpendicular to the transmission direction of ultrasonic waves, and the Y direction means the direction perpendicular to the Z direction and the X direction. The same applies to Figs. 2 and 3 described below. 1, the ultrasonic flaw detection device 100 according to this embodiment includes an ultrasonic probe 1, a flaw detector 2, and a calculation processing unit 3. The calculation processing unit 3 includes a three-dimensional voxel data generation unit 31, a scope data generation unit 32, and a determination unit 33. Each of the components included in the ultrasonic flaw detection device 100 will be described below in order.

[0021] [Ultrasonic probe 1] The ultrasonic probe 1 is disposed opposite a material S to be detected, such as steel, and transmits ultrasonic waves U from a plurality of transducers 11 provided in the ultrasonic probe 1 toward the inside of the material S to be detected, and echoes reflected inside the material S to be detected are received by the plurality of transducers 11. The ultrasonic probe 1 shown in Fig. 1 is a one-dimensional array ultrasonic probe in which the transducers 11 are arranged one-dimensionally along the X direction perpendicular to the transmission direction of the ultrasonic waves U, but the present invention is not limited to this, and it is also possible to use an ultrasonic probe having a single transducer or a two-dimensional array ultrasonic probe in which the transducers are arranged in a two-dimensional plane (arranged along the XY plane).

[0022] [Flaw detector 2] The flaw detector 2 is electrically connected to the ultrasonic probe 1 and is equipped with well-known components similar to those of flaw detectors equipped in general ultrasonic flaw detection devices, such as a pulser for transmitting ultrasonic waves U from the transducer 11, a receiver for allowing the transducer 11 to receive echoes, an amplifier for amplifying the analog flaw detection signal output from the receiver, and an A / D converter for converting the analog signal output from the amplifier into a digital signal.

[0023] [Calculation processing unit 3] The arithmetic processing unit 3 is configured, for example, by a general-purpose computer electrically connected to the flaw detector 2, and stores programs for causing the arithmetic processing unit 3 to function as the 3D voxel data generating unit 31, the scope data generating unit 32, and the determining unit 33. The arithmetic processing unit 3 can then function as the 3D voxel data generating unit 31, the scope data generating unit 32, and the determining unit 33 by executing these programs.

[0024] [3D Voxel Data Generation Unit 31] The three-dimensional voxel data generation unit 31 receives a flaw detection signal obtained from the echo received by the ultrasonic probe 1, i.e., a digital flaw detection signal output from the flaw detector 2 (an A / D converter provided in the flaw detector 2). Based on the input flaw detection signal, the three-dimensional voxel data generation unit 31 generates three-dimensional voxel data of the material S to be detected, which is three-dimensional intensity data (data in which the value of each voxel in XYZ space is represented by the intensity of the flaw detection signal) composed of the intensity of the flaw detection signal.

[0025] Specifically, in this embodiment, the 3D voxel data generation unit 31 generates 3D voxel data based on flaw detection signals sequentially obtained by scanning the ultrasonic probe 1 in the Y direction (a direction perpendicular to the arrangement direction of the transducers 11) while transmitting ultrasonic waves U from the multiple transducers 11 included in the ultrasonic probe (one-dimensional array ultrasonic probe) 1. Note that if the ultrasonic probe 1 is an ultrasonic probe having a single transducer 11, the 3D voxel data is generated based on flaw detection signals sequentially obtained by scanning the ultrasonic probe 1 in the X direction and the Y direction while transmitting ultrasonic waves U from the transducer 11 included in the ultrasonic probe 1. Also, if the ultrasonic probe 1 is an ultrasonic probe in which the transducers 11 are arranged two-dimensionally, the 3D voxel data is generated based on flaw detection signals sequentially obtained by transmitting ultrasonic waves U from the multiple transducers 11 without scanning the ultrasonic probe 1.

[0026] [Scope data generation unit 32] Based on the three-dimensional voxel data, the scope data generating unit 32 generates C-scope data, which is two-dimensional intensity data composed of the intensity of the inspection signal, for each of a plurality of different cross sections of the material S to be inspected that are at least two voxels apart in the transmission direction (Z direction) of the ultrasonic waves U (generating a plurality of discrete C-scope data in the Z direction). Alternatively, the scope data generating unit 32 generates B-scope data, which is two-dimensional intensity data composed of the intensity of the inspection signal, for each of a plurality of different cross sections of the material S to be inspected that are at least two voxels apart in the direction orthogonal to the transmission direction of the ultrasonic waves U (Y direction) (generating a plurality of discrete B-scope data in the Y direction).

