JP7615067B2 - Ultrasound data evaluation system, ultrasound data evaluation method, and judgment model generation method - Google Patents

Description

Embodiments of the present invention relate to ultrasound data evaluation techniques.

Conventionally, ultrasonic inspection technology has been widely used for non-destructive inspection of structures. Inspectors detect the presence or absence of defects inside the structure from cross-sectional images obtained by ultrasonic inspection, and measure their location and depth. One method for measuring the depth of a defect is the tip echo method. This tip echo method measures the depth of a defect from the beam path and refraction angle of the echo from the upper or lower end of the defect. Inspectors make judgments based on their own specialized knowledge and experience, but evaluation results may vary depending on the skill of the inspector. Therefore, technology is being developed for automatic evaluation in ultrasonic inspection using signal processing or artificial intelligence (AI).

As a first example, a technique is known that uses ultrasound images as input for machine learning to detect defects. This technique determines whether an ultrasound image contains a defect, but it cannot identify the location and size of the defect in the image.

As a second example, a technique is known in which a flaw detection image is generated from an ultrasonic response waveform, and this flaw detection image is compared with a normal pattern image using a pattern recognition algorithm. With this technique, if a flaw detection image differs from a normal pattern image, the position of the defect can be detected by determining the shape echo and defect echo in the flaw detection image. With this technique, it is necessary to extract an image containing individual echoes based on the flaw detection image. However, if the image has a high noise intensity and the defect echo is unclear, or if the shape echo and the defect echo overlap, the extraction work may be difficult and there is a risk of overlooking the defect. In addition, the extracted echoes are classified into shape echo and defect echo, and training data is created. However, the determination of whether an echo is a defect echo or a shape echo is always made based on whether it is displayed or not, and there is a risk that echoes originating from the structure of the weld or inclusions may also be determined to be defect echoes.

In recent years, attempts have been made to detect objects in images using AI for object recognition, mainly in the fields of autonomous driving technology and medicine. One well-known object recognition technique is segmentation. Segmentation is a machine learning algorithm that performs class classification on images at the pixel level, and can detect objects at the pixel level. Supervised learning is used to build a learning model for this segmentation. For example, learning is performed using teacher labels that are colored for each object on a pixel-by-pixel basis. When segmentation is applied to defect detection in ultrasound images, it takes more effort and cost to create teacher labels than when it is applied to camera images. When comparing camera images and ultrasound images, camera images often have a clear boundary between the object and the background. On the other hand, ultrasound images contain noise echoes such as scattered waves generated from weld metal in addition to ultrasonic echoes from objects. Therefore, ultrasound images often have unclear boundaries between objects and other things compared to general camera images. In other words, it is very difficult to distinguish objects on a pixel-by-pixel basis in ultrasound images.

Several methods have been proposed to automate the creation of teacher labels in the segmentation of ultrasound images. For example, there is a method in which a threshold is set to separate echoes to be detected from others. However, with this method, separation using a threshold becomes difficult when the noise intensity is high and the defect echo is unclear in the image, or when shape echoes and defect echoes overlap. There is also a method that separates defect echoes from noise echoes based on whether they are always displayed. However, with this method, there is a risk that echoes originating from the structure or inclusions of the weld will also be classified as defect echoes.

Currently, the mainstream segmentation method is to perform segmentation on a single image. With this method, an image is input into a learning model, and an image segmented for each object is output. In contrast to this, there is a method that performs segmentation on multiple continuous images, such as a video.

On the other hand, in actual ultrasonic flaw detection, an inspector scans the image of interest manually or mechanically with an ultrasonic probe, and identifies defects by reading how the brightness and shape of the defect echo in the image changes with the scan. For example, when an inspector moves the ultrasonic probe closer to or farther away from the defect position, the intensity of the defect echo changes accordingly. Also, when an ultrasonic probe is scanned so that ultrasonic waves are incident on the defect from various angles, it is not possible to determine the type of defect as in an X-ray transmission image, but it is possible to determine whether the defect is a spherical defect or a planar defect. The echo on the screen changes in response to any scan of the ultrasonic probe, and by reading the amount of change, the inspector can determine whether the echo is from a defect or not. On the other hand, if an echo appears at the same position on the screen even when the ultrasonic probe is scanned, it is highly likely that the echo on the screen is not from a defect, but is from the surface or bottom of the object.

When segmenting defect areas in a single image, it is not possible to make a judgment that takes into account changes in defect echoes in response to the scanning of an ultrasonic probe, making it difficult to distinguish between noise and defect echoes. In addition, when applying segmentation to detect defects in ultrasonic images, creating training labels requires more effort and cost than applying it to general camera images. In addition, when detecting defects using ultrasonic images as input for machine learning, it is difficult to automatically classify defect echoes from echoes originating from the structure or inclusions of the weld.

Special Publication No. 2020-503509 International Publication No. 2015/001624

The problem that the present invention aims to solve is to provide an ultrasonic data evaluation technique that can improve the detection accuracy of the detection target position contained in an object when evaluating ultrasonic data used in non-destructive testing.

An ultrasonic data evaluation system according to an embodiment of the present invention includes an ultrasonic probe that performs flaw detection on a detection target included in an object to be inspected and acquires ultrasonic data including at least one of an image or voxel data , a position information acquisition device that acquires flaw detection position data indicating the position of the ultrasonic probe when the ultrasonic data was acquired , and a computer that inputs at least one of the ultrasonic data to be judged and the flaw detection position data into a machine -learned judgment model and outputs a judgment value that can determine whether or not the detection target exists for each unit indicating a position in the ultrasonic data, and the judgment model has been machine-learned using a plurality of training ultrasonic data, training flaw detection position data, and teacher labels, the training ultrasonic data includes at least one of the ultrasonic data or data generated by simulating it, the training flaw detection position data includes at least one of the flaw detection position data or data generated by simulating it, and the teacher label sets a specific common area indicating a specific area of the detection target present in the training ultrasonic data as the position of the detection target .

Embodiments of the present invention provide an ultrasonic data evaluation technique that can improve the detection accuracy of the detection target position contained in an object when evaluating ultrasonic data used in non-destructive testing.

FIG. 1 is a block diagram illustrating an ultrasound data evaluation system. FIG. 1 is a block diagram showing an ultrasonic image evaluation device. FIG. FIG. 2 is a perspective view showing a scanning mode of an ultrasonic probe. FIG. FIG. FIG. 4 is a plan view showing an ultrasonic probe when scanning a spherical defect. FIG. 13 is a plan view showing an ultrasonic probe when scanning a spherical defect by oscillating scanning. FIG. 4 is a plan view showing an ultrasonic probe when scanning a planar defect. FIG. 11 is a plan view showing an ultrasonic probe when scanning a planar defect by oscillating scanning. FIG. 4 is a plan view showing a mode in which the ultrasonic probe is scanned in the X direction. FIG. 4 is a side view showing a mode in which the ultrasonic probe is scanned in the X direction. FIG. 4 is an image diagram showing an ultrasound image acquired by scanning an ultrasound probe in the X direction. FIG. 4 is a plan view showing a mode in which the ultrasonic probe is scanned in the Y direction. FIG. 4 is a side view showing a mode in which the ultrasonic probe is scanned in the Y direction. FIG. 4 is an image diagram showing an ultrasound image acquired by scanning an ultrasound probe in the Y direction. FIG. 4 is a plan view showing a mode in which an ultrasonic probe is subjected to zigzag scanning. FIG. 4 is a side view showing a mode in which the ultrasonic probe is subjected to zigzag scanning. FIG. 2 is an image diagram showing an ultrasound image acquired by zigzag scanning with an ultrasound probe. FIG. 4 is a plan view showing a mode in which the ultrasonic probe is oscillated for scanning. FIG. 4 is a side view showing a mode in which the ultrasonic probe is oscillated for scanning. FIG. 4 is an image diagram showing an ultrasound image acquired by performing oscillating scanning with an ultrasound probe. FIG. 13 is an image diagram showing a learning ultrasound image. FIG. 11 is an image showing teacher labels. FIG. 4 is an explanatory diagram showing a flow of selecting a channel from a learning ultrasound image. 6 is a graph showing a received waveform of an echo that detects a detection target. 6 is a graph showing the absolute value of the received waveform of an echo that detected a detection target. FIG. 13 is an explanatory diagram showing pixels of teacher labels. FIG. 11 is an image diagram showing an ultrasonic image for determination. FIG. FIG. 11 is an image diagram showing a binarized determined image. FIG. 13 is an explanatory diagram showing an image that appears in a determined image. FIG. 11 is an explanatory diagram showing the coordinates of a specific common area. FIG. 4 is a plan view showing a mode in which an ultrasonic probe is subjected to zigzag scanning. 11 is a graph showing position information in the X direction and the Y direction and a determination certainty degree. FIG. 4 is a plan view showing a mode in which a planar defect is oscillated and scanned with an ultrasonic probe. 13 is a graph showing the rotation angle of an ultrasonic probe scanning a planar defect and the determination certainty. FIG. 4 is a plan view showing a manner in which a spherical defect is scanned by a swiveling ultrasonic probe. 13 is a graph showing the rotation angle of an ultrasonic probe scanning a spherical defect and the degree of certainty of the determination. FIG. 1 is a perspective view showing a 3D CAD model on which the positions of detection targets are plotted. 1 is a flowchart showing a method for generating a teacher label. 11 is a flowchart showing a method for generating a determination model. 13 is a flowchart showing a method for evaluating a determination ultrasound image.

