JP7465522B2 - Fracture surface analysis device and fracture surface analysis method - Google Patents

Description

The present invention relates to a technique for analyzing the fracture surfaces of structures, and in particular to a fracture surface analysis device and a fracture surface analysis method.

When a structural damage accident occurs, fracture surface analysis is carried out to investigate the cause. Fracture surface analysis involves macroanalysis, which focuses on observation at low magnification, and microanalysis, which focuses on observation at high magnification. Macroanalysis mainly involves analysis of the fracture origin and the direction of crack propagation, while microanalysis mainly involves analysis of the fracture pattern (see, for example, Patent Documents 1 and 2).

Fracture surface analysis examines the similarities between the characteristics of previously observed fracture surfaces, so extensive experience is required to derive correct analytical results. For this reason, young, inexperienced engineers are often unsure of the fracture origin, crack propagation direction, and fracture pattern, and must make up for their lack of experience by consulting experienced engineers or researching literature.

Japanese Patent Application Laid-Open No. 11-101625 JP 2009-085737 A

In recent years, a combination of factors such as a "decrease in experienced engineers due to retirement, etc." and a "demand to shorten the time required for fracture surface analysis in order to quickly formulate measures to prevent recurrence" has created a tough situation for young engineers. Furthermore, if measures to prevent recurrence are based on incorrect fracture surface analysis results, there is a high possibility that damage accidents will occur again, which places a lot of mental pressure on young engineers. As a technology to support young engineers in this environment, there is a demand for the development of an analysis method that does not require extensive experience in fracture surface analysis.

However, because research and development into fracture surface analysis is carried out by researchers with advanced analytical skills, conventional fracture surface analysis techniques are intended to be used by engineers with the same level of analytical ability as the researchers. This makes it difficult for young engineers to master them.

In view of the above, the present invention aims to enable even beginners in fracture surface analysis to determine the location of the fracture origin from a fracture surface image.

To achieve the above objective, this invention estimates the fracture origin from a fracture surface image based on the learning results of the machine learning section.

A first aspect of the present disclosure is a fracture surface analysis device that includes a machine learning unit that performs a first learning process using a first data set that includes a combination of fracture surface image data and fracture origin data, and an estimation unit that estimates the fracture origin of an input fracture surface image based on a first trained model obtained by the first learning process.

In this first aspect, the estimation unit estimates the fracture origin of the input fracture surface image based on the first trained model, so that even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image.

In a second aspect of the present disclosure, in the first aspect, the learning model for the first learning can be selected from a regression model that learns to be able to estimate the coordinates of the fracture origin, and a classification model that learns to be able to estimate the probability distribution where the fracture origin exists, and the estimation unit estimates the coordinates of the fracture origin in the input fracture surface image when the learning model is the regression model, and estimates the probability distribution where the fracture origin exists in the input fracture surface image when the learning model is the classification model.

In this second aspect, a specific configuration that achieves the effects of the first aspect is obtained, and even a beginner in fracture surface analysis can easily determine the location of the fracture origin from the fracture surface image.

A third aspect of the present disclosure is the second aspect, in which the learning model of the first learning includes the regression model and the classification model, and the estimation unit estimates the coordinates of the fracture origin of the input fracture surface image and the probability distribution of the existence of the fracture origin in the input fracture surface image, and further includes a result display unit that displays at least two of the input fracture surface image, the estimated coordinates of the fracture origin, and the estimated probability distribution as images superimposed on the same screen.

In this third aspect, the estimation unit estimates both the coordinates of the fracture origin in the input fracture surface image and the probability distribution of the existence of the fracture origin in the input fracture surface image, and the result display unit displays at least two of the input fracture surface image, the coordinates of the estimated fracture origin, and the estimated probability distribution, superimposed on the same screen as images, so that the validity of the estimated fracture origin can be verified. Therefore, even a beginner in fracture surface analysis can not only determine the position of the fracture origin from the fracture surface image, but can also improve the accuracy through verification.

A fourth aspect of the present disclosure is the third aspect, further comprising a divided image generation unit that determines the number n of fracture origins included in the estimated probability distribution, divides the input fracture surface image into n regions such that each region contains one of the fracture origins, and generates n divided images, each of which shows an image of one of the n regions, and the estimation unit further estimates the coordinates of the fracture origin in each of the n divided images and the probability distribution in which the fracture origin exists.

In this fourth aspect, the divided image generation unit generates n divided images, each of which contains one fracture origin, and the estimation unit estimates the coordinates of the fracture origin in each of the n divided images and the probability distribution in which the fracture origin exists. Therefore, even if the input fracture surface image contains multiple fracture origins, the position of the fracture origin can be correctly estimated. Therefore, even if there are multiple fracture origins in the fracture surface image, a beginner in fracture surface analysis can easily determine the position of the fracture origin from the fracture surface image. In addition, the estimated coordinates of the fracture origin and the probability distribution in which the fracture origin exists can be compared with each other to verify the validity of the position of the fracture origin.

A fifth aspect of the present disclosure is any one of the first to fourth aspects, in which the machine learning unit further performs a second learning process using a second data set including a combination of fracture surface image data and fracture surface pattern data, and the estimation unit further estimates the fracture surface pattern of the input fracture surface image based on a second trained model obtained by the second learning process.

