Method for manufacturing honeycomb structure and inspection apparatus

Abstract

To reduce excessive detection of defects of a prescribed defect type about an end surface of a honeycomb structure.SOLUTION: A manufacturing method of a honeycomb structure extracts multiple fragment images from an inspection object image that shows at least a portion of a first end surface of a honeycomb structure that has a porous partition wall that partitions multiple cells from each other, which extend from the first end surface to a second end surface, and classifies the fragment image for each of the multiple fragment images by inputting the multiple fragment images to a learning model that takes images as input and types as output. The manufacturing method determines whether or not the first end surface is defective on the basis of the number of defective cells that are cells that belong to the fragment image classified into a prescribed defect type.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention generally relates to a manufacturing technique of a honeycomb structure, and specifically relates to, for example, an inspection technique for an end face of a honeycomb structure.

Background Art

[0002] It is known that an inspection of a honeycomb structure can be included in a method for manufacturing a honeycomb structure. As a technique related to the inspection of a honeycomb structure, for example, an inspection apparatus disclosed in Patent Document 1 is known. Patent Document 1 discloses the following. That is, the inspection apparatus obtains first image data for processing of the end face while irradiating the end face of the honeycomb structure with light at an angle of 40° or more, and obtains second image data for processing of the end face while irradiating the end face with light at an angle of less than 40°, and compares the obtained first image data for processing with the second image data for processing to detect a crack.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the inspection of the end face of a honeycomb structure, for the purpose of further improving the detection accuracy of cracks, a method of inputting the obtained image into a learning model can be considered. However, in that method, there is a concern about over-detection of cracks. One of the reasons is as follows.

[0005] That is, the image input to the learning model is an image of a portion where a defect appears, that is, an image in which the entire defect appears. Therefore, when the defect is a crack, an image of the entire crack is input to the learning model.

[0006] However, it is difficult to train a learning model to accurately determine that the defect type is a crack. One of the reasons is that the lengths of cracks vary, and depending on the crack, the image in which the crack appears may cover a wide range of the image. In a wide-range image, the area where the crack does not appear may be larger than the area where the crack appears. Therefore, the features of the crack are small, and as a result, the accuracy of the learning model decreases. In addition, in the area where the crack does not appear, there may be defects other than cracks (for example, chipping). If an image contains both a crack and other defects, the features of the crack become inaccurate, and as a result, the accuracy of the learning model decreases.

[0007] For these reasons, it is difficult to train a learning model accurately. Therefore, even if an image of a defect other than a crack is input to the learning model, there is a risk of over-detection that the defect type is a crack. Also, such problems can occur for defects of a predetermined type different from cracks (for example, a predetermined type of cell deformation).

Means for Solving the Problem

[0008] A method for manufacturing a honeycomb structure includes a model-based classification step and a rule-based determination step. The model-based classification step includes extracting a plurality of fragment images from a test image in which at least a part of a first end face of a honeycomb structure having a porous partition wall that partitions a plurality of cells extending from the first end face to the second end face is shown, and inputting each of the plurality of fragment images to a learning model that takes an image as an input and outputs a type, thereby classifying each of the plurality of fragment images. The rule-based determination step includes determining whether there is a defect on the first end face based on the number of defective cells that are cells belonging to the fragment images classified into a predetermined defect type.

Effect of the Invention

[0009] According to the present invention, since the image input to the learning model is a fragmentary image, a learning model with high determination accuracy can be prepared, thereby reducing the over-detection of defects of a predetermined defect type with respect to the end face of the honeycomb structure.

Brief Description of the Drawings

[0010]

Figure 1

Figure 2A

Figure 2B

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10A

Figure 10B

Figure 10C

Figure 11

Modes for Carrying Out the Invention

[0011] In the following description, the "interface device" may be one or more interface devices. The one or more interface devices may be at least one of the following. · One or more I / O (Input / Output) interface devices. The I / O (Input / Output) interface device is an interface device for at least one of an I / O device and a remote display computer. The I / O interface device for the display computer may be a communication interface device. At least one I / O device may be either an input device such as a user interface device, for example, a keyboard and a pointing device, or an output device such as a display device. · One or more communication interface devices. The one or more communication interface devices may be one or more homogeneous communication interface devices (for example, one or more NICs (Network Interface Cards)) or two or more heterogeneous communication interface devices (for example, a NIC and an HBA (Host Bus Adapter)).

[0012] Also, in the following description, "memory" is one or more memory devices which are an example of one or more storage devices and may typically be a main memory device. At least one memory device in the memory may be a volatile memory device or a non-volatile memory device.

[0013] Also, in the following description, "persistent storage device" may be one or more persistent storage devices which are an example of one or more storage devices. The persistent storage device may typically be a non-volatile storage device (for example, an auxiliary storage device), specifically, for example, an HDD (Hard Disk Drive), an SSD (Solid State Drive), an NVME (Non-Volatile Memory Express) drive, or an SCM (Storage Class Memory).

[0014] Also, in the following description, "storage device" may be at least the memory of the memory and the persistent storage device.

[0015] Also, in the following description, a "processor" may be one or more processor devices. At least one processor device may typically be a microprocessor device such as a CPU (Central Processing Unit), but may also be other types of processor devices such as a GPU (Graphics Processing Unit). At least one processor device may be single-core or multi-core. At least one processor device may be a processor core. At least one processor device may be a circuit that is an aggregate of gate arrays (e.g., an FPGA (Field-Programmable Gate Array), a CPLD (Complex Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit)) described in a hardware description language that performs part or all of the processing, which is a processor device in a broad sense.

