Metal strip manufacturing method

Abstract

To provide a generation method and apparatus of a surface defect determination model of a metal band, a success / failure determination method and apparatus of the metal band and a manufacturing method of the metal band which can perform success / failure determination on a surface defect of the metal band according to the demand of a user when discriminating types and grades of the surface defect by using a surface defect meter in a manufacture line, an inspection line and the like of the metal band, and can ship the metal band at a quality level according to the demand of the user for each user.SOLUTION: A generation method of a surface defect determination model of a metal band performs machine learning on the basis of a data set including defect discrimination data D2, defect image data D3, user identification data D4 and user determination data D5, and generates a surface defect determination model using user determination prediction D6 being a success / failure determination result of the surface defect of the metal band at a current time point by the user that is predicted from input data including the defect discrimination data D2, the defect image data D3 and the user identification data D4 at a current time point as output data.SELECTED DRAWING: Figure 8

Description

[Technical field]

[0001] The present invention relates to a method and an apparatus for generating a surface defect judgment model for a metal strip, a method and an apparatus for judging the acceptability of a metal strip, and a method for manufacturing a metal strip. [Background technology]

[0002] An example of a metal strip is a strip-shaped steel plate. There are many types of surface defects that occur in the steel plate, and in order to guarantee the surface quality of the steel plate, it is important to detect the surface defects and determine their level and type. This is because it is practically difficult to manufacture a steel plate that does not contain any surface defects, and it is unavoidable to have some surface defects at some positions when manufacturing a large number of steel plates. Since steel plates shipped as products (hereinafter referred to as steel plate products) are made into metal parts, etc. through secondary processing at consumers, even if a small number of minor surface defects are contained in a part of the metal strip, if the rest of the metal strip is sound, it is economical to remove such parts, i.e., parts with surface defects, before performing secondary processing, etc. For this reason, for steel plate products, the surface quality level of the steel plate is determined depending on the number and type of surface defects contained therein, and the use of the steel plate may be changed depending on the result of the determination.

[0003] For detecting surface defects of steel plates, a surface defect meter equipped with a light source unit that irradiates light onto the surface of the steel plate being transported and an imaging unit that receives and images the regular reflected light or diffuse reflected light reflected by the surface of the steel plate is widely used. A line camera or a two-dimensional camera is often used as the imaging unit of the surface defect meter. When a line camera is used, a line scan is performed on the transported steel plate at a predetermined cycle, and the scan data is usually combined to obtain two-dimensional image data. On the other hand, when a two-dimensional camera is used, the images captured at a predetermined cycle are converted into two-dimensional image data by performing image processing to remove overlaps between images. Regardless of which camera is used, two-dimensional image data is obtained by capturing an image of the surface of the steel plate, and surface defects are judged based on the two-dimensional image data. Therefore, a surface defect meter is usually equipped with a defect discrimination unit for discriminating surface defects from the acquired two-dimensional image data. The defect discrimination unit judges the type of surface defects of the steel plate and also judges the grade indicating the severity of the surface defects.

[0004] Patent Document 1 discloses a surface inspection device for identifying the type and grade of a surface defect in real time even when the image data of the surface defect is not clear. The device is equipped with an identification means for determining an area where a surface defect is present and then identifying the type and grade of the surface defect contained in the area. Patent Document 2 discloses a surface defect inspection method using an optical surface defect meter. In this method, image processing is used on image data of a steel plate to detect defects on the front and back surfaces, and a primary judgment is made as to whether the detected surface defects are of a minor or major grade. Furthermore, when a defect is detected at approximately the same position on the front and back surfaces among those that are primarily judged to be minor, the surface defect is secondarily judged to be of a major grade. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 8-105842 [Patent Document 2] JP 2014-32058 A Summary of the Invention [Problem to be solved by the invention]

[0006] As described above, optical surface defect meters that detect surface defects in steel plates are often equipped with a defect discrimination unit for discriminating the type and grade of the detected surface defects. However, the discrimination criteria of the discrimination methods for the type and grade of surface defects described in Patent Document 1 and Patent Document 2 are set for each inspection line or each manufacturing process of the steel plate, and do not necessarily match the criteria for pass / fail judgment as a steel plate product. For example, even in the case of steel plate products used as thin steel plates for automobiles, the criteria for pass / fail judgment for surface defects are often different between steel plate products used as inner plates of automobiles and steel plate products used as outer plates. In addition, the criteria for pass / fail judgment of acceptable surface defects differ depending on the difference in the press processing process, painting process, etc. when a customer performs secondary processing, etc. at the shipping destination of the steel plate product. Therefore, even if the type and grade of surface defects are discriminated using a surface defect meter in the manufacturing line or inspection line of the steel plate, the pass / fail judgment of the steel plate product needs to be performed again for each customer or each use of the steel plate product. On the other hand, if the criteria for determining surface defects in steel sheet production and inspection lines are set to match the strictest steel sheet applications and customer requirements, unnecessarily strict quality control will be required, which may result in a decrease in product yield and an increase in manufacturing costs.

[0007] The present invention has been made to solve the above-mentioned problems, and aims to provide a method and apparatus for generating a surface defect judgment model for metal strip, a method and apparatus for judging the pass / fail of metal strip, and a method for manufacturing metal strip, which enable a pass / fail judgment of surface defects of metal strip in accordance with the needs of each customer when a surface defect meter is used to determine the type and grade of surface defects in metal strip production lines or inspection lines, thereby enabling metal strip to be shipped to each customer at a quality level that meets the customer's needs. [Means for solving the problem]

[0008] In order to achieve the above object, the present invention provides [1] A method for generating a surface defect determination model for a metal strip, comprising the steps of: forming a determination record database that accumulates a data set including defect determination data that determines surface defects on a metal strip, defect image data that images the surface of the metal strip including the surface defects, consumer identification data that identifies a consumer to which the metal strip is shipped, and consumer determination data that is a pass / fail determination result of the consumer for the surface defects on the metal strip; and performing machine learning on the correlation between the defect determination data, the defect image data, the consumer identification data, and the consumer determination data based on the data set accumulated in the determination record database, thereby generating a surface defect determination model that uses data including the current defect determination data for the metal strip, the defect image data for the metal strip at the current time, and the consumer identification data for the metal strip at the current time as input data, and a consumer determination prediction that is the pass / fail result of the consumer's determination of the surface defects on the metal strip at the current time as output data. [2] A method for generating a surface defect judgment model for steel plate described in [1] above, wherein the defect discrimination data includes a defect type representing a type of the surface defect, a defect grade representing a grade of the surface defect, and defect position information representing position information of the surface defect. [3] The method for generating a surface defect determination model for steel plate described in [1] above, wherein the machine learning uses at least one of a neural network, decision tree learning, random forest, and support vector regression. [4] The method for generating a surface defect determination model for steel plate described in [2] above, wherein the machine learning uses at least one of a neural network, a decision tree learning, a random forest, and a support vector regression. [5] A method for determining whether or not a metal strip is acceptable, comprising the steps of: acquiring defect determination data that identifies surface defects on a metal strip; defect image data that images the surface of the metal strip including the surface defects; and consumer identification data that identifies a consumer of the metal strip; inputting the acquired defect determination data, the defect image data, and the consumer identification data as input data into the surface defect determination model generated by the method for generating a surface defect determination model for a metal strip described in any one of [1] to [4] above; outputting the consumer determination prediction, which is the output data; and determining whether or not the metal strip is acceptable based on the output consumer determination prediction. [6] A method for manufacturing a metal strip, comprising the steps of: manufacturing a metal strip that is judged as pass / fail by the method for judging pass / fail of a metal strip described in [5] above, over the entire length of the metal strip. [7] A surface defect determination model generation device for metal strip, comprising: a machine learning unit that performs machine learning based on a dataset including defect determination data that determines surface defects of a metal strip based on image data of the surface defects of the metal strip, defect image data that is image data including the surface defects among the image data, consumer identification data that identifies a consumer to which the metal strip is to be shipped, and consumer determination data that is a pass / fail determination result of the consumer for the surface defects of the metal strip, and generates a surface defect determination model for the correlation between the defect determination data, the defect image data, the consumer identification data, and the consumer determination data, wherein the surface defect determination model has as input data the defect determination data of the metal strip at the current time, the defect image data of the metal strip at the current time, and the consumer identification data of the metal strip at the current time, and has as output data a consumer determination prediction that is the pass / fail determination result of the consumer for the surface defects of the metal strip at the current time. [8] A surface defect judgment model generation device for a metal strip as described in [7] above, comprising: a camera that images surface defects of the metal strip and acquires the image data; a defect judgment unit that creates the defect judgment data based on the image data; a quality information acquisition unit that acquires the consumer identification data and the consumer judgment data from the consumer; and a judgment performance database that accumulates a data set including the defect judgment data, the defect image data, the consumer identification data, and the consumer judgment data, wherein the machine learning unit generates the surface defect judgment model based on the data set accumulated in the judgment performance database. [9] A metal strip pass / fail judgment device comprising a surface defect judgment unit that inputs the current metal strip defect judgment data, the current defect image data, and the current customer identification data as input data into the surface defect judgment model generated by the metal strip surface defect judgment model generation device described in [7] above, and outputs the customer judgment prediction as the output data. Effect of the Invention