In this fifth aspect, the machine learning unit performs a second learning in addition to the first learning, and the estimation unit further estimates the fracture surface pattern of the fracture surface image in addition to the fracture origin of the fracture surface image, so that the estimated fracture origin can be compared with the estimated fracture surface pattern to verify whether the estimated fracture origin is correct. Therefore, even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image, and the accuracy can be improved by verification.

A sixth aspect of the present disclosure is the fifth aspect, in which the learning model of the second learning includes a semantic segmentation model that is trained to be able to estimate the fracture surface pattern.

In this sixth aspect, a specific configuration that achieves the effects of the fifth aspect is obtained.

A seventh aspect of the present disclosure is the fifth or sixth aspect, further comprising a result display unit that displays at least two of the input fracture surface image, the fracture origin estimated based on the first trained model, and the fracture surface pattern estimated based on the second trained model as images superimposed on the same screen.

In this seventh aspect, the result display unit displays at least two of the input fracture surface image, the fracture origin estimated based on the first trained model, and the fracture surface pattern estimated based on the second trained model as images superimposed on the same screen, so that, for example, it is easy to verify whether the estimated fracture origin is correct based on the relationship between the estimated fracture origin and the estimated fracture surface pattern. Therefore, even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image, and the accuracy can be improved by verification.

However, there may be cases where the fracture origin is not present in the input fracture surface image. In such cases, the engineer must look at a wide-area fracture surface image that includes a wider area of fracture than the fracture surface shown in the fracture surface image, and re-capture the fracture surface image so that the fracture origin is included. In this case, by obtaining a region segmentation image in which the fracture surface shown in the wide-area fracture surface image is divided into different areas based on the difference in the shape of the fracture surface, it becomes easier to estimate the region of the fracture surface to be photographed in the damaged object.

Here, the eighth aspect is any one of the first to seventh aspects, in which the machine learning unit further performs a third learning using a third data set including a combination of wide-area image data including a fracture surface wider than the fracture surface included in the fracture surface image data and region division data in which the wide-area fracture surface is divided into two or more regions according to the shape of the fracture surface, and the estimation unit further estimates a region division image in which the fracture surface region is divided in the input wide-area fracture surface image based on a third trained model obtained by the third learning.

According to this eighth aspect, the estimation unit further estimates an area division image in which the area of the fracture surface is divided in the input wide-area fracture surface image based on the third learned model obtained by the third learning, so that by looking at this area division image, it becomes easier to consider the area of the fracture surface to be photographed in the damaged object. Therefore, even a beginner in fracture surface analysis can avoid a situation in which a fracture surface image that does not include the fracture origin is mistakenly used, and can easily determine the position of the fracture origin.

A ninth aspect of the present disclosure relates to a fracture surface analysis method, and includes a first learning step of performing a first learning process using a first data set including a combination of fracture surface image data and fracture origin data, and a first estimation step of estimating the fracture origin of the input fracture surface image based on the first trained model obtained in the first learning step.

In this ninth aspect, in the first estimation step, the fracture origin of the input fracture surface image is estimated based on the first trained model, so that even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image.

A tenth aspect of the present disclosure is the ninth aspect, further comprising a second learning step of performing a second learning using a second data set including a combination of fracture surface image data and fracture surface pattern data, and a second estimation step of estimating the fracture surface pattern of the input fracture surface image based on the second trained model obtained in the second learning step.

In this tenth aspect, in addition to the first learning step, a second learning step is further performed, and in addition to estimating the fracture origin of the fracture surface image in the first estimation step, the fracture surface pattern of the fracture surface image is further estimated in the second estimation step, so that the estimated fracture origin can be compared with the estimated fracture surface pattern to verify whether the estimated fracture origin is correct. Therefore, even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image, and the accuracy can be improved by verification.

An eleventh aspect of the present disclosure is the ninth or tenth aspect, further comprising a third learning step of performing a third learning using a third data set including a combination of wide-area image data including a fracture surface wider than the fracture surface included in the fracture surface image data and area division data in which the wide-area fracture surface is divided into two or more areas according to the shape of the fracture surface, and a third estimation step of estimating an area division image in which the fracture surface area is divided in the input wide-area fracture surface image based on a third trained model obtained in the third learning.

In this eleventh aspect, in addition to the first and second learning steps, a third learning step is further performed, and in the third estimation step, a region division image in which the region of the fracture surface of the wide-area fracture surface image is further estimated, so that, as in the eighth aspect, it becomes easier to consider the region of the fracture surface to be photographed in the damaged object. In other words, even a beginner in fracture surface analysis can avoid a situation in which a fracture surface image that does not include the fracture origin is mistakenly used, and it becomes easier to determine the position of the fracture origin.

A twelfth aspect of the present disclosure is a program for causing a computer to execute the fracture surface analysis method according to any one of the ninth to eleventh aspects.

In this twelfth aspect, the method in any one of the ninth to eleventh aspects can be implemented by a computer.

As explained above, with this invention, even a beginner in fracture surface analysis can determine the location of the fracture origin from a fracture surface image.

