Apparatus for detecting defects in steel materials and method for detecting defects in steel materials

Abstract

To provide a steel material flaw detection device and a steel material flaw detection method which can easily inspect a flaw of a steel material conveyed by a conveyance device.SOLUTION: A flaw detection device of a steel material conveyed by a conveyance device includes: an imaging part provided facing a lower surface of a steel material having the lower surface; an imaging data acquisition part for acquiring imaging data including the lower surface of the steel material imaged by the imaging part; and a detection data generation part for generating detection data on detection of a flaw on the lower surface of the steel material on the basis of the imaging data.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a steel material flaw detection device for detecting flaws in steel materials conveyed by a conveying device and a steel material flaw detection method.

Background Art

[0002] In the process of manufacturing steel materials such as thick steel plates, in order to confirm the presence or absence of surface flaws generated during manufacturing, the flaws in the steel materials are inspected. As a device used for such inspection, for example, a flaw inspection device that irradiates laser light toward the surface of a steel material with a laser projector and receives the reflected light with a light receiver and converts it into a voltage intensity signal is disclosed in Patent Document 1. In Patent Document 1, the flaw inspection device is installed on a production line, and flaw inspection is performed on the steel materials on the production line.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] The steel materials that have undergone flaw inspection on the production line are conveyed to a storage location by a conveying device and stored. In order to ensure that the steel materials have a certain quality, it is necessary to confirm whether new flaws have been formed in the steel materials between the time when flaw inspection is performed on the production line and the time when they arrive at the storage location.

[0005] For this reason, conventionally, when the steel materials arrive at the storage location, they are inspected again to check whether new flaws have been formed. The inspection of the flaws in the steel materials at the storage location is carried out with the lower surface of the steel material being conveyed while in contact with the conveying device. Therefore, it is easy to inspect the upper surface, but there is a problem that the lower surface can only be inspected by inverting the steel material.

[0006] The reversing of steel materials is carried out using a machine called a reversing machine. However, this process takes a lot of time to move the steel materials to the reversing machine, which prolongs the inspection of the steel materials for defects. Consequently, the so-called lead time, which is the time or number of days from when an order is placed until the product is delivered, also becomes longer.

[0007] Furthermore, due to limited storage space, it is difficult to install large-scale inspection equipment. For this reason, users visually inspect the steel materials for defects at the storage site, which presents a problem as the accuracy of the inspection can be low depending on the user's skill level. In addition, visual inspection by users presents a problem as defects can be difficult to see depending on the surface properties of the steel material (color, faint scratches, etc.).

[0008] The present invention has been made in view of the above problems, and aims to provide a steel defect detection device and a steel defect detection method that can easily inspect steel materials for defects transported by a transport device. [Means for solving the problem]

[0009] To solve the above problems, the present invention has the following features.

[0010] [1] A defect detection device for steel materials transported by a transport device, An imaging unit provided so as to face the lower surface of a steel material having a lower surface, An imaging data acquisition unit acquires imaging data including the lower surface of the steel material that has been imaged by the imaging unit, A detection data generation unit generates detection data related to the detection of defects on the lower surface of the steel material based on the imaging data, A defect detection device for steel materials having the following features. [2] It has a reference data acquisition unit that acquires reference data that serves as the standard for defects in the steel material, The steel defect detection device according to [1], wherein the detection data generation unit generates the detection data based on the imaging data and the reference data. [3] The aforementioned reference data has a priority set according to the type of defect in which it is detected. The steel defect detection device according to [2], wherein the detection data generation unit generates the detection data according to the degree of agreement between the imaging data and the reference data, and determines the degree of agreement in order from the reference data with the highest priority. [4] The steel material is conveyed in one direction by the conveying device. The steel defect detection device according to [3], wherein the priority is set such that the defect formed along the one direction is given a higher priority than the other forms of the defect. [5] The steel material is transported on a transport table having multiple table rolls. The steel defect detection device according to [3], wherein the priority is set such that the defect having a shape corresponding to the shape of the table roll is set higher than the defect having other forms. [6] The steel material defect detection device according to [2] or [3], wherein the reference data is generated using the imaging data in which the defect is detected on the lower surface of the steel material, from among the imaging data used to generate the detection data. [7] The detection data generation unit, A steel defect detection device according to [1], comprising a machine learning model that takes the aforementioned imaging data as input and outputs the aforementioned detection data. [8] The aforementioned machine learning model, A first machine learning model takes the aforementioned imaging data as input and outputs an image of the lower surface of the steel material, A defect detection device for steel materials according to [7], comprising a second machine learning model that takes an image of the lower surface of the steel material as input and outputs the detection data. [9] The machine learning model is generated using images of the lower surface of the steel material having the defect and images of the lower surface of the steel material not having the defect as training data. The steel material flaw detection device according to [7], which is generated using an image including the form of the lower surface of the steel material as the teacher data of the image in which the lower surface of the steel material has no flaw.

