Processed surface state estimation device, estimation program and processed surface state estimation method

Abstract

To shorten and reduce time and labor required for work for setting a learned learning device.SOLUTION: A processed surface state estimation device 30 includes a storage part 56, an acquisition part, a specification part 58, and an output part. The storage part 56 stores a plurality of classification learning devices 62 defined to each of a plurality of kinds of grinding wheels, and a learned specific learning device 67 learning to specify one of the plurality of classification learning devices 62. The acquisition part acquires a photographic image 44 which is an image related to the abrasive plane of the grinding wheel as an object to be estimated. The specification part 58 utilizes the photographic image 44 and the specific learning device 67, and specifies one of the plurality of classification learning devices 62. The output part utilizes the photographic image 44, and the classification learning device 62 specified by the specification part 58, and outputs a classification result of a determination image 65 as processed surface information indicating the state of the abrasive plane of the grinding wheel as the object to be estimated.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a machined surface condition estimating device that estimates the state of a machined surface of a tool, an estimation program used in the device, and a machined surface condition estimating method that estimates the state of a machined surface of a tool. [Background technology]

[0002] In grinding processing using a grinding device, the finish quality of the machined surface of a workpiece depends on the condition of the machining surface (so-called grinding surface) of a grinding wheel serving as a grinding tool. Conventionally, in order to grasp the condition of the grinding surface of a grinding wheel, a measuring device has been proposed for measuring the condition of the grinding surface (specifically, the amount of protrusion of abrasive grains) (for example, Patent Document 1). In the device described in Patent Document 1, the grinding surface is imaged using a metallurgical microscope equipped with a camera, and the condition of the grinding surface is measured based on the image data. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2004-45078 A Summary of the Invention [Problem to be solved by the invention]

[0004] In a device using an electron microscope, the surface properties of the grinding surface can be observed in detail using the electron microscope. However, since the observation range of the electron microscope is very narrow, it takes a long time to measure the condition of the grinding surface over the entire surface using the electron microscope. Moreover, if the condition of the grinding surface is estimated for each observation range of the electron microscope, i.e., each measurement range of the measuring device, the estimation also takes a long time.

[0005] This situation is generally the same when estimating the condition of the machined surface of a saw wire, which is the tool of a wire saw, or when estimating the condition of the machined surface of a wire electrode, which is the tool of a wire electric discharge machine. [Means for solving the problem]

[0006] A machining surface condition estimation device for solving the above problem is a machining surface condition estimation device that estimates the state of a machining surface of a tool, and includes a first learning device memory unit that stores a plurality of first learning devices that are trained to use an image of the machining surface of the tool as first learning data and to output machining surface information indicating the state of the machining surface of the tool, and that are respectively determined for a plurality of types of tools; a second learning device memory unit that stores a trained second learning device that uses an image of the machining surface of the tool as second learning data and has been trained to identify one of the plurality of first learning devices; an acquisition unit that acquires an estimated image that is an image of the machining surface of the tool to be estimated; an identification unit that uses the estimated image and the second learning device to identify one of the plurality of first learning devices; and an output unit that uses the estimated image and the first learning device identified by the identification unit to output the machining surface information indicating the state of the machining surface of the tool to be estimated.

[0007] The estimation program for solving the above problem is an estimation program that causes an electronic control device provided in the machining surface condition estimation device to execute the processing of the first learning device memory unit, the processing of the second learning device memory unit, the processing of the acquisition unit, the processing of the identification unit, and the processing of the output unit, which are provided in the machining surface condition estimation device.