[0016] Also, in the following description, the function may be described in terms of a "yyy section", but the function may be realized by one or more computer programs being executed by a processor, or may be realized by one or more hardware circuits (e.g., an FPGA or an ASIC), or may be realized by a combination thereof. When the function is realized by a program being executed by a processor, since the defined processing is performed while appropriately using a storage device and / or an interface device, etc., the function may be regarded as at least part of the processor. The processing described with the function as the subject may also be the processing performed by the processor or a device having the processor. The program may be installed from a program source. The program source may be, for example, a program distribution computer or a computer-readable recording medium (e.g., a non-transitory recording medium). The description of each function is an example, and a plurality of functions may be combined into one function, or one function may be divided into a plurality of functions.

[0017] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0018] FIG. 1 schematically shows the configuration of an inspection system 500 according to an embodiment of the present invention.

[0019] The inspection system 500 includes a stage 14, a first light source 11a, a second light source 11b, an imaging device 13, and an inspection device 450.

[0020] The stage 14 is a stage for placing the honeycomb structure 1 made of ceramics. The honeycomb structure 1 has a first end face 2a, a second end face 2b, and side faces. The second end face 2b of the honeycomb structure 1 is placed on the stage 14.

[0021] The first light source 11a is a high-angle illumination light source. The first light source 11a irradiates the first end face 2a of the honeycomb structure 1 with light 12 whose angle formed with respect to the first end face 2a is a high angle. This irradiation can be referred to as "high-angle irradiation".

[0022] The second light source 11b is a low-angle illumination light source. The second light source 11b irradiates the first end face 2a of the honeycomb structure 1 with light 12 whose angle formed with respect to the first end face 2a is a low angle. This irradiation can be referred to as "low-angle irradiation".

[0023] The imaging device 13 obtains first processing target image data (e.g., captured image data) of the first end face 2a that receives the light 12 irradiated at a high angle by the first light source 11a. Further, the imaging device 13 obtains second processing target image data (e.g., captured image data) of the first end face 2a that receives the light 12 irradiated at a low angle by the second light source 11b. The first processing target image data can be referred to as "high-angle illumination image data", and the first processing target image represented by the first processing target image data can be referred to as a "high-angle illumination image". Further, the second processing target image data can be referred to as "low-angle illumination image data", and the second processing target image represented by the second processing target image data can be referred to as a "low-angle illumination image". In the present embodiment, the "high angle" may be an angle higher than the "low angle" or an angle equal to or greater than a first angle threshold value. Further, the "low angle" may be an angle lower than the "high angle" or an angle less than a second angle threshold value. The second angle threshold value may be the same as or smaller than the first angle threshold value.

[0024] The imaging device 13 may be provided in a direction perpendicular to the mounting surface of the honeycomb structure 1. Further, at least a pair of the first light sources 11a may be provided at symmetric positions with the imaging device 13 interposed therebetween in a top view (plan view). At least a pair of the second light sources 11b may be provided at symmetric positions with the imaging device 13 interposed therebetween in a top view (plan view).

[0025] An area camera may be used as the imaging device 13. Further, the imaging device 13 may include a telecentric optical system. The entire first end face 2a of the honeycomb structure 1 can be imaged by the area camera. The imaging device 13 may include a plurality of area cameras and image the first end face 2a separately. Alternatively, a line sensor may be used as the imaging device 13. The line sensor may be longer than the length in the direction perpendicular to the moving direction of the first end face 2a of the imaging target.

[0026] As the light sources 11a and / or 11b, a line illumination source or a point illumination source may be used. Also, as the light sources 11a and / or 11b, an LED light source, a laser, a halogen, a metal halide light source, etc. may be used.

[0027] Either the high-angle illumination image data or the low-angle illumination image data may be obtained first, or those data may be obtained simultaneously. For example, the light 12 of the first light source 11a and the light 12 of the second light source 11b may be lights of different wavelengths. In this case, the high-angle illumination image data and the low-angle illumination image data may be obtained simultaneously. As lights of different wavelengths, for example, a blue light source and a red light source can be used. When the high-angle illumination image data and the low-angle illumination image data are obtained simultaneously, the high-angle illumination image data and the low-angle illumination image data will be included in one piece of processed image data. In this case, the inspection device 450 can separate the high-angle illumination image data (for example, the image data of the red component) and the low-angle illumination image data (for example, the image data of the blue component) from the RGB image data as one piece of processed image data.

[0028] The inspection device 450 may be a computer such as a personal computer, and includes an input device 572, a display device 540, and a control device 570 connected thereto. The input device 572 and the display device 540 may be integrated like a touch panel.

[0029] Processed image data (captured image) from the imaging device 13 that is capturing the first end face 2a of the honeycomb structure 1 is input to the control device 570. The control device 570 detects a defect of a predetermined type if there is a defect in the first end face 2a represented by the processed image data.

[0030] FIG. 2A is a schematic plan view of the first end face 2a of the honeycomb structure 1. FIG. 2B is a perspective view of the honeycomb structure.

[0031] The honeycomb structure 1 is a structure having a porous partition wall 4 that partitions and forms a plurality of cells 5 extending from the first end face 2a to the second end face 2b. Specifically, for example, the honeycomb structure 1 is formed in a columnar shape and has a flow path that penetrates from the first end face 2a to the second end face 2b in the longitudinal direction (axial direction) 9. A large number of cells 5 that serve as flow paths are partitioned and formed by the partition wall 4.

[0032] The shape, size, material, etc. of the honeycomb structure 1 may not be limited. For example, the honeycomb structure 1 may be prismatic instead of cylindrical. Also, for example, the thickness (rib thickness) of the partition wall 4 of the cell 5 and the density of the partition wall 4 may be appropriately designed according to the purpose. Further, the honeycomb structure 1 may be a structure mainly composed of ceramics (for example, containing 50% by mass or more of ceramics).

[0033] Figure 3 shows the configuration of the control device 570.

[0034] The control device 570 has an interface device 10, a storage device 20, and a processor 30 connected thereto.

[0035] The interface device 10 is communicably connected to an input device 572, a display device 540, and a photographing device 13.

[0036] The storage device 20 stores computer programs and information. For example, the storage device 20 stores a deep learning model 260, work specification information 270, and inspection result information 280.

[0037] The deep learning model 260 is an example of a learning model that takes an image as an input and outputs the type of defect. The deep learning model 260 is typically a neural network. The deep learning model 260 is used for learning and inference by a model-based processing unit 230 described later.

[0038] The work specification information 270 represents information on the work specifications for each customer. A "customer" is the recipient of the honeycomb structure 1. For each customer, the "work specification" is the specification of the honeycomb structure 1 and includes conditions regarded as defects for each predetermined type of defect.

[0039] The inspection result information 280 represents information on the inspection results for each honeycomb structure 1. The inspection results may include the presence or absence of defects, the types of detected defects, and the inspected images in which the defects appear. The inspection results may also include result details such as the position of the defect (for example, coordinates when a predetermined position on the first end face 2a is used as a reference coordinate (origin)).

[0040] When the processor 30 executes the computer program stored in the storage device 20, functions such as the rule-based processing unit 220, the model-based processing unit 230, the display control unit 240, and the model management unit 250 are realized. Also, a control unit (not shown) that controls various devices such as the stage 14 and the imaging device 13 may be realized. At least some of the rules applied by the rule-based processing unit 220 may be rules based on the work specifications represented by the work specification information 270.

[0041] FIG. 4 shows an overview of the process of end face inspection. In the present embodiment, predetermined types of defects are roughly classified into "cracks" and "cell deformation". A "crack" is a break in the partition walls 4 (ribs) and outer walls in the longitudinal direction 9 of the honeycomb structure 1. "Cell deformation" is deformation of the cells as flow paths. Therefore, in the present embodiment, as the deep learning model 260, there are a crack model 260A and a cell deformation model 260B. Each of the crack model 260A and the cell deformation model 260B may include one or more deep learning models.

[0042] First, the control device 570 performs rule-based classification (S401). The rule-based classification applied in S401 may be based on the work specifications represented by the work specification information 270. Specifically, for example, the rule-based processing unit 220 performs overall crack determination (S401A) and overall cell deformation determination (S401B).

[0043] S401A (Overall crack determination) is a rule-based crack determination, which is a determination of whether there is a crack range image. The "crack range image" is an image including the range in which a crack is shown. In S401A, the rule-based processing unit 220 determines whether the target image is an image including a range that satisfies the rule representing the condition as a crack, that is, a crack range image. The "target image" in this paragraph may be at least a part of the two-dimensional captured image (the image represented by the processing target image data from the imaging device 130), or may be at least a part of the two-dimensional captured image subjected to rule-based image processing for crack determination (for example, image processing for increasing the detection probability of a crack). Also, in S401, the rule-based processing unit 220 may specify a crack range image based on the result of comparison between a high-angle illumination image and a low-angle illumination image based on the technique disclosed in Patent Document 1.

[0044] S401B (Overall cell deformation determination) is a rule-based cell deformation determination, which is a determination of whether there is a cell deformation range image. The "cell deformation range image" is an image including the range in which a cell deformation is shown. In S401B, the rule-based processing unit 220 determines whether the target image is an image including a range that satisfies the rule representing the condition as a cell deformation, that is, a cell deformation range image. The "target image" in this paragraph may be at least a part of the two-dimensional captured image, or may be at least a part of the two-dimensional captured image subjected to rule-based image processing for cell deformation determination (for example, image processing for increasing the detection probability of a cell deformation).

[0045] The crack range image and the cell deformation range image can be collectively referred to as the "image to be inspected". The image to be inspected may be an image including both the crack range and the cell deformation range.

[0046] Next, the control device 570 performs model-based classification (S402). Specifically, the model-based processing unit 230 performs crack individual determination (S402A) and cell deformation individual determination (S402B).

[0047] S402A (Crack Individual Judgment) is a model-based crack judgment. In S402A, the model-based processing unit 230 extracts each fragment image from the test image (crack range image). The fragment image may typically be a square image, for example, an image of 200 pixels in length and width. Also, the fragment images may be extracted in a predetermined order (for example, sequentially from the beginning to the end of the test image). For example, the fragment image extracted from the crack range image is the fragment image 411 illustrated in FIG. 5 (in FIG. 5, reference numeral 500 represents a high-angle illumination image, and reference numeral 410 represents a crack). In S402A, the model-based processing unit 230 infers the type of defect shown in the fragment image. Specifically, the model-based processing unit 230 inputs the extracted fragment image into the crack model 260A, and obtains the judgment result (judged defect type (class)) and the reliability (reliability of the judgment) for the fragment image from the crack model 260A. In S402A, the extraction of the fragment image may be performed for one or both of the test image with high-angle illumination and the test image with low-angle illumination, or may be performed for one test image including the component of high-angle illumination and the component of low-angle illumination. When fragment images are extracted from each of the test image with high-angle illumination and the test image with low-angle illumination, two fragment images extracted from the same position of the two test images are input into the crack model 260A, and one judgment result may be output for the two fragment images (for example, based on the result of comparison of the two fragment images).

[0048] S402B (Cell deformation individual determination) is a model-based cell deformation determination. In S402B, the model-based processing unit 230 extracts each fragment image from the test image (cell deformation range image). Here too, the fragment image may typically be a square image, for example, an image of 200 pixels in both vertical and horizontal directions. For example, the fragment image extracted from the cell deformation range image is the fragment image 412 illustrated in FIG. 5 (in FIG. 5, reference numeral 420 represents the cell deformation range image). In S402B, the model-based processing unit 230 infers the type of defect shown in the fragment image. Specifically, the model-based processing unit 230 inputs the extracted fragment image into the cell deformation model 260B, and obtains the determination result (determined defect type (class)) and the confidence level (confidence level of the determination) for the fragment image from the crack model 260A. In S402B, the extraction of the fragment image may be performed for one or both of the test image with high-angle illumination and the test image with low-angle illumination, or may be performed for one test image including the component of high-angle illumination and the component of low-angle illumination. However, in the present embodiment, as will be described later, it is performed for the test image with high-angle illumination.

[0049] In at least one of S402A and S402B, a part of the fragment image may overlap with a part of another fragment image, or each fragment image may be adjacent to another fragment image without overlap. Also, in at least one of S402A and S402B, the model-based processing unit 230 may rotate the test image and extract a plurality of fragment images from each of the test images at a plurality of rotation angles (for example, the test image at 0° (unrotated), the test image after 90° rotation,...), and input each of the plurality of fragment images at different rotation angles into the deep learning model 260.

[0050] In this way, for each of the fragment images extracted from the test image, the type of defect is determined by inputting the fragment image into the deep learning model 260. Note that the deep learning model 260 has been trained using the teacher data 700 for each type of defect, as shown in FIG. 6. The teacher data 700 may include a plurality of fragment images belonging to the type of defect (or non-defect type) corresponding to the teacher data 700. Further, the teacher data 700 may include a plurality of fragment images with different rotation angles. Also, the sizes of the fragment images may be uniform, or may vary depending on whether a given type of defect is "crack" or "cell deformation".

[0051] Referring again to FIG. 4. Also, in the following description, a fragment image classified as "ABC" may be expressed as "fragment image 'ABC'".

[0052] After the model-based classification, the control device 570 performs rule-based determination (S403). Specifically, the model-based processing unit 230 performs crack final determination (S403A), cell deformation final determination (S403B), and comprehensive determination (S403C). S403A (crack final determination) is a rule-based determination based on the determination result in S402A (the determination result for each extracted fragment image). S403B (cell deformation final determination) is a rule-based determination based on the determination result in S402B (the determination result for each extracted fragment image). The comprehensive determination (S403C) is a determination based on the determination results of S403A and S403B. For example, the comprehensive determination (S403C) may be a determination of whether the final result of the end face inspection is OK or NG based on the number of defective cells. A "defective cell" is a cell belonging to the fragment image "NG". For example, when one cell belongs to a plurality of fragment images, if a predetermined ratio or more of the plurality of fragment images are fragment images "NG", the cell may be regarded as a "defective cell". One cell may be entirely shown in one fragment image, or one cell may be entirely shown in a plurality of consecutive fragment images (for example, overlapping fragment images) (that is, only a part of the cell may be shown in each fragment image). The model-based processing unit 230 may create information representing the association between the fragment image and the cell shown in the fragment image (for example, information representing the relationship between the serial number of the fragment image and the serial number of the cell), and store the information in the storage device 20 (for example, it may be included in the inspection result information 280). Note that the determination in S403C may be, for example, a determination based on the work specification represented by the work specification information 270. Also, S403A may not be necessary.

[0053] The work ID of the honeycomb structure 1 may be input to the inspection device 450 together with the two-dimensional photographed image data. Information representing the results of S401 to S403 may be stored in the inspection result information 280 and associated with the work ID.

[0054] The display control unit 240 displays an inspection result screen on the display device 540 based on the inspection result information 280 (S404). On the inspection result screen, the inspection result represented by the inspection result information 280 and the work ID may be displayed.

[0055] FIG. 7 schematically shows S402A (crack individual determination).

[0056] In S402A (crack individual determination), the model-based processing unit 230 extracts a plurality of fragment images from the test image, and classifies each fragment image by inputting each fragment image into the crack model 260A. The classification destinations are roughly divided into two: "NG" (defect type "crack") and "OK" (defect type "non-crack").

[0057] In FIG. 7, each image displayed in the balloon associated with the frame of the crack model 260A is an example of a fragment image. In FIG. 7, each image shown within the frame of the crack model 260A is an image representing examples of "NG" or "OK" over a wider range than the fragment image.

[0058] "NG" belongs to three defect types: "first crack" (cut of the end face), "second crack" (absence of partition 4), and "third crack" (cut of partition 4). Being classified into any of these three defect types means being classified as "crack" as a predetermined defect type. For each of the three defect types, teacher data 700 is prepared, and the crack model 260A may be trained using those teacher data. Note that the defect types belonging to "NG" may not be limited to the defect types illustrated in FIG. 7. For example, "fourth crack" may adopt a cut at the intersection of cells.

[0059] "OK" belongs to three non-defect types: "tipping", "fiber" (adhesion of fiber), and "foreign matter" (adhesion of foreign matter other than fiber). Being classified into any of these three non-defect types means not being classified as "crack" as a predetermined defect type. For each of the three non-defect types, teacher data 700 is prepared, and the crack model 260A may be trained using those teacher data. Note that the non-defect types belonging to "OK" may not be limited to the non-defect types illustrated in FIG. 7.

[0060] Note that in S403A (final crack determination) after this S402A, for example, the following may be performed. That is, when a predetermined number of fragment images "OK" are sandwiched between fragment images "NG", the model-based processing unit 230 may change each of the predetermined number of fragment images "OK" to a fragment image "NG" (the classification destination of the fragment image (the label associated with the fragment image) may be changed from "OK" to "NG"). Alternatively, when there are a predetermined number of fragment images "NG", the model-based processing unit 230 may determine that there is a crack on the first end face 2a.

[0061] FIG. 8 schematically shows S402B (individual cell deformation determination).

[0062] In S402B (individual cell deformation determination), the model-based processing unit 230 extracts a plurality of fragment images from the test image and classifies each fragment image by inputting each fragment image into the cell deformation model 260B. As classification destinations, in addition to "NG" (defect type "cell deformation") and "OK" (defect type "non-cell deformation"), there is "deformation candidate". That is, as classification destinations other than "OK", in addition to "NG", there is "deformation candidate".

[0063] In FIG. 8, each image displayed in the balloon associated with the frame of the cell deformation model 260B is an example of a fragment image. In FIG. 8, each image shown within the frame of the cell deformation model 260B is an image representing examples of "NG", "OK", or "deformation candidate" over a wider range than the fragment image.

[0064] "NG" includes three types of defect categories: "First Deformation" (deformation at the back of the cell) belonging to "NG1", "Second Deformation" (cross-sectional deformation A of the entire cell) and "Third Deformation" (cross-sectional deformation B of the entire cell) belonging to "NG2". Being classified into any of these three defect categories means being classified as "Cell Deformation" as a predetermined defect category. For each of the three defect categories, teacher data 700 is prepared, and using these teacher data, the learning of the cell deformation model 260B may be performed. Note that the defect categories belonging to "NG" may not be limited to the defect categories illustrated in FIG. 8.

[0065] "OK" includes two non-defect categories: "Chipping" and "Fiber" (adhesion of fibers). Being classified into any of these two non-defect categories means not being classified as "Cell Deformation" as a predetermined defect category. For each of the two non-defect categories, teacher data 700 is prepared, and using these teacher data, the learning of the cell deformation model 260B may be performed. Note that the non-defect categories belonging to "OK" may not be limited to the non-defect categories illustrated in FIG. 8.

[0066] "Deformation candidate" means a candidate corresponding to "NG". The determination of whether the "deformation candidate" becomes "NG" is made in S403B (Final cell deformation determination). For the "deformation candidate", teacher data 700 is prepared, and using this teacher data, the learning of the cell deformation model 260B may be performed.

[0067] FIG. 9 shows the details of S401B (Overall cell deformation determination), S402B (Individual cell deformation determination), and S403B (Final cell deformation determination).

[0068] In S401B, the rule-based processing unit 220 classifies all cells in the high-angle illumination image based on rules. Specifically, for example, in S401B, the rule-based processing unit 220 determines, for each of all the cells shown in the high-angle illumination image, whether the diameter of the inscribed circle of the cell opening is less than a first length (for example, the sum of the standard pin gauge size and the first margin). A cell for which this determination result is true is a good cell, and a cell for which this determination result is false is a defective candidate cell. An image in which one or more defective candidate cells are shown in the high-angle illumination image is the test image obtained in S401B.

[0069] In S402B, the model-based processing unit 230 extracts a plurality of fragment images from the test image, and classifies each fragment image by inputting each fragment image into the cell deformation model 260B (S901).

[0070] In S403B, the rule-based processing unit 220 refers to the classification results for each fragment image (S910), and determines whether each cell is a good cell or a defective cell.

[0071] Specifically, the rule-based processing unit 220 determines that the cells belonging to the fragment image "OK" are good cells (S911). The rule-based processing unit 220 determines that the fragment image "First deformation" is a defective cell (S912).

[0072] The rule-based processing unit 220 classifies, based on rules, the final candidate cells that are the cells belonging to the fragment image "NG2" ("Second deformation" or "Third deformation") or "Deformation candidate" (S913). Specifically, for example, in S913, the rule-based processing unit 220 determines, for the final candidate cells shown in the low-angle illumination image, whether the diameter of the inscribed circle of the cell opening is less than a second length (for example, the sum of the standard pin gauge size and the second margin). A cell for which this determination result is true is a good cell, and a cell for which this determination result is false is a defective cell. Note that the second margin may be shorter than the first margin. Also, classifications that require S913 (classification based on rules), such as "Second deformation" and "Third deformation", may be included in "Deformation candidate".

[0073] In S403C (overall determination), the rule-based processing unit 220 may determine whether the final result of the end face inspection is OK or NG based on the number of defective cells.

[0074] In at least one of S401B and S913, instead of determining whether the diameter of the inscribed circle of the cell opening is less than a predetermined size, the rule-based determination may be another determination, for example, determining whether the area of the cell opening is less than a predetermined size. The rule serving as the basis for the determination may be based on the work specification represented by the work specification information 270.

[0075] Although one embodiment has been described above, this is an exemplification for explaining the present invention, and is not intended to limit the scope of the present invention only to this embodiment. The present invention can be implemented in various other forms.

[0076] The above embodiments can be summarized as follows, for example. The following summary may include supplementary explanations and explanations of modification examples of the above description.

[0077] The manufacturing method of the honeycomb structure 1 includes a model-based classification step (for example, S402) and a rule-based determination step (for example, S403) (for example, the inspection apparatus 450 includes a model-based processing unit 230 and a rule-based processing unit 220). The model-based classification step (for example, the model-based processing unit 230) extracts a plurality of fragment images from the image to be inspected, and inputs each of the plurality of fragment images to a learning model that takes the image as an input and outputs a type, thereby classifying each of the plurality of fragment images. The honeycomb structure 1 has a porous partition wall 4 that partitions and forms a plurality of cells 5 extending from the first end face 2a to the second end face 2b. The image to be inspected is an image in which at least a part of the first end face 2a of the honeycomb structure 1 is shown. The rule-based determination step (for example, the rule-based processing unit 220) determines whether there is a defect in the first end face 2a based on the number of defective cells (cells belonging to the fragment images classified into a predetermined defect type).

[0078] Since the image input to the learning model is a fragmented image, a learning model with high determination accuracy can be prepared, and thus, over-detection of defects of a predetermined defect type regarding the end face of the honeycomb structure 1 can be reduced.

[0079] The rule-based determination step (for example, the rule-based processing unit 220) performs a rule-based determination as to whether or not a final candidate cell (a cell belonging to a fragmented image classified as a candidate of a predetermined defect type) satisfies a predetermined rule, and the final candidate cell for which the result of the rule-based determination is true may be regarded as a defective cell. Thereby, it can be expected that only by the model-based processing, a cell that is difficult to determine whether it is of a predetermined defect type or not can be accurately distinguished as a non-defective cell or a defective cell, and thus, an improvement in inspection accuracy can be expected.

[0080] The image to be inspected may be at least a part of the image for processing. The image for processing may be a high-angle illumination image out of a high-angle illumination image of the first end face 2a and a low-angle illumination image of the first end face 2a. The high-angle illumination image may be an image represented by the captured image data of the first end face 2a irradiated with light whose angle formed with respect to the first end face 2a is a high angle. The low-angle illumination image may be an image represented by the captured image data of the first end face 2a irradiated with light whose angle formed with respect to the first end face 2a is a low angle. The predetermined defect type may be cell deformation. The learning model may be a cell deformation model 260B (a deep learning model for cell deformation). The predetermined rule applied to the final candidate cell may be a rule regarding the cell opening of the final candidate cell shown in the low-angle illumination image.

[0081] In this way, by extracting the fragmented image from the high-angle illumination image and setting the predetermined rule applied to the final candidate cell as a rule regarding the cell opening of the cell shown in the low-angle illumination image, it can be expected to efficiently and accurately detect defective cells. Specifically, for example, it is as follows.

[0082] That is, regarding cell deformation, there are the above-described first deformation to third deformation and deformation candidates. Among them, some of them, specifically, the first deformation is shown in FIG. 10A, the second deformation is shown in FIG. 10B, and the deformation candidates are shown in FIG. 10C. In each of FIGS. 10A to 10C, the left-side figure on the paper surface is a schematic diagram of the cell opening (the opening of cell 5) in the high-angle illumination image and a schematic diagram of the cell opening in the low-angle illumination image, and the right-side figure on the paper surface is a schematic diagram of the cell cross-section along the longitudinal direction of the honeycomb structure 1. Also, in the schematic diagram of the cell cross-section, there are two types of broken lines. The upper broken line corresponds to the broken line of the inscribed circle described in the schematic diagram of the cell opening in the low-angle illumination image, and the lower broken line corresponds to the broken line of the inscribed circle described in the schematic diagram of the cell opening in the high-angle illumination image. That is, according to FIGS. 10A to 10C, according to the high-angle illumination image, the inscribed circle of the cell opening depends on the shape of the flow path cross-section (the cross-section along the direction orthogonal to the longitudinal direction of the honeycomb structure 1) deeper (inside) than the cell opening. This is because, due to the high-angle illumination, it can capture up to the inside of the cell opening. On the other hand, according to the low-angle illumination image, the inscribed circle of the cell opening depends on the shape of the cell opening.

[0083] According to FIG. 10A, in the first deformation and the deformation candidates, the inscribed circle of the cross-section inside the cell is smaller than the inscribed circle of the cell opening. On the other hand, according to FIG. 10B, in the second deformation, the inscribed circle of the cross-section inside the cell is substantially the same size as the inscribed circle of the cell opening.

[0084] In view of such optical reasons, as described above, by extracting a fragment image from a high-angle illumination image and inputting the extracted fragment image into the cell deformation model 260B, the fragment image belonging to the cell 5 having the first deformation can be accurately classified as the "first deformation". For cells having a deformation in which the inscribed circle of the cross-section inside the cell is substantially the same as the inscribed circle of the cell opening, it is possible to determine whether it is a good cell or a defective cell in the subsequent rule-based processing. Note that when extracting a fragment image from low-angle illumination, it may be difficult to accurately classify the fragment image belonging to the cell having the first deformation as the "first deformation". According to the high-angle illumination image, in both the first deformation and the deformation candidate, the inscribed circle of the cross-section inside the cell is smaller than the inscribed circle of the cell opening. However, in the first deformation, the partition wall 4 looks thick, and in the deformation candidate, the partition wall 4 is tilted in the same direction. From these differences, in the model-based classification, the fragment image extracted from the high-angle illumination image of the first deformation or the deformation candidate is accurately classified as the first deformation or the deformation candidate.

[0085] Note that, as described above, since the fragment image for cell deformation is extracted from the high-angle illumination image, the cell deformation model 260B may be learned using the training data 700 including a plurality of fragment images of the high-angle illumination image.

[0086] The rule-based classification step (for example, S401, in other words, for example, the rule-based processing unit 220) may perform another rule-based determination that is a determination as to whether each cell shown in the high-angle illumination image satisfies a predetermined other rule regarding the cell opening. The test image may be a part of the processing image and an image in which a defective candidate cell, which is a cell for which the result of another rule-based determination is true, is shown. Thereby, the extraction basis of the fragment image can be narrowed down to an appropriate range, and thus it is expected to reduce the time required for the end face inspection.

[0087] Another predetermined rule (the rule used in the previous rule-based process) may be that in the high-angle illumination image, the diameter of the inscribed circle of the cell opening is less than a first diameter size, or the area of the cell opening is less than a first area. A predetermined rule (the rule used in the subsequent rule-based process) may be that in the low-angle illumination image, the diameter of the inscribed circle of the cell opening is less than a second diameter size, or the area of the cell opening is less than a second area. Thereby, cells can be accurately classified into good cells and bad cells.

[0088] The predetermined defect type may be a crack instead of or in addition to cell deformation. The learning model may be a crack model 260A (a deep learning model for cracks). Thereby, cracks can be accurately detected instead of or in place of cell deformation. That is, the inspection device may be a device that detects cell deformation, a device that detects cracks, or a device that detects both cell deformation and cracks. Note that the rule-based processing unit 220 may obtain, as a test image (a part of the processing image) as an image of a range where a crack appears, by performing rule-based processing on the processing image. Also, for cracks, the processing image may be one or both of a high-angle illumination image of the first end face and a low-angle illumination image of the first end face.

[0089] Also, as one method for reducing over-detection, instead of using the deep learning model 260, adoption of a filtering process with a high computational load such as performing processing using a bilateral filter on the test image is conceivable. However, since the computational load of the process is high, a long time is required for inspection. In the above-described embodiment, by adopting the process using the deep learning model 260, even if some image processing is performed in the rule-based process before that process, the image processing may be a process with a low computational load. As a result, it can be expected to shorten the inspection time while ensuring inspection accuracy.

[0090] Instead of the deep learning model 260, other types of learning models, such as decision trees, may be adopted. However, it is difficult to perform inferences with input images with high accuracy using other types of learning models. The deep learning model 260 is suitable for inferences with input images, and high-accuracy inferences are expected.

[0091] In addition, the deep learning model 260 is a so-called black box type model. That is, even if the type of the input fragment image is determined (output), there is no output of the reason for the determination. In other words, it has no explainability. Therefore, if the image input to the deep learning model 260 is a test image, it is not possible to provide explainability to the determination result for the test image. In the above-described embodiment, the type is determined by the model-based processing unit 230 for each fragment image, and then, for the final candidate cell belonging to the fragment images classified into the deformation candidates, it is determined by the rule-based processing unit 220 whether it is a good cell or a defective cell. For this reason, both high accuracy regarding inferences with input images and providing explainability to the determination result for the test image (for example, an explanation that it was determined to be a good cell (or a defective cell) because the inscribed circle satisfied (or did not satisfy) the conditions in the low-angle illumination image) can be realized. That is, the inspection device 450 may further include a display control unit 240 that displays the inspection result based on the inspection result information 280 including information representing the result of determining whether there is a defect in the first end face 2a. The inspection result may include the result of determining whether there is a defect of a predetermined defect type in the first end face 2a and the reason for the determination result (the reason including whether the condition for having a defect of a predetermined defect type is satisfied). The information representing the reason for the determination result may include information representing the number of defective cells and information representing the reason for determining each defective cell as a defective cell. The display control unit 240 may display the inspection result on the display device 540 included in the inspection device 450, or may display it on a remote computer connected to the inspection device 450 (for example, a server).

[0092] In addition, in the present embodiment, when the amount of teacher data including the fragment image and the type corresponding to the fragment image is less than a certain amount for each fragment image, in the teacher data, for each fragment image, the fragment image is classified into any one of two or more detailed types belonging to the predetermined defect type or any one of two or more detailed types belonging to the non-defect type, and the model management unit 250 may use the teacher data to train the deep learning model 260. This may be performed, for example, for each of the models 260A and 260B. The process performed by the model management unit 250 may be referred to as the process performed in the model management step.

[0093] Specifically, for example, for each of the models 260A and 260B, the process shown in FIG. 11 may be performed. Taking one deep learning model as an example. The model management unit 250 determines whether the amount of teacher data of the deep learning model 260 is sufficient (equal to or greater than the threshold value), that is, whether the amount of teacher data is equal to or greater than a certain amount (S1101).

[0094] If the determination result of S1101 is true (S1101: YES), the model management unit 250 performs learning of less classification (S1102). "Learning of less classification" means that in the teacher data used for learning, for both the predetermined defect type and the non-defect type (types other than the predetermined defect type), the types prepared are either the predetermined defect type itself or the non-defect type itself, or a small number of types.

[0095] On the other hand, when the determination result of S1101 is false (S1101: NO), the model management unit 250 performs multi-classification learning (S1103). "Multi-classification learning" means that, in the teacher data used for learning, for both a predetermined defect type and a non-defect type (types other than the predetermined defect type), the number of prepared types is larger than that of the teacher data used for binary classification learning. That is, in the teacher data for multi-classification learning, for each fragment image in the teacher data, the fragment image is classified into any one of two or more detailed types belonging to the predetermined defect type or any one of two or more detailed types belonging to the non-defect type. When the amount of teacher data is insufficient, for both the predetermined defect type and the non-defect type, by classifying the fragment images into more detailed types, it is expected that the deep learning model will be a highly accurate model considering the amount of teacher data.

[0096] In the description of the above embodiment, the inspection in the manufacturing method of the honeycomb structure 1 is mainly described. However, the manufacturing method of the honeycomb structure 1 includes steps other than the steps belonging to the inspection. Specifically, for example, the manufacturing method may include a step of mixing ceramic raw materials with a sintering aid or the like to prepare a green soil, a step of extruding the green soil with a die, and a step of firing the extruded green soil. In the honeycomb structure produced by these steps, defects may appear on its end face in the extrusion molding process or the firing process or the like. The manufacturing method may include an inspection for the presence or absence of such defects. The end face of the honeycomb structure for which the final result is OK in the above-described inspection has a number of defective cells less than a predetermined number.

Explanation of Reference Numerals

[0097] 450…Inspection device

Claims

1. In a method for manufacturing a honeycomb structure having a porous partition wall that defines a plurality of cells extending from a first end face to a second end face, a model-based classification step of classifying each of the plurality of fragment images by inputting each of the plurality of fragment images into a learning model that takes an image as an input and outputs a type, by extracting the plurality of fragment images from a test image in which at least a part of the first end face of the honeycomb structure is shown; a rule-based determination step of determining whether there is a defect in the first end face based on the number of defective cells that are cells belonging to the fragment images classified into a predetermined defect type; A manufacturing method having the above.

2. The rule-based determination step includes a rule-based determination step of determining whether a final candidate cell, which is a cell belonging to a fragment image classified into a candidate of a predetermined defect type, satisfies a predetermined rule; a step of regarding the final candidate cell for which the result of the rule-based determination is true as a defective cell; Including The manufacturing method according to claim 1.

3. The test image is at least a part of a processing image, The processing image is a high-angle illumination image among a high-angle illumination image of the first end face and a low-angle illumination image of the first end face, The high-angle illumination image is an image represented by captured image data of the first end face irradiated with light having a high angle with respect to the first end face, The low-angle illumination image is an image represented by captured image data of the first end face irradiated with light having a low angle with respect to the first end face, The predetermined defect type is cell deformation, The learning model is a deep learning model for cell deformation, The predetermined rule is a rule regarding the cell opening of the cells shown in the low-angle illumination image. The manufacturing method according to claim 2.

4. Further having a rule-based classification step of performing another rule-based determination, which is a determination as to whether each cell shown in the high-angle illumination image satisfies another predetermined rule regarding the cell opening, The test image is a part of the processing image, and is an image in which defective candidate cells, which are cells for which the result of the another rule-based determination is true, are shown. The manufacturing method according to claim 3.

5. The another predetermined rule is that in the high-angle illumination image, the diameter of the inscribed circle of the cell opening is less than a first diameter size, or the area of the cell opening is less than a first area. The predetermined rule is that in the low-angle illumination image, the diameter of the inscribed circle of the cell opening is less than a second diameter size, or the area of the cell opening is less than a second area. The manufacturing method according to claim 4.

6. The predetermined defect type is a crack. The learning model is a deep learning model for cracks. The manufacturing method according to claim 1.

7. The method further includes a display control step of displaying an inspection result based on inspection result information including information representing a result of determining whether there is a defect on the first end face. The inspection result is a determination result of whether there is a defect of the predetermined defect type on the first end face, and a reason including whether a condition that there is a defect of the predetermined defect type, which is the reason for the determination result, is satisfied. including The manufacturing method according to claim 1.

8. The method further includes a model management step. The learning model is a deep learning model. When the amount of teacher data including the fragment image and the type corresponding to the fragment image is less than a certain amount for each fragment image, in the teacher data, for each fragment image in the teacher data, the fragment image is classified into any one of two or more detailed types belonging to the predetermined defect type or any one of two or more detailed types belonging to the non-defect type. The model management step includes learning the deep learning model using the teacher data. The manufacturing method according to claim 1.

9. Extracting a plurality of fragment images from a test image in which at least a part of the first end face of a honeycomb structure having a porous partition wall that partitions and forms a plurality of cells extending from the first end face to the second end face is shown, and inputting each of the plurality of fragment images to a learning model that takes an image as an input and outputs a type, thereby classifying each of the plurality of fragment images; a model-based processing unit for classifying the fragment image; a rule-based processing unit that determines whether there is a defect on the first end face based on the number of defective cells that are cells belonging to the fragment images classified into a predetermined defect type. An inspection apparatus comprising.

10. Extracting a plurality of fragment images from a test image in which at least a part of the first end face of a honeycomb structure having a porous partition wall that partitions and forms a plurality of cells extending from the first end face to the second end face is shown, and inputting each of the plurality of fragment images to a learning model that takes an image as an input and outputs a type, thereby classifying each of the plurality of fragment images. A computer program that causes a computer to determine whether there is a defect on the first end face based on the number of defective cells that are cells belonging to a fragment image classified into a predetermined defect type. A computer program that causes a computer to execute this. **Claim 11** Extract a plurality of fragment images from a test image in which at least a part of the first end face of a honeycomb structure having a porous partition wall that partitions a plurality of cells extending from the first end face to the second end face is shown, and input the images into a learning model that takes an image as input and outputs a type. By inputting each of the plurality of fragment images, each of the plurality of fragment images is classified, A computer program that causes a computer to determine whether there is a defect on the first end face based on the number of defective cells that are cells belonging to a fragment image classified into a predetermined defect type. A recording medium that records a computer program that causes a computer to execute this.

Images