FIG. 1 is a configuration diagram of a fracture surface analysis device according to a first embodiment and a second embodiment. A schematic diagram showing first learning when using a regression model and estimation based on a learned regression model in the first and second embodiments. A schematic diagram showing first learning when a classification model is used and estimation based on a learned classification model according to the first and second embodiments. FIG. 2 is a configuration diagram of a dataset storage unit according to the first embodiment. 5A to 5C are diagrams illustrating an example of an image displayed on a display screen of a result display unit according to the first embodiment. FIG. 2 is a flowchart showing a fracture surface analysis method according to the first embodiment. A schematic diagram showing second learning when using a semantic segmentation model and estimation based on a learned semantic segmentation model in the second embodiment. FIG. 5 is a view corresponding to FIG. 4 according to the second embodiment. FIG. 6 is a view corresponding to FIG. 5 according to a second embodiment. 9B is a diagram corresponding to FIG. 5 according to the second embodiment, showing an example different from FIG. 9A . FIG. FIG. 7 is a view corresponding to FIG. 6 according to the second embodiment. A schematic diagram showing third learning and estimation based on a learned semantic segmentation model according to a modified example of the second embodiment.

The fracture surface analysis device and fracture surface analysis method according to the embodiment will be described below with reference to the drawings.

First Embodiment
(Functional configuration of fracture surface analysis device)
1 is a configuration diagram of a fracture surface analysis device 10 according to the present invention. This fracture surface analysis device 10 is for determining the position of the fracture origin of a fracture surface from an image of the fracture surface of a damaged object 1 (a structure in which the position of the fracture origin is to be determined). This fracture surface analysis device 10 includes a data set storage unit 11, a machine learning unit 12, an imaging unit 13, an estimation unit 14, a result display unit 15, and an output data storage unit 16. Note that each component of the fracture surface analysis device 10 may be distributed and arranged as a separate device part.

2 and 3 are schematic diagrams showing the functions of the machine learning unit 12 and the estimation unit 14. As shown in the upper part of each of FIGS. 2 and 3, the fracture surface analysis device 10 is configured such that the machine learning unit 12 performs a first learning process described below, and as shown in the lower part of each of FIGS. 2 and 3, the estimation unit 14 estimates the fracture origin of the fracture surface image Im captured and input by the imaging unit 13 based on the first trained model M1 obtained by this first learning. The result display unit 15 displays the estimation result of the estimation unit 14, and the engineer performing the fracture surface analysis determines the position of the fracture origin of the fracture surface of the damaged object 1 by looking at the estimation result.

In addition, in the fracture surface analysis device 10, the machine learning unit 12 and the estimation unit 14 are mainly provided with a storage medium such as a memory in which a program for executing each function is stored, and a processor that operates according to the program, as hardware components.

The following describes in detail the functional configurations of the dataset storage unit 11, machine learning unit 12, imaging unit 13, estimation unit 14, and result display unit 15.

(Data Set Storage Unit)
The dataset storage unit 11 is a storage medium such as a hard disk or an SD card. Fig. 4 is a configuration diagram of the dataset storage unit 11. The dataset storage unit 11 stores a first dataset D1 for the machine learning unit 12 to perform first learning. This first dataset D1 is composed of a combination of fracture surface image data D11 and fracture origin data D12 indicating the position of the fracture origin of the fracture surface shown in the fracture surface image data D11.

As the fracture surface image data D11, multiple types of images prepared using various photographing and generation methods can be used. Examples of "multiple types of images" include SEM images and images taken with a digital camera (digital camera images), as well as binary images in which these photographed images are displayed using only black and white pixels. It is preferable to use at least two of the SEM images, digital camera images, and binary images for the fracture surface image data D11, and it is even more preferable to use three of these.

The fracture origin data D12 includes coordinate data indicating the position of the fracture origin on the fracture surface of the fracture surface image data D11 as coordinates, and probability distribution data describing the probability of the existence of the fracture origin for each small region described later. The coordinate data is composed of a combination of (X, Y) representing the coordinates of the fracture origin on each fracture surface indicated by the fracture surface image data D11, as shown in the upper right of FIG. 2. The probability distribution data is composed of a numerical value indicating the probability of the existence of the fracture origin in each small region, which is obtained by dividing each fracture surface into a plurality of small regions A ₁ , A ₂ , ..., as shown in the upper right of FIG. 3. The small region may be an area occupied by each pixel of the fracture surface image Im, or may be an area occupied by a set of a plurality of adjacent pixels. The probability distribution data shown in the upper right of FIG. 3 includes 144 small regions A ₁ to A ₁₄₄ .

The above-mentioned coordinate data and probability distribution data are prepared before the first learning is performed. The method of preparing the coordinate data may be, for example, a method in which an engineer looks at the fracture surface shown in the fracture surface image data D11, identifies the fracture origin, and records the coordinates. The method of preparing the probability distribution data may be, for example, a method in which an engineer looks at the fracture surface shown in the fracture surface image data D11, and determines the rank of the probability that the fracture origin exists for each small area of the fracture surface (for example, rank 1: probability 50% or more, rank 2: probability 10% or more but less than 50%, rank 3: probability 0% or more but less than 10%, etc.). More specifically, for example, a method may be used in which all small areas in the fracture surface image Im are painted in different colors according to the rank of the probability that the fracture origin exists, using image editing software that can edit images on a pixel-by-pixel basis. In the example shown in FIG. 3, small area _A115 is assigned rank 1 (probability 50%), small areas _A102 , _A104 , _A116 , _A127 , and _A128 are assigned rank 2 (probability 10%), and the other small areas are assigned rank 3 (probability of existence 0%).