[0008] A machining surface condition estimation method for solving the above problem is a machining surface condition estimation method for estimating the condition of a machining surface of a tool, comprising: a first learning device storage process for storing a plurality of first learning devices that are trained to use an image of the machining surface of the tool as first learning data and to output information indicating the condition of the machining surface of the tool, the first learning devices being determined separately for a plurality of types of tools; a second learning device storage process for storing a trained second learning device that is trained to use an image of the machining surface of the tool as second learning data and to identify one of the plurality of first learning devices; an acquisition process for acquiring an estimated image that is an image of the machining surface of the tool to be estimated; an identification process for identifying one of the plurality of first learning devices using the estimated image and the second learning device; and an output process for outputting machining surface information indicating the condition of the machining surface of the tool to be estimated, using the estimated image and the first learning device identified in the identification process.

[0009] According to the above-mentioned machined surface state estimation device, estimation program, and machined surface state estimation method, when estimating the state of the machined surface of a tool, it is possible to identify a first learner suitable for estimating the state of the tool to be estimated, for example, by identifying a first learner corresponding to a tool having a similar basic structure to the tool to be estimated from among a plurality of first learners stored in advance, using an estimated image and a second learner. Then, by using the identified first learner and an estimated image of the tool to be estimated, it is possible to obtain information indicating the state of the machined surface of the tool to be estimated.

[0010] In this way, according to the above-mentioned machined surface condition estimating device, the estimation program, and the machined surface condition estimating method, it is possible to estimate the state of the tool to be estimated by using one of the multiple first learning devices stored in advance, rather than storing a first learning device for each type of tool in advance. This makes it possible to reduce the number of first learning devices stored in advance, thereby shortening the time and reducing the effort required for setting the first learning device. Effect of the Invention

[0011] According to the present invention, it is possible to reduce the time and effort required for setting up a trained learning device. [Brief description of the drawings]

[0012] [Figure 1] 1 is a schematic diagram showing a schematic configuration of an automatic grinding device to which a machined surface state estimating device of a first embodiment is applied. [Diagram 2] 1 is a block diagram showing a schematic configuration of a machined surface state estimating device according to a first embodiment. [Diagram 3] 1 is a schematic diagram showing a structure of a camera unit according to a first embodiment. [Figure 4] 4 is a schematic diagram showing how an imaging target area of ​​the grinding surface is illuminated by an illumination unit according to the first embodiment. [Diagram 5] 1 is a schematic diagram conceptually showing the data structure of learning data of an extraction learning device. [Figure 6] FIG. 13 is a diagram showing an example of learning data corresponding to an evenly abrasive-fell surface; [Figure 7] FIG. 13 is a diagram showing an example of learning data corresponding to an unbalanced abrasive surface; [Figure 8] FIG. 13 is a diagram showing an example of learning data corresponding to a chip welding grinding surface. [Figure 9] 5 is a flowchart showing a procedure for identifying an extraction learning device and a classification learning device according to the first embodiment. [Figure 10] 5 is a flowchart showing the execution procedure of a determination process according to the first embodiment. [Figure 11] FIG. 11 is a block diagram showing a schematic configuration of a machined surface state estimating device according to a second embodiment. [Figure 12] 13 is a flowchart showing a procedure for identifying an extraction learning device and a classification learning device according to the second embodiment. [Figure 13] 10 is a flowchart showing the execution procedure of a determination process according to a second embodiment. [Figure 14] 13 is a flowchart showing an execution procedure of a determination process according to another embodiment. [Figure 15] 13 is a graph showing the relationship between the rotational phase of the grindstone and the integrated value when deterioration of the grindstone progresses. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] (First embodiment) Hereinafter, a first embodiment of a machined surface condition estimating device, an estimating program, and a machined surface condition estimating method will be described with reference to FIGS.

[0014] <Automatic grinding device> As shown in FIG. 1, the machined surface condition estimation device 30 of this embodiment is applied to an automatic grinding device 20. This automatic grinding device 20 is a numerically controlled (NC) type surface grinding device that grinds the surface of a workpiece (not shown) using a grinding wheel 21 as a tool. The grinding wheel 21 has a structure in which a large number of abrasive grains are fixed to the outer peripheral surface of a cylindrical base material by a binder. The grinding wheel 21 is supported by a grinding wheel support part 22 in a rotatable state. The grinding wheel 21 and the grinding wheel support part 22 are movable relative to a base 23 in the vertical direction (hereinafter, Y direction) and the horizontal direction, specifically, the depth direction in FIG. 1 (hereinafter, Z direction). In the automatic grinding device 20, the grinding wheel 21 is brought into contact with the workpiece while being rotationally driven, thereby grinding the surface of the workpiece.