[0027] FIG. 2 is a diagram illustrating the C-scope data and B-scope data generated by the scope data generating unit 32. FIGS. 2(a) and 2(b) illustrate the C-scope data, and FIGS. 2(c) and 2(d) illustrate the B-scope data. As shown in FIGS. 2(a) and 2(b), the C-scope data is two-dimensional data (two-dimensional intensity data) in which the value of each voxel constituting the XY cross section is represented by the intensity of a flaw detection signal obtained from an echo reflected in a cross section perpendicular to the transmission direction (Z direction) of the ultrasonic wave U (for example, the XY cross section at Z=Z1 hatched in FIG. 2(a)). The scope data generating unit 32 generates this C-scope data for each of a plurality of discrete XY cross sections in the transmission direction (Z direction) of the ultrasonic wave U. 2(c) and 2(d), the B-scan data is two-dimensional data (two-dimensional intensity data) in which the value of each voxel constituting the XZ cross section is expressed as the intensity of a flaw detection signal obtained from an echo reflected in a cross section (for example, the XZ cross section at Y=Y1 hatched in FIG. 2(c)) parallel to the transmission direction (Z direction) of the ultrasonic waves U. The scope data generating unit 32 generates this B-scan data for each of a plurality of discrete XZ cross sections in the direction (Y direction) orthogonal to the transmission direction of the ultrasonic waves U.

[0028] Figure 3 is a diagram illustrating how differences in echo diffusion patterns due to differences in defect morphology become apparent in multiple discrete C-scope data. Figure 3(a) shows an example of multiple C-scope data (three XY cross sections, Z=Z1, Z2, Z3) generated when a spherical defect F1 is present inside the test material S. Figure 3(b) shows an example of multiple C-scope data (three XY cross sections, Z=Z1, Z2, Z3) generated when a cylindrical defect F2 extending in the Y direction is present inside the test material S. Figure 3(c) shows an example of multiple C-scope data (three XY cross sections, Z=Z1, Z2, Z3) generated when a slit-shaped defect F3 tilted with respect to the Z direction is present inside the test material S. In the C-scope data shown in Figure 3, the darker the shaded areas, the greater the intensity of the inspection signal obtained from the echo. As can be seen from Figure 3, differences in defect morphology (different three-dimensional shapes in the example shown in Figure 3) cause differences in the diffusion pattern of echoes reflected from the defect, which become apparent in the discrete multiple C-scope data. The same is true for the discrete multiple B-scope data. Therefore, the morphology of the defect can be determined by the determination unit 33, which will be described later, using this discrete multiple C-scope data or B-scope data.

[0029] [Judgment section 33] The judgment unit 33 is equipped with one machine learning model that takes as input data obtained based on a plurality of discrete C-scope data or data obtained based on a plurality of discrete B-scope data, and outputs the form of a defect present inside the material S to be inspected, or another machine learning model that takes as input data obtained based on a plurality of discrete C-scope data or data obtained based on a plurality of discrete B-scope data, and outputs the degree of harmfulness based on the form of a defect present inside the material S to be inspected. As data obtained based on C-scope data, for example, the C-scope data itself is used. Similarly, as data obtained based on B-scope data, for example, the B-scope data itself is used. The defect morphology includes at least the three-dimensional shape of the defect, but may also include size, type, and position. One machine learning model is generated by machine learning using data obtained based on a plurality of known discrete C-scope data, etc., and the morphology of the known defect, such as the three-dimensional shape, as training data, and the other machine learning model is generated by machine learning using data obtained based on a plurality of known discrete C-scope data, etc., and the harmfulness level based on the morphology of the known defect, such as the three-dimensional shape, as training data.