(Machine Learning Department)
The machine learning unit 12 reads out the first data set D1 from the data set storage unit 11 and performs first learning. The learning models of the first learning are a regression model that learns to be able to estimate the coordinates of the fracture origin as shown in FIG. 2, and a classification model that learns to be able to estimate the probability that the fracture origin exists as shown in FIG. 3. The machine learning unit 12 performs two patterns of first learning to construct these two learning models. Hereinafter, the first trained model M1 obtained in the first learning when the learning model is a regression model and when the learning model is a classification model will be referred to as a trained regression model MR and a trained classification model MC, respectively.

The learning method of the machine learning unit 12 can be, for example, a convolutional neural network or a support vector machine.

(Photography Department)
The photographing unit 13 photographs a fracture surface image Im of the broken object 1. As the photographing unit 13, various photographing means such as an electron microscope such as an SEM, a stereo microscope, a microscope, and a digital camera can be used depending on the type of the broken object 1. The fracture surface analysis device 10 may include an image storage unit (not shown) that stores the fracture surface image Im photographed by the photographing unit 13.

(Estimation section)
The estimation unit 14 receives the fracture surface image Im captured by the imaging unit 13 and estimates the fracture origin. The result of the estimation is output by the result display unit 15. As shown in the lower part of each of Figs. 2 and 3, the estimation unit 14 is equipped with a trained regression model MR and a trained classification model MC obtained by the first learning of the machine learning unit 12, and the estimation unit 14 estimates the fracture origin based on the trained regression model MR and the trained classification model MC. Specifically, as shown in the lower part of Fig. 2, the estimation unit 14 estimates the coordinates of the fracture origin of the fracture surface image Im based on the trained regression model MR, and as a result, the coordinates Pt are obtained. In addition, as shown in the lower part of Fig. 3, the estimation unit 14 estimates the probability of each small area in which the fracture origin of the fracture surface image Im exists based on the trained classification model MC, and as a result, a fracture origin contour diagram PD (probability distribution) indicating the probability that the fracture origin exists is obtained. In addition, in the example of the fracture origin contour diagram PD shown in the lower right of FIG. 3, the probability of the existence of a fracture origin in each small region is shown in gray scale, with the darker the color, the higher the probability.

(Result display section)
The result display unit 15 has a display surface (not shown) configured by a liquid crystal display or the like. The result display unit 15 displays, as an image, the coordinates Pt of the fracture origin and the fracture origin contour map PD obtained as a result of the estimation of the fracture origin by the estimation unit 14 using two patterns of the first trained model M1 (i.e., the trained regression model MR and the trained classification model MC) on the display surface (same screen) in a superimposed manner. Fig. 5 is a diagram showing an example of an image displayed on the display surface of the result display unit 15.

In the example shown in Figure 5, it can be seen that the coordinates Pt of the fracture origin almost coincide with the most probable area of the fracture origin contour map PD. If the fracture origin estimation results based on the trained regression model MR and the trained classification model MC differ significantly, it is possible that at least one of the estimation results is incorrect. Only in such a case can it be verified which fracture origin is more reliable by comparing the two different estimated fracture origins with the input fracture surface image Im.

The display method by the result display unit 15 may be different from the example shown in FIG. 5. For example, the result display unit 15 may display the fracture origin coordinate Pt and fracture surface image Im, or the fracture origin contour diagram PD and fracture surface image Im, as an image superimposed on the display surface, or may display all of the fracture origin coordinate Pt, fracture origin contour diagram PD, and fracture surface image Im as an image superimposed on the display surface.

The estimation results displayed by the result display unit 15 are stored in the output data storage unit 16. The output data storage unit 16 is a storage medium such as a hard disk.

(Fracture surface analysis method)
6 is a flow diagram showing a fracture surface analysis method S10 according to the first embodiment. In the fracture surface analysis method S10, the functions of the fracture surface analysis device 10 described above are executed in sequence. Before starting the fracture surface analysis method S10, a first data set D1 consisting of a combination of fracture surface image data D11 and fracture origin data D12 is prepared and stored in the data set storage unit 11.

First, in the first step S11, the machine learning unit 12 performs the first learning using the first data set D1 (first learning step). In the subsequent second step S12, the imager 13 captures a fracture surface image Im of the broken object 1. In the subsequent third step S13, the fracture surface image Im captured in the second step S12 is input, and the fracture origin of the input fracture surface image Im is estimated based on the first trained model M1 (i.e., the trained regression model MR and the trained classification model MC) obtained in the first step S11, and as a result, the coordinates Pt of the fracture origin and the fracture origin contour diagram PD are obtained (first estimation step). In the subsequent fourth step S14, the result display unit 15 displays the estimated fracture origin. Note that the second step S12 may be performed prior to the first step S11.

The above fracture surface analysis method can be executed by a computer as a computer program recorded on a recording medium.

(Action and Effects)
In this embodiment, the estimation unit 14 estimates the fracture origin of the input fracture surface image Im based on the first learned model M1, so that even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image Im.

The learned regression model MR of the estimation unit 14 estimates the coordinates of the fracture origin in the fracture surface image Im, and as a result, the coordinates Pt are obtained, so that even a beginner in fracture surface analysis can easily determine the position of the fracture origin in the fracture surface image Im.

However, there are cases where it is desired to verify whether the coordinates Pt of the fracture origin obtained by the estimation unit 14 are correct. Here, the estimation unit 14 can display the position of the fracture origin in the fracture surface image Im as a fracture origin contour diagram PD in addition to the coordinates Pt of the fracture origin, so that the coordinates Pt of the fracture origin and the fracture origin contour diagram PD can be compared with each other. This makes it possible to verify the validity of each estimation result. Therefore, even a beginner in fracture surface analysis can not only determine the position of the fracture origin from the fracture surface image Im, but can also improve the accuracy by verification.