[0009] According to the method for generating a surface defect judgment model for metal strip of the present invention, it is possible to judge the pass / fail of surface defects of metal strip according to the needs of each customer, and metal strip of a quality level according to the customer's needs can be shipped to each customer. [Brief description of the drawings]

[0010] [Figure 1] These are diagrams showing an example of the configuration of a surface defect meter to which the present invention can be applied, where (a) of Figure 1 is a diagram in which a camera is positioned to receive the specularly reflected light of light irradiated from a light source unit, and (b) of Figure 1 is a diagram in which a camera is positioned to receive the diffusely reflected light of light irradiated from the light source unit. [Diagram 2] 2(a) is a photograph substituting a drawing showing an example of a dent, and FIG. 2(b) is a photograph substituting a drawing showing an example of a scale defect. [Diagram 3] 4 is a block diagram showing an example of data input / output to / from a defect determination unit; FIG. [Figure 4] 11 is a block diagram showing an example of data stored in a judgment record database by a surface defect meter. FIG. [Diagram 5] FIG. 13 is a diagram showing an example of a data set stored in a judgment result database. [Figure 6] FIG. 2 is a diagram illustrating an example of a surface defect determination model generating unit that generates a surface defect determination model according to an embodiment of the present invention. [Figure 7] FIG. 4 is a diagram showing an example of a structure of a surface defect determination model. [Figure 8] FIG. 11 is a block diagram for explaining an example of a method for determining pass / fail of a metal strip according to another embodiment of the present invention. [Figure 9] FIG. 1 is a diagram showing an example of a surface defect determination model generating device for a metal strip according to the present invention. [Figure 10] 1 is a diagram showing an example of a metal strip pass / fail determination device according to the present invention; DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] [Surface Defect Meter] An example of the metal strip according to the embodiment of the present invention is a strip-shaped steel plate that is continuously manufactured. The surface defect determination model according to the embodiment of the present invention is generated when surface defects occurring in the above-mentioned steel plates, particularly in thin steel plates whose plate thickness is thinner than a predetermined value, are to be inspected. In addition, in the manufacturing process of thin steel plates, thin steel plates processed in any process after the hot rolling process are to be inspected for surface defects. Furthermore, at least after the hot rolling process, the thin steel plate is wound into a coil shape and then processed in each process, so in the embodiment of the present invention, the thin steel plate may be described as a "metal strip". Note that the surface defect determination model of the steel plate according to the embodiment of the present invention is not limited to the above-mentioned thin steel plate, but can be applied to steel plates that are subjected to various processes in the form of a sheet, such as thick steel plates that are thicker than the thin steel plate. The surface defect determination model according to the embodiment of the present invention is not limited to metal strips made of steel materials, but can be applied to metal strips including aluminum, copper, etc.

[0012] Specific examples of manufacturing processes in which surface defect inspection is performed include a hot rolling process, a pickling process, a cold rolling process, a continuous annealing process, a plating process, and a temper rolling process. The present invention can be applied to surface defect inspection at any position in the equipment in each process from the upstream side, i.e., the entrance side, to the downstream side, i.e., the exit side, in the conveying direction of the steel sheet. In particular, it is preferable to apply the present invention to a final inspection process in which quality assurance of the metal strip is performed.

[0013] Fig. 1 is a diagram showing an example of a surface defect meter to which the present invention can be applied. The surface defect meter 1 shown in Fig. 1 is installed at a location determined by design in any of the manufacturing processes described above. The surface defect meter 1 includes an image data acquisition unit, i.e., a camera 3, which acquires image data D1 by capturing an image of at least one of the front and back surfaces (hereinafter referred to as the steel plate surface) of a steel plate 2 transported by a transport device (not shown), and a defect discrimination unit 4 which discriminates surface defects of the steel plate 2 based on the image data D1 captured by the camera 3.

[0014] Moreover, the surface defect meter 1 of the steel sheet 2 shown in Fig. 1 further includes a light source unit 5 that irradiates light onto the steel sheet surface. The light source unit 5 may be a slit light source, a surface light source, or the like. The light irradiated from the light source unit 5 may not only be white light, but also light of a specific wavelength appropriately selected according to the surface defects of the steel sheet 2 detected by the camera 3. Alternatively, the light source unit 5 may be configured to irradiate light other than visible light, such as infrared light.

[0015] A line camera or a two-dimensional camera can be used as the camera 3 that receives the light reflected from the steel plate surface. When a line camera is used, the system is configured to perform line scanning on the transported steel plate 2 at a predetermined cycle, and combine successively acquired line scan data to convert it into two-dimensional image data for the surface of the steel plate 2. When a two-dimensional camera is used, the system is configured to perform image processing on the images captured at each predetermined cycle to remove overlapping portions between images, and convert it into two-dimensional image data for the steel plate surface. In this way, two-dimensional image data capturing the surface of the steel plate 2 can be obtained.

[0016] Regarding the installation position of the camera 3 relative to the light source unit 5, the camera 3 is installed at a position where it can receive at least one of specularly reflected light and diffusely reflected light that is generated when light irradiated from the light source unit 5 is reflected on the surface of the steel plate. The example shown in FIG. 1(a) is an example where the camera 3 is positioned so as to receive specularly reflected light of the light irradiated from the light source unit 5. The positioning of the camera 3 shown in FIG. 1(a) is suitable for detecting uneven defects on the steel plate surface. This is because it utilizes the characteristic that the intensity of the specularly reflected light captured by the camera 3 is reduced compared to when the surface of the steel plate 2 is flat, since the optical path of the specularly reflected light changes due to the uneven surface of the steel plate.

[0017] The example shown in FIG. 1(b) is an example in which the camera 3 is arranged so as to receive diffusely reflected light of light irradiated from the light source unit 5. The arrangement of the camera 3 shown in FIG. 1(b) is suitable for detecting pattern defects on the surface of the steel sheet. Pattern defects may be composed of fine irregularities close to the wavelength band of light, and when the camera 3 is arranged as shown in FIG. 1(b), scattered light due to the irregularities close to the wavelength band of light can be detected, and the detection sensitivity of the pattern defects can be improved. In the embodiment of the present invention, one of the above two arrangements of the camera 3 may be selected according to the surface defect to be judged. Alternatively, although not shown in detail, a configuration may be adopted in which multiple cameras are prepared and the multiple cameras receive both regular reflected light and diffuse reflected light. In that case, a single piece of image data can be generated by superimposing multiple images acquired at the same time by multiple cameras on each other.

[0018] [Surface defect inspection target] The surface defects of the steel sheet 2 to be inspected in the embodiment of the present invention are mainly uneven defects and pattern defects. The uneven defect is a defect having a shape composed of a concave or convex part or a plurality of concave and convex shapes on the surface of the steel sheet 2, and the uneven shape can be recognized to some extent by visual inspection. Specifically, the uneven defect is a circular, elliptical, streak-like or irregular surface defect having a diameter of 0.1 mm or more on the surface of the metal strip, and the depth of the concave or the height of the convex is approximately 5 μm or more and 1000 μm or less. The size of the uneven defect, that is, the diameter of the concave or convex part, the depth of the concave or the height of the convex part, may be the above-mentioned diameter, depth or height. Conventionally, the uneven defect is further divided into finer categories, and examples thereof include dents, dull baldness, scale defects, compression defects, scratches, and scab defects. Dents are bite-like dent defects generated when a solid foreign object is caught during the sheet passing through the line. Dulled baldness is a defect in which the unevenness (dull grain) of the rolling roll is not transferred to a part of the metal strip during the rolling process using dulled rolls, i.e., the cold rolling process or the temper rolling process. It is caused by wear of the rolling roll or adhesion of foreign matter to the rolling roll. Scale defects are defects in which mill scale or the like is embedded in the surface in stripes or grains. They can occur when scale removal during hot rolling is insufficient. Pressed defects are defects that occur when a part of the interlayer of the coil adheres to the metal strip when it is subjected to batch annealing in the coil state, and the adhered part is peeled off. Scratched defects are defects that are observed as sliding marks on the surface of the metal strip when the metal strip comes into contact with foreign matter such as equipment parts. Scaling defects are defects that are formed in a wrap shape due to surface peeling caused by internal inclusions in the metal strip. Scratches and scabs are often observed as defects that extend continuously or intermittently in the longitudinal direction of a metal strip. Examples of uneven defects are shown in Figure 2, where (a) of Figure 2 is a photograph substituting a drawing showing an example of a dent, and (b) of Figure 2 is a photograph substituting a drawing showing an example of a scale defect. Both are images taken by the camera of a surface defect meter.

[0019] A pattern defect is a defect that can be visually recognized as an uneven color tone on the surface of a metal strip. However, even a pattern defect may be formed by fine irregularities when observed microscopically. In an embodiment of the present invention, a pattern defect refers to an irregularity on the surface of a metal strip having a height of less than about 5 μm. Specifically, a pattern defect is a defect that is recognized when, for example, a plating process is performed on the surface of a metal strip, because the deposition state of metal particles in the plating film is different from other parts. The fine irregularities caused by such precipitated particles scatter light of a specific wavelength or less, for example, 600 nm or less, and are observed as a defect colored brown, i.e., a brown pattern, depending on the viewing angle.

[0020] There are many causes of occurrence of pattern defects, and pattern defects can be classified according to the cause. For example, a rust pattern is a pattern formed on the surface of a metal strip by oxidation of a part of the surface of the metal strip during rolling or threading, which is a rust-colored pattern (a part whose color tone is different from other areas) although the uneven shape cannot be distinguished from other parts of the metal strip. A temper color is a change in color, i.e., a pattern, depending on the thickness of the oxide film on the surface of the metal strip. This is caused by the fact that the thickness of the oxide film on the surface changes partially during annealing of the metal strip, and the interference color generated between the surface of the oxide film and the bare surface of the metal strip differs depending on the location on the surface of the metal strip. In addition to the above, other examples of pattern defects include scale patterns and dirt. In the embodiment of the present invention, the classification of surface defects such as the uneven defects and pattern defects, i.e., the classification of types of surface defects, is based on a standard set in advance for each manufacturing process of the metal strip.

[0021] [Defect detection section] The defect discrimination unit 4 is configured, for example, mainly by a microcomputer, and is configured to perform calculations based on input data, pre-stored data, calculation expressions, and the like, and to output the calculation results to a database described later. FIG. 3 is a block diagram showing an example of data input / output to the defect discrimination unit 4. As shown in FIG. 3, when image data D1 of the steel plate 2 is acquired by the camera 3, the image data D1 is input to the defect discrimination unit 4. In the defect discrimination unit 4, surface defects of the steel plate 2 are discriminated based on the acquired image data D1, and defect discrimination data D2 is calculated. The defect discrimination data D2 is data including the type (type) and grade (severity) of the surface defect, and position information of the surface defect. The defect discrimination data D2 may also include information regarding the size of the surface defect, such as the length and width of the detected surface defect, and information such as the luminance value and color tone of the image in the area discriminated to have a surface defect in the acquired image data D1. The type and grade of the surface defect performed by the defect discrimination unit 4 can be discriminated by a conventionally known method described in the above-mentioned Patent Document 1, Patent Document 2, and the like.

[0022] Regarding the classification of surface defects (defect types), any classification to be determined as the quality standard of the steel sheet 2 may be set. For example, the defects may be broadly classified into uneven defects and pattern defects. As described above, the uneven defects may be classified into more detailed classifications, and classified using commonly used classifications such as dents, dull baldness, compression defects, scratches, scabs, etc. Similarly, the pattern defects may be classified using commonly used classifications such as rust patterns, scale patterns, brown patterns, temper colors, stains, etc.

[0023] In addition, with regard to the classification of surface defect grades (defect grades), the classification to be determined as the quality standard of the steel sheet 2 may be set arbitrarily. For example, the classification of surface defects may be set according to the size, depth, or height of each surface defect. Also, the classification may be based on the number of surface defects per unit area or the area ratio. Furthermore, the classification of surface defects may be determined by performing image processing on the image data D1 acquired by the camera 3. Classification of surface defects based on image data D1 will be specifically described. For example, from among the multiple acquired image data D1, image data D1 as a candidate for detecting a surface defect is identified or selected by the defect discrimination unit 4. Then, the image data D1 identified as a candidate for detecting a surface defect is referred to, and a region in which unevenness in brightness distribution occurs is identified in the image data D1. Also, the size of the surface defect is identified from the spread of the identified region (the range of pixels showing similar brightness). In this way, the class of the surface defect is classified. Also, in the identified image data D1, the size of the surface defect, i.e., the class of the surface defect, may be classified according to the difference in brightness tone, i.e., the difference from the background.

[0024] The defect discrimination data D2 in the embodiment of the present invention may include position information within the steel plate surface of the area determined to have a surface defect. The position information within the steel plate surface is the coordinates of the surface defect or the area determined to have a surface defect. Specifically, the position information may be specified by the distance from the tip of the metal strip to the surface defect or the area in the conveying direction of the metal strip in the manufacturing process, and the distance from one of both ends of the metal strip in the sheet width direction to the surface defect or the area. In addition, the above-mentioned position information preferably includes information specifying whether the surface defect or the area is on the front surface or the back surface of the steel plate 2.

[0025] In the embodiment of the present invention, the surface defect meter 1 constantly acquires image data D1 of the transported steel plate 2 by the camera 3. The surface defect meter 1 accumulates defect discrimination data (type, grade, and position information of the surface defect) D2 about the surface defects of the steel plate 2 that are discriminated by the defect discrimination unit 4 based on the image data D1 in the judgment record database 6. The surface defect meter 1 also accumulates, in the judgment record database 6, image data D1 that includes an area that is discriminated or recognized by the defect discrimination unit 4 as having a surface defect, as defect image data D3.

[0026] The judgment record database 6 can be provided in a control computer for controlling the operation of the defect detection process by the surface defect meter 1 and the manufacturing process of the steel plate 2. The judgment record database 6 may be provided in a higher-level computer than the control computer that outputs operation instructions (control command signals) to the control computer, or may be provided in an independent computer configured separately from the surface defect meter 1. Alternatively, the judgment record database 6 may be provided in a data center or on the cloud. The surface defect meter 1 stores the defect discrimination data D2 determined to be a surface defect in the judgment record database 6 in association with the defect image data D3.

[0027] [Customer Determination Data] In this embodiment, in addition to the defect discrimination data D2 and defect image data D3 acquired during the manufacturing process of the steel plate 2, the surface defect meter 1 is configured to accumulate customer identification data D4 and customer judgment data D5 in the judgment performance database 6, as shown in Figure 4.

[0028] The customer identification data D4 is data for identifying a customer to which the steel plate product is shipped, and is, for example, data such as a code or symbol set and stored for each customer in the above-mentioned computer to identify the customer. The customer identification data D4 may be set to identify not only a company or organization that purchases the steel plate product, but also different factories within the same company. For example, different factories within the same company manufacture different processed parts (products) and require different quality evaluation standards for steel plate products. Note that, usually, when the steel plate 2 is manufactured, the customer to which the steel plate product is shipped is determined, and the customer identification data D4 for the customer to which the steel plate product is shipped is saved or stored in the upper computer. Therefore, when the defect discrimination data D2 and defect image data D3 of the surface defect are accumulated in the evaluation record database 6, it is preferable to associate the defect discrimination data D2, the defect image data D3, and the customer identification data D4.

[0029] The consumer judgment data D5 may include, in addition to the consumer's pass / fail judgment result on the surface defects, information on complaints from consumers about surface defects of steel plate products (hereinafter referred to as complaint information), information on requests and demands regarding surface defects obtained from consumers even if they do not amount to complaints, and position information on the steel plate surface of the surface defects that are the subject of the complaint. The consumer judgment data D5 is saved or stored in a higher-level computer, similar to the consumer identification data D4. The consumer judgment data D5 is data that is associated with position information on the steel plate surface for surface defects judged by consumers to be "fail" regarding surface defects of steel plate products. In other words, it is data on surface defects fed back from consumers. Note that the position information on the steel plate surface of the surface defects does not necessarily need to be expressed by coordinates on the steel plate surface, and may be information that identifies within a certain range the area including the surface defects judged to be fail. In this case, to specify the area containing the surface defect, it is sufficient that the position information of the surface defect that the customer judged to be unacceptable is specified with an accuracy of about ±10 m in the longitudinal direction of the steel plate, i.e., the conveying direction, and preferably within a range of ±5 m. The customer judgment data D5 does not need to include information on the type or grade of the surface defect that the customer judged to be unacceptable.

[0030] In addition, defect discrimination data D2 acquired during the manufacturing process of the steel plate 2 is provided to the consumer, accompanying the steel plate product. Information on "pass" or "fail" regarding surface defects newly judged by the consumer based on the provided defect discrimination data D2 may be used as consumer judgment data D5 in this embodiment. This is because the defect discrimination data D2 accompanying the steel plate product includes position information of the surface defects acquired during the manufacturing process, making it easy to individually identify surface defects judged as failing by the consumer. This allows a highly accurate judgment record database 6 to be constructed.

[0031] The above-mentioned customer judgment data D5 is stored in the judgment results database 6 in correspondence with the defect discrimination data D2 and defect image data D3 obtained during the manufacturing process of the steel plate 2 using the product number (coil number) of the steel plate product and customer identification data D4, etc.

[0032] [Judgment performance database] In this manner, the defect discrimination data (defect type, defect grade, defect position information) D2 acquired in the steel plate manufacturing process, the defect image data D3, the consumer identification data D4, and the consumer judgment data D5 are stored in the judgment record database 6 in a mutually associated manner. At this time, among the consumer judgment data D5, the surface defects of the steel plate product that are judged as "rejected" are accompanied by position information within the steel plate surface. Therefore, the defect discrimination data D2 stored in the judgment record database 6 may be associated based on such position information. In addition, when the position information of the surface defects in the steel plate product is specified within a certain range as the consumer judgment data D5, the defect discrimination data D2 of the steel plate included in that range is extracted. Then, the surface defect judged to be the lowest grade among them may be regarded as the defect judged by the consumer as "rejected". In addition, a case where image data of the surface defects judged as "rejected" by the consumer can be acquired as the consumer judgment data D5 will be described. In this case, the similarity between the surface defect of the steel plate product in the image data and the surface defect of the steel plate in the defect image data D3 stored in the judgment record database 6 including the area specified by the image data is evaluated. As the above-mentioned consumer judgment data D5, the image data of the surface defect judged as unacceptable by the consumer means image data acquired by the consumer by photographing the surface of the steel plate product with a camera. Then, among the surface defects of the steel plate in the defect image data D3 used in the evaluation, the surface defect having the highest similarity to the surface defect of the steel plate product in the image data acquired from the consumer may be judged to be approximately equal to the surface defect judged as unacceptable by the consumer. The evaluation of the similarity may be a conventionally known method of extracting feature points and feature amounts of the image data and performing feature point matching. Alternatively, since a tool for similar image search trained by deep learning is generally available, such a known evaluation method may be applied.

[0033] Furthermore, the consumer judgment data D5 acquired from the consumer may not necessarily include judgment data classified as "pass". Therefore, for the defect judgment data D2 that has not been judged as "fail" by the consumer, the consumer judgment data D5 may be assigned information of "pass" as if the consumer had judged the data as "pass". In this way, a data set is formed for each consumer (consumer identification data D4) in which the defect judgment data D2, defect image data D3, and consumer judgment data D5 are linked to each other, and the data set is accumulated in the judgment record database 6.

[0034] FIG. 5 is a diagram showing an example of a data set stored in the judgment record database 6. Specifically, the data set shown in FIG. 5 is a data set constructed for each consumer CM regarding surface defects acquired in the inspection process of the steel plate 2. In the data set shown in FIG. 5, for each consumer CM, the product number of the steel plate 2 inspected for surface defects, the defect type of the surface defect found in the steel plate 2, the defect grade of the surface defect, the defect image of the surface defect, the defect position information of the surface defect, and the consumer judgment data D5 regarding the surface defect are associated with each other. In FIG. 5, the symbols "△" and "■" indicate defect images included in the defect image data D3. In addition, the defect position information is, as described above, the distance from the tip of the steel plate to the surface defect, the position in the plate width direction, and information regarding the surface on which the surface defect is formed, and is capable of identifying these. Specifically, in the defect position information in FIG. 5, for example, in the description "a1, b1, 1," "a1, b1" indicates the coordinates of the surface defect, and the description "1" indicates that the defect is formed on the surface of the steel plate 2. The description "2" indicates that the defect is formed on the back surface of the steel plate 2. The defect types "A" and "B" are the types of defects determined by the defect discrimination unit 4, and represent identification codes assigned in advance to, for example, uneven defects, pattern-like defects, and the like. The defect grades "1," "2," and "3" are the severity of the defects determined by the defect discrimination unit 4, and represent, for example, identification codes corresponding to categories preset according to the area of ​​the defects.

[0035] Moreover, the description of "pass" in the consumer judgment data D5 in FIG. 5 indicates the defect judgment data D2 that was not judged as "fail" by the consumer. Therefore, the description of "fail" indicates the defect judgment data D2 that was judged as "fail" by the consumer. Note that, for example, a metal strip wound in a roll may be rewound between the time when the defect judgment data D2 and the defect image data D3 of the steel plate are acquired and the time when the consumer judges the surface defect of the steel plate product. In this case, the leading end and the trailing end of the metal strip when unwound from the roll are reversed from the time of shipment. Therefore, in such a case, the defect position information may be converted to the position information attached to the consumer judgment data D5 by referring to the rewinding process described above.

[0036] In the judgment record database 6, it is preferable that 50 or more data sets for product standards that can be classified into the same category and for each customer are stored as the product standard information of the steel sheet 2. The number of the above data sets stored in the judgment record database 6 is more preferably 100 or more, and even more preferably 500 or more.

[0037] It is also preferable to perform screening on the data stored in the judgment record database 6 as necessary. This is to avoid the possibility that the accuracy of the predicted data may decrease when the defect image data D3 is unclear, for example. In other words, this is to improve the surface defect judgment accuracy of the surface defect judgment model by removing unclear defect image data D3 and accumulating highly reliable data. On the other hand, the number of data sets accumulated in the judgment record database 6 may be set to a preset upper limit, and the data sets accumulated in the judgment record database 6 may be updated as appropriate within this upper limit.

[0038] [Surface defect judgment model] A surface defect judgment model generating unit 7 that generates a surface defect judgment model according to an embodiment of the present invention will be described with reference to FIG. 6. The surface defect judgment model generating unit 7 shown in FIG. 6 includes a machine learning unit 8. The machine learning unit 8 performs machine learning on the correlation between the defect discrimination data D2, the defect image data D3, the consumer identification data D4, and the consumer judgment data D5 based on the data set accumulated in the judgment record database 6. The machine learning unit 8 is configured to generate a surface defect judgment model that is a machine learning model based on the correlation between the data D2, D3, D4, and D5 as a result of the machine learning. Note that the machine learning model is, for example, a model that performs data analysis on a plurality of input data to find a relationship between the plurality of input data. Therefore, when any of the plurality of input data is input to the machine learning model, data corresponding to the input data can be output as output data. In this embodiment, the defect discrimination data D2n, the defect image data D3n, the consumer identification data D4n, and the like for the steel plate 2n that has not yet been shipped to the consumer at present are input as input data to the surface defect judgment model. As a result, a consumer judgment prediction, which is a consumer's pass / fail judgment result for the surface defects of the steel plate 2n predicted from the input data, is output as output data. The surface defect judgment model generation unit 7 shown in Fig. 6 generates a surface defect judgment model in which the above input data and output data are linked to each other by performing machine learning based on the data set for each consumer.

[0039] The surface defect judgment model generating unit 7 can be provided in a control computer that controls a manufacturing process in which surface defects are detected by the surface defect meter 1. It may also be provided in a higher-level computer that outputs operation instructions (control command signals) to the control computer, or in an independently installed computer that can communicate with other devices. The surface defect judgment model generating unit 7 may be configured in a device separate from the judgment record database 6, using a device that can receive the data set accumulated in the judgment record database 6. In other words, the surface defect judgment model generating unit 7 may be provided in a data center or on the cloud.

[0040] Here, a machine learning model for generating a surface defect judgment model executed in the machine learning unit 8 will be described. The machine learning model is not particularly limited as long as it can obtain a prediction accuracy of consumer judgment prediction data sufficient for practical use. For example, commonly used neural networks (including deep learning and convolutional neural networks, etc.), decision tree learning, random forest, support vector regression, etc. may be used as the machine learning model. Alternatively, an ensemble model combining multiple models may be used as the machine learning model. Furthermore, a classification model such as k-nearest neighbor method or logistic regression may be used as the machine learning model.

[0041] FIG. 7 is a diagram showing an example of the structure of a surface defect determination model. For example, the defect image data D3 in this embodiment is two-dimensional image data. Therefore, as the surface defect determination model in this embodiment, it is preferable to use a machine learning model of a network structure including a convolutional neural network having a convolutional layer FL, a pooling layer PL, and a coupling layer CL as the first layer L1 to which the defect image data D3 is input. This makes it possible to compress the defect image data D3 into one-dimensional information while maintaining the feature amount of the defect image data D3. Then, the defect image data D3 compressed into one-dimensional information by the coupling layer CL is input to the second input layer L2 shown in FIG. 7. The defect discrimination data (defect type, defect grade, defect position information, etc.) D2 and the consumer identification data D4 are input to the second input layer L2 together with the defect image data D3 compressed into one-dimensional information. Then, the second input layer L2 is connected to the intermediate layer ML and the output layer OL in the same manner as in a conventionally known neural network.

[0042] When the defect image data D3 is a color image, the defect image data D3 is converted into image data for each RGB channel (defect image data in which the luminance value of the defect image is converted into numerical information in the range of 0 to 255). Then, the defect image data D3 of three channels may be input to the first input layer L1. When the defect image data D3 is grayscale, the defect image data D3 of one channel in which the luminance value of the defect image data D3 is converted into numerical information in the range of 0 to 255 may be input to the first input layer L1. The defect image data D3 may be subjected to a data compression process to compress the number of pixels in the horizontal and vertical directions of the defect image data D3 before being input to the first input layer L1. In the convolution layer FL arranged downstream of the first input layer L1, a filtering process using a filter called a kernel is performed on the input data to generate a first feature map. Here, convolution refers to a calculation process in which a filter is applied to the input data, that is, a filtering process is performed, to generate output data called a feature map. The filter used in the convolution layer FL may be, for example, a filter of 3 pixels vertically by 3 pixels horizontally, and it is preferable to set the stride for moving the position of the filter within the image to "1" and to apply padding for filling the periphery of the image data with "0". In addition, it is preferable to use a nonlinear function as the activation function in the convolution layer FL, and it is preferable to use the Relu function so that the gradient vanishing problem during machine learning can be suppressed.

[0043] The pooling layer PL compresses the information of the first feature map by using the first feature map output by the convolution layer FL as input. Maximum pooling or average pooling can be applied to the pooling layer PL. Here, maximum pooling is a process of dividing the first feature map, which is the input of the pooling layer PL, into a certain area (pool size), extracting the maximum value in the area, and outputting it as a new feature map. Average pooling is a process of extracting an average value instead of a maximum value. Such a pooling layer PL reduces the amount of information while maintaining the features of the input defect image data D3, and generates a dimensionally compressed second feature map. The size of the filter used in the pooling layer PL can be, for example, 3 pixels vertically by 3 pixels horizontally. Since the pooling layer PL aims to reduce the size of the output data while retaining the features of the image, padding to fill the periphery of the output data with "0" is not particularly performed.

[0044] The combining layer CL converts the second feature map generated by the pooling layer PL, and is arranged to arrange the values ​​of the second feature map in a row and consolidate the output from the pooling layer PL. An example of a preferred form of the combining layer CL is one in which the number of nodes is 32 or more and 2048 or less. Note that in the configuration of the convolutional neural network shown in FIG. 7, a configuration may be made in which a plurality of convolutional layers FL and pooling layers PL are arranged to further compress the information of the defect image data D3 input from the first input layer L1.

[0045] The defect image data D3 compressed into one-dimensional information by the coupling layer CL in this way is input to the second input layer L2 together with the defect discrimination data (defect type, defect grade, defect position information) D2 and the customer identification data D4. The middle layer ML arranged between the second input layer L2 and the output layer OL is composed of a plurality of neurons constituting a conventionally known normal neural network. The middle layer ML is composed of a plurality of hidden layers, and a plurality of neurons are arranged in each hidden layer. The number of hidden layers arranged in the middle layer ML is not particularly limited, but it has been empirically found that if the number of hidden layers is too large, the prediction accuracy decreases. Therefore, it is preferable that the number of layers is 5 or less. In addition, it is preferable that the number of neurons arranged in each hidden layer is in the range of 1 to 10 times the number of data input to the second input layer L2. In the middle layer ML, the transmission of a neuron from a certain neuron to the following hidden layer is performed via an activation function together with weighting of variables by weighting coefficients. A sigmoid function, a hyperbolic tangent function, or a ramp function can be used as the activation function.

[0046] The output layer OL combines the neuron information transmitted by the intermediate layer ML, and outputs the combined information as the final consumer judgment prediction data D6. That is, for the steel plate 2n that has not yet been shipped to the consumer, prediction data of "pass" or "fail" by the consumer is output. In addition, the probability that the consumer will be judged as "fail" by the consumer may be output as the consumer judgment prediction data D6 using a softmax function. The above-mentioned consumer judgment prediction data D6 corresponds to the consumer judgment prediction in the embodiment of the present invention.

[0047] Furthermore, the machine learning unit 8 may improve the estimation accuracy of the consumer judgment prediction data D6 by dividing the data set accumulated in the judgment record database 6 into training data and test data and performing learning. The training data is data for generating a surface defect judgment model, also called teacher data. The test data is data used for evaluating and verifying the generated surface defect judgment model. For example, the weighting coefficients of the neural network are learned using the training data. Then, the surface defect judgment model may be obtained while appropriately changing the structure of the neural network (the number of convolution layers and pooling layers, the filter size, etc.) so that the accuracy rate of the consumer judgment prediction data D6 in the test data is increased, that is, the prediction accuracy. The weighting coefficients may be updated using an error propagation method. Note that the surface defect judgment model may be updated to a new model by re-learning, for example, every six months or every year. This is because the more data stored in the judgment record database 6, the more accurate the consumer judgment prediction data D6 can be estimated. By updating the surface defect judgment model based on the latest data, a surface defect judgment model that reflects the change in the quality judgment standard by the consumer can be generated for each consumer.

[0048] [Method of determining pass / fail of steel plate] In another embodiment of the present invention, the surface defect judgment model generated as described above is used to predict the pass / fail judgment of a steel plate by a customer in a surface defect judgment unit. FIG. 8 is a block diagram for explaining an example of a method for judging the pass / fail of a steel plate 2 according to another embodiment of the present invention. The surface defect judgment unit 9 of the steel plate shown in FIG. 8 may be a part of a host computer that gives an operation instruction (control command signal) to a manufacturing process equipped with a surface defect meter 1, or may be configured by a device capable of communicating with the host computer. It may also be installed in a tablet terminal equipped with a communication function with the host computer. Hereinafter, the method for judging the pass / fail of a steel plate according to another embodiment of the present invention will be described with reference to FIG. 8.

[0049] The operation of the surface defect judgment unit 9 shown in FIG. 8 is executed after the surface defect of the steel plate to be judged is inspected by the surface defect meter 1 and the defect discrimination data D2 discriminating the surface defect of the steel plate 2 and the defect image data D3 are acquired. The steel plate to be judged is a steel plate that has not been shipped to a consumer. That is, it is a steel plate that has not become a steel plate product. Therefore, the consumer judgment data D5 of the steel plate to be judged is not stored in the judgment record database 6 at the time when the surface defect judgment unit 9 operates. In the following description, the time when the surface defect judgment unit 9 operates is referred to as the present time. In addition, the image data D1 of the steel plate 2 captured by the camera 3 at the present time is referred to as image data D1n, the defect discrimination data D2 acquired at the present time is referred to as defect discrimination data D2n, and the defect image data D3 is referred to as defect image data D3n.

[0050] In the example shown in FIG. 8, first, the defect discrimination data D2n and the defect image data D3n at the present time are input from the defect discrimination unit 4 to the surface defect judgment unit 9. The surface defect judgment unit 9 also refers to a product number or the like for identifying the steel plate 2 to be inspected, and acquires the consumer identification data D4 stored in the above-mentioned computer corresponding to the product number. In the surface defect judgment unit 9, the defect discrimination data D2n, the defect image data D3n, and the consumer identification data D4 are input as input data to the surface defect judgment model generated by the above-mentioned method. Then, the consumer judgment prediction data D6, which is prediction data of the pass / fail judgment result as a steel plate product, is output as output data from the surface defect judgment unit 9 to the pass / fail judgment output unit 10. Note that the "probability" of being "failed" by the consumer may be output as output data of the surface defect judgment model. In that case, a threshold value may be set in advance, and the "pass" or "fail" may be determined based on the threshold value. The pass / fail judgment output unit 10 may be a monitor, for example.

[0051] The pass / fail judgment information of the steel plate 2 outputted as described above, that is, the customer judgment prediction data D6, is displayed on a pass / fail judgment output unit 10 such as a monitor. In this way, based on the customer judgment prediction data D6 displayed on the pass / fail judgment output unit 10, the operator can confirm the position of the surface defect in the steel plate 2 that may be rejected by the customer, and set an additional treatment process for the steel plate 2. Here, the treatment process includes a cutting process for removing from the steel plate 2 an area including a surface defect that may be judged as a failure. The treatment process determined by the operator is appropriately processed for the target steel plate 2 by inputting the additional process to the upper computer. On the other hand, steel plate products are often shipped as steel strip coils, and when pass / fail judgment is performed along the longitudinal direction of the steel plate, a steel plate judged as pass along the entire length of the steel plate has a high quality level for the customer of the steel plate product. Therefore, it is preferable to manufacture a steel plate judged as pass along the entire length of the steel plate using the pass / fail judgment means described above, and ship it as a steel plate product. This is because the steel sheet does not contain any surface defects that would cause customers to deem the product unacceptable.

[0052] Therefore, according to the present embodiment, it is possible to judge surface defects according to the needs of each customer of the steel sheet product, more than ever before. In addition, by using the customer judgment prediction data D6 to judge the pass / fail of the steel sheet product, it is possible to meet the individual needs of each customer at a high quality level, and to provide high quality steel sheets to each customer.

[0053] [Surface defect judgment model generation device] FIG. 9 is a diagram showing an example of a surface defect judgment model generating device for a metal strip according to the present invention. The surface defect judgment model generating device 11 shown in FIG. 9 includes a quality information acquiring unit 12 that acquires quality information on shipped steel plate products directly or indirectly from each of a plurality of consumers CM1, CM2, and CM3, for example. The quality information acquiring unit 12 is connected to each of a plurality of consumers CM1, CM2, and CM3 via an electric communication line 13, and acquires the quality information from each of the consumers CM1, CM2, and CM3 via the electric communication line 13. The quality information acquiring unit 12 is also connected to a judgment record database 6 via the electric communication line 13, and outputs the quality information acquired via the electric communication line 13 to the judgment record database 6. In FIG. 9, three consumers CM1, CM2, and CM3 are described as the shipping destinations of the steel plate products.

[0054] Specifically, the electric communication line 13 may be, for example, an Internet line. The quality information acquisition unit 12 may be a server connected to each of the consumers CM1, CM2, and CM3 and the judgment record database 6 via the Internet line. The surface defect judgment model generation device 11 is configured to perform machine learning based on the quality information acquired by the quality information acquisition unit 12 and various data D2, D3, etc. stored in the judgment record database 6 to generate a surface defect judgment model. Since the other configurations are similar to those shown in Figures 3, 4, and 8, the same reference numerals as those in Figures 3, 4, and 8 are used for the configurations similar to those shown in Figures 3, 4, and 8, and the description thereof will be omitted.

[0055] The above-mentioned quality information will be explained. The quality information may be consumer identification data D4 about a consumer to which the steel plate product is shipped, and consumer judgment data D5 which is a result of the consumer's judgment of the pass / fail of the surface defects of the steel plate product. The consumer judgment data D5 may include, in addition to the consumer's judgment of the pass / fail of the surface defects of the steel plate product, complaint information about the surface defects of the steel plate product and information about the requests and demands regarding the surface defects obtained from the consumer. The consumer judgment data D5 may also include position information of the surface defects judged as "fail" by the consumer, and information identifying the steel plate product that was the subject of the judgment of the surface defects, i.e., information such as the product number of the steel plate product.

[0056] [Metal strip pass / fail judgment device] Fig. 10 is a diagram showing an example of a metal strip pass / fail judgment device 14. The pass / fail judgment device 14 shown in Fig. 10 includes a model acquisition unit 15, an input data acquisition unit 16, a storage unit 17, a surface defect judgment unit 18, and a pass / fail judgment output unit 10.

[0057] The model acquisition unit 15 acquires the surface defect judgment model generated by the machine learning unit 8. For example, the model acquisition unit 15 is connected to the machine learning unit 8 via an electric communication line, and is configured so that the surface defect judgment model is input from the machine learning unit 8 to the model acquisition unit 15. Alternatively, the model acquisition unit 15 is configured so as to acquire the surface defect judgment model from the machine learning unit 8. The model acquisition unit 15 may be a communication interface configured so as to be able to communicate with the machine learning unit 8 via an electric communication line. The communication interface may be a conventionally known one.

[0058] The input data acquisition unit 16 is configured to acquire defect discrimination data D2n, defect image data D3n, and consumer identification data D4 for the steel plate 2n inspected by the surface defect meter 1 and prior to shipping as a product. The input data acquisition unit 16 may be a communication interface configured to be capable of communicating with the defect discrimination unit 4 and the quality information acquisition unit 12 via an electric communication line. The input data acquisition unit 16 may be a conventionally known communication interface, similar to the model acquisition unit 15. Note that the time prior to shipping as a product refers to the time when the surface defect determination unit 18 described below operates, that is, the time when the surface defects of the steel plate 2n are determined.

[0059] The storage unit 17 at least temporarily stores information used in the operation of the pass / fail judgment device 14. Examples of the information used in the operation of the pass / fail judgment device 14 include a surface defect judgment model, defect discrimination data D2n, defect image data D3n, and consumer identification data D4. That is, the storage unit 17 temporarily stores information used in the pass / fail judgment of the surface defects of the steel sheet 2 by the pass / fail judgment device 14. The storage unit 17 may be composed of a conventionally known semiconductor memory, magnetic memory, or optical memory. Specifically, the storage unit 17 may be composed of at least one semiconductor memory, at least one magnetic memory, or at least one optical memory. Alternatively, the storage unit 17 may be composed of a combination of at least two types of memories among the various types of memories described above.

[0060] The surface defect judgment unit 18 predicts the pass / fail judgment of the steel plate by the customer in a manner similar to the surface defect judgment unit 9 shown in FIG. 8. The surface defect judgment unit 18 is mainly composed of at least one processor, and is communicably connected to each unit 15, 16, 17, and 10 of the pass / fail judgment device 14. In other words, the surface defect judgment unit 18 is configured to perform calculations based on data input to the pass / fail judgment device 14 and data stored in advance, and to output the calculation results as output data to the pass / fail judgment output unit 10. Specifically, the surface defect judgment unit 18 inputs the defect discrimination data D2n, the defect image data D3n, and the customer identification data D4 as input data to the surface defect judgment model stored in the storage unit 17. Then, the surface defect judgment unit 18 outputs the customer judgment prediction data D6 linked to the input data as output data. The surface defect determination unit 18 may be, for example, an electronic device such as a PC (Personal Computer) or a smartphone, or may be at least one server device capable of communicating with the units 15, 16, 17, and 10 of the pass / fail determination device 14.

[0061] As described above, the pass / fail judgment output unit 10 may be a monitor. Since it is sufficient for the pass / fail judgment output unit 10 to notify the operator of the consumer judgment prediction data D6, the pass / fail judgment output unit 10 may be provided integrally with the pass / fail judgment device 14, or may be configured as a separate item from the pass / fail judgment device 14. Instead of a monitor, the pass / fail judgment output unit 10 may be a speaker that outputs the consumer judgment prediction data D6 by voice.

[0062] [Functions of the surface defect judgment model generating device and the pass / fail judgment device] When the steel sheet products are shipped to the consumers CM1, CM2, and CM3, surface defects of the steel sheet products are judged at each of the consumers CM1, CM2, and CM3. In addition, quality information including pass / fail judgment results, complaints, requests, etc. by each of the consumers CM1, CM2, and CM3 is transmitted to the quality information acquisition unit 12 from each of the consumers CM1, CM2, and CM3 via the electric communication line 13. Alternatively, the quality information is stored by each of the consumers CM1, CM2, and CM3 at a predetermined location connected to the quality information acquisition unit 12 via the electric communication line 13, and the quality information is read by the quality information acquisition unit 12. Next, the quality information acquisition unit 12 associates the consumer identification data D4 of each of the consumers CM1, CM2, and CM3 that created the quality information with the consumer judgment data D5 of each of the consumers CM1, CM2, and CM3. Then, the quality information acquisition unit 12 outputs the consumer identification data D4 and the consumer judgment data D5 that are associated with each other to the judgment record database 6 via the electric communication line 13.

[0063] At the time when the quality information acquisition unit 12 acquires the quality information from each of the consumers CM1, CM2, CM3, each of the data D2, D3, D4 of the steel plate products that were the subject of the surface defect judgment at each of the consumers CM1, CM2, CM3 has already been stored in the judgment record database 6. In the surface defect judgment model generation unit 7 shown in FIG. 9, a data set is created in which each of the data D4, D5 fed back from each of the consumers CM1, CM2, CM3 and each of the data D2, D3 stored in the judgment record database 6 are linked to each other. This data set is accumulated in the judgment record database 6 as a learning data set for machine learning in the machine learning unit 8. Then, as described above, the machine learning unit 8 performs machine learning based on the data set accumulated in the judgment record database 6, and generates a surface defect judgment model as a result of the machine learning.

[0064] The above-mentioned surface defect judgment model is acquired by a model acquisition unit 15 of the pass / fail judgment device 14 and temporarily stored in a memory unit 17. In addition, defect discrimination data D2n, defect image data D3n, and consumer identification data D4 for a steel plate 2n inspected by the surface defect meter 1 and prior to shipping as a product are acquired by an input data acquisition unit 16 and temporarily stored in the memory unit 17.

[0065] The surface defect determination unit 18 inputs the defect discrimination data D2n, the defect image data D3n, and the consumer identification data D4 as input data to the surface defect determination model stored in the storage unit 17. Then, the consumer determination prediction data D6 linked to the input data is output as output data. The consumer determination prediction data D6 is displayed on the pass / fail determination output unit 10. That is, the pass / fail determination output unit 10 displays the prediction data on "pass" or "fail" judged by the consumer at the shipping destination, and the operator takes measures to prohibit the shipment of the surface defect presented as "fail" as a steel plate product. In addition, the operator instructs an additional process to remove, by cutting or the like, a portion including a surface defect that is predicted to be judged as "fail" by the consumer. This makes it possible to prevent the shipment of steel plate products that may be judged as "fail" or that are highly likely to be judged as "fail" to each consumer CM1, CM2, CM3. That is, even with an apparatus having such a configuration, it is possible to judge surface defects according to the requirements of each of the customers CM1, CM2, and CM3 of steel plate products, more than ever before. Also, it is possible to meet the individual requirements of each customer CM1, CM2, and CM3 at a high quality level, and to provide high quality steel plates to each customer CM1, CM2, and CM3.

[0066] [Example] In this embodiment, a metal strip having a thickness of 1.8 mm to 4.2 mm and a width of 1000 mm to 1500 mm was prepared as an inspection target, and surface defects of the metal strip were inspected by a surface defect meter 1 installed on an inspection line. A judgment record database 6 was formed based on the surface defect discrimination results by the surface defect meter 1, and then a surface defect discrimination model was generated in a surface defect discrimination model generation unit 7. The judgment record database 6 was stored with defect discrimination data D2 acquired by the surface defect meter 1, defect image data D3, customer identification data D4 of the steel plate product, and customer judgment data D5 by customers. The defect image data D3 was converted from a color image into grayscale, and then stored in the judgment record database 6 as an image with 1200 vertical pixels and 2000 horizontal pixels.

[0067] On the other hand, the consumer judgment data D5 was set based on the complaint information received from the consumer within six months after the metal strip was shipped. That is, the complaint information from the consumer includes information on the surface defects of the metal strip and includes the position information of the surface defects that were judged as unacceptable by the consumer. The surface defect judgment model was generated as described above when 30 complaints from the consumer were collected. Note that although the consumer judgment data D5 was not judged as "unacceptable", the surface defect meter 1 may judge it as a surface defect. In that case, when the surface defect judgment model was generated, data representing "pass" was associated with the consumer judgment data D5 of the data set accumulated in the judgment record database 6 corresponding to the above-mentioned situation. Then, by performing machine learning based on the data set of the judgment record database 6, a surface defect judgment model was generated in which the defect discrimination data D2, the defect image data D3, and the consumer identification data D4 were used as input data and the consumer judgment prediction data D6 was used as output data. The number of data sets of learning data used when generating the surface defect judgment model was 1000.

[0068] The structure of the neural network used in the machine learning unit 8 includes a convolutional neural network for defect image data and a normal neural network from the second input layer L2 to the output layer OL, as shown in Fig. 7. Although not shown in detail, the convolutional neural network has a structure in which a first convolutional layer, a first pooling layer (maximum pooling), a second convolutional layer, and a second pooling layer (maximum pooling) are arranged in series. From the second input layer L2 to the output layer OL, there are three intermediate layers ML, and a sigmoid function is used as the activation function.

[0069] The surface defect judgment model thus generated was installed on a tablet terminal capable of communicating with a host computer, and surface defects of the metal strip were judged. The consumer identification data D4, which is input data for the surface defect judgment model, was acquired from the host computer by referring to the serial number of the metal strip to be judged. For the metal strip to be inspected, the current defect discrimination data D2n and defect image data D3n of the metal strip including surface defects were acquired from the surface defect meter 1 shown in FIG. 1(a) and transmitted to the tablet terminal equipped with the surface defect judgment model. In the tablet terminal, the acquired defect discrimination data D2n, defect image data D3n, and consumer identification data D4 were input to the surface defect judgment model. In this way, consumer judgment prediction data D6 was obtained as output data of the surface defect judgment model based on the input data and the surface defect judgment model. The consumer judgment prediction data D6 was then displayed on the screen of the tablet terminal. In this embodiment, when the pass / fail judgment of the metal strip was performed based on the consumer judgment prediction data D6 predicted in this way, the number of complaints from consumers was reduced by 20% compared to the conventional method. [Explanation of symbols]

[0070] 1. Surface Defect Meter 2 steel plate 3. Camera 4. Defect detection section 5 Light source section 6. Judgment results database 7. Surface defect judgment model generation unit 8. Machine Learning Department 9. Surface defect judgment section 10 Pass / fail judgment output section 11. Surface defect judgment model generation device 12 Quality information acquisition department 13 Telecommunications Lines 14 Pass / fail judgment device 15 Model Acquisition Department 16 Input data acquisition section 17 Memory section 18 Surface Defect Judgment Section

Claims

1. forming an evaluation record database that accumulates a data set including defect evaluation data that identifies surface defects of a metal strip, defect image data that images the surface of the metal strip including the surface defects, customer identification data that identifies a customer to which the metal strip is to be shipped, and customer evaluation data that is a pass / fail evaluation result of the surface defects of the metal strip by the customer; By performing machine learning on the correlation between the defect discrimination data, the defect image data, the customer identification data, and the customer judgment data based on the data set accumulated in the judgment record database, a surface defect judgment model is generated in which data including the current metal strip defect discrimination data, the current metal strip defect image data, and the current metal strip customer identification data are used as input data, and a customer judgment prediction, which is the pass / fail judgment result of the consumer on the surface defects of the metal strip at the current time, is used as output data; Obtaining defect discrimination data that discriminates surface defects of a metal strip, defect image data that captures an image of the surface of the metal strip including the surface defects, and customer identification data that identifies a customer of the metal strip; The acquired defect discrimination data, the defect image data, and the consumer identification data are input as input data to the surface defect determination model, and the consumer determination prediction, which is the output data, is output; A pass / fail judgment is made for the metal strip based on the outputted customer judgment prediction, and the metal strip that is judged to be pass over the entire length of the metal strip is shipped, and surface defects of the metal strip that is judged to be failing are removed. A method for manufacturing a metal strip, wherein the customer judgment data includes judgment results of the surface defects that have been judged as pass by the customer and judgment results of the surface defects that have not been judged as fail by the customer and have been assigned pass information.

2. The method for manufacturing a metal strip according to claim 1 , wherein the defect discrimination data includes a defect type indicating a type of the surface defect, a defect grade indicating a grade of the surface defect, and defect position information indicating position information of the surface defect.

3. The method for manufacturing a metal strip according to claim 1 or 2, wherein the machine learning uses at least one of a neural network, a decision tree learning, a random forest, and a support vector regression.

Images