[10] The second machine learning model is generated using, as teacher data, an image of the lower surface of the steel material having a flaw and an image assuming that the lower surface of the steel material has no flaw. The steel material flaw detection device according to [8], which is generated using an image including the form of the lower surface of the steel material as the teacher data assuming that the lower surface of the steel material has no flaw.

[11] The steel material flaw detection device according to any one of [1] to

[10] , wherein the imaging unit has a drive unit that moves along the shape of the steel material.

[12] The steel material flaw detection device according to any one of [1] to

[11] , which has a display unit that displays the imaging data in which a flaw is detected on the lower surface of the steel material among the imaging data used for generating the detection data.

[13] The steel material flaw detection device according to

[12] , which has a display unit that displays the imaging data in which a flaw is detected on the lower surface of the steel material among the imaging data used for generating the detection data.

[14] A method for detecting flaws in a steel material conveyed by a conveying device, comprising: an imaging step of imaging the lower surface of a steel material having a lower surface; an imaging data acquisition step of acquiring imaging data including the lower surface of the steel material imaged in the imaging step; a detection data generation step of generating detection data regarding the detection of flaws on the lower surface of the steel material based on the imaging data. A method for detecting flaws in a steel material, comprising the above steps.

[15] The method for detecting flaws in a steel material according to

[14] , further comprising a reference data acquisition step of acquiring reference data serving as a reference for the flaws in the steel material, wherein the detection data generation step generates the detection data based on the imaging data and the reference data. ​

[16] In the detection data generation step, using a machine learning model that takes the imaging data as input and outputs the detection data, the steel material flaw detection method according to

[14] , in which the detection data is generated.

[17] The machine learning model includes a first machine learning model that takes the imaging data as input and outputs an image of the lower surface of the steel material, and a second machine learning model that takes the image of the lower surface of the steel material as input and outputs the detection data, the steel material flaw detection method according to

[16] .

[18] The machine learning model is generated using an image including the shape of the lower surface of the steel material as teacher data for the detection data on the assumption that the lower surface of the steel material does not have the flaw, the steel material flaw detection method according to

[16] .

[19] The second machine learning model is generated using an image including the shape of the lower surface of the steel material as teacher data for the detection data on the assumption that the lower surface of the steel material does not have the flaw, the steel material flaw detection method according to

[17] .

Advantages of the Invention

[0011] According to the steel material flaw detection device of the present invention, based on imaging data including the lower surface of the steel material imaged by an imaging unit provided so as to face the lower surface of the steel material, detection data regarding the detection of flaws on the lower surface of the steel material is generated. For this reason, the user can perform a flaw inspection without inverting the steel material. Therefore, the flaw inspection of the steel material can be easily performed, and the inspection time for flaws in the steel material can be shortened. As a result, the so-called lead time can be shortened. Further, since the flaw inspection of the steel material is not performed only by visual inspection by the user, it is possible to prevent the detection accuracy of flaws from being biased. Furthermore, since the detection data is generated using the imaging data, it is possible to detect flaws regardless of the surface properties (color, thin flaws, etc.) of the steel material.

Brief Description of the Drawings

[0012] [Figure 1] This is an explanatory diagram showing an overview of a defect detection device for steel materials. [Figure 2] This is a block diagram showing the functional blocks of a defect detection device for steel materials. [Figure 3] This is a flowchart illustrating the process for detecting defects in steel materials. [Figure 4] This is an explanatory diagram showing how imaging data is displayed on the display unit. [Figure 5] Figure 1 is an explanatory diagram showing the configuration of the imaging unit. [Figure 6] This is a block diagram showing the functional block of the steel defect detection device according to the second embodiment. [Figure 7] This is an explanatory diagram showing an example of training data to be used to train the second machine learning model. [Figure 8] This is an explanatory diagram showing an example of training data to be used to train the second machine learning model. [Figure 9] This is an explanatory diagram showing an example of training data to be used to train the second machine learning model. [Figure 10] This is a flowchart showing the processing steps for the steel material defect detection method according to the second embodiment. [Figure 11] This graph shows the time required for the defect detection test on the underside of the steel materials in the examples and comparative examples. [Modes for carrying out the invention]

[0013] (First Embodiment) Embodiments of the present invention will be described below with reference to the drawings. Figure 1 shows an overview of the steel material defect detection device. As shown in Figure 1, the steel material defect detection device 100 detects defects in the steel material 20 that has been transported by the transport device 10.

[0014] The steel material 20 is not particularly limited as long as it has a bottom surface. Examples of the steel material 20 include steel plates, structural steel, rails, steel bars, steel pipes, etc. In this embodiment, an example in which a steel plate having an upper surface and a lower surface is used as the steel material 20 will be described.

[0015] The conveying device 10 is installed, for example, between the end of the production line PL and a designated storage location. A device commonly referred to as a transfer device can be used as the conveying device 10.

[0016] The conveying device 10 has a base 11 that is formed in a rectangular shape when viewed from above, and a conveying drive unit 12 formed on the base 11 along one direction D1. The conveying drive unit 12 is formed using an annular member such as a chain and a plurality of gears (not shown) that fit with the annular member, and is provided to be able to move circulatingly in one direction D1. Multiple conveying drive units 12 are laid in the width direction of the base 11 parallel to one direction D1. Therefore, when the conveying drive unit 12 is operated, the steel material 20 placed on the conveying drive unit 12 is conveyed along one direction D1.

[0017] At the end of the transport device 10, a lifting magnet 30 is provided as a moving device for moving the steel material 20 upward. Below the lifting magnet 30, an imaging unit 40 is provided. In other words, the imaging unit 40 is positioned to face the lower surface of the steel material 20.

[0018] The steel material 20 may be placed on a stand (not shown) having a frame-shaped support portion that surrounds the steel material 20. In such a case, the imaging unit 40 may be positioned below the support portion of the stand, facing the lower surface of the steel material 20.

[0019] The imaging unit 40 can use, for example, a camera capable of generating still images as image data. The imaging unit 40 is not limited to a camera capable of capturing still images; a camera capable of capturing video may also be used.

[0020] The steel material defect detection device 100 has a display unit 50 that displays imaging data generated by the imaging unit 40. The display unit 50 can be any device capable of displaying imaging data, such as a liquid crystal display or an organic EL display.

[0021] The steel material defect detection device 100 has a control unit 60 that controls the entire device. The control unit 60 is communicated with the imaging unit 40 and the display unit 50.

[0022] Figure 2 shows the functional blocks of the steel defect detection device 100. As shown in Figure 2, the steel defect detection device 100 includes an imaging unit 40, a display unit 50, a control unit 60, and an image database (hereinafter referred to as DB) 70. The imaging unit 40, the display unit 50, the control unit 60, and the image DB 70 are connected to each other via a transmission line 80 so as to be able to communicate with each other.

[0023] The image DB 70 is not particularly limited, but can use known storage means such as an HDD (hard disk drive) or SSD (solid state drive). The image DB 70 stores the imaging data captured by the imaging unit 40.

[0024] The imaging data should be generated in a way that allows for the identification of the steel material 20. For example, the imaging data should be linked to the time when it was captured by the imaging unit 40. Furthermore, the imaging data should be recorded in conjunction with lighting environment information such as the number of lighting fixtures and the illuminance of the lighting fixtures at the time the imaging data was generated.

[0025] Furthermore, the image DB70 stores reference data that serves as a standard for defects in the steel material 20. The reference data is image data in which defects are formed on the steel material 20. For example, the reference data can be imaging data in which defects on the steel material 20 are detected. Reference data is provided for each type of defect in the steel material 20.

[0026] Reference data should ideally be linked to the time when it was generated. Furthermore, reference data should ideally be recorded in conjunction with lighting environment information and the white balance of the image itself at the time the reference data was generated.

[0027] The reference data should have a priority set for detection according to the type of defect. For example, the priority should be set so that defects in the steel material 20 formed along one direction D1 are given a higher priority than defects of other types. Linear defects are an example of such defects. Linear defects are formed, for example, when there is a defect in the operation of at least one of the multiple transport drive units 12 of the transfer.

[0028] Furthermore, when the steel material 20 is transported on a transport table (not shown) having multiple table rolls (not shown), it is preferable that the priority of defects corresponding to the shape of the table rolls be set higher than that of other types of defects. Examples of such defects include indentations. For example, if the steel material 20 strikes the surface of the table roll and is gouged, forming a convex portion, then as the steel material 20 is transported, the convex portion is pressed against the surface of the steel material 20, forming an indentation.

[0029] The control unit 60 is a computer including a CPU, ROM, and RAM. The control unit 60 includes an imaging data acquisition unit 61 that acquires imaging data of the lower surface of the steel material 20 captured by the imaging unit 40, and a reference data acquisition unit 62 that acquires reference data that serves as a standard for defects in the steel material 20. The control unit 60 also includes a detection data generation unit 63 that generates detection data related to the detection of defects on the lower surface of the steel material 20 based on the imaging data and the reference data.

[0030] The imaging data acquisition unit 61, the reference data acquisition unit 62, and the detection data generation unit 63 are realized by reading data and programs (computer software) stored in ROM, and performing calculation processing based on the data and according to the programs.

[0031] The imaging data acquisition unit 61 acquires imaging data by reading it from the image DB 70. The imaging data acquisition unit 61 may acquire the time the imaging data was captured and lighting environment information recorded in the imaging data.