In addition, in this embodiment, the machine learning unit 12 uses multiple types of images prepared by various shooting and generation methods as the fracture surface image data D11, so the number of first data sets D1 can be increased to improve the accuracy of the first learning. Therefore, even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image Im with high accuracy based on the obtained first trained model M1.

<Modification of the First Embodiment>
Incidentally, when only one fracture origin exists in the fracture surface image Im, the trained regression model MR estimates one fracture origin, and the trained classification model MC also estimates one fracture origin. Therefore, in the case of a fracture surface image Im in which only one fracture origin exists, as described above, the validity of the position of the fracture origin can be verified and its accuracy can be improved by comparing the coordinate Pt with the fracture origin contour map PD.
However, on the other hand, when a plurality of fracture origins exist in the fracture surface image Im, the trained classification model MC can estimate the plurality of fracture origins as the fracture origin contour diagram PD, but the trained regression model MR estimates only one fracture origin, so it becomes difficult to verify the validity of each estimation result and improve its accuracy. Therefore, it is desired to enable beginners in fracture surface analysis to verify the validity of the positions of the fracture origins even when a plurality of fracture origins exist in the fracture surface image Im.

In view of the above, the fracture surface analysis device 10 may further include a divided image generation unit (not shown) that determines the number n of fracture origins included in the fracture origin contour diagram PD obtained by estimation, and divides the input fracture surface image Im into n fracture surface images so that each image contains one fracture origin. The estimation unit 14 may then input each of the n divided images and estimate the coordinates Pt of the fracture origin and the fracture origin contour diagram PD for each of the n divided images.

According to this modified example, the divided image generation unit generates n divided images that are divided so that each image contains one fracture origin indicated by the estimation result by the trained classification model MC, and the estimation unit 14 estimates the fracture origin for each of the n divided images, so that the coordinates Pt and fracture origin contour diagram PD for each of the n images are obtained. As a result, even if the input fracture surface image Im contains multiple fracture origins, the output coordinates Pt and fracture origin contour diagram PD can be compared with each other. Therefore, even if there are multiple fracture origins in the fracture surface image Im, a beginner in fracture surface analysis can verify the validity of the fracture origin position and improve its accuracy.

Second Embodiment
The functional configuration of a fracture surface analysis device 20 according to the second embodiment will be described below. This fracture surface analysis device 20 differs from the first embodiment in the functions of a machine learning unit 22, an estimation unit 24, and a result display unit 25. Note that components having the same functions as those in the first embodiment are designated by the same reference numerals as those in the first embodiment, and descriptions thereof will be omitted.

The machine learning unit 22 performs further machine learning (second learning) in addition to the first learning shown in Figures 2 and 3. Figure 7 is a schematic diagram showing the functions of the machine learning unit 22 and the estimation unit 24 regarding the second learning. That is, the estimation unit 24 estimates the fracture origin of the fracture surface image Im based on the first trained model M1, and also estimates the fracture surface pattern Ptn captured in the fracture surface image Im based on the second trained model M2 obtained by the second learning.

Specifically, the fracture surface pattern Ptn is a pattern that reflects unevenness such as steps on the fracture surface. In the lower right of Figure 7, an example of this fracture surface pattern Ptn is shown, in which a radial pattern in which the propagation path of a crack as it expands appears as steps on the fracture surface. This radial pattern serves as a foothold for identifying the fracture origin, as the approximate position of the fracture origin can be identified from its convergence point. Note that the fracture surface pattern Ptn is not limited to a radial pattern in which the propagation path of a crack as it expands appears as steps on the fracture surface, but may be any pattern on the fracture surface that serves as a foothold for identifying the fracture origin.

Figure 8 is a configuration diagram of the dataset storage unit 21. In addition to the first dataset D1, the dataset storage unit 21 stores a second dataset D2. The second dataset D2 consists of a combination of fracture surface image data D21 (D11) and fracture surface pattern data D22 that shows the radial pattern of the fracture surface image data D21.

The fracture surface pattern data D22 is, for example, the hatched portion shown in the upper right of FIG. 7. This fracture surface pattern data D22 is prepared before the second learning. The fracture surface pattern data D22 may be prepared, for example, by an engineer looking at the fracture surface shown in the fracture surface image data D21 and drawing it using image editing software or the like.

The machine learning unit 22 reads out the first data set D1 and the second data set D2 from the data set storage unit 21, and performs the first learning and the second learning. In the second learning, as shown in FIG. 7, a semantic segmentation model (hereinafter referred to as the SS model), which is a learning model for estimating the fracture surface pattern, is constructed. Hereinafter, the second trained model M2 obtained in the second learning is referred to as the trained SS model MS.

As shown in the lower part of Figure 7, the estimation unit 24 receives the fracture surface image Im captured by the imaging unit 13 and estimates the fracture surface pattern Ptn, thereby obtaining the fracture surface pattern Ptn.

The result display unit 25 displays the fracture origin (coordinate Pt or fracture origin contour diagram PD) estimated based on the first trained model M1 and the fracture surface pattern Ptn estimated based on the second trained model M2 as an image superimposed on its display surface (same screen).

9A and 9B are diagrams showing two examples of images displayed on the display surface of the result display unit 25. In Fig. 9A and Fig. 9B, the fracture surface images Im input to the estimation unit 24 are different from each other. On the left side of each of Fig. 9A and Fig. 9B, the coordinates Pt of the estimated fracture origin and the fracture surface pattern Ptn are displayed superimposed, and on the right side of each of Fig. 9A and Fig. 9B, the fracture origin contour diagram PD (excluding the area with a probability of 0%) of the estimated fracture origin and the fracture surface pattern Ptn are displayed superimposed. In Fig. 9A and Fig. 9B, the position where the fracture origin can be determined from the radial pattern shown by the fracture surface pattern Ptn (i.e., the position where the radial pattern converges) is surrounded by a dashed line.

In the example shown in Figure 9A, the position of the coordinate Pt of the estimated fracture origin (see the left side of Figure 9A) and the most probable position of the fracture origin contour map PD (see the right side of Figure 9B) are both in the area surrounded by the dashed line. In other words, the estimation results of the fracture origin based on the trained regression model MR and the trained classification model MC (coordinate Pt and fracture origin contour map PD) are both consistent with the estimation results based on the trained SS model MS (fracture surface pattern Ptn). Therefore, the fracture origin can be determined without hesitation from the estimation results based on the three trained models.

On the other hand, in the example shown in Figure 9B, although the most probable area of the estimated fracture origin contour map PD is within the area surrounded by the dashed line (see the right side of Figure 9B), it can be seen that the coordinates Pt of the estimated fracture origin are outside the area surrounded by the dashed line (see the left side of Figure 9B). In other words, while the estimation result based on the trained classification model MC matches the estimation result based on the trained SS model MS, the estimation result based on the trained regression model MR does not match the estimation result based on the trained SS model MS. In this case, it can be said that the estimation result based on the trained classification model MR (right side of Figure 9B), which matches the estimation result based on the trained SS model MS, is more reliable than the estimation result based on the trained regression model MR (left side of Figure 9B).

Figure 10 is a flow diagram showing a fracture surface analysis method S20 according to the second embodiment. In the fracture surface analysis method S20, following the first step S21 (S11) of performing the first learning, a second step S22 of performing the second learning is performed using a combination of the fracture surface image data D21 (D11) and the fracture surface pattern data D22 as the second data set D2 (second learning step). In the subsequent third step S23 (S12), the imager 13 captures a fracture surface image Im of the broken object 1. In the subsequent fourth step S24 (S13), the position of the fracture origin is estimated based on the first trained model M1, and the coordinates Pt of the fracture origin and the fracture origin contour diagram PD are obtained (first estimation step). In the subsequent fifth step S25, the fracture surface pattern is estimated based on the second trained model M2, and the fracture surface pattern Ptn is obtained (second estimation step). In the subsequent sixth step S26, the result display unit 25 displays the estimated coordinates Pt of the fracture origin, the fracture origin contour diagram PD, and the fracture surface pattern Ptn. The order in which the first step S21 to the fifth step S25 are performed is not limited to this, and it is sufficient that the first step S21 precedes the fourth step S24, the second step S22 precedes the fifth step S25, the third step S23 precedes both the fourth step S24 and the fifth step S25, and both the fourth step S24 and the fifth step S25 precede the sixth step S26.

In this embodiment, the machine learning unit 22 performs a second learning in addition to the first learning, and the estimation unit 24 further estimates the fracture surface pattern Ptn of the fracture surface image Im in addition to the fracture origin of the fracture surface image Im, and as a result, the fracture surface pattern Ptn is obtained. Therefore, by comparing the estimated fracture origin with the fracture surface pattern Ptn, it is possible to verify whether the estimated fracture origin is correct. Therefore, even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image Im, and the accuracy can be improved by verification.

In addition, in this embodiment, the result display unit 25 displays the estimated fracture origin and fracture surface pattern Ptn as an image superimposed on the same screen, so that it is easy to verify whether the estimated fracture origin is correct based on the relationship between the estimated fracture origin and the fracture surface pattern Ptn. Therefore, even a beginner in fracture surface analysis can determine the position of the fracture origin from the fracture surface image, and the accuracy can be improved by verification.

The result display unit 25 may display the coordinates Pt of the fracture origin obtained by estimation, the fracture origin contour diagram PD, and the fracture surface pattern Ptn on the same screen.

<Modification of the second embodiment>
In the second embodiment described above, when all three of the estimated fracture origin coordinates Pt, fracture origin contour diagram PD, and fracture surface pattern Ptn are different, it is necessary to consider the possibility that the fracture origin does not exist in the input fracture surface image Im. In such a case, the engineer needs to look at the wide-area fracture surface image WI including a fracture surface wider than the fracture surface shown in this fracture surface image Im, and re-take the fracture surface image so that the fracture origin is included. At this time, by obtaining a region division image SA described later in which the fracture surface shown in this wide-area fracture surface image WI is divided into different regions depending on the difference in the shape of the fracture surface, it becomes easier to guess the area of the fracture surface to be photographed in the broken object 1. The shape of the fracture surface refers to, for example, the unevenness of the fracture surface as described later.

In view of the above, in addition to the first learning and the second learning, the machine learning unit 32 may perform a third learning to learn so as to estimate the region-divided image SA. FIG. 11 is a schematic diagram showing the functions of the machine learning unit 32 and the estimation unit 34 regarding the third learning. The region-divided image SA is an image in which, as shown in the lower right of FIG. 11, for example, the wide-area fracture surface image WI is divided into two regions, an uneven region having a relatively large amount of unevenness on the fracture surface (see the dot-hatched portion in the lower right of FIG. 11) and a flat region excluding the uneven region (see the hatched portion in the lower right of FIG. 11). In addition to the uneven region and the flat region, the region-divided image SA may be further divided into regions with different shapes (for example, a region where a clear radial pattern is recognized) (i.e., it may be divided into three), or may be further divided into other regions (i.e., it may be divided into four or more).

11, the machine learning unit 32 further performs a third learning process using a third data set including a combination of wide-area image data D31 including a fracture surface wider than the fracture surface included in the fracture surface image data D21 initially input, and area division data D32 in which the wide-area fracture surface is divided into two or more areas according to the shape of the fracture surface. The area division data D32 may be prepared, for example, by an engineer viewing the fracture surface shown in the wide-area image data D31 and coloring the fracture surface into uneven areas and flat areas, similar to the method of preparing the fracture surface pattern data D22. The learning model for the third learning is, for example, an SS model, similar to the second learning process.

As shown in the lower left of FIG. 11, the wide-area fracture surface image WI captured by the imaging unit 13 is input to the estimation unit 34. This wide-area fracture surface image WI may be captured when the initially input fracture surface image Im is captured, or may be captured after being appropriately adjusted so as to include an appropriate area of the broken object 1 (at an appropriate magnification) only when necessary.

As shown in the lower part of FIG. 11, the estimation unit 34 estimates an area division image SA obtained by dividing the input wide-area fracture surface image WI based on the third learned model M3 (learned SS model) obtained by the third learning, and as a result, an area division image SA is obtained. The result display unit 25 displays the obtained area division image SA.

In the fracture surface analysis method according to this modification, the fracture surface analysis method S20 according to the second embodiment described above further includes a third learning step of performing a third learning using a third data set, and a third estimation step of estimating an area division image SA based on the third learned model obtained in the third learning. Specifically, in the sixth step S26 shown in FIG. 10, when it is considered that the fracture origin does not exist in the input fracture surface image Im by looking at the coordinates Pt of the fracture origin, the fracture origin contour diagram PD, and the fracture surface pattern Pt displayed by the result display unit 25, the third estimation step is performed to obtain an area division image SA. Then, after estimating the area of the fracture surface to be photographed in the broken object 1, the fracture surface image Im is photographed again so that the fracture origin is included, and the fourth step S24 to the sixth step S26 are performed using the re-photographed fracture surface image Im as an input.

According to this modified example, the estimation unit 34 estimates an area division image SA in which the area of the fracture surface is divided in the input wide-area fracture surface image WI based on the third trained model M3 obtained by the third learning, so that the engineer can easily estimate the area of the fracture surface to be photographed in the damaged object 1 by looking at this area division image SA. For example, if the initially input fracture surface image Im belongs to an uneven area, the engineer can determine that the fracture origin is not present in the uneven area and consider photographing a flat area.

This makes it easier for even beginners to analyze fracture surfaces to avoid accidentally using a fracture surface image Im that does not contain the fracture origin and to determine the location of the fracture origin.

<Other embodiments>
In each of the above-described embodiments, the learning model of the first learning includes both the regression model and the classification model, but it is also possible to select either the regression model or the classification model. In response to this, the estimation unit 14 (24) may estimate the coordinates Pt of the fracture origin of the input fracture surface image Im when the learning model is a regression model, and may estimate the fracture origin contour diagram PD in the input fracture surface image Im when the learning model is a classification model.

In addition, in each of the above-mentioned embodiments, as an example of a method for preparing the fracture origin data D12 and the fracture surface pattern data D22, an engineer has looked at the fracture surface image Im and prepared the data, but this is not limited to the above. For example, the fracture surface analysis device 10 (20) may be equipped with a data generation unit (not shown) that analyzes the fracture surface of the fracture surface image data D11 (D21) and automatically generates the fracture origin data D12 and the fracture surface pattern data D22. The data generation unit may, for example, perform a binarization process on the fracture surface image Im, extract straight lines along the crack pattern (radial pattern) from the binarized image, and automatically generate the fracture origin data D12 and the fracture surface pattern data D22 by determining the fracture origin and the propagation direction of the crack on the fracture surface from the extracted straight lines. This can avoid the burden required for an engineer to prepare all the fracture origin data D12 and the fracture surface pattern data D22.

1 Broken object 10 Fracture surface analysis device according to the first embodiment 11 Data set storage unit 12 Machine learning unit 13 Photography unit 14 Estimation unit 15 Result display unit 16 Output data storage unit Im Input fracture surface image D1 First dataset D11 Fracture surface image data (first dataset)
D12 Fracture origin data (first data set)
M1 First trained model MR Trained regression model (first trained model)
MC Trained classification model (first trained model)
Pt: Coordinates of the fracture origin obtained by estimation PD: Contour map of the fracture origin obtained by estimation (probability distribution)
S10 Fracture surface analysis method according to the first embodiment S11 First step (first learning step)
S12: Second step S13: Third step (first estimation step)
S14 Fourth step 20 Fracture surface analysis device according to the second embodiment 21 Data set storage unit 22 Machine learning unit 24 Estimation unit 25 Result display unit D2 Second data set D21 Fracture surface image data (second data set)
D22 Fracture surface pattern data (second data set)
M2 Second trained model MS Trained semantic segmentation model (second trained model)
Fracture surface pattern obtained by Ptn estimation S20 Fracture surface analysis method according to the second embodiment S21 First step (first learning step)
S22 Second step (second learning step)
S23: Third step S24: Fourth step (first estimation step)
S25 Fifth step (second estimation step)
S26 Sixth step 31 Machine learning unit 34 according to the modified example of the second embodiment Estimation unit D31 according to the modified example of the second embodiment Wide area image data D32 Region division data M3 Third trained model SA Region division image