[0015] The automatic grinding device 20 has a table 24 as a support base for supporting a workpiece. The table 24 is movable in a horizontal direction perpendicular to both the Z direction and the Y direction, specifically, in the left-right direction in FIG. 1 (hereinafter, referred to as the X direction). In the automatic grinding device 20, the workpiece on the table 24 can be moved by moving the table 24 in the X direction. This makes it possible to change the relative positions of the grinding wheel 21 and the workpiece in the X direction.

[0016] 1 and 2, the automatic grinding device 20 has an operation control unit 25 that controls the operation of the automatic grinding device 20. The operation control unit 25 has an X-axis drive unit 26, a Y-axis drive unit 27, a Z-axis drive unit 28, and a grindstone drive unit 29.

[0017] The automatic grinding device 20 can change the relative positions of the grindstone 21 and the workpiece in the X direction, the Y direction, and the Z direction, respectively, by moving the grindstone 21 in the Y direction and the Z direction and by moving the table 24 in the X direction. The X-axis drive unit 26 is a device for moving the table 24 in the X direction relative to the base 23. The X-axis drive unit 26 has a ball screw mechanism, a servo motor, and the like, which are not shown. The X-direction position of the table 24 relative to the base 23 is controlled through the operation control of the X-axis drive unit 26. The Y-axis drive unit 27 is a device for moving the grindstone 21 and the grindstone support unit 22 in the Y direction relative to the base 23. The Y-axis drive unit 27 has a ball screw mechanism, a servo motor, and the like, which are not shown. The Y-direction position of the grindstone 21 relative to the base 23 is controlled through the operation control of the Y-axis drive unit 27. The Z-axis drive unit 28 is a device for moving the grindstone 21 and the grindstone support unit 22 in the Z direction relative to the base 23. The Y-axis driving unit 27 has a ball screw mechanism, a servo motor, etc. (not shown). Through the operation control of the Y-axis driving unit 27, the position of the grindstone 21 in the Z direction relative to the base 23 is controlled.

[0018] The grindstone driving unit 29 has an electric motor connected to the rotation shaft of the grindstone 21, a rotation phase sensor that detects the rotation phase of the grindstone 21, and the like. In the operation control of the grindstone driving unit 29 (more specifically, the electric motor), the grindstone 21 is basically rotated at a constant speed. The rotation speed of the grindstone 21 is set based on the processing conditions input to the automatic grinding device 20 and is set to a speed that matches the processing conditions.

[0019] <Machine surface condition estimation device> The machined surface condition estimation device 30 of this embodiment is a device that estimates the state of the grinding surface 211 of the grinding wheel 21 to be estimated, using a captured image 44 obtained by capturing an image of the machined surface (hereinafter, grinding surface 211) of the grinding wheel 21 to be estimated, and trained learning devices 61, 62. Furthermore, the machined surface condition estimation device 30 of this embodiment determines the occurrence of a deterioration abnormality in the grinding wheel 21 to be estimated, based on the estimated state of the grinding surface 211.

[0020] <Camera unit> 1 to 3, the machined surface condition estimating device 30 has a camera unit 40. The camera unit 40 captures an image of the grinding surface 211 of the grindstone 21. The camera unit 40 has a camera 41 as an imaging section, and an illumination section 43.