[0032] The reference data acquisition unit 62 acquires reference data by reading it from the image DB 70. Specifically, the reference data acquisition unit 62 should acquire reference data that is close to the time the image data was captured and the lighting environment information recorded in the image data.

[0033] The detection data generation unit 63 extracts feature points from, for example, the imaging data and reference data, calculates the degree of similarity, and generates detection data according to the degree of similarity. Feature points are set, for example, based on the amount of change of each pixel in the imaging data relative to its surroundings. Specifically, a pixel can be used as a feature point if its value changes by a larger amount than the values ​​of the surrounding pixels. For feature point extraction, algorithms such as SIFT (Scale Invariant Feature Transform) can be used.

[0034] The detection data generation unit 63 generates detection data indicating that a defect has been formed in the steel material 20 if the degree of agreement exceeds a predetermined standard. The detection data generation unit 63 also performs the same determination for other standard data if the degree of agreement is below the predetermined standard. If the degree of agreement for all standard data is below the predetermined standard, the detection data generation unit 63 generates detection data indicating that no defect has been formed in the steel material 20.

[0035] The detection data generation unit 63 should determine the degree of agreement starting with the reference data with the highest priority. By determining the degree of agreement in this order of priority, defects in the steel material 20 can be detected efficiently and accurately.

[0036] The detection data generation unit 63 should generate detection data that includes the type and extent of the defect. Specifically, the detection data generation unit 63 should generate detection data that includes the shape of the defect (width, length, area, direction, contour, perimeter of the contour), the location of the defect on the steel material 20, and the distribution of the density (luminance) of the pixels determined to be defective. For example, image recognition software can be used as the detection data generation unit 63.

[0037] Figure 3 shows the flow of the method for detecting defects in steel materials. The flow of the method for detecting defects in steel materials starts, for example, when the lifting magnet 30 is activated. As shown in Figure 3, the imaging unit 40 images the lower surface of the steel material 20 and generates imaging data (step S11). The imaging unit 40 transmits the imaging data generated in the imaging process of step S11 to the control unit 60.

[0038] The control unit 60 stores the image data in the image database 70 upon receiving it. The image data acquisition unit 61 reads and acquires the image data from the image database 70 (step S21).

[0039] Furthermore, it is preferable that the imaging data acquisition step in step S21 acquires imaging data that has undergone prior image processing. Examples of image processing include adaptive binarization and closing / opening processing.

[0040] Adaptive binarization is an image processing method that performs binarization using a threshold value defined for each pixel. In adaptive binarization, for example, the average value of the surrounding pixels can be set as the threshold value. With adaptive binarization, by setting a threshold in this way, even in images where the brightness changes in certain areas, the effect of these brightness fluctuations can be reduced, and the image can be binarized.

[0041] In other words, in general binarization processes, one threshold is used for each image. Therefore, in images where the brightness changes in certain areas, there is a risk of inappropriate binarization occurring, where areas other than defects are emphasized due to the brightness fluctuations. Consequently, by using adaptive binarization, a good binarized image can be obtained, making it possible to improve the accuracy of defect detection in steel materials.

[0042] Closing-opening processing is an image processing technique in which a binarized image undergoes multiple dilation and condensation processes. Specifically, black noise can be removed during the dilation-to-condensation process, and white noise can be removed during the condensation-to-dilation process. By performing closing-opening processing, such noise can be removed, thereby improving the accuracy of defect detection in steel materials.

[0043] The reference data acquisition unit 62 reads and acquires reference data from the image DB 70 (step S22). For example, the reference data acquisition unit 62 acquires reference data that is close to the imaging time and lighting environment information of the imaging data acquired in the imaging data acquisition process of step S21.

[0044] The detection data generation unit 63 generates detection data based on the imaging data acquired in the imaging data acquisition step of step S21 and the reference data acquired in the reference data acquisition step of step S22 (step S23). The detection data generation unit 63 determines whether the generated detection data indicates that a defect has been formed on the steel material 20 (step S24).

[0045] In the determination in step S24, if the generated detection data indicates that no defects have been formed on the steel material 20 (step S24: NO), the process is terminated.

[0046] In the determination in step S24, if the generated detection data indicates that a defect has been formed on the steel material 20 (step S24: YES), the detection data generation unit 63 transmits the imaging data used to generate the detection data to the display unit 50.

[0047] When the display unit 50 receives imaging data, it displays the data (step S31). In other words, the display unit 50 displays the imaging data used to generate the detection data in which a defect was detected on the lower surface of the steel material 20. The display unit 50, for example, displays a confirmation completion button on its display surface, and when the user selects this button, the process ends.

[0048] In this way, by displaying the imaging data on the display unit 50, the user can visually inspect the imaging data of the steel material 20 in which defects were detected. Therefore, the user only needs to confirm whether the detection is correct or incorrect, significantly reducing the inspection burden. Furthermore, the user's confirmation results should be fed back into the reference data. By doing so, the defect detection accuracy of the steel material defect detection device 100 can be improved.

[0049] In other words, the reference data should be generated using the imaging data used in the detection data generation process in step S23, specifically the imaging data in which a defect was detected on the lower surface of the steel material 20, i.e., the imaging data that was determined to have a defect in the judgment in step S24. By generating the reference data in this way, the reliability of the reference data can be increased, and the accuracy of defect detection can be improved.

[0050] Figure 4 shows the display unit 50 in the display step S31 of Figure 3, where the image data is displayed. As shown in Figure 4, the display unit 50 displays image data of the lower surface of the steel material 20.

[0051] In the image data, for example, the area indicated by the white dashed circle in the figure is shown as the area where the defect is formed. Furthermore, a magnified image of the area where the defect is formed is displayed below the image data. This display allows the user to easily confirm the location, shape, and surface characteristics of the defect.

[0052] Furthermore, the way the underside of the steel material 20 is captured in the image data differs depending on the lighting environment. Figure 4 shows an example where image data was generated using four lighting devices. The left side of the figure shows image data generated during daytime hours, and the right side shows image data generated during nighttime hours. As shown, the way the background of the image data is captured differs depending on the lighting environment.

[0053] Furthermore, depending on the size and shape of the steel material 20 and the field of view of the imaging unit 40, it may not be possible to capture the entire material at a single location. In such cases, the imaging unit 40 may be provided at multiple locations, and the imaging data from each location may be combined to obtain a single set of imaging data.

[0054] Alternatively, the imaging unit 40 may be moved along the shape of the steel material 20 to combine imaging data captured at multiple positions to obtain a single image data. For example, the imaging unit 40 may be moved by providing a drive unit to it.

[0055] Figure 5 shows the configuration of the imaging unit 40 having a drive unit. As shown in Figure 5, a drive unit 41 is provided below the imaging unit 40. The drive unit 41 has, for example, a roller 42 that fits into a groove in the rail 90. Therefore, when the drive unit 41 is operated, the roller 42 rotates and the imaging unit 40 moves in the direction in which the rail 90 is laid.

[0056] The rails 90 should be laid according to the shape of the steel material 20. For example, if the steel material 20 is a rectangular plate, the rails 90 should be laid along the longitudinal direction of the steel material 20. If the steel material 20 is a circular disc, the rails 90 should be laid in a ring shape. By laying the rails 90 in this way, the imaging unit 40 can be moved along the shape of the steel material 20.

[0057] The drive unit 41 is not limited to this configuration, and may, for example, be composed of multiple wheels. In the case of such a configuration, instead of rails 90, walls may be provided along the direction in which the drive unit 41 moves to form a movement course, and the movement of the imaging unit 40 may be guided by the movement course.

[0058] As described above, with the steel defect detection device 100 of the present invention, detection data is generated based on the imaging data of the lower surface of the steel material 20 captured by the imaging unit 40 and reference data that serves as a standard for defects in the steel material 20. Therefore, the user can inspect for defects without inverting the steel material 20. Consequently, the inspection of defects in the steel material 20 can be easily performed, and the inspection time for defects in the steel material 20 can be shortened. As a result, the so-called lead time can be shortened.

[0059] Specifically, the steel defect detection device 100 can reduce the inspection time by more than half a day compared to conventional methods. Because the steel defect detection device 100 can detect defects in the steel material 20, the presence or absence of defects can be confirmed regardless of the user's skill level.

[0060] Furthermore, with the steel defect detection device 100 of the present invention, if the imaging unit 40 is positioned to face the underside of the steel material 20, other components do not need to be positioned in close proximity to the imaging unit 40. Therefore, the storage space required for defect detection of the steel material 20 can be minimized. Moreover, since the inspection of defects in the steel material 20 is no longer performed solely by the user's visual inspection, bias in the defect detection accuracy can be suppressed. Furthermore, since detection data is generated using imaging data, it is possible to detect defects regardless of the surface properties of the steel material 20 (color, faint defects, etc.). Specifically, the imaging data can be processed such as enlargement, reduction, and correction of pixel values ​​according to the surface properties of the steel material 20 (color, faint defects, etc.), making it possible to detect defects with higher accuracy than visual inspection by the user. It is also preferable to adjust the shooting conditions of the imaging unit 40 (lighting, aperture, etc.) as appropriate according to the surface properties of the steel material 20 (color, faint defects). For example, if the color of the steel material 20 is dark, it is advisable to adjust the shooting conditions by lowering the aperture value (f-number) or increasing the brightness of the lighting.

[0061] Furthermore, the display of the imaging data by the display unit 50 is an optional step. For example, the display of the imaging data by the display unit 50 may be omitted if the detection accuracy of defects on the lower surface of the steel material 20 by the steel material defect detection device 100 is high.