Below, embodiments of the ultrasound data evaluation system, ultrasound data evaluation method, and judgment model generation method are described in detail with reference to the drawings.

In this embodiment, an ultrasonic inspection method generally called a phased array ultrasonic inspection is used as an example. Among them, a linear scan inspection method in which an ultrasonic element is electronically scanned while forming an ultrasonic beam in a fixed direction, a sector scan inspection method in which an angle at which an ultrasonic beam is formed is changed in a fan shape while the driven ultrasonic element is fixed or electronically scanned, a total focusing method (TFM) in which a beam is focused by comprehensively setting a focus in an arbitrary coordinate region, or an aperture synthesis method or other imaging method using ultrasonic waves can be used. Note that an ultrasonic inspection method in which scanning is performed manually or mechanically using a single probe may also be used. In the following explanation, an example of an inspection image obtained by a typical linear scan inspection method is shown.

Reference numeral 1 in FIG. 1 denotes an ultrasonic data evaluation system according to this embodiment. This ultrasonic data evaluation system 1 evaluates the presence or absence of a defect (detection target K) in an object M to be inspected. In this embodiment, ultrasonic data acquired by a predetermined ultrasonic flaw detection method is automatically analyzed using a computer equipped with artificial intelligence (AI) to evaluate the presence or absence of a defect in the object M. By evaluating the presence or absence of this defect, it is possible to utilize the result for the diagnosis of the soundness of the object M. In other words, the ultrasonic data evaluation system 1 analyzes the object M using artificial intelligence obtained by machine learning.

The term "ultrasound data" that is the subject of processing in this embodiment includes at least one of an image (two-dimensional data, pixel data) and voxel data (three-dimensional volume data). For example, the ultrasound data acquired by the ultrasonic flaw detection method and the ultrasound data that is the subject of processing during judgment are at least one of an image and voxel data. Furthermore, the ultrasound data (image or voxel data) may be either a still image or a video. Note that when the ultrasound data is a video of voxel data, it can be treated as three-dimensional data plus fourth-dimensional data indicating the passage of time. In the following explanation, to facilitate understanding, ultrasound data of a still image (pixel data) is exemplified.

The ultrasonic data evaluation system 1 includes an array probe 10, which is an example of an ultrasonic probe, an ultrasonic flaw detection device 20, an ultrasonic image processing device 30, a position information acquisition device 90, and an ultrasonic image evaluation device 100.

The ultrasonic flaw detection device 20, the ultrasonic image processing device 30, the position information acquisition device 90, and the ultrasonic image evaluation device 100 are configured as a computer having hardware resources such as a CPU, ROM, RAM, and HDD, and in which software-based information processing is realized using the hardware resources as the CPU executes various programs. Furthermore, the ultrasonic data evaluation method of this embodiment is realized by having the computer execute various programs.

The components of the ultrasound data evaluation system 1 do not necessarily need to be provided on multiple computers. For example, the components of the ultrasound data evaluation system 1 may be provided on a single computer.

As shown in FIG. 3, the array probe 10 emits ultrasonic waves U to the object M to be inspected and detects the reflected waves R reflected by the object K. In the following description, the reflected waves R may be referred to as echoes. The array probe 10 includes a probe 11 in which a plurality of ultrasonic elements 11A that emit ultrasonic waves U are arranged. These ultrasonic elements 11A are arranged linearly in a direction perpendicular to the scanning direction of the array probe 10. The ultrasonic image (two-dimensional ultrasonic data) acquired by the array probe 10 is an image showing a cross section of the object M. In this embodiment, the array probe 10 is shown as an example of an ultrasonic probe, but it is of course applicable to ultrasonic probes arranged two-dimensionally or single probes.

An example of the object M is a component made of a metal material. In particular, the detection object K occurring at or near the weld W (Figure 4) is evaluated by the ultrasonic data evaluation system 1. Note that the object M may also be a material other than a metal material, for example, a composite material such as reinforced concrete, mortar, or fiber-reinforced plastic. The detection object K occurring inside or on the surface of such an object M is evaluated by the ultrasonic data evaluation system 1.

The detection target K can be arbitrarily determined by the user of the ultrasonic data evaluation system 1. The detection target K (defect) can be arbitrarily determined to be a crack, flaw, cavity, round hole, peeling, thinning, inclusion, etc. occurring inside or on the surface of the object M.

The array probe 10 is controlled by applying a voltage to it by the ultrasonic inspection device 20. This array probe 10 exemplifies a phased array ultrasonic inspection method in which a beam of ultrasonic waves U is formed in a fixed direction while electronically scanning the ultrasonic element 11A to be driven. Note that this embodiment can be applied to other methods. For example, it may be applied to a sector imaging method, a TFM, or an aperture synthesis method. Furthermore, this embodiment may be applied to an ultrasonic inspection method in which mechanical scanning is performed using a single probe as an ultrasonic probe.

When inspecting the object M, an acoustic propagation medium 2 called a wedge is provided between the array probe 10 and the object M. This acoustic propagation medium 2 is used to make the ultrasonic waves U incident on the object M at a highly directional angle.

As the acoustic propagation medium 2, an isotropic material is used that allows the ultrasonic waves U to propagate and has a known acoustic impedance. If the surface of the target object M is flat, it is not necessary to use the acoustic propagation medium 2.

Isotropic materials used as the acoustic propagation medium 2 include, for example, acrylic, polyimide, gel, and other polymers. The acoustic propagation medium 2 may be made of a material that has an acoustic impedance similar to or the same as that of the front panel (not shown) of the ultrasonic element 11A. It may also be made of a material that has an acoustic impedance similar to or the same as that of the object M. A composite material that changes the acoustic impedance stepwise or gradually may also be used.

In addition, to prevent multiple reflected waves inside the acoustic propagation medium 2 from affecting the flaw detection results, damping material may be placed inside and outside the acoustic propagation medium 2. A mountain-shaped wave-eliminating shape may also be provided. A multiple reflection reduction mechanism may also be provided.

In the following explanation, the acoustic propagation medium 2 may be omitted when explaining how ultrasonic waves U are incident on the target M from the array probe 10.

An acoustic contact medium (not shown) for propagating ultrasonic waves U is used at the contact portion of the path from the array probe 10 to the object M. For example, when the acoustic propagation medium 2 is used, an acoustic contact medium (not shown) is used at the contact portion between the array probe 10 and the acoustic propagation medium 2, and at the contact portion between the acoustic propagation medium 2 and the object M. When the acoustic propagation medium 2 is not used, an acoustic contact medium (not shown) is used at the contact portion between the array probe 10 and the object M. For this acoustic contact medium, a medium capable of propagating ultrasonic waves U, such as water, glycerin, machine oil, castor oil, acrylic, polystyrene, or gel, is used.

As shown in FIG. 1, a detection signal containing information on the reflected wave R detected by the array probe 10 is input to an ultrasonic flaw detector 20. The detection signal obtained by the ultrasonic flaw detector 20 is then processed by an ultrasonic image processing device 30.

The ultrasonic image processing device 30 acquires ultrasonic images using ultrasonic waves U to detect flaws in a detection target K contained in an object M. This ultrasonic image processing device 30 can generate ultrasonic images along the scanning direction of the ultrasonic element 11A based on the reflected waves of the ultrasonic waves U.

In this embodiment, the position (displacement) and rotation angle θ (Figure 5) of the array probe 10 when scanning are obtained, and an evaluation is made of how the brightness or shape of the detection target K (defect echo) changes according to the position or angle of the array probe 10. By performing this evaluation, the position and depth of the detection target K can be more reliably identified in the ultrasound image. Note that a machine learning judgment model is generated by combining two types of data, the ultrasound image and information indicating the position of the array probe 10 when the ultrasound image was acquired.

In conventional technology, the detection target K is determined using only image data. In contrast, in this embodiment, the determination is made using two types of data: image and position information. In this way, the detection target K can be determined using information that an inspector actually uses in ultrasonic flaw detection, and the accuracy is improved dramatically.

For example, as shown in FIG. 7, assume that the detection target K is a spherical defect. Ultrasonic waves U are emitted from the array probe 10 toward the detection target K, and the reflected waves R reflected by the detection target K are detected. Next, as shown in FIG. 8, the position and orientation of the array probe 10 are changed, and ultrasonic waves U are emitted toward the detection target K from a different direction. Here, if the detection target K is a spherical defect, the intensity of the reflected waves R does not change even if the position and orientation of the array probe 10 is changed.

On the other hand, when the detection target K is a planar defect, changing the position and orientation of the array probe 10 changes the direction in which the reflected wave R is reflected. In other words, the intensity of the reflected wave R changes. For example, as shown in FIG. 9, when ultrasonic waves U are incident perpendicularly to the surface on which the detection target K extends, the reflected wave R becomes stronger. On the other hand, as shown in FIG. 10, when ultrasonic waves U are incident obliquely to the surface on which the detection target K extends, the reflected wave R becomes weaker. In this way, by grasping the change in the reflected wave R when the array probe 10 is moved, it is possible to determine whether the detection target K is a spherical defect or a planar defect. When detecting internal defects in the weld W, the main spherical defects expected to be found include blowholes, and the main planar defects include slag inclusion, poor fusion, poor penetration, and cracks.

As shown in FIG. 1, the position information acquisition device 90 includes a displacement acquisition unit 91 that acquires the position of the array probe 10, and an angle acquisition unit 92 that acquires the rotation angle θ of the array probe 10. This position information acquisition device 90 acquires flaw detection position data including the position (displacement) and rotation angle θ (FIG. 5) of the array probe 10 when the target M is scanned with the array probe 10.

In this embodiment, for example, an XY axis scanner with an encoder is used as the displacement acquisition unit 91. Also, an angle sensor with an encoder is used as the angle acquisition unit 92.

The displacement acquisition unit 91 and the angle acquisition unit 92 are not limited to this, and may be a combination of multiple types of position information acquisition methods, or other position information acquisition methods may be used. For example, when the array probe 10 is attached to the end of a robot arm (not shown), the position (displacement) and rotation angle θ of the array probe 10 may be acquired based on information indicating control of the position and rotation of the array probe 10. The position and rotation angle θ of the array probe 10 may also be acquired based on image information acquired by a specified camera (not shown) that captures the array probe 10.

The inspection position data includes at least two coordinate values expressed in Cartesian coordinates that represent the position of the array probe 10. In this way, when the object M is flat, it becomes easier to represent the position of the array probe 10 on the surface of the object M. The Cartesian coordinates are, for example, a two-dimensional coordinate system that follows the surface of the flat object M.

The inspection position data may include at least two coordinate values represented by polar coordinates or cylindrical coordinates that represent the position of the array probe 10. In this way, when the object M is spherical, cylindrical, or cylindrical, it becomes easier to represent the position of the array probe 10 on the surface of the object M. Polar coordinates are, for example, a two-dimensional coordinate system along the surface of the object M that is spherical, cylindrical, or cylindrical. For example, the inspection position data includes a radius and a declination that are represented by polar coordinates. The inspection position data also includes an angular position and an axial position that are represented by cylindrical coordinates.

The flaw detection position data also includes an angle representing the rotation of the array probe 10. In this way, the orientation of the array probe 10 can be reflected in the determination of the detection target K.

As shown in FIG. 4, the two directions along the surface of the object M are represented by the X-axis and the Y-axis, and the depth direction of the object M is represented by the Z-axis. In FIG. 4, a flat object M is shown as an example.

The position (displacement) of the array probe 10 is, for example, information indicating the XY coordinates or changes of the array probe 10 moving along the surface of the object M. The position of the array probe 10 is a position with a predetermined position as the origin of the XY coordinates. The position of the array probe 10 may also be the amount of change in the XY coordinates that change at regular intervals.

The rotation angle θ of the array probe 10 is information indicating the orientation of the array probe 10 rotating along the surface of the object M. In other words, the rotation angle θ is the angle when the array probe 10 rotates around an axis that extends perpendicularly from the surface of the object M as the center Q ( FIG. 5 ). The rotation angle θ is the rotation angle θ of the array probe 10 with an angle along a predetermined direction (e.g., the Y-axis direction) as the reference θ ₀ ( FIG. 20 ). The rotation angle θ may be the amount of change in the rotation angle θ of the array probe 10 that changes at regular intervals.

As shown in FIG. 6, the position information acquisition device 90 acquires a group of position information, which is a collection of flaw detection position data. In FIG. 6, the group of position information is illustrated as a predetermined table. For example, when an ultrasound image is acquired, an acquisition number is assigned as identification information that allows each ultrasound image to be individually identified. Various information is registered using these acquisition numbers as primary keys. For example, the X coordinate, Y coordinate, and rotation angle θ are registered in association with the acquisition number.

Next, the manner in which the array probe 10 is scanned in the X direction will be described with reference to Figs. 11 to 13. For example, assume that a weld W extending linearly is provided on the object M (Figs. 11 and 12). Assume that a defect such as a crack occurring near the weld W along the weld W is the detection object K.

As shown in Figures 11 and 12, when the array probe 10 is scanned in the X direction (linearly) along the surface of the object M and along the direction in which the weld W extends, ultrasonic images showing the cross-section of the object M at each position are obtained. For example, ultrasonic images ( _Xn to Xn ₊₄ ) are obtained corresponding to each position ( _Xn to Xn ₊₄ ) of the object M (Figure 13). These ultrasonic images ( _Xn to Xn ₊₄ ) are the learning ultrasonic images 31 or the judgment ultrasonic images 35. This scanning is called left-right scanning. It is used to estimate the length of the detection object K in the scanning direction.

13, for example, in the first ultrasound image ( _Xn ), the outgoing light (echo) of the detection target K is not displayed. However, in the next ultrasound image (Xn ₊₁ ) and the next ultrasound image ( _Xn+2 ), the outgoing light of the detection target K gradually increases. Furthermore, when approaching the end of the detection target K, the outgoing light is displayed weakly in the ultrasound image ( _Xn+4 ). This is a change in intensity specific to the detection target K.

Next, the manner in which the array probe 10 is scanned in the Y direction will be described with reference to Figs. 14 to 16. For example, assume that a weld W extending linearly is provided on the object M (Figs. 14 and 15). Assume that a defect such as a crack occurring along and in the vicinity of this weld W is the detection object K.

14 and 15, when the array probe 10 is scanned in the Y direction (linearly) along the surface of the object M and along the direction in which the weld W extends, ultrasonic images showing the cross sections of the object M at each position are obtained. For example, ultrasonic images (Y _n to Y _n+4 ) are obtained corresponding to each position (Y _n to Y _n+4 ) of the object M ( FIG. 16 ). These ultrasonic images (Y _n to Y _n+4 ) are the learning ultrasonic images 31 or the determination ultrasonic images 35.

16, for example, the echoes (echoes) of the detection target K are displayed from the first ultrasound image (Y _n ) to the last ultrasound image (Y _n+4 ). However, the positions of the echoes of the detection target K displayed in each ultrasound image (Y _n to Y _n+4 ) differ depending on the position of the array probe 10 in the Y direction. This scanning is called forward and backward scanning. The position of the detection target K in the Y direction is estimated by measuring the distance from the initial position Y ₀ to Y _n .

Next, the manner in which the array probe 10 is scanned in a zigzag manner will be described with reference to Figs. 17 to 19. For example, assume that a weld W extending linearly is provided on the object M (Figs. 17 and 18). Assume that a defect such as a crack occurring along and in the vicinity of this weld W is the detection object K.

As shown in FIG. 17, the array probe 10 is moved in the weld line direction (X direction) while scanning back and forth. This scanning method is called zigzag scanning because the trajectory of the array probe 10 follows a zigzag course in a planar view. This zigzag scanning is used for rough flaw detection (preliminary flaw detection).

In this zigzag scanning, the position ( _Xn , _Yn ) of the array probe 10 is changed while the rotation angle (θ) of the array probe 10 is fixed to obtain ultrasound images. These ultrasound images ( _Xn , _Yn , _θn ) are the learning ultrasound images 31 or the determination ultrasound images 35.

19, an ultrasound image ( _Xn , _Yn , _θ0 ) acquired when the array probe 10 is at a predetermined position displays an echo (echo) of the detection target K. However, other ultrasound images do not display the echo of the detection target K.

Next, the manner in which the array probe 10 is swivelled for scanning will be described with reference to Figs. 20 to 22. For example, assume that a weld W extending linearly is provided on the object M (Figs. 20 and 21). Assume that a defect such as a crack occurring near the weld W along the weld W is the detection object K.

As shown in Figure 20, after rough flaw detection by zigzag scanning, a scan is performed at a specific position on the XY plane where the detection target K is obtained, by swinging the array probe 10 to change the incident angle of the ultrasonic waves U. This scanning method is called swing scanning (pendulum scanning). This swing scanning is performed to more precisely estimate the shape, dimensions, etc. of the detection target K inside the object M. Changing the incident angle of the ultrasonic waves U changes the intensity of the reflected waves R (Figures 7 to 10) of the detection target K.

For example, it is possible to determine whether the detection target K is a spherical defect such as a blowhole, or a planar defect such as a crack. For example, in the case of a spherical defect, no change in the intensity of the reflected wave R is observed even when a swiveling scan is performed (Figures 7 and 8). On the other hand, in the case of a planar defect such as a crack, the reflected wave R cannot be received unless the ultrasonic wave U is incident perpendicular to the direction in which the crack runs (Figures 9 and 10). Therefore, when a swiveling scan is performed, a change in the intensity of the reflected wave R is observed. Furthermore, with swiveling scanning, in the case of a planar defect, the angle at which the crack runs relative to the weld line direction can be roughly examined.

In this embodiment, each time an ultrasound image is acquired, the position information acquisition device 90 acquires flaw detection position data indicating the position and angle of the array probe 10 at that time. For example, three parameters including the two-dimensional plane coordinate values XY and the rotation angle θ of the array probe 10 are acquired for each ultrasound image.

The timing of acquiring ultrasound images may be at equal spatial intervals or at irregular intervals. The timing of acquisition may also be specified in terms of time.

As shown in FIG. 5, the rotation angle of the array probe 10 acquired by the position information acquisition device 90 is expressed as θ = ∠AOB on the XY plane with the center Q of the array probe 10 as the reference.

Figure 23 illustrates an example of an ultrasound image (ultrasound data) as a typical linear scan inspection image. This ultrasound image is used as a training ultrasound image 31 as the training ultrasound data of this embodiment. In this training ultrasound image 31, the background B is displayed in black, the incidence range 32 where the ultrasound U is incident on the object M is displayed in gray, the surface 33 of the object M is displayed as a white line, and the detection object K is displayed in white. Note that this embodiment can also be implemented with an ultrasound image that is a color image represented by RGB or the like, but for clarity of explanation, the ultrasound image will be described as a grayscale image. In the example of Figure 23, one detection object K is shown in the training ultrasound image 31. Note that the ultrasound image is also used as a judgment ultrasound image 35 (Figure 29) as the judgment ultrasound data of this embodiment.

As shown in FIG. 3, the ultrasonic waves U incident on the object M from the array probe 10 have their angles of incidence and refraction determined according to the delay time, and propagate through the inside of the object M. The propagated ultrasonic waves U are reflected or scattered by the detection object K and the like, and reach the array probe 10 again. In addition to the detection object K, the ultrasonic waves U are also reflected or scattered at the bottom surface or corners of the object M, and reach the array probe 10.

The reflected waves (scattered waves) that reach the array probe 10 are visualized based on the arrival time and the information of the ultrasonic elements 11A. At this time, if a detection target K is present in the object M, an echo (echo) of the detection target K appears in the training ultrasonic image 31.

In this embodiment, the ultrasound image processing device 30 generates a training ultrasound image 31 (FIG. 23) or a judgment ultrasound image 35 (FIG. 29). The training ultrasound image 31 or the judgment ultrasound image 35 is input from the ultrasound image processing device 30 to the ultrasound image evaluation device 100.

The training ultrasonic image 31 of this embodiment may be a CG image generated using computer graphics (CG) to imitate an ultrasonic image obtained by the ultrasonic flaw detection device 20. Also, a wide variety of training ultrasonic images 31 may be generated by manually or automatically editing the ultrasonic image obtained by the ultrasonic flaw detection device 20.

Furthermore, an ultrasound image obtained by a numerical simulation of ultrasound propagation may be used. For example, an ultrasound image may be generated using a simulation technique that visualizes the propagation of ultrasound U inside the object M, and the generated ultrasound image may be used as the training ultrasound image 31.

Furthermore, when using a simulated ultrasound image, the flaw detection position data is also simulated. For example, the positional relationship between the detection target K and the array probe 10 in a virtual three-dimensional space is simulated, and the position and rotation angle θ of the array probe 10 in this three-dimensional space are generated as flaw detection position data.

As shown in FIG. 2, the ultrasound image evaluation device 100 performs machine learning using a training ultrasound image 31, and identifies the position of the detection target K contained in the judgment ultrasound image 35 used for judgment based on the results of this machine learning.

In other words, analysis techniques based on artificial intelligence learning can be used for the computer-based analysis of this embodiment. For example, learning models generated by machine learning using neural networks, learning models generated by other machine learning, deep learning algorithms, regression analysis, and other mathematical algorithms can be used.

The system of this embodiment includes a computer equipped with artificial intelligence that performs machine learning. For example, the system may be configured with one computer equipped with a neural network, or may be configured with multiple computers equipped with neural networks.

Here, a neural network is a mathematical model that represents the characteristics of brain functions through computer simulation. For example, this model shows that artificial neurons (nodes) that form a network through synaptic connections change the strength of synaptic connections through learning, and acquire problem-solving abilities. Furthermore, neural networks acquire problem-solving abilities through deep learning.

In addition, deep reinforcement learning may be used in the neural network, in which a reward function is set for each information item to be learned, and the information item with the highest value is extracted based on the reward function.

In this embodiment, a supervised machine learning method is used. For example, semantic segmentation is used. However, this is not limited to this, and multiple types of machine learning methods may be combined, or any one machine learning method may be adopted. For example, supervised learning that handles time-series data, such as a recurrent neural network or a recurrent neural network with a long short-term memory structure, may be used.

Other machine learning methods that can be used besides semantic segmentation include the Fully Convolutional Neural Network, R-CNN (Region Convolutional Neural Network) which has a proven track record in image recognition, k-nearest neighbors, logistic regression, support vector machine (SVM), and deep learning.

Note that there are various deep learning techniques, such as autoencoder, RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), and GAN (Generative Adversarial Network). These techniques may be applied to this embodiment, or other machine learning techniques may be used.

As shown in FIG. 29, the ultrasonic image for determination 35 has a background B displayed in black, an incident range 32 where ultrasonic waves U are incident on the object M displayed in gray, a surface 33 of the object M displayed as a white line, and a detection object K displayed in white. In the example of FIG. 29, two detection objects K are shown in the ultrasonic image for determination 35. Then, a determined image 36 is generated from this ultrasonic image for determination 35 as determined data using a determination model.

As shown in FIG. 30, in the determined image 36, the background B is displayed in black, and the pixels in an area 37 where the edge of the detection target K is more likely to exist are displayed brighter. In other words, the pixels are displayed in colors ranging from gray to white. In the example of FIG. 30, the area 37 where two edges are more likely to exist is shown in the determined image 36. A binary image 38 is then generated from this determined image 36.

As shown in FIG. 31, in the binarized image 38, the background B is displayed in black, and the specific common area 39 where the edge of the detection target K is located is displayed in white. In the example of FIG. 31, two specific common areas 39 are shown in the binarized image 38. The position and depth of the detection target K can then be determined from this binarized image 38.

Next, the system configuration of the ultrasound image evaluation device 100 will be described with reference to the block diagram shown in FIG. 2. Furthermore, the ultrasound image evaluation device 100 may include components other than those shown in FIG. 2, and some of the components shown in FIG. 2 may be omitted.

The ultrasound image evaluation device 100 includes an input unit 110, a calculation unit 120, a judgment model generation unit 130, a memory unit 140, and an output unit 150.

An ultrasonic image obtained by irradiating ultrasonic waves U on an object M is input from the outside to the input unit 110. In this embodiment, an example is shown in which an ultrasonic image is input from an ultrasonic image processing device 30, but an ultrasonic image may be input from another device. Also, an example is shown in which flaw detection position data is input from a position information acquisition device 90, but flaw detection position data may be input from another device.

The input unit 110 includes an image input unit 111 and a position information input unit 112. An ultrasound image (ultrasound data) is input to the image input unit 111 from the ultrasound image processing device 30. Flaw detection position data is input to the position information input unit 112 from the position information acquisition device 90.

There are two types of ultrasound images input to the image input unit 111: a training ultrasound image 31 (FIG. 23) used for machine learning and a judgment ultrasound image 35 (FIG. 29). This image input unit 111 serves as the ultrasound data acquisition unit of this embodiment.

In this embodiment, when machine learning is performed, multiple training ultrasound images 31 and the corresponding flaw detection position data are input to the input unit 110. Then, a group of teacher labels is generated based on this group of training ultrasound images. Furthermore, a judgment model is generated using the group of training ultrasound images, the group of teacher labels, and the group of flaw detection position data (group of position information). On the other hand, when the detection target K contained in the judgment ultrasound image 35 is judged using this judgment model, the judgment ultrasound image 35 and the corresponding flaw detection position data are input to the input unit 110. Note that the number of judgment ultrasound images 35 input may be one or multiple.

The machine-learned judgment model includes an input layer to which ultrasonic data and flaw detection position data are input, an output layer that outputs a judgment value, and an intermediate layer in which parameters are learned using training data (teacher data) including ultrasonic data, flaw detection position data, and teacher labels. This judgment model causes a computer to function by inputting ultrasonic data and flaw detection position data to the input layer, performing calculations in the intermediate layer, and outputting a judgment value from the output layer.

As the learning data, at least one of ultrasonic data obtained using the array probe 10 or ultrasonic data simulating this, and at least one of flaw detection position data of the array probe 10 when the ultrasonic data was acquired or flaw detection position data simulating this are used.

There are two types of flaw detection position data input to the position information input unit 112: flaw detection position data for learning used in machine learning, and flaw detection position data for judgment. This position information input unit 112 serves as the flaw detection position data acquisition unit of this embodiment.

The calculation unit 120 includes a teacher label generation unit 121, a judgment unit 122, an extraction unit 123, a detection target position calculation unit 124, and a judgment confidence level calculation unit 125. These are realized by the CPU executing a program stored in the memory or HDD.

The storage unit 140 stores images obtained by ultrasonic inspection, inspection position data, judgment models, etc. This storage unit 140 includes a learning ultrasound image storage unit 141, a learning position information storage unit 142, a teacher label storage unit 143, a judgment model storage unit 144, a judgment ultrasound image storage unit 145, a judgment position information storage unit 146, and a detection target position storage unit 147. These are a collection of information stored in memory, HDD, or cloud, and organized so that they can be searched or accumulated.

The output unit 150 outputs predetermined information. For example, an image showing the position of the detection target K in the evaluation ultrasound image 35 is displayed. This output unit 150 may be anything that can display digital data, such as a PC display, television, projector, or head-mounted display. It may also be a device that first converts the signal into an analog signal, such as a cathode ray tube, and then displays the image on the screen. In other words, the ultrasound image evaluation device 100 includes a device that displays an image that outputs the analysis result. Furthermore, the output unit 150 may be a printer that prints predetermined information on a paper medium. The output unit 150 may also be a device that writes predetermined information to a removable and portable storage medium, such as a USB memory.

The output unit 150 may display the ultrasonic echo composite signal, the visualization results, the coordinates of the array probe 10 and its relative position to the detection target K, the delay time, the focal depth, the inspection conditions such as the inspection refraction angle, etc. The output unit 150 may also have a user interface function that generates an alarm by sound or light emission according to the set conditions, or inputs operations as a touch panel.

The ultrasound image evaluation device 100 may have other configurations. For example, it may have a communication unit (not shown). This communication unit communicates with other computers via a communication line such as the Internet.

The components of the ultrasound image evaluation device 100 do not necessarily have to be provided on a single computer. For example, a single ultrasound image evaluation device 100 may be realized using multiple computers connected to each other via a network.

The teacher label generation unit 121 generates a teacher label 40 in which a specific common area 39 (FIG. 31) indicating a specific area of the detection target K present in each training ultrasound image 31 is set as the position of the detection target K. One teacher label 40 is generated corresponding to one training ultrasound image 31. Then, a group of teacher labels corresponding to the group of training ultrasound images is generated.

As shown in FIG. 24, the teacher label 40 is an image (two-dimensional data) of the same size as the corresponding training ultrasound image 31. In other words, the teacher label 40 is an image having the same number of pixels as the training ultrasound image 31. In this teacher label 40, the background B is displayed in black, and the specific common area 39 of the detection target K is displayed in white. In other words, the teacher label 40 is a binary image in which the pixel value of the specific common area 39 is "1" and the other pixel values are "0". The position (coordinates) of the specific common area 39 in this teacher label 40 corresponds to the position of the upper end of the detection target K shown in the training ultrasound image 31.

A specific common area 39 is set in the teacher label 40, which corresponds to the part where the echo of the detection target K is present. The specific common area 39 can be set arbitrarily according to the type of detection target K, for example, the crack end in the case of a crack, or the center in the case of a blowhole. This setting may be performed by the user or automatically by the ultrasound image evaluation device 100.

When the detection target K is a crack, the end of the echo of the detection target K is determined as the specific common area 39. Then, the specific common area 39 of the training ultrasound image 31 corresponds to the position of the end of the crack. Then, the ultrasound image evaluation device 100 detects the specific common area 39 from the training ultrasound image 31. Furthermore, the depth of the crack can be found by calculating the distance between the specific common area 39 and the echo of the surface or bottom surface of the object M.

Next, the method for generating the teacher label 40 will be described using the flowchart in FIG. 41. Note that the detection target K is a crack, and its edge is a specific common region 39.

First, in step S11, a plurality of training ultrasound images 31 are acquired by the ultrasound image processing device 30. These training ultrasound images 31 are input to the input unit 110 of the ultrasound image evaluation device 100. The plurality of training ultrasound images 31 input to the input unit 110 are stored in the training ultrasound image storage unit 141.

In the next step S12, a user, such as an inspector, researcher, or expert, checks the training ultrasound images 31. The user carefully examines each training ultrasound image 31 to see whether or not a crack is visible.

In the next step S13, the teacher label generation unit 121 selects a channel of the array probe 10 when a crack is captured in the training ultrasound image 31. For example, among the multiple (e.g., N) ultrasonic elements 11A provided in the array probe 10, a received waveform 41 (Figure 26) that includes an echo from the end of the crack is identified. Then, the channel of the array probe 10 that received this received waveform 41 is selected. Here, the number of that channel is designated as m (Figure 25).

In the next step S14, the teacher label generating unit 121 acquires an echo of the end of the crack from the received waveform 41 of channel number m, and calculates the beam path distance Xd at the peak position of this echo. Note that the channel number m may be selected by the teacher label generating unit 121 or by the user.

For example, as shown in Figure 26, the received waveform 41 of channel number m includes a waveform 42 of an echo from the surface 33 of the object M and a waveform 44 of an echo from the end of a crack. The amplitude value of the received waveform 41 of channel number m is Am(X). Furthermore, a waveform 45 obtained by converting the received waveform 41 into an absolute value is shown in Figure 27.

Here, the teacher label generation unit 121 calculates the peak value |Am(Xd)| of the echo at the end of the crack from |Am(X)|, which is the absolute value of Am(X). The teacher label generation unit 121 can then calculate the beam path distance Xd to this position (Figure 25).

Returning to FIG. 41, in the next step S15, the teacher label generation unit 121 determines the coordinates Ex, Ey of the center of the specific common area 39 of the teacher label 40 from the channel number m, the beam path distance Xd, and the refraction angles α, β (FIG. 25).

In the next step S16, the teacher label generation unit 121 sets a square with a side length of d and centered at the coordinates Ex, Ey (Figure 28). Here, a binary image is generated in which squares are represented as "1" and other areas are represented as "0". This binary image becomes the teacher label 40 (Figure 24).

In the next step S17, the truth label generation unit 121 stores the generated truth label 40 in the truth label storage unit 143. Then, the method for generating the truth label 40 is completed.

As shown in FIG. 28, the size (maximum length) of the specific common area 39 in this embodiment is represented by d. For example, it is a square of d×d. Here, the diagonal of the specific common area 39 is set to be equal to or greater than one-quarter of the wavelength λ of the echo of the crack end.

As shown in FIG. 26, a length of one-quarter of the wavelength λ corresponds to the length from the rising edge of the waveform to the peak, and is the minimum length that includes the echo features from the end of the crack. If the size (maximum length) of the specific common area 39 is less than one-quarter of the wavelength λ, the echo features will not be fully included in the specific common area 39, and there is a possibility that the end of the crack will not be detected correctly.

Even if the echo is made up of several successive waves, it is necessary to extract a representative waveform from the waves and determine the wavelength.

If the specific common area 39 is a circle, the diameter of the circle is set to a size equal to or greater than a quarter of the wavelength λ. If the specific common area 39 is an ellipse, the major axis of the ellipse is set to a size equal to or greater than a quarter of the wavelength λ. The specific common area 39 can also be set to a polygon or an indefinite shape other than a circle. In that case, the maximum length of each is set to a size equal to or greater than a quarter of the wavelength λ.

In this embodiment, the size d of the specific common area 39 is equal to or greater than one-fourth the wavelength of the frequency of the ultrasonic wave U used by the ultrasonic flaw detection device 20, so that the specific common area 39 sufficiently contains the characteristics of the detection target K, and the detection target K can be correctly detected. Furthermore, the detection target K can be distinguished from noise.

As shown in FIG. 2, the judgment model generation unit 130 uses the training ultrasound images 31 and the teacher labels 40 and inspection position data corresponding to each training ultrasound image 31 to generate a judgment model that can calculate a judgment value that can determine whether or not a detection target K exists for each unit indicating a position in the image.

In this embodiment, the unit indicating a position in an image is, for example, a pixel. The position (coordinates) at which a specific pixel exists in an image is managed by a conventionally known method.

In this embodiment, the unit indicating a position in an image is a pixel, but other configurations are also possible. For example, a unit may consist of multiple adjacent pixels. Also, the image may be divided into multiple grids, and each square of the grid may be considered as a unit.

In the case of voxel data (three-dimensional volume data), the unit indicating position is, for example, a voxel. Alternatively, multiple adjacent voxels may be considered as one unit. Alternatively, voxels may be divided into multiple grids, and one square (division) of the grid may be considered as one unit.

The judgment model generating unit 130 of this embodiment performs supervised learning based on a group of training ultrasound images (group of ultrasound data), a group of corresponding teacher labels, and a group of training flaw detection position data. By performing machine learning, this judgment model is able to obtain the ability to determine whether or not a specific common region 39 of the detection target K is included in an ultrasound image and output the judgment value for each pixel.

In the supervised learning of this embodiment, in addition to the learning data of two-dimensional data (images), flaw detection position data indicating the position and angle of the array probe 10 is used as learning data. Therefore, when the detection target K is continuous, it becomes possible to newly learn the defect-specific tendency of the luminance value gradually changing. The input/output data used for this learning may be input/output one sheet at a time, or multiple sheets may be input/output together.

Next, we will explain how to generate a judgment model using the flowchart in Figure 42.

First, in step S21, the judgment model generation unit 130 acquires multiple training ultrasound images 31 that are input to the input unit 110 (image input unit 111) and stored in the training ultrasound image storage unit 141.

In the next step S22, the judgment model generating unit 130 acquires a plurality of pieces of training flaw detection position data that are input to the input unit 110 (position information input unit 112) and stored in the training position information storage unit 142. Each piece of training flaw detection position data is acquired corresponding to each training ultrasound image 31.

In the next step S23, the judgment model generation unit 130 performs supervised learning based on multiple training ultrasound images 31, teacher labels 40 corresponding to each training ultrasound image 31, and training flaw detection position data corresponding to each training ultrasound image 31.

In the next step S24, the judgment model generation unit 130 generates a judgment model by supervised learning. By performing supervised learning, a judgment model is generated that has the ability to determine whether or not a specific common region 39 of the detection target K is included in an image and output the judgment value for each pixel. This judgment model makes it possible to output a judgment value, that is, a judged image 36 (Figure 30), as an output value when a judgment ultrasound image 35 (Figure 29) is accepted as an input value.

In addition, when generating a judgment model (supervised learning), it is desirable to prepare as many training ultrasound images 31 as possible and use them to generate a judgment model. It is desirable that at least two of the training ultrasound images 31 are images in which the detection target K is present. Furthermore, the training ultrasound images 31 may be images in which multiple detection targets K are captured per image. Furthermore, the training ultrasound images 31 may include images in which the detection target K is not present and only echoes of the surface 33 are present.

In the next step S25, the judgment model generation unit 130 stores the generated judgment model in the judgment model storage unit 144. Then, the judgment model generation method ends.

As shown in FIG. 2, the determination unit 122 uses a determination model to calculate a determination value indicating whether or not a detection target K exists for each pixel of the determination ultrasound image 35.

Next, the method for evaluating the evaluation ultrasound image 35 will be explained using the flowchart in Figure 43.

First, in step S31, a plurality of judgment ultrasound images 35 (judgment ultrasound data) are acquired by the ultrasound image processing device 30. These judgment ultrasound images 35 are input to the input unit 110 (image input unit 111) of the ultrasound image evaluation device 100. The judgment ultrasound images 35 input to the input unit 110 are stored in the judgment ultrasound image storage unit 145.

In the next step S32, the position information acquisition device 90 acquires flaw detection position data of the ultrasound image processing device 30 when each of the judgment ultrasonic images 35 was acquired. These multiple flaw detection position data are input to the input unit 110 (position information input unit 112) of the ultrasound image evaluation device 100. The judgment flaw detection position data input to the input unit 110 is stored in the judgment position information storage unit 146.

In the next step S33, the judgment unit 122 inputs the multiple judgment ultrasound images 35 stored in the judgment ultrasound image memory unit 145 and the multiple judgment inspection position data stored in the judgment position information memory unit 146 into a machine-learned judgment model, and outputs (generates) multiple judged images 36 (judgment values).

As shown in FIG. 30, the determined image 36 is an image (two-dimensional data) of the same size as the determination ultrasound image 35. Each pixel of the determined image 36 is represented by a continuous value from "0" to "1" as a determination value.

As shown in FIG. 32, in the determined image 36, the pixels are displayed brighter in areas 37 where the edge of the detection target K is more likely to exist. For example, when the determination value is "0", the pixel value is black. When the determination value is "1", the pixel value is white. When the determination value is greater than "0" and less than "1", the pixel value is gray. In other words, in the determined image 36, the closer the determination value is to "1", the higher the possibility that a specific common area 39 exists.

Returning to FIG. 43, in the next step S34, the extraction unit 123 extracts the position of the detection target K from the determined image 36 (FIG. 30). This extraction unit 123 performs a binarization process on the determined image 36 to generate a binarized image 38 (FIG. 31) from which the specific common region 39 has been extracted.