Claims (11)

A machine learning unit that performs a first learning process using a first data set including a combination of fracture surface image data and fracture origin data;
An estimation unit that estimates a fracture origin of an input fracture surface image based on a first learned model obtained by the first learning ,
The learning model for the first learning can be selected from a regression model that learns to estimate the coordinates of the fracture origin and a classification model that learns to estimate the probability distribution of the existence of the fracture origin, or both of them;
The estimation unit is
When the learning model is the regression model, the coordinates of the fracture origin of the input fracture surface image are estimated;
A fracture surface analysis device that, when the learning model is the classification model, estimates a probability distribution of the presence of a fracture origin in the input fracture surface image . 2. The fracture surface analysis apparatus according to claim 1 ,
a learning model for the first learning includes the regression model and the classification model;
The estimation unit estimates coordinates of a fracture origin in the input fracture surface image and a probability distribution of a fracture origin in the input fracture surface image,
The fracture surface analysis device further comprises a result display unit that displays at least two of the input fracture surface image, the estimated coordinates of the fracture origin, and the estimated probability distribution as images superimposed on the same screen. 3. The fracture surface analysis apparatus according to claim 2 ,
A divided image generating unit determines the number n of fracture origins included in the estimated probability distribution, divides the input fracture surface image into n regions so that each region includes one of the fracture origins, and generates n divided images each showing an image of one of the n regions;
The estimation unit further estimates the coordinates of the fracture origin in each of the n divided images and a probability distribution of the existence of the fracture origin. A machine learning unit that performs a first learning process using a first data set including a combination of fracture surface image data and fracture origin data;
An estimation unit that estimates a fracture origin of an input fracture surface image based on a first learned model obtained by the first learning,
The machine learning unit further performs second learning using a second data set including a combination of fracture surface image data and fracture surface pattern data;
The estimation unit further estimates a fracture surface pattern of the input fracture surface image based on a second learned model obtained by the second learning. The fracture surface analysis apparatus according to claim 4 ,
A fracture surface analysis device, wherein the learning model for the second learning includes a semantic segmentation model that learns to estimate a fracture surface pattern. The fracture surface analysis apparatus according to claim 4 or 5 ,
The fracture surface analysis device further includes a result display unit that displays at least two of the input fracture surface image, the fracture origin estimated based on the first trained model, and the fracture surface pattern estimated based on the second trained model as images superimposed on the same screen. A machine learning unit that performs a first learning process using a first data set including a combination of fracture surface image data and fracture origin data;
An estimation unit that estimates a fracture origin of an input fracture surface image based on a first learned model obtained by the first learning,
The machine learning unit further performs a third learning using a third data set including a combination of wide-area image data including a fracture surface wider than the fracture surface included in the fracture surface image data and area division data in which the wide-area fracture surface is divided into two or more areas for each shape of the fracture surface;
The estimation unit further estimates a region division image in which the region of the fracture surface is divided in the input wide-area fracture surface image based on a third learned model obtained by the third learning. A first learning step of performing a first learning using a first data set including a combination of fracture surface image data and fracture origin data;
A first estimation step of estimating a fracture origin of the input fracture surface image based on the first trained model obtained in the first learning step ,
The learning model for the first learning can be selected from a regression model that learns to estimate the coordinates of the fracture origin and a classification model that learns to estimate the probability distribution of the existence of the fracture origin, or both of them;
The first estimation step includes:
When the learning model is the regression model, the coordinates of the fracture origin of the input fracture surface image are estimated;
When the learning model is the classification model, a probability distribution of the existence of a fracture origin in the input fracture surface image is estimated.
Fracture surface analysis method. A first learning step of performing a first learning using a first data set including a combination of fracture surface image data and fracture origin data;
A first estimation step of estimating a fracture origin of the input fracture surface image based on the first trained model obtained in the first learning step;
a second learning step of performing second learning using a second data set including a combination of the fracture surface image data and the fracture surface pattern data;
a second estimation step of estimating a fracture surface pattern of the input fracture surface image based on the second trained model obtained in the second learning step;
A fracture surface analysis method comprising : A first learning step of performing a first learning using a first data set including a combination of fracture surface image data and fracture origin data;
A first estimation step of estimating a fracture origin of the input fracture surface image based on the first trained model obtained in the first learning step;
a third learning step of performing a third learning using a third data set including a combination of wide-area image data including a fracture surface wider than the fracture surface included in the fracture surface image data and area division data in which the wide-area fracture surface is divided into two or more areas according to the shape of the fracture surface;
A third estimation step of estimating a region segmentation image in which the region of the fracture surface is segmented in the input wide-area fracture surface image based on the third learned model obtained by the third learning ;
A fracture surface analysis method comprising : A program for causing a computer to execute the fracture surface analysis method according to any one of claims 8 to 10 .

Images