[0062] If the detection accuracy of defects on the lower surface of the steel material 20 is high, in the judgment of step S24, when it is determined that "defects exist", the system may instead output sound from the speaker to notify the user by displaying the image data on the display unit 50.

[0063] (Second Embodiment) In the above-described embodiment, an example was explained in which the detection data generation unit 63 generates detection data based on imaging data and reference data. The detection data generation unit is not limited to this embodiment and may include a machine learning model that takes imaging data as input and outputs detection data. Components identical to those in the first embodiment are denoted by the same reference numerals and their description is omitted.

[0064] Figure 6 shows the configuration of the steel defect detection device 200 according to the second embodiment. As shown in Figure 6, the control unit 60 has a detection data generation unit 65. The detection data generation unit 65 has a first machine learning model 65a that takes imaging data as input and outputs an image of the underside of the steel material, and a second machine learning model 65b that takes the image of the underside of the steel material as input and outputs detection data.

[0065] The first machine learning model 65a and the second machine learning model 65b can be known machine learning models such as neural networks and deep learning. The first machine learning model 65a and the second machine learning model 65b are trained using, for example, a dataset stored in a memory device as training data.

[0066] The first machine learning model 65a is generated using images containing the underside of a steel plate and images not containing the underside of a steel material as training data. By being trained using these images as training data, the first machine learning model 65a can learn the morphological features of the underside of a steel material.

[0067] Therefore, the first machine learning model 65a recognizes the region in the image that matches the features of the underside of the steel material as the underside, and can output an image of the underside extracted from the input imaging data.

[0068] In images that capture only the underside of a steel material, it is difficult to capture the boundary between the steel and non-steel materials. In other words, by outputting an image in which the first machine learning model 65a extracts the underside, it becomes possible to output an image that includes the boundary between the steel and non-steel materials. This makes it possible to detect defects formed on the ends of the steel material.

[0069] The second machine learning model 65b is generated using images of steel materials with defects on their underside and images of steel materials without defects as training data. By training with these images as training data, the second machine learning model 65b can learn the characteristics of defects on the underside of steel materials.

[0070] Examples of training data for steel materials with defects on the underside include images selected by operators in past inspections that meet arbitrarily defined criteria for defects.

[0071] Training data that indicates the underside of steel materials is free of defects can include, for example, images selected by operators in past inspections that meet arbitrarily defined criteria and are free of defects. Alternatively, images that do not meet the criteria but have visible patterns or shapes can also be used as training data. The criteria for defects can be defined, for example, by a depth of N (where N is an arbitrary number) mm or more, a width of M (where M is an arbitrary number) mm or more, etc.

[0072] Figure 7 shows an example of training data used to train the second machine learning model 65b. In Figure 7, the area enclosed by the white dotted line shows that a dotted pattern PT1 has been generated on the underside of the steel material. The dotted pattern PT1 is an example of an image in which a pattern or shape is visible to the naked eye, although it does not meet the arbitrarily defined criteria for defects. The dotted pattern PT1 is thought to be caused, for example, by contact with conveying equipment or by scale peeling off the steel material.

[0073] Figure 8 shows an example of training data used to train the second machine learning model 65b. In Figure 8, the area enclosed by the black dotted line shows that a mottled pattern PT2 has formed on the underside of the steel material. The mottled pattern PT2 is an example of an image in which a pattern or shape is visible to the naked eye, although it does not meet the arbitrarily defined criteria for defects. The mottled pattern PT2 is thought to be caused, for example, by contact with conveying equipment or by scale peeling off the steel material.

[0074] Figure 9 shows an example of training data used to train the second machine learning model 65b. In Figure 9, linear shapes PT3 are generated on the underside of the steel material in the area enclosed by the white dotted line. Linear shapes PT3 are examples of patterns or shapes that are visible to the naked eye, although they do not meet the arbitrarily defined criteria for defects. Linear shapes PT3 are thought to be caused by contact with conveying equipment or by scale peeling off the steel material.

[0075] Figure 10 shows the processing of the defect detection method according to the second embodiment. As shown in Figure 10, the imaging unit 40 images the lower surface of the steel material 20 and generates imaging data (step S11). The imaging unit 40 transmits the imaging data generated in the imaging step of step S11 to the control unit 60.

[0076] The control unit 60 stores the image data in the image database 70 upon receiving it. The image data acquisition unit 61 reads and acquires the image data from the image database 70 (step S41).

[0077] The detection data generation unit 65 generates detection data based on the imaging data acquired in the imaging data acquisition step of step S41 (step S42).

[0078] Specifically, the detection data generation unit 65 uses the first machine learning model 65a as input to the imaging data acquired in step S41 and outputs an image of the underside of the steel material. The detection data generation unit 65 also uses the second machine learning model 65b as input to the image of the underside of the steel material output by the first machine learning model 65a and outputs detection data.