As shown in Figures 31 and 33, the determined image 36 is converted to two gradations to generate a binary image 38. For example, a threshold value is set in advance by the user. Then, if the value of each pixel in the determined image 36 exceeds the threshold value, it is replaced with a white value of "1", and if it is below the threshold value, it is replaced with a black value of "0". In this way, an image of a specific common area 39 can be made to appear in the binary image 38.

Returning to FIG. 43, in the next step S35, the detection target position calculation unit 124 calculates the position of the detection target K (position of the crack) based on the specific common region 39 extracted by the extraction unit 123. Furthermore, the depth dimension of the detection target K is calculated based on the position of the detection target K.

For example, the detection target position calculation unit 124 calculates the coordinates (Gx1, Gy1 and Gx2, Gy2) of the center position of the specific common area 39 of the binarized image 38 (FIG. 33). This coordinate does not necessarily have to be the center of the specific common area 39 in the binarized image 38. For example, it may be the coordinate with the maximum intensity in the specific common area 39 of the binarized image 38, or the coordinate may be the center of gravity of a pixel group exceeding a predetermined threshold. In other words, it is sufficient if it can be uniquely determined by a specific logic.

The detection target position calculation unit 124 calculates the coordinates of the center position of the specific common region 39 in the binarized image 38 and the relative distance (depth of the crack) to the echo of the surface 33 or bottom surface of the object M. For example, it calculates the distance from the surface or bottom surface of the object M to the position of the edge of the detection target K (Figure 29). When calculating the distance, a conversion is made from the number of pixels to the dimension of length.

The position of the surface 33 or bottom surface of the object M is assumed to be determined in advance by a predetermined method. For example, the coordinates of the surface 33 or bottom surface of the object M may be calculated using the maximum value of each echo as the coordinate, or the center of an echo exceeding a predetermined threshold value.

In addition, the detection target position calculation unit 124 stores the calculated position and depth of the detection target K (crack) in the detection target position memory unit 147.

In the next step S36, the detection target position calculation unit 124 outputs the calculated position and depth of the detection target K (crack) using the output unit 150. Then, the evaluation method for the evaluation ultrasound image 35 is completed.

For example, as shown in FIG. 40, three-dimensional data 46 indicating the defect coordinates of detection target K is displayed on the display. The three-dimensional data 46 visualizes detection target K on 3D-CAD. The detection target position calculation unit 124 calculates the absolute coordinates of the detection target from the calculated position of detection target K (crack), and plots the absolute coordinates of detection target K on 3D-CAD. By visualizing the defect occurrence position of detection target K in three dimensions, it becomes easier to grasp the deterioration state of target M, and can be used for repair proposals after inspection, etc.

In this embodiment, in addition to visualization as three-dimensional data 46, visualization in other forms is performed. For example, steps S37 to S38 are executed in parallel with steps S34 to S36.

In step S37, the judgment confidence calculation unit 125 calculates a judgment confidence C indicating the likelihood of the judgment from the judged image 36 (judgment value). This judgment confidence C is a value indicating the degree of possibility that the detection target K is included.

In the next step S38, the determination certainty calculation unit 125 outputs the calculated determination certainty C using the output unit 150. Then, the method for evaluating the determination ultrasound image 35 is terminated.

For example, a specific common area 39 is extracted from the judgment value. The specific common area 39 may be rectangular, elliptical, or irregular. Then, a judgment confidence level C is calculated for each specific common area 39. The judgment confidence level C is calculated by summing up the values of the specific common areas 39 represented in a two-dimensional array. If this judgment confidence level C is high, it can be said that there is a high possibility that the detection target K is included in the judgment value.

As shown in Figure 34, when the detection target K is a planar defect, zigzag scanning is performed with the array probe 10. Next, the resulting determination confidence C is shown in the graph of Figure 35. This graph shows the determination confidence C as a color map against the XY coordinates indicating the position of the array probe 10.

For example, in the graph of FIG. 35, the value of the judgment confidence C is visualized on the XY plane in correspondence with a color map. When the value of the judgment confidence C is high, the plot (point) is white, and when the value of the judgment confidence C is low, the plot is black. When representing values between white and black, the plot is displayed using a color such as gray. By outputting this graph, the output unit 150 can visualize the defect distribution and reduce judgment errors.

In addition, the reflection intensity of the ultrasonic waves U due to changes in the rotation angle θ of the array probe 10 differs depending on whether the detection target K is a planar defect or a spherical defect. Therefore, the shape of the detection target K can be determined based on the difference in reflection intensity.

As shown in Figure 36, when the detection target K is a planar defect, a swiveling scan is performed with the array probe 10. Next, the resultant determination confidence C is shown in the graph of Figure 37. This graph plots the determination confidence C on the vertical axis against the rotation angle θ of the array probe 10 on the horizontal axis.

37, for example, the graph shows a pulse-like graph in which the judgment confidence C suddenly rises when the array probe 10 reaches a specific rotation angle θ _C. From the shape of this graph, it can be seen that the detection target K is a planar defect. In addition, the degree to which the planar defect is inclined can be estimated from the rotation angle θ _C.

As shown in Figure 38, when the detection target K is a spherical defect, a swiveling scan is performed with the array probe 10. Next, the resultant determination confidence C is shown in the graph of Figure 39. This graph plots the determination confidence C on the vertical axis against the rotation angle θ of the array probe 10 on the horizontal axis.

38, for example, the graph is a mountain-shaped graph depicting a gentle curve in which the determination confidence C is increased not only at a specific rotation angle θ C of the array probe 10 but also around the specific rotation angle θ _C. From the shape of this graph, it can be seen that the detection target K is a spherical defect.

In this embodiment, instead of labeling the entire echo of the detection target K, only the ends of the detection target K are labeled. This makes it possible to significantly reduce the effort required to create the teacher labels 40. In addition, because only the ends of the detection target K are detected, rather than the entire target K, even if the shape of the detection target K differs in each ultrasound image, the ends (specific common areas 39) that are commonly required for measuring its depth can be detected stably.

In addition, since the specific common area 39 is the end of the detection target K, the accuracy of determining the depth at which the detection target K exists can be improved.

In addition, the extraction unit 123 extracts the position of the detection target K from the determined image 36 (Figure 30) generated based on the determination value calculated from the determination ultrasound image 35 (Figure 29), thereby extracting the position of the detection target K from the determined image 36 and detecting the detection target K present in the determination ultrasound image 35.

In addition, the extraction unit 123 extracts the position of the detection target K by binarizing the judgment value for each unit indicating the position in the image based on a threshold value, thereby making it possible to extract the part of the judged image 36 where the detection target K is likely to exist.

In addition, by using pixels as the unit in this embodiment, processing can be performed for each pixel, which is the smallest unit of an image, and the evaluation ultrasound image 35 can be evaluated in detail.

In addition, by calculating the position of the detection target K in the image calculated by the judgment model, it becomes possible to display the position of the detection target K in three dimensions, making it easier to understand the deterioration state of the object M being inspected. Furthermore, the evaluation results can be used for proposing repairs after the inspection.

In this embodiment, the judgment of an arbitrary value (pixel value, judgment value) using a reference value (threshold value) may be a judgment of "whether the arbitrary value is equal to or greater than the reference value", a judgment of "whether the arbitrary value exceeds the reference value", a judgment of "whether the arbitrary value is equal to or less than the reference value", or a judgment of "whether the arbitrary value is less than the reference value". The reference value may not be fixed, but may be variable. Therefore, a value within a predetermined range may be used instead of the reference value, and a judgment of whether the arbitrary value falls within the predetermined range may be made. Furthermore, the error that occurs in the device may be analyzed in advance, and a predetermined range including the error range centered on the reference value may be used for the judgment.

In the flowchart of this embodiment, an example is shown in which each step is executed in series, but the order of steps is not necessarily fixed, and the order of some steps may be interchanged. Also, some steps may be executed in parallel with other steps.

The system of this embodiment includes a control device that is a highly integrated processor such as a dedicated chip, FPGA (Field Programmable Gate Array), GPU (Graphics Processing Unit), or CPU (Central Processing Unit), a storage device such as ROM (Read Only Memory) or RAM (Random Access Memory), an external storage device such as HDD (Hard Disk Drive) or SSD (Solid State Drive), a display device such as a display, an input device such as a mouse or keyboard, and a communication interface. This system can be realized with a hardware configuration that uses a normal computer.

The program executed by the system of this embodiment is provided in advance in a ROM or the like. Alternatively, this program may be provided stored in an installable or executable file format on a non-transitory computer-readable storage medium such as a CD-ROM, CD-R, memory card, DVD, or flexible disk (FD).

The programs executed by this system may also be stored on a computer connected to a network such as the Internet and provided by downloading them via the network. This system may also be configured by combining separate modules that independently perform the functions of the components and connect them together via a network or dedicated lines.

In addition, in this embodiment, the ultrasound image (or voxel data) has been described as being a grayscale image, but other configurations are also possible. For example, the ultrasound image may be a color image expressed in RGB or the like. In that case, for example, pixel values above a threshold value may be colored red, and pixel values below the threshold value may be colored blue, to give the display form a distinctive feature.

In this embodiment, the training ultrasonic data is a plurality of training ultrasonic images 31 obtained by the ultrasonic inspection device 20, but may be in other forms. Furthermore, in this embodiment, the judgment ultrasonic data is a plurality of judgment ultrasonic images 35 obtained by the ultrasonic inspection device 20, but may be in other forms. For example, the array probe 10 may be configured as a matrix array probe, and ultrasonic data of the object M may be acquired as training three-dimensional ultrasonic data or judgment three-dimensional ultrasonic data, which are voxel data from the beginning. In this case, the specific common area 39 ( FIG. 31 ) may be a cube.

Furthermore, the learning ultrasonic data in this embodiment may be at least one of voxel data obtained by the ultrasonic flaw detection device 20 and voxel data generated to imitate the voxel data. In this way, the detection target K can be learned using three-dimensional information, improving the efficiency of learning.

In this embodiment, voxel data can be generated by superimposing multiple ultrasound images. For example, voxel data generated from multiple training ultrasound images 31 can be used as training 3D ultrasound data. Furthermore, voxel data generated from multiple determination ultrasound images 35 can also be used as determination 3D ultrasound data.

According to the embodiment described above, by providing a computer that inputs at least one ultrasonic data to be judged and flaw detection position data into the judgment model and outputs a judgment value that can judge whether or not a detection target exists for each unit indicating a position in the ultrasonic data, it is possible to improve the detection accuracy of the detection target position contained in the target in the evaluation of ultrasonic data used in non-destructive testing. In particular, it becomes possible to reliably identify the position and depth of a defect, which can be used for the soundness diagnosis of a structure, etc.

Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, modifications, and combinations can be made without departing from the spirit of the invention. These embodiments or modifications thereof are within the scope of the invention and its equivalents as set forth in the claims, as well as the scope and spirit of the invention.

1...Ultrasonic data evaluation system, 2...Acoustic propagation medium, 10...Array probe, 11...Probe, 11A...Ultrasonic element, 20...Ultrasonic flaw detection device, 30...Ultrasonic image processing device, 31...Learning ultrasonic image, 32...Incidence range, 33...Surface of object, 35...Ultrasonic image for judgment, 36...Judgmented image, 37...Area where edge is likely to exist, 38...Binarized image, 39...Specific common area, 40...Teaching label, 41...Received waveform, 42...Waveform of echo from surface of object, 44...Waveform of echo from edge of crack, 45...Absolutized waveform, 46...Three-dimensional data, 90...Position information acquisition device, 91...Displacement acquisition unit, 92... Angle acquisition unit, 100...ultrasound image evaluation device, 110...input unit, 120...calculation unit, 121...teacher label generation unit, 122...judgment unit, 123...extraction unit, 124...detection target position calculation unit, 125...judgment confidence level calculation unit, 130...judgment model generation unit, 140...storage unit, 141...learning ultrasound image storage unit, 142...learning position information storage unit, 143...teacher label storage unit, 144...judgment model storage unit, 145...judgment ultrasound image storage unit, 146...judgment position information storage unit, 147...detection target position storage unit, 150...output unit, B...background, K...detection target, M...object, R...reflected wave, U...ultrasound, W...weld.

Claims (11)

an ultrasonic probe for detecting a flaw in a detection target included in an object to be inspected and acquiring ultrasonic data including at least one of an image and voxel data ;
a position information acquisition device that acquires flaw detection position data indicating a position of the ultrasonic probe when the ultrasonic data is acquired;
A computer that inputs at least one of the ultrasonic data to be judged and the flaw detection position data to a machine learning judgment model, and outputs a judgment value that can judge whether or not the detection target exists for each unit indicating a position in the ultrasonic data;
Equipped with
The judgment model has been machine-learned using a plurality of learning ultrasonic data, learning flaw detection position data, and teacher labels;
The learning ultrasound data includes at least one of the ultrasound data and data generated by imitating the ultrasound data,
The learning inspection position data includes at least one of the inspection position data and data generated by simulating the inspection position data,
The teacher label is a specific common area indicating a specific area of the detection target present in the learning ultrasound data, which is set as the position of the detection target.
Ultrasound data evaluation system. The maximum length of the specific common area is equal to or greater than a quarter of the wavelength of the ultrasonic frequency used by the ultrasonic probe.
2. The ultrasound data evaluation system of claim 1. The inspection position data includes an angle representing a rotation of the ultrasonic probe.
3. An ultrasound data evaluation system according to claim 1 or 2. The flaw detection position data includes at least two coordinate values represented by orthogonal coordinates representing the position of the ultrasonic probe.
4. An ultrasound data evaluation system according to claim 1. The flaw detection position data includes at least two coordinate values represented by polar coordinates or cylindrical coordinates representing the position of the ultrasonic probe.
4. An ultrasound data evaluation system according to claim 1. The specific common area is an edge of the detection target.
6. An ultrasound data evaluation system according to any one of claims 1 to 5. An extraction unit is further provided which extracts a position of the detection target based on the determination value.
7. An ultrasound data evaluation system according to any one of claims 1 to 6. The determination value is a continuous value calculated for each unit,
The extraction unit binarizes the determination value for each unit based on a threshold value to extract a position of the detection target.
8. The ultrasound data evaluation system of claim 7. the unit is at least one of a pixel or a voxel,
The judgment model is a machine-learned model of segmentation.
An ultrasound data evaluation system according to any one of claims 1 to 8. Obtaining ultrasonic data including at least one of an image and voxel data obtained by detecting a flaw in a detection target included in an object to be inspected using an ultrasonic probe;
acquiring flaw detection position data indicating a position of the ultrasonic probe when the ultrasonic data was acquired;
At least one of the ultrasonic data to be subjected to the judgment and the flaw detection position data is input to a machine learning-processed judgment model, and a judgment value capable of judging whether or not the detection target exists for each unit indicating a position in the ultrasonic data is output.
The processing is performed by a computer,
The judgment model has been machine-learned using a plurality of learning ultrasonic data, learning flaw detection position data, and teacher labels;
The learning ultrasound data includes at least one of the ultrasound data and data generated by imitating the ultrasound data,
The learning inspection position data includes at least one of the inspection position data and data generated by simulating the inspection position data,
The teacher label is a specific common area indicating a specific area of the detection target present in the learning ultrasound data, which is set as the position of the detection target.
Methods for evaluating ultrasound data. Acquire ultrasonic data including at least one of an image or voxel data obtained by performing flaw detection on a detection target included in an object to be inspected using an ultrasonic probe , or acquire training ultrasonic data including at least one of data generated by simulating the same;
Acquire learning inspection position data including at least one of inspection position data indicating the position of the ultrasonic probe when the ultrasonic data is acquired using the ultrasonic probe and data generated by simulating the same ;
generating a teacher label that sets a specific common area indicating a specific area of the detection target present in the learning ultrasound data as the position of the detection target;
Performing machine learning using a plurality of the learning ultrasonic data, the learning flaw detection position data, and the teacher label;
At least one of the ultrasonic data to be determined and the flaw detection position data is input, and a determination value capable of determining whether or not the detection target is present for each unit indicating a position in the ultrasonic data is output.
Generate a decision model,
The processing is carried out by a computer.
A method for generating decision models.

Images