[0030] Then, the determination unit 33 uses the above-mentioned machine learning model to determine whether defects present inside the material to be detected S are good or bad. Specifically, when data obtained based on a plurality of discrete C-scope data or the like of the material to be detected S to be determined is input, the determination unit 33 determines the harmfulness of the defects present inside the material to be detected S based on the defect form output from one machine learning model and the correspondence relationship between the defect form and the harmfulness (e.g., correspondence table data) between the defect form and the harmfulness. More specifically, for example, the determination unit 33 stores harmfulness degree values ​​corresponding to each form for each defect form (e.g., for each of the spherical defect F1, cylindrical defect F2, and slit-like defect F3 shown in FIG. 3 above) in the form of a table such as correspondence table data, and the determination unit 33 performs processing to convert the defect form output from one machine learning model into the corresponding harmfulness degree. A possible embodiment is that the judgment unit 33 stores a threshold value, and if the obtained harmfulness value is equal to or greater than the threshold value, the judgment unit 33 judges the defect to be bad (harmful), and if the obtained harmfulness value is less than the threshold value, the judgment unit 33 judges the defect to be good (harmless). Furthermore, when data obtained based on a plurality of discrete C-scope data of the material S to be inspected is input, the judgment unit 33 judges whether a defect present inside the material S to be inspected is good or bad based on the harmfulness level based on the shape of the defect output from another machine learning model. More specifically, for example, the judgment unit 33 may judge that the defect is bad (harmful) if the value of the harmfulness level based on the defect output from another machine learning model is equal to or greater than a pre-stored threshold value, and judge that the defect is good (harmless) if it is less than the threshold value. Furthermore, the determination unit 33 may be configured to include a plurality of machine learning models that output different defect morphologies (e.g., four types of machine learning models that output the three-dimensional shape, size, type, and position of the defect, respectively), or a plurality of other machine learning models that output different defect morphologies that serve as criteria for setting the harmfulness levels (e.g., four types of other machine learning models that use the three-dimensional shape, size, type, and position of the defect as criteria for setting the harmfulness levels), and the determination unit 33 may store threshold values ​​for the harmfulness levels obtained from the morphologies output from the plurality of machine learning models or the harmfulness levels output from the plurality of other machine learning models. In this case, a conceivable configuration is one in which the defect is determined to be bad (harmful) when the harmfulness level value based on any of the machine learning models is equal to or greater than the corresponding threshold, and the defect is determined to be good (harmless) when the harmfulness level values ​​based on all of the machine learning models are less than the corresponding threshold.

[0031] According to the ultrasonic flaw detection device 100 of this embodiment described above, by inputting data obtained based on multiple discrete C-scope data of the material to be detected S, in which differences in the diffusion patterns of echoes reflected by defects are apparent, into a machine learning model provided in the judgment unit 33, it is possible to accurately judge whether a defect is good or bad based on the shape of the defect.

[0032] The data obtained based on the C-scope data input to the machine learning model provided in the determination unit 33 is not limited to the C-scope data itself, and it is also possible to use first two-dimensional amplitude data (two-dimensional data in which an amplitude value on the frequency axis obtained by two-dimensional Fourier transform of the C-scope data is assigned to each voxel constituting the XY cross section) composed of amplitudes (amplitude values ​​on the frequency axis obtained by two-dimensional Fourier transform of the C-scope data) obtained by two-dimensional Fourier transform of the C-scope data, or first two-dimensional phase data (two-dimensional data in which a phase value on the frequency axis obtained by two-dimensional Fourier transform of the C-scope data is assigned to each voxel constituting the XY cross section) composed of phases (phase values ​​on the frequency axis) obtained by two-dimensional Fourier transform of the C-scope data. Similarly, the data obtained based on the B-scope data input to the machine learning model provided in the determination unit 33 is not limited to the B-scope data itself, but may also be second two-dimensional amplitude data (two-dimensional data in which an amplitude value on the frequency axis obtained by two-dimensional Fourier transform of the B-scope data is assigned to each voxel constituting the XZ cross section) composed of amplitudes (amplitude values ​​on the frequency axis obtained by two-dimensional Fourier transform of the B-scope data) obtained by two-dimensional Fourier transform of the B-scope data, or second two-dimensional phase data (two-dimensional data in which a phase value on the frequency axis obtained by two-dimensional Fourier transform of the B-scope data is assigned to each voxel constituting the XZ cross section) composed of phases (phase values ​​on the frequency axis) obtained by two-dimensional Fourier transform of the B-scope data).

[0033] Furthermore, it is also possible to use differential data as data obtained based on C-scope data or data obtained based on B-scope data that is input to the machine learning model included in the determination unit 33. By using differential data, it is possible to make the difference in the diffusion pattern of echoes reflected from defects even more apparent. That is, as data obtained based on the C-scope data input to the machine learning model, it is possible to use differential data of C-scope data adjacent to the transmission direction (Z direction) of the ultrasonic wave U (for example, differential data between the C-scope data at Z=Z1 and the C-scope data at Z=Z2 shown in FIG. 3, or differential data between the C-scope data at Z=Z2 and the C-scope data at Z=Z3), differential data of first two-dimensional amplitude data composed of amplitude obtained by subjecting C-scope data adjacent to the transmission direction of the ultrasonic wave U to a two-dimensional Fourier transform, or differential data of first two-dimensional phase data composed of phase obtained by subjecting C-scope data adjacent to the transmission direction of the ultrasonic wave U to a two-dimensional Fourier transform. Similarly, as data obtained based on the B-scope data input into the machine learning model, it is also possible to use differential data of B-scope data adjacent in a direction perpendicular to the transmission direction of the ultrasound U (Y direction), differential data of second two-dimensional amplitude data composed of amplitude obtained by two-dimensional Fourier transform of B-scope data adjacent in a direction perpendicular to the transmission direction of the ultrasound U, or differential data of second two-dimensional phase data composed of phase obtained by two-dimensional Fourier transform of B-scope data adjacent in a direction perpendicular to the transmission direction of the ultrasound U. [Explanation of symbols]

[0034] 1...Ultrasonic probe 2...Flaw detector 3. Processing unit 11. Oscillator 31. 3D voxel data generation unit 32 Scope data generation unit 33... Judgment section 100...Ultrasonic flaw detection equipment S...Material to be tested U...Ultrasonic

Claims

1. an ultrasonic probe disposed opposite to the material to be detected, transmitting ultrasonic waves toward the inside of the material to be detected and receiving echoes reflected from the inside of the material to be detected; a three-dimensional voxel data generating unit that generates three-dimensional voxel data of the test material, which is three-dimensional intensity data composed of the intensity of the flaw detection signal, based on the flaw detection signal obtained from the echo received by the ultrasonic probe; a scope data generating unit that generates C-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the material to be detected that are at least two voxels apart in the direction of transmission of the ultrasonic waves, based on the three-dimensional voxel data, or that generates B-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the material to be detected that are at least two voxels apart in a direction perpendicular to the direction of transmission of the ultrasonic waves; The present invention includes a machine learning model that receives data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data as an input and outputs the form of a defect present inside the material to be inspected, or a machine learning model that receives data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data as an input and outputs a harmfulness level based on the form of a defect present inside the material to be inspected, a determination unit that determines whether a defect present inside the material to be flaw-detected is good or bad using the machine learning model; An ultrasonic flaw detection device comprising:

2. the data obtained based on the C-scope data input to the machine learning model is the C-scope data, first two-dimensional amplitude data configured from amplitudes obtained by two-dimensional Fourier transform of the C-scope data, or first two-dimensional phase data configured from phases obtained by two-dimensional Fourier transform of the C-scope data; 2. The ultrasonic flaw detection device according to claim 1, wherein the data obtained based on the B-scope data input to the machine learning model is the B-scope data, second two-dimensional amplitude data composed of amplitudes obtained by performing a two-dimensional Fourier transform on the B-scope data, or second two-dimensional phase data composed of phases obtained by performing a two-dimensional Fourier transform on the B-scope data.

3. The data obtained based on the C-scope data input to the machine learning model is differential data of the C-scope data adjacent to the transmission direction of the ultrasonic waves, differential data of first two-dimensional amplitude data composed of amplitudes obtained by two-dimensional Fourier transform of the C-scope data adjacent to the transmission direction of the ultrasonic waves, or differential data of first two-dimensional phase data composed of phases obtained by two-dimensional Fourier transform of the C-scope data adjacent to the transmission direction of the ultrasonic waves, 2. The ultrasonic flaw detection device according to claim 1, wherein the data obtained based on the B-scope data input to the machine learning model is difference data of the B-scope data adjacent in a direction perpendicular to the transmission direction of the ultrasonic waves, difference data of second two-dimensional amplitude data composed of amplitudes obtained by performing a two-dimensional Fourier transform on the B-scope data adjacent in a direction perpendicular to the transmission direction of the ultrasonic waves, or difference data of second two-dimensional phase data composed of phases obtained by performing a two-dimensional Fourier transform on the B-scope data adjacent in a direction perpendicular to the transmission direction of the ultrasonic waves.

4. The ultrasonic flaw detector according to claim 1 , wherein the defect morphology includes at least a three-dimensional shape of the defect.

5. an ultrasonic transmitting / receiving step of placing an ultrasonic probe opposite to the material to be detected, transmitting ultrasonic waves from the ultrasonic probe toward the inside of the material to be detected, and receiving echoes reflected from the inside of the material to be detected by the ultrasonic probe; a three-dimensional voxel data generation step of generating three-dimensional voxel data of the test material, which is three-dimensional intensity data composed of the intensity of the flaw detection signal, based on the flaw detection signal obtained from the echo received by the ultrasonic probe; a scope data generation step of generating, based on the three-dimensional voxel data, C-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the material to be detected that are at least two voxels apart in the direction of transmission of the ultrasonic waves, or generating B-scope data, which is two-dimensional intensity data composed of the intensity of the flaw detection signal, for each of a plurality of different cross sections of the material to be detected that are at least two voxels apart in a direction perpendicular to the direction of transmission of the ultrasonic waves; a machine learning model that inputs data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data and outputs the form of a defect present inside the material to be inspected, or a machine learning model that inputs data obtained based on a plurality of C-scope data or data obtained based on a plurality of B-scope data and outputs the degree of harmfulness based on the form of a defect present inside the material to be inspected, a determination step of determining whether or not a defect present inside the test object is acceptable; The ultrasonic flaw detection method has the following features.

Images