[0079] The detection data generation unit 65 determines whether the detection data generated in step S42 indicates that a defect has been formed on the steel material 20 (step S43).

[0080] In the determination in step S43, if the data indicates that no defects have been formed on the steel material 20 (step S43: NO), the process is terminated.

[0081] In the determination in step S43, if the generated detection data indicates that a defect has been formed on the steel material 20 (step S43: YES), the detection data generation unit 65 transmits the imaging data used to generate the detection data to the display unit 50.

[0082] When the display unit 50 receives imaging data, it displays the data (step S31). The display unit 50, for example, displays a confirmation completion button on its display surface, and when the user selects this button, it terminates the process.

[0083] Furthermore, the first machine learning model 65a described in the above embodiment can be used arbitrarily. That is, detection data may be generated using only the second machine learning model 65b without using the first machine learning model 65a.

[0084] (Test Example 1) A defect detection test was conducted on the underside of the steel material. An example was given using the steel material defect detection device 100 described in the first embodiment. A comparative example was given in which the steel material was inverted using an inverting machine and inspected by the user's visual inspection.

[0085] Figure 11 shows the time required for the defect detection test on the underside of the steel material in the example and comparative example. As shown in Figure 11, the time for the example was 11 minutes, and the time for the comparative example was 251 minutes. Thus, the inspection time in the example can be reduced by 240 minutes compared to the comparative example.

[0086] (Test Example 2) The detection rate of defects on the underside of steel materials was confirmed using the steel material defect detection device 200 described in the second embodiment. The steel material defect detection devices 200 of Examples 2 to 4 were fabricated by changing the training data used to generate the first machine learning model and the second machine learning model.

[0087] Example 2 is an example in which a first machine learning model 65a was generated using images in which only the lower surface of the steel material was captured as training data, as the containing image that includes the lower surface of the steel plate.

[0088] Example 3 is an example in which a first machine learning model 65a was generated using images containing the lower surface of steel materials and images containing non-steel materials as training data, as the containing image containing the lower surface of the steel plate.

[0089] Example 4 is an example in which a first machine learning model 65a was generated using images containing the lower surface of steel materials and images containing non-steel materials as training data, as the containing image containing the lower surface of the steel plate.

[0090] Furthermore, Example 4 is an example in which a second machine learning model 65b was generated using training data that, while not meeting arbitrarily defined criteria for defects on the underside of the steel material, had visible patterns or shapes.

[0091] For the image data of the underside of the steel material captured by the steel material defect detection device 200 of Examples 2 to 4, a defect detection test was performed, and the defect detection rate (%) was determined for each steel material defect detection device 200. The results are shown in Table 1. The defect detection rate (%) for Examples 2 to 4 was calculated by dividing the number of defects detected by the steel material defect detection device 200 by the number of defects that should be detected as defects according to the conditions specified in the specifications of the target steel material.

[0092] Furthermore, in Examples 2 to 4, the percentage of non-defective forms (hereinafter also referred to as non-defective forms) that were detected as defects (over-detected) according to the conditions specified in the specifications of the target steel material was determined as the defect over-detection rate (%). The defect over-detection rate (%) was calculated by dividing the number of images in which over-detection occurred by the number of images that included non-defective forms.

[0093] [Table 1]

[0094] The improvement rate (%) in detection rate was calculated using the detection rate (%) of Example 2 as the baseline, and the percentage increase (%) of the detection rate (%) of Example 3 or Example 4 relative to that baseline. Furthermore, the improvement rate (%) in overdetection was calculated using the overdetection rate (%) of Example 2 as a baseline, and the percentage reduction (%) of the overdetection rate (%) of Example 3 or Example 4 relative to that baseline.

[0095] As shown in Table 1, it was found that the detection rate in Examples 3 and 4 was higher than that of Example 2. In the first machine learning models of Examples 3 and 4, images including images of non-steel materials were also used as training data, in addition to the images including the underside of the steel plate. Therefore, the amount of data is larger compared to when only images including the underside of steel materials are used as training data. For this reason, the detection accuracy improved in Examples 3 and 4, and it is thought that the detection rate of Examples 3 and 4 was higher than that of Example 2.