[0021] <Camera> As shown in FIG. 1 and FIG. 3, the camera 41 is disposed below the grinding wheel 21. The camera 41 captures an image of the grinding surface 211 of the grinding wheel 21 from a direction perpendicular to the grinding surface 211 (specifically, the tangent plane of the grinding surface 211 shown by the dashed line in FIG. 3). In this embodiment, the camera 41 captures the image of the grinding surface 211 at a first time point immediately after the dressing of the grinding wheel 21 is performed. The captured image 44 captured at the first time point is stored in the electronic control device 50, which will be described later, as a basic image 44B. In this embodiment, the camera 41 also captures the image of the grinding surface 211 at a second time point after the first time point, specifically, at any timing during the use of the grinding wheel 21. The captured image 44 captured at the second time point is stored in the electronic control device 50 as a latest image 44N. Among the captured images 44 captured at any timing during use of the grinding wheel 21, previously captured images 44 may be stored as N (N is a natural number) previous images, for example, the captured image 44 captured immediately before may be stored as the previous image. In this embodiment, the grinding wheel 21 attached to the automatic grinding device 20 when capturing an image by the camera 41 corresponds to the tool to be estimated. Also, in this embodiment, the basic image 44B, the latest image 44N, and the Nth previous image correspond to estimated images that are images related to the machining surface of the tool to be estimated.

[0022] <Lighting Department> 3, the illumination unit 43 illuminates the lower end portion of the grinding surface 211, i.e., the area imaged by the camera 41 (hereinafter, image capture target area AR). The illumination unit 43 has a first irradiating unit 45 and a second irradiating unit .

[0023] <1st irradiation section> The first irradiation unit 45 irradiates the above-mentioned imaging target area AR of the grinding surface 211 with a first illumination light 45L having a first emission color (for example, blue) from a direction perpendicular to the grinding surface 211 (specifically, the tangent plane of the grinding surface 211 shown by the dashed line in FIG. 3). Specifically, the first irradiation unit 45 has a first lamp 451 and a reflective translucent member 452. The first lamp 451 is composed of a light-emitting diode. The first lamp 451 is disposed in such a manner that the first illumination light 45L is emitted from a direction perpendicular to the optical axis of the camera 41 toward the optical axis. The reflective translucent member 452 is disposed at a portion where the optical axis of the camera 41 and the optical axis of the first lamp 451 intersect. The reflective translucent member 452 reflects the first illumination light 45L and transmits the reflected light reflected by the grinding surface 211. In this embodiment, the first illumination light 45L emitted from the first lamp 451 is reflected by the reflective light-transmitting member 452 and travels toward the grinding surface 211, thereby irradiating the imaging target area AR on the grinding surface 211. A portion of the reflected light of the first illumination light 45L thus irradiated to the grinding surface 211 is transmitted through the reflective light-transmitting member 452 and enters the camera 41.

[0024] <Second irradiation section> The second irradiating unit 46 irradiates the imaging target area AR of the grinding surface 211 from an oblique direction with respect to the grinding surface 211 (specifically, the tangent plane of the grinding surface 211 shown by the dashed line in FIG. 3) with the second illumination light 46L having a second emission color (for example, white color) different from the first emission color. Specifically, the second irradiating unit 46 has an illumination dome 461 and a second lamp 462. The illumination dome 461 is dome-shaped with an open top, and is disposed so as to cover the lower surface of the grinding wheel 21 (specifically, the imaging target area AR). The second lamp 462 is constituted by a light-emitting diode. A plurality of the second lamps 462 are provided in a line on the inner surface of the illumination dome 461. Each second lamp 462 irradiates the imaging target area AR of the grinding surface 211 with the second illumination light 46L. The second lamp 462 includes a lamp that irradiates the second illumination light 46L onto the imaging target area AR of the grinding surface 211 from a direction along the grinding surface 211 (more specifically, the tangent plane of the grinding surface 211 shown by the dotted line in Figure 3), specifically, from a direction that forms "5 degrees" with respect to the tangent plane of the grinding surface 211.

[0025] <Action by lighting unit> In this embodiment, the camera unit 40 is detachable from the automatic grinding device 20. When estimating the state of the grinding surface 211, the camera unit 40 is attached to the automatic grinding device 20 to obtain a captured image 44 used for the estimation. In the camera unit 40, when the camera 41 captures an image of the grinding surface 211, the illumination unit 43 illuminates an imaging target area AR of the grinding surface 211. Note that in this embodiment, when grinding is performed by the automatic grinding device 20, the camera unit 40 is detached from the automatic grinding device 20.

[0026] 4, in this embodiment, the first illumination light 45L irradiated from the first irradiating unit 45 in a direction perpendicular to the grinding surface 211 can be substantially specularly reflected by the top surface of the grinding wheel 21 (including the top surfaces of the abrasive grains 212) and can be made to enter the camera 41. This allows the color of the portion corresponding to the top surface of the grinding wheel 21 in the captured image 44 captured by the camera 41 to be the first emission color (e.g., blue) of the first illumination light 45L.

[0027] On the other hand, the first illumination light 45L from the first irradiating unit 45 and the second illumination light 46L from the second irradiating unit 46 are diffusely reflected by portions of the grindstone 21 other than the top surface (specifically, the side surfaces of the abrasive grains 212, the surface of the binder, etc.). Then, a portion of the diffusely reflected first illumination light 45L and second illumination light 46L enters the camera 41. As a result, the color of the portions of the grindstone 21 other than the top surface in the captured image 44 captured by the camera 41 becomes a mixture of the first emission color of the first illumination light 45L and the second emission color of the second illumination light 46L.

[0028] According to this embodiment, since the color of the top surface of the grinding wheel 21 in the captured image 44 can be made different from the color of the portion other than the top surface, the boundary between the top surface of the grinding wheel 21 and the portion other than the top surface can be made clear. As a result, an image clearly showing the top surface of the grinding wheel 21 (including the tip portion of the abrasive grains 212), which is a portion that influences the finishing quality of the workpiece, can be obtained as the captured image 44 captured by the camera 41. Therefore, the state of the grinding surface 211 of the grinding wheel 21, specifically the tip portion of the abrasive grains 212, can be estimated with high accuracy using the captured image 44. In addition, the tip portion of the grinding surface 211 (specifically the portion where the abrasive grains 212 function as a cutting edge) can be easily extracted from the captured image 44. Therefore, an extracted image, which is an image in which the tip portion of the grinding surface 211 is extracted from the image of the grinding surface 211, can be easily generated.

[0029] The machined surface condition estimating device 30 of this embodiment includes an electronic control device 50. The electronic control device 50 executes various controls for estimating the condition of the grinding surface 211 of the grinding wheel 21. The electronic control device 50 has a CPU 51 , a ROM 52 , a RAM 53 , an input unit 54 , a display unit 55 , and a storage unit 56 .

[0030] The input unit 54 is an input device such as a mouse, a keyboard, a touch panel, or a button. The input unit 54 is operated when performing various settings of the machined surface state estimation device 30 or when operating the machined surface state estimation device 30. In this embodiment, particle size information 66 (specifically, average particle size) regarding the particle size of the grinding wheel 21 to be estimated is input through the operation of the input unit 54. In this embodiment, the operation of the input unit 54 for inputting the particle size information 66 is performed when estimating the grinding wheel 21 (for example, when the grinding wheel 21 to be estimated is attached to the automatic grinding device 20 or when the camera unit 40 is attached to the automatic grinding device 20). In this embodiment, the time when dressing was performed on the grinding wheel 21 to be estimated (dressing information) is input through the operation of the input unit 54.

[0031] The display unit 55 has a display device such as a liquid crystal display, an organic EL display, etc. The display unit 55 displays various information related to the judgment, such as the judgment result of the state of the grinding surface 211. The storage unit 56 is configured with a non-volatile memory that can be written to and read from at any time, such as a hard disk drive or a solid state drