Classification device, classification method, program, and trained model

Abstract

To provide a classification device capable of classifying landing positions of foreign objects on a glass plate on the basis of shapes (e.g., irregularities) formed on the surface of the glass plate.SOLUTION: A classification device includes a determination unit that inputs image data of a glass plate manufactured in a predetermined flow direction in a plate-like form into a trained machine learning model, and acquires, as a determination result regarding the position in the flow direction where a foreign object has fallen onto the glass plate, an output result from the learned model based on a shape caused by the foreign object on the glass plate.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present disclosure relates to a classification device, a classification method, a program, and a trained model. [Background technology]

[0002] Glass sheets are manufactured by a float process or the like. In the float process, for example, molten glass continuously supplied onto molten tin in a bath is caused to flow on the molten tin and formed into a strip shape.

[0003] Patent Document 1 describes an inspection device that detects defects in optical materials that result in irregular colors and shapes (see Patent Document 1). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2022-56389 Summary of the Invention [Problem to be solved by the invention]

[0005] However, with conventional techniques, there are cases where it is not possible to adequately determine the position where a foreign object has fallen onto a glass plate during the manufacturing process of the glass plate.

[0006] The present disclosure has been made in consideration of these circumstances, and aims to provide a classification device, classification method, program, and trained model that can classify the position at which foreign objects fall onto a glass plate based on the shape (e.g., shape of bumps or recesses) that occurs on the surface of the glass plate. [Means for solving the problem]

[0007] One aspect of the present disclosure is a classification device that includes a determination unit that inputs image data of a glass plate manufactured into a plate shape in a predetermined flow direction into a machine learning trained model, and acquires an output result from the trained model as a determination result regarding the position in the flow direction at which a foreign object fell onto the glass plate, based on a shape caused by the foreign object on the glass plate.

[0008] One aspect of the present disclosure is a classification method in which a classification device inputs image data of a glass plate manufactured into a plate shape in a predetermined flow direction into a trained model of machine learning, and obtains an output result from the trained model as a determination result regarding the position in the flow direction at which a foreign object fell onto the glass plate, based on a shape caused by the foreign object on the glass plate.

[0009] One aspect of the present disclosure is a program for enabling a computer to implement the following functions: inputting image data of a glass plate, which is manufactured into a plate shape in a predetermined flow direction, into a trained model of machine learning; and acquiring the output result from the trained model as a determination result regarding the position in the flow direction at which a foreign object has fallen onto the glass plate, based on the shape of the foreign object on the glass plate due to the foreign object.

[0010] One aspect of the present disclosure is a trained machine learning model that inputs image data of a glass plate manufactured into a plate shape in a predetermined flow direction into a trained machine learning model, and outputs determination result data regarding the position in the flow direction at which a foreign object has fallen onto the glass plate, based on a shape caused by the foreign object on the glass plate. [Effects of the Invention]

[0011] According to the information processing device, classification device, classification method, program, and trained model disclosed herein, it is possible to classify the position at which foreign matter falls on a glass plate based on the shape (e.g., shape such as unevenness) that occurs on the surface of the glass plate. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a diagram illustrating a schematic configuration example of a glass manufacturing apparatus according to an embodiment. [Figure 2] FIG. 1 is a cross-sectional view showing an example of the configuration of a float bath according to an embodiment. [Figure 3] FIG. 2 is a diagram illustrating an example of the arrangement of an imaging unit according to the embodiment. [Figure 4] FIG. 1 is a diagram illustrating an example of the configuration of a classification device according to an embodiment. [Figure 5] 10A to 10I are diagrams showing examples of images in which a foreign object falls upstream (low viscosity) according to an embodiment. [Figure 6] 10A to 10I are diagrams showing examples of images taken when a foreign object falls in the midstream (moderate viscosity) according to the embodiment. [Figure 7] 10A to 10I are diagrams showing examples of images taken when a foreign object falls downstream (high viscosity) according to an embodiment. [Figure 8A] 10A and 10B are diagrams illustrating an example of the shape of a defect in the height direction when a foreign object falls upstream (low viscosity) according to the embodiment. [Figure 8B] 10A and 10B are diagrams illustrating an example of the shape of a defect in the height direction when a foreign object falls in the midstream (medium viscosity) according to the embodiment. [Figure 8C] 10A and 10B are diagrams illustrating an example of the shape of a defect in the height direction when a foreign object falls downstream (high viscosity) according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0014] [Glass manufacturing equipment] FIG. 1 is a diagram showing a schematic configuration example of a glass manufacturing apparatus 1 according to an embodiment. For convenience of explanation, FIG. 1 shows an XYZ Cartesian coordinate system, which is a three-dimensional Cartesian coordinate system. The glass manufacturing apparatus 1 includes a melting furnace 11, a float bath 12, an annealing furnace 13, and an inspection and cutting section 14.

[0015] In the melting furnace 11, the melting process is carried out. In the melting process, glass raw materials prepared by mixing multiple types of raw materials are melted to obtain molten glass. The glass raw materials are charged into the melting furnace 11 and then melted by heating means such as radiant heat from a flame sprayed from a burner or electric melting to become molten glass.

[0016] In the float bath 12, the molding process is carried out. In the forming process, the molten glass obtained in the melting process is continuously supplied onto molten tin in a bath, and the molten glass is caused to flow and form on the molten tin to obtain a glass sheet (so-called glass ribbon), which is a sheet of glass. The glass sheet is cooled while flowing in a predetermined direction and is pulled up from the molten tin.

[0017] In the annealing furnace 13, a annealing process is carried out. In the annealing step, the glass sheet obtained in the forming step is annealed inside the annealing furnace 13. Inside the annealing furnace 13, the glass sheet is annealed while being transported horizontally on rolls from the entrance to the exit of the annealing furnace 13.

[0018] In the inspection and cutting section 14, the inspection process and the cutting process are carried out. In the inspection step, the glass sheet that has been slowly cooled in the slow cooling step is inspected. In the cutting process, the glass plate that has been annealed in the annealing process is cut into predetermined dimensions using a cutting machine. In the cutting process, both widthwise edge portions (so-called edge portions) of the glass plate are cut off. Both widthwise edge portions of the glass plate are cut off because they become thick due to the influence of surface tension, etc. After the cutting process, the glass sheets are shipped.

[0019] [Float Bus] FIG. 2 is a cross-sectional view showing an example of the configuration of the float bath 12 according to the embodiment. FIG. 2 shows an XYZ Cartesian coordinate system similar to that shown in FIG. The cross section shown in FIG. 2 is a side cross section of the float bath 12 as seen from the negative side to the positive side of the X axis. In this embodiment, for convenience of explanation, the upper surface (the surface on the positive side of the Z axis) of the molten glass G (glass sheet) in the float bath 12 will be described as the front surface, and the lower surface (the surface on the negative side of the Z axis) will be described as the back surface, but these surfaces may be called by other names.

[0020] The float bath 12 forms molten glass G, which is continuously supplied onto molten tin M in a bath 22, by causing it to flow on the molten tin M. After being supplied onto the molten tin M near the inlet 21 of the float bath 12, the molten glass G is cooled while flowing in a predetermined direction and is pulled up from the molten tin M near the outlet 23 of the float bath 12. The float bath 12 is composed of a bath 22 that contains molten tin M, a sidewall 24 that is installed along the upper outer periphery of the bath 22, and a ceiling 26 that is connected to the sidewall 24 and covers the upper part of the bath 22. A gas supply path 30 is provided in the ceiling 26 to supply a reducing gas to a space 28 formed between the bath 22 and the ceiling 26. A heater 32 serving as a heat source is inserted into the gas supply path 30.

[0021] The gas supply passage 30 supplies a reducing gas to the space 28 in the float bath 12 to prevent oxidation of the molten tin M. The reducing gas contains, for example, 1 to 15 volume % hydrogen gas and 85 to 99 volume % nitrogen gas. The space 28 in the float bath 12 is set to a pressure higher than atmospheric pressure to prevent air from entering through gaps between the bricks that make up the side wall 24. A plurality of heaters 32 are provided at intervals in the flow direction (Y direction) and width direction (X direction) of the molten glass G to adjust the temperature distribution in the float bath 12. The output of the heaters 32 is controlled so that the temperature of the molten glass G decreases from the inlet 21 toward the outlet 23 of the float bath 12. The output of the heaters 32 is controlled so that the thickness of the molten glass G becomes uniform in the width direction (X direction).

[0022] The bath 22 is composed of a box-shaped metal case 34 that is open at the top, and bottom bricks 36 and side bricks 38 that are installed inside the metal case 34. The metal case 34 prevents air from entering the bath 22 from the sides or below. The bottom bricks 36 are arranged two-dimensionally at small intervals that prevent them from coming into contact with each other due to thermal expansion. The bottom bricks 36 are surrounded by side bricks 38 that are arranged in a ring shape. Also shown in FIG. 2 is the top surface 36a of the bottom brick 36.

[0023] <Example of imaging unit placement> FIG. 3 is a diagram illustrating an example of the arrangement of the imaging unit according to the embodiment. For convenience of explanation, FIG. 3 shows XY coordinates in an XYZ orthogonal coordinate system similar to that shown in FIG. Fig. 3 schematically shows the state of the molten glass G when viewed from above (the positive side of the Z axis) of the float bath 12. Specifically, Fig. 3 shows a glass width region E that represents the width region of the molten glass G in the float bath 12. The glass width region E in the flow direction represents the change in glass width in the flow direction.

[0024] Tensile forces F1 to F4 directed outward in the width direction (to both outer sides) are applied to the molten glass G. As a result, the width of the glass width region E gradually increases from the upstream side to the downstream side of the flow. Although an example in which the width gradually increases is shown as an example, the arrangement from upstream to downstream in the flow is important, and the shape is not limited to this.

[0025] FIG. 3 shows a schematic diagram of a defect inspection section 311 (the inspection section of the inspection and cutting section 14 in FIG. 1). The defect may also be called, for example, a flaw. In this embodiment, in the manufacturing process of a glass plate, a foreign substance falls from above and adheres to the surface (upper surface) of the glass plate, and the shape of the defect, including the position of the foreign substance and its surroundings, changes over time.

[0026] In the example of FIG. 3, the inspection unit 311 has three imaging units C1 to C3 arranged in the width direction. The imaging unit C1 captures an image of the inspection region R1. The imaging unit C2 captures an image of the inspection region R2. The imaging unit C3 captures an image of the inspection region R3. Here, the inspection region R1, the inspection region R2, and the inspection region R3 are aligned in the width direction corresponding to the imaging units C1 to C3, respectively. 3, a defect H12 exists in inspection area R1, and a defect H11 exists in inspection area R3. Here, the defect H11 corresponds to the defect H1 and the defect H1a, and the defect H12 corresponds to the defect H2 and the defect H2a. In the example of FIG. 3, there is no defect in the inspection area R2. In this way, each of the imaging units C1 to C3 captures an image of a partial area of ​​the glass plate. If the area contains a defect, the image contains a portion where the defect is captured (defective portion).

[0027] Here, in this embodiment, three imaging units C1 to C3 are shown, but the number of imaging units may be arbitrary. 3, the inspection regions R1 to R3 arranged in the width direction are shown spaced apart from each other to make the inspection regions R1 to R3 easier to see, but they may be adjacent to each other, that is, they may be capable of inspecting the entire width direction of the molten glass G. Alternatively, as shown in FIG. 3, a plurality of inspection regions spaced apart from each other may be used. Furthermore, although each of the imaging units C1 to C3 is shown simply in FIG. 3, each of the imaging units C1 to C3 may be a combination of devices such as a plurality of cameras.

[0028] <Flow direction classification> FIG. 3 shows three regions in the flow direction: upstream P1, midstream P2, and downstream P3. Such flow direction division may or may not be used to estimate the location of defect occurrence. In the example of FIG. 3, the flow direction is divided into three regions, but the number of divisions and the size of each region may be arbitrary.

[0029] In the example of Figure 3, with respect to the direction of the Y axis, which corresponds to the flow direction of the float bath 12, the temperature of the molten glass G is higher and the viscosity of the molten glass G is lower on the upstream side of the flow of the float bath 12 (negative side of the Y axis) than on the downstream side (positive side of the Y axis). Conversely, the temperature of the molten glass G is lower and the viscosity of the molten glass G is higher on the downstream side of the flow of the float bath 12 than on the upstream side.

[0030] [Classification device] FIG. 4 is a diagram illustrating an example of the configuration of the classification device 111 according to the embodiment. The classification device 111 is, for example, configured by a computer. The classification device 111 includes an input unit 131 , an output unit 132 , a communication unit 133 , a storage unit 134 , and a control unit 135 . The output unit 132 includes a display unit 151 . The control unit 135 includes a learning control unit 191 and a determination unit 192 .

[0031] The input unit 131 has a function of inputting information in response to an operation by a user (for example, a person), and a function of inputting information from an external device. The output unit 132 has a function of outputting information to the outside. For example, the display unit 151 has a function of displaying information to be displayed on a screen. The communication unit 133 has a function of communicating with an external device.

[0032] The storage unit 134 stores various types of information. In this embodiment, the storage unit 134 stores a learning model 171, training data 172, and the like. The learning model 171 may be set in advance in the storage unit 134, or may be set externally at any timing. The teacher data 172 may be set in advance in the storage unit 134, or may be set externally at any timing.

[0033] Here, any model may be used as the learning model 171, for example, a neural network model may be used. The learning model 171 extracts, for example, feature quantities (image feature quantities) of input image data and outputs an estimation result (determination result) based on the feature quantities. In this embodiment, the feature quantities are information representing the positional status of a foreign object falling onto the glass plate in the flow direction.

[0034] In this embodiment, the function of the feature extraction unit that extracts the features of image data and the function of the estimation unit that outputs an estimation result (judgment result) regarding the position in the flow direction where a foreign object has fallen onto a glass plate based on the extracted features are integrated and included in the learning model 171. As another example, the function of the feature extraction unit and the function of the estimation unit may be configured using separate learning models. In this case, image data is input to a learning model having the function of the feature extraction unit, and the output result (feature extraction result) from the learning model is input to a learning model having the function of the estimation unit, and the output from the learning model is used as the estimation result (determination result).

[0035] The control unit 135 performs various processes and controls. The learning control unit 191 uses the teacher data 172 to learn the learning model 171. This learning is machine learning. Note that, in the learning stage, for example, instructions based on the judgment results of the user may be used for learning.

[0036] The determination unit 192 inputs image data to be determined to the trained learning model 171, and outputs the output from the learning model 171 as a determination result via the output unit 132. For example, the judgment unit 192 may directly adopt the output (estimation result) from the trained learning model 171 as the judgment result by the judgment unit 192, or may make a further judgment (for example, trend judgment) based on the output (estimation result) from the trained learning model 171, and use the result of that judgment as the judgment result by the judgment unit 192.

[0037] Here, the image data to be determined may be input from the input unit 131, or may be stored in advance in the storage unit . For example, the judgment unit 192 may always perform judgment processing in real time by inputting image data acquired by the inspection unit 311 or image data processed from the image data in real time from the input unit 131. The determination process may be performed automatically in the classifier 111, for example.

[0038] <Example of an image with unevenness> Referring to Figures 5(A) to 5(I), Figures 6(A) to 6(I), and Figures 7(A) to 7(I), examples of images of positions where foreign objects fell in the flow direction of the glass plate manufacturing process are shown. 5(A) to 5(I) are diagrams showing examples of images when a foreign object falls upstream (low viscosity) according to the embodiment. 6(A) to 6(I) are diagrams showing examples of images when a foreign object falls in the midstream (medium viscosity) according to the embodiment. 7(A) to 7(I) are diagrams showing examples of images when foreign matter falls downstream (high viscosity) according to the embodiment.

[0039] 5(A) to 5(I) show nine types of images Q1 to Q9, respectively. The viscosity of the glass plate is low upstream. Of Figures 5(A) to 5(I), Figure 5(I) shows an XYZ Cartesian coordinate system similar to that of Figure 1. The orientation of the images in Figures 5(A) to 5(H) is the same as that in Figure 5(I). In the examples of FIGS. 5(A) to 5(I), the vertical direction in the drawings corresponds to the flow direction in the manufacturing process of a glass sheet by the float method. These images Q1 to Q9 are images captured in the same imaging range, but at least one of the imaging conditions and the image processing is different. Images Q1, Q2, and Q3 are images captured by a transmission optical system camera. Images Q4, Q5, and Q6 are images captured by a reflective optical system camera. Images Q7, Q8, and Q9 are images captured by a diffuse transmission optical system camera.

[0040] Moreover, the images Q1, Q4, and Q7 are images as they are captured (raw images). Moreover, images Q2, Q5, and Q8 are images resulting from contrast-uniform image processing. Moreover, images Q2, Q5, and Q8 are images resulting from image processing for contrast enhancement.

[0041] Here, as the contrast equalization image processing, for example, histogram equalization image processing may be used. Furthermore, the image processing for contrast enhancement is not particularly limited, but may, for example, be processing for enhancing gradation values ​​in image data captured by an imaging device. As one example, this processing may be processing for converting gradation values ​​so that the original gradation values ​​and the changed gradation values ​​match at a predetermined boundary value (predetermined gradation value), and the changed gradation values ​​are smaller than the original gradation values ​​at gradation values ​​smaller than the boundary value, and the changed gradation values ​​are larger than the original gradation values ​​at gradation values ​​larger than the boundary value.

[0042] 6(A) to 6(I) show nine types of images Q11 to Q19, respectively. In the midstream, the viscosity of the glass plate is medium. Of Figures 6(A) to 6(I), Figure 6(I) shows an XYZ Cartesian coordinate system similar to that of Figure 1. The orientation of the images in Figures 6(A) to 6(H) is the same as that in Figure 6(I). In the examples of FIGS. 6(A) to 6(I), the vertical direction in the drawings corresponds to the flow direction in the manufacturing process of a glass sheet by the float method. These images Q11 to Q19 are images captured in the same imaging range, but at least one of the imaging conditions and the image processing is different. The types of these images Q11 to Q19 (types of combinations of imaging conditions and image processing) are the same as the types of the images Q1 to Q9 shown in FIGS. 5(A) to 5(I), respectively.

[0043] 7(A) to 7(I) show nine types of images Q21 to Q29, respectively. The viscosity of the glass plate is high downstream. Of Figures 7(A) to 7(I), Figure 7(I) shows an XYZ Cartesian coordinate system similar to that of Figure 1. The orientation of the images in Figures 7(A) to 7(H) is the same as that in Figure 7(I). In the examples of FIGS. 7(A) to 7(I), the vertical direction in the drawings corresponds to the flow direction in the manufacturing process of a glass sheet by the float method. These images Q21 to Q29 are images captured in the same imaging range, but at least one of the imaging conditions and the image processing is different. The types of these images Q21 to Q29 (types of combinations of imaging conditions and image processing) are the same as the types of the images Q1 to Q9 shown in FIGS. 5(A) to 5(I), respectively.

[0044] In the machine learning and determination in this embodiment, for example, one type of image may be used as an input image, or two or more types of images (number of types) may be used. When one type of image is used as the input image, for example, any of the nine types of images shown may be used. Note that an image of a type not shown in this example may also be used as the input image. When two or more types of images are used as input images, any two or more types of images from the nine types of images shown as examples may be used. Note that images of types not shown in this example may be used as part or all of the two or more types of input images. Here, in machine learning and determination, it is believed that the accuracy of determination (classification accuracy) increases as the number of types of input images increases, for example.

[0045] <Example of the height direction shape of the defect> 8A, 8B, and 8C show examples of the shape of the defect in the height direction. FIG. 8A is a diagram showing an example of the shape of a defect in the height direction when a foreign object falls upstream (low viscosity) according to the embodiment. FIG. 8B is a diagram showing an example of the shape of a defect in the height direction when a foreign object falls in the midstream (moderate viscosity) according to the embodiment. FIG. 8C is a diagram showing an example of the shape of a defect in the height direction when a foreign object falls downstream (high viscosity) according to the embodiment. Here, these shapes represent shapes that appear in images captured by, for example, an imaging device installed at the same predetermined location (in the example of FIG. 3, any of the imaging devices that make up the imaging units C1 to C3). 8A, 8B, and 8C each show an XYZ Cartesian coordinate system similar to that shown in FIG.

[0046] In each of Figures 8A, 8B and 8C, the horizontal axis represents the flow direction of the glass plate in the manufacturing process, and the vertical axis represents the position of one defect and its periphery in the height direction on the upper surface of the glass plate. The examples of FIGS. 8A, 8B, and 8C show a case where a foreign substance falls onto the upper surface of a glass plate from above, and the glass plate reacts with the foreign substance to generate a defect.

[0047] FIG. 8A shows the characteristic 1011 of the height direction shape of the defect with respect to the position in the flow direction. As shown in Figure 8A, when a foreign object falls upstream (where the glass plate has low viscosity), the image captured by the imaging device installed downstream shows a concave shape around the defect, meaning that there are areas on both the upstream and downstream sides of the defect that are lower in height.

[0048] FIG. 8B shows the characteristic 1021 of the height direction shape of the defect with respect to the position in the flow direction. As shown in Figure 8B, when a foreign object falls in the midstream (when the glass plate has medium viscosity), the image captured by the imaging device installed downstream shows a raised shape around the defect, meaning that there are areas that are higher in height both upstream and downstream of the defect.

[0049] FIG. 8C shows the profile 1031 of the height direction of the defect with respect to the position in the flow direction. As shown in Figure 8C, when a foreign object falls downstream (where the glass plate has high viscosity), the image captured by an imaging device installed further downstream shows a shape in which the depression in the defective area (the depression in the center of the defective area) and the raised area around it do not exist.

[0050] <First example of learning and judgment> As input data (teaching data 172, data to be determined) of the learning model 171, image data of a glass plate containing foreign matter is used. The output data of the learning model 171 is used as data of the determination result regarding the position in the flow direction of the foreign matter that fell on the glass plate.

[0051] Here, the determination result regarding the position in the flow direction may be, for example, one or more of information identifying the position (itself), or information identifying the area (division area) in which the position is included among multiple areas (division areas) in the flow direction. As such regions, for example, regions such as upstream P1, midstream P2, and downstream P3 shown in FIG. 3 may be used.

[0052] As a specific example, the determination unit 192 may determine an index of the position in the flow direction where the foreign object fell from an image captured (scanned) of information related to the position in the flow direction where the foreign object fell, using a trained model (trained learning model 171). The determination unit 192 may, for example, determine an index of the position in the flow direction where the foreign matter fell for a large number of images that change over time in the same inspection areas R1 to R3, and perform trend aggregation of the determination results, thereby more accurately identifying (determining) the index.

[0053] Here, in this example, in machine learning, if the shape has the characteristic that there is no depression at the position of the defect caused by the foreign object (for example, the position of the center of the defect) and no bulge around the position of the defect, the result of a judgment (for convenience of explanation, referred to as judgment A1) that the position where the foreign object fell is downstream may be used as training data 172. Furthermore, in this example, in machine learning, if the shape has a characteristic that there is a depression at the position of the defect caused by the foreign object (for example, the position of the center of the defect) and a bulge around the position of the defect, the result of a judgment (for convenience of explanation, referred to as judgment A2) that the position where the foreign object fell is midstream may be used as training data 172. Furthermore, in this example, in machine learning, for example, if the shape has a characteristic that there is a depression at the position of the defect caused by the foreign object (for example, the position of the center of the defect) and there is no bulge around the position of the defect, the result of a judgment (for convenience of explanation, referred to as judgment A3) that the position where the foreign object fell is upstream may be used as training data 172.

[0054] Furthermore, for example, the result of determining that one of a plurality of discrete or continuous positions in the flow direction is the position where the foreign object fell may be used as training data 172 based on one or more of the degree to which the shape features focused on in the above-mentioned judgment A1 appear, the degree to which the shape features focused on in the above-mentioned judgment A2 appear, or the degree to which the shape features focused on in the above-mentioned judgment A3 appear.

[0055] In the trained model (trained learning model 171), for example, one or more of the following may be used as features: a depression (concave portion) at the location of the foreign object, a depression (concave portion) around the location of the foreign object, or a protrusion (convex portion) around the location of the foreign object.

[0056] <Second example of learning and judgment> As input data (teacher data 172, data to be determined) to the learning model 171, data of an image of a glass plate containing foreign matter and auxiliary data (auxiliary data) may be used. The output data of the learning model 171 is used as data of the determination result regarding the position in the flow direction of the foreign matter that fell on the glass plate. Here, data representing various characteristics may be used as the auxiliary data.

[0057] <Example of the result of determining the position of a foreign object in the flow direction when it fell onto a glass plate> In this embodiment, the determination result regarding the position in the flow direction where the foreign object fell onto the glass plate is an index of that position, and may be, for example, a value representing the position itself, or a value representing an area including that position, or information used to derive such a value.

[0058] Here, as information relating to the position in the flow direction where a foreign particle has fallen onto the glass plate, for example, information for each individual foreign particle may be used, or information for a group of multiple foreign particles may be used. For example, if a unit area to be judged (judgment unit area) includes multiple foreign objects, a judgment result regarding the position in the flow direction of the foreign objects that fell onto the glass plate may be acquired for a collection of these multiple foreign objects. As a specific example, image data obtained by capturing an image of the unit area to be judged (judgment unit area) may be used as image data to be used during machine learning learning and judgment (image data to be input to the learning model 171).

[0059] <Example of imaging device> In this embodiment, a preferred example is a mode in which a transmission camera (transmission optical system camera) is used as the imaging device (the imaging device constituting the imaging units C1 to C3 shown in FIG. 3). However, the imaging device is not limited to this, and other devices such as a reflective optical system camera or a diffuse transmission camera may also be used. Furthermore, in this embodiment, it is particularly preferable to use the results of contrast-uniform image processing performed on an image captured by a transmission camera, although other image processing may also be performed as the image processing, or image processing may not necessarily be performed.

[0060] The positions at which the imaging devices are installed (the positions of the imaging units C1 to C3 shown in FIG. 3) are not necessarily limited to the example in FIG. 3, and they may be installed at other positions (other locations) in the manufacturing process of the glass plate.

[0061] As described above, the classification device 111 according to this embodiment can classify the positions at which foreign matter has fallen on a glass plate based on the shapes that appear on the surface of the glass plate. The shape is not particularly limited, and for example, the shape of part or all of a recess, the shape of part or all of a protrusion, or one or more of other shapes may be used.

[0062] It should be noted that the classification device 111 may be, for example, a non-online device or an online device (for example, a cloud server or an on-premise server). Furthermore, in the present embodiment, the float method is used as a method for manufacturing a flat glass sheet. However, as another example, a fusion method may also be used.

[0063] As one configuration example, in the classification device 111, the judgment unit 192 inputs image data of a glass plate manufactured into a plate shape in a predetermined flow direction into a machine learning trained model, and obtains the output result from the trained model as a judgment result regarding the position in the flow direction at which a foreign object fell onto the glass plate, based on the shape caused by the foreign object on the glass plate. Therefore, the classification device 111 can classify the positions at which foreign matter has fallen on the glass plate based on the shape (for example, the shape of irregularities, etc.) that appears on the surface of the glass plate.

[0064] As one configuration example, in the classification device 111, the trained model uses one or more of the following features: a depression at the position of a foreign object, a depression around the position of a foreign object, or a protrusion around the position of a foreign object. Therefore, the classification device 111 can improve the accuracy of classification.

[0065] As an example configuration, the classification device 111 includes an imaging device (in the example of FIG. 2, imaging devices constituting the imaging units C1 to C3) that captures images of the image data. The imaging device is a see-through camera. Therefore, the classification device 111 can easily recognize shapes resulting from foreign matter in image data, and the accuracy of classification can be further improved.

[0066] As an example of configuration, the classification device 111 acquires, as a determination result, information specifying one of a plurality of regions in the flow direction as a position in the flow direction. Therefore, the classification device 111 can perform classification using regions such as upstream, midstream, and downstream.

[0067] As one configuration example, in the classification device 111, the image data is data of an image of a part of a glass plate, and includes a part in which a foreign object is captured. Therefore, the classification device 111 can classify the positions at which the foreign matter falls on the glass plate based on the shape (such as the shape of irregularities) that appears on the surface of the glass plate for each portion where the foreign matter exists on the glass plate.

[0068] As an example configuration, in the classification device 111, the learning control unit 191 generates a trained model using training data. Therefore, the classification device 111 generates a trained model and uses the trained model to classify the positions at which foreign objects fall on the glass plate based on the shapes (e.g., shapes such as bumps and depressions) that occur on the surface of the glass plate.

[0069] In this embodiment, the classification device 111 is shown to have both the learning control unit 191 and the judgment unit 192. However, as another example, the learning control unit 191 may not be provided in the classification device 111 but in another device, and the classification device 111 may use a trained model obtained by the other device.

[0070] A program for implementing the functions of any of the components of any of the above-described devices may be recorded on a computer-readable recording medium and loaded into a computer system for execution. The term "computer system" as used herein includes hardware such as an operating system or peripheral devices. The term "computer-readable recording medium" refers to portable media such as flexible disks, optical magnetic disks, ROMs, and compact discs (CDs) or read-only memories (ROMs), as well as storage devices such as hard disks built into computer systems. The term "computer-readable recording medium" also includes devices that retain a program for a certain period of time, such as volatile memory within a computer system that acts as a server or client when a program is transmitted over a network such as the Internet or a communication line such as a telephone line. Such volatile memory may be, for example, random access memory (RAM). The recording medium may also be, for example, a non-transitory recording medium.

[0071] The above program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by transmission waves in the transmission medium. Here, the "transmission medium" that transmits the program refers to a medium that has the function of transmitting information, such as a network such as the Internet or a communication line such as a telephone line. The above program may also be one that realizes part of the above-mentioned functions. Furthermore, the above program may be a so-called differential file that can realize the above-mentioned functions in combination with a program already recorded in a computer system. A differential file may also be called a differential program.

[0072] Furthermore, the functions of any of the components in any of the above-described devices may be implemented by a processor. For example, each process in the embodiments may be implemented by a processor operating based on information such as a program and a computer-readable recording medium storing information such as the program. Here, the functions of each unit of the processor may be implemented by, for example, individual hardware, or may be implemented by integrated hardware. For example, the processor may include hardware, and the hardware may include at least one of a circuit for processing digital signals and a circuit for processing analog signals. For example, the processor may be configured using one or more circuit devices mounted on a circuit board, or one or both of one or more circuit elements. An integrated circuit (IC) or the like may be used as the circuit device, and a resistor or a capacitor may be used as the circuit element.

[0073] Here, the processor may be, for example, a CPU. However, the processor is not limited to a CPU, and various types of processors such as a GPU (Graphics Processing Unit) or a DSP (Digital Signal Processor) may be used. The processor may also be, for example, a hardware circuit such as an ASIC (Application Specific Integrated Circuit). The processor may also be, for example, composed of multiple CPUs, or may be, for example, composed of a hardware circuit such as a multiple ASIC. The processor may also be, for example, composed of a combination of multiple CPUs and a hardware circuit such as a multiple ASIC. The processor may also include, for example, one or more of an amplifier circuit or a filter circuit that processes analog signals.

[0074] The embodiments of this disclosure have been described in detail above with reference to the drawings, but the specific configuration is not limited to this embodiment, and includes designs within the scope that do not deviate from the gist of this disclosure.

[0075] [Note] (Configuration example 1) to (Configuration example 9) are shown.

[0076] (Configuration example 1) a determination unit that inputs image data of a glass plate manufactured in a plate shape in a predetermined flow direction into a machine learning trained model, and acquires an output result from the trained model as a determination result regarding a position in the flow direction at which a foreign object has fallen on the glass plate, based on a shape of the foreign object on the glass plate due to the foreign object; Classification device.

[0077] (Configuration example 2) In the trained model, one or more of a depression at the position of the foreign object, a depression around the position of the foreign object, or a protrusion around the position of the foreign object is used as a feature. The classification device according to (Configuration Example 1).

[0078] (Configuration example 3) an imaging device that captures an image of the image data; The imaging device is a transmission camera. The classification device according to (Configuration Example 1) or (Configuration Example 2).

[0079] (Configuration Example 4) The determination result acquires information that identifies one area among a plurality of areas in the flow direction as the position in the flow direction. The classification device according to any one of (Configuration Example 1) to (Configuration Example 3).

[0080] (Configuration Example 5) The image data is data of an image of a part of the glass plate, and includes a part in which the foreign matter is captured. The classification device according to any one of (Configuration Example 1) to (Configuration Example 4).

[0081] (Configuration Example 6) A learning control unit that generates the trained model using training data. The classification device according to any one of (Configuration Example 1) to (Configuration Example 5).

[0082] (Configuration Example 7) It is also possible to provide a method of processing (classification method) performed by the classification device. A classification device inputs image data of a glass plate, which is manufactured into a plate shape in a predetermined flow direction, into a machine learning trained model, and acquires an output result from the trained model as a determination result regarding a position in the flow direction at which a foreign object has fallen onto the glass plate, based on a shape of the foreign object on the glass plate. Classification method.

[0083] (Configuration Example 8) It is also possible to provide a program that is executed by a computer (computer program). On the computer, A function to input image data of glass sheets manufactured into a plate shape in a predetermined flow direction into a machine learning trained model; A function of acquiring an output result from the trained model as a determination result regarding a position in the flow direction at which the foreign object fell on the glass plate, based on a shape caused by the foreign object on the glass plate; A program to achieve this.

[0084] (Configuration Example 9) It is also possible to provide pre-trained machine learning models. Image data of a glass plate manufactured in a plate shape in a predetermined flow direction is input into a machine learning trained model, and determination result data regarding a position in the flow direction at which the foreign object has fallen on the glass plate is output based on a shape of the foreign object on the glass plate due to the foreign object. A trained machine learning model. [Explanation of symbols]

[0085] 1...glass manufacturing apparatus, 11...melting furnace, 12...float bath, 13...lehr, 14...inspection and cutting section, 21...inlet, 22...bathtub, 23...outlet, 24...side wall, 26...ceiling, 28...space, 30...gas supply path, 32...heater, 34...metal case, 36...bottom brick, 36a...top surface, 38...side brick, 111...sorting device, 131...input section, 132...output section, 133...communication section, 134...storage section, 135...control section, 151...display section, 1 71...Learning model, 172...Teaching data, 191...Learning control unit, 192...Determination unit, 311...Inspection unit, 1011, 1021, 1031...Characteristics, C1 to C3...Imaging unit, E...Glass width area, F1 to F4...Tensile force, G...Molten glass, H1, H1a, H2, H2a, H11, H12...Defect area, M...Molten tin, P1...Upstream, P2...Midstream, P3...Downstream, Q1 to Q9, Q11 to Q19, Q21 to Q29...Image, R1 to R3...Inspection area

Claims

1. a determination unit that inputs image data of a glass plate manufactured in a plate shape in a predetermined flow direction into a machine learning trained model, and acquires an output result from the trained model as a determination result regarding a position in the flow direction at which a foreign object has fallen on the glass plate, based on a shape of the foreign object on the glass plate due to the foreign object; Classification device.

2. In the trained model, one or more of a depression at the position of the foreign object, a depression around the position of the foreign object, or a protrusion around the position of the foreign object is used as a feature. The classification device of claim 1 .

3. an imaging device that captures an image of the image data; The imaging device is a transmission camera. The classification device according to claim 1 or 2.

4. the determination result acquires information that identifies one area among a plurality of areas in the flow direction as the position in the flow direction. The classification device according to claim 1 or 2.

5. The image data is data of an image of a part of the glass plate, and includes a part in which the foreign matter is captured. The classification device according to claim 1 or 2.

6. A learning control unit that generates the trained model using training data. The classification device according to claim 1 or 2.

7. A classification device inputs image data of a glass plate, which is manufactured into a plate shape in a predetermined flow direction, into a machine learning trained model, and acquires an output result from the trained model as a determination result regarding a position in the flow direction at which a foreign object has fallen onto the glass plate, based on a shape of the foreign object on the glass plate. Classification method.

8. On the computer, A function to input image data of glass sheets manufactured into a plate shape in a predetermined flow direction into a machine learning trained model; A function of acquiring an output result from the trained model as a determination result regarding a position in the flow direction at which the foreign object fell on the glass plate, based on a shape caused by the foreign object on the glass plate; A program to achieve this.

9. Image data of a glass plate manufactured in a plate shape in a predetermined flow direction is input into a machine learning trained model, and determination result data regarding a position in the flow direction at which the foreign object has fallen on the glass plate is output based on a shape of the foreign object on the glass plate due to the foreign object. A trained machine learning model.

Images