[0096] Furthermore, it was found that Example 4 showed an improved false positive rate compared to Examples 2 and 3. The second machine learning model in Example 4 was generated using training data that did not meet the arbitrarily defined criteria for defects on the underside of the steel material, but still showed visible patterns or shapes. Therefore, it is believed that the second machine learning model was able to distinguish and recognize defects from patterns and other forms, thus improving the false positive rate. [Explanation of symbols]

[0097] 100 Steel material defect detection device 200 Steel material defect detection device 10 Conveying device 20 Steel 40 Imaging Unit 41 Drive unit 50 displays 60 Control Unit 61 Imaging data acquisition unit 62 Reference Data Acquisition Unit 63 Detection Data Generation Unit 65 Detection Data Generation Unit 65a First Machine Learning Model 65b Second Machine Learning Model

Claims

1. A defect detection device for steel materials transported by a transport device, An imaging unit provided so as to face the lower surface of a steel material having a lower surface, An imaging data acquisition unit acquires imaging data including the lower surface of the steel material that has been imaged by the imaging unit, The system includes a detection data generation unit that generates detection data related to the detection of defects on the lower surface of the steel material based on the imaging data, The aforementioned steel material has a standard data acquisition unit that acquires a plurality of standard data sets provided for each type of defect, which serve as the standard for defects in the steel material. The detection data generation unit generates the detection data based on the imaging data and the reference data. The multiple reference data sets are configured such that the priority of detection as a defect is set according to the form of the defect. The detection data generation unit generates the detection data according to the degree of agreement between the imaging data and the reference data, and determines the degree of agreement in order from the reference data with the highest priority. The aforementioned priority is set according to the method of transporting the steel material, and is a defect detection device for steel material.

2. The steel material is conveyed in one direction by the conveying device. The steel defect detection device according to claim 1, wherein the priority is set higher for defects formed along one direction than for other forms of defects.

3. The steel material is transported on a transport table having multiple table rolls. The steel defect detection device according to claim 1, wherein the priority is set higher for defects of a shape corresponding to the shape of the table roll than for defects of other shapes.

4. The steel material defect detection device according to claim 1, wherein the reference data is generated using the imaging data in which the defect is detected on the lower surface of the steel material, from among the imaging data used to generate the detection data.

5. The detection data generation unit, The steel defect detection device according to claim 1, comprising a machine learning model that takes the aforementioned imaging data as input and outputs the aforementioned detection data.

6. The aforementioned machine learning model, A first machine learning model takes the aforementioned imaging data as input and outputs an image of the lower surface of the steel material, A defect detection device for steel materials according to claim 5, comprising a second machine learning model that takes an image of the lower surface of the steel material as input and outputs the detection data.

7. The machine learning model is generated using images of the lower surface of the steel material having the defect and images of the lower surface of the steel material not having the defect as training data. A defect detection device for steel materials according to claim 5, wherein the training data for the image of the lower surface of the steel material not having the defect is generated using an image that includes the shape of the lower surface of the steel material.

8. The second machine learning model is generated using images of the lower surface of the steel material having the defect and images of the lower surface of the steel material not having the defect as training data. A defect detection device for steel materials according to claim 6, wherein the training data used to determine whether the lower surface of the steel material is free from defects is an image that includes the shape of the lower surface of the steel material.

9. The steel material defect detection device according to any one of claims 1 to 8, wherein the imaging unit has a drive unit that moves along the shape of the steel material.

10. A steel material defect detection device according to any one of claims 1 to 8, further comprising a display unit that displays the imaging data used to generate the detection data in which the defect is detected on the lower surface of the steel material.

11. The steel material defect detection device according to claim 9, further comprising a display unit that displays the imaging data used to generate the aforementioned detection data, wherein the imaging data in which the defect is detected on the lower surface of the steel material.

12. A method for detecting defects in steel materials transported by a transport device, An imaging step of imaging the lower surface of a steel material having a lower surface, An imaging data acquisition step is to acquire imaging data including the lower surface of the steel material that was imaged in the imaging step, A detection data generation step that generates detection data related to the detection of defects on the lower surface of the steel material based on the imaging data, The process includes acquiring standard data, which serves as a standard for defects in the steel material and acquires a plurality of standard data sets provided for each type of defect, The detection data generation process generates the detection data based on the imaging data and the reference data. The multiple reference data sets are configured such that the priority of detection as a defect is set according to the form of the defect. The detection data generation step generates the detection data according to the degree of agreement between the imaging data and the reference data, and determines the degree of agreement in order from the reference data with the highest priority. The priority is set according to the manner in which the steel material is transported, and is a method for detecting defects in steel material.

13. In the detection data generation process, The method for detecting defects in steel materials according to claim 12, wherein the detection data is generated using a machine learning model that takes the imaging data as input and outputs the detection data.

14. The aforementioned machine learning model, A first machine learning model takes the aforementioned imaging data as input and outputs an image of the lower surface of the steel material, A method for detecting defects in steel materials according to claim 13, comprising a second machine learning model that takes an image of the lower surface of the steel material as input and outputs the detection data.

15. The method for detecting defects in steel materials according to claim 13, wherein the machine learning model is generated using an image that includes the shape of the lower surface of the steel material as training data for the detection data, which assumes that the lower surface of the steel material does not have the defect.

16. The method for detecting defects in steel materials according to claim 14, wherein the second machine learning model is generated using an image that includes the shape of the lower surface of the steel material as training data for the detection data, assuming that the lower surface of the steel material does not have the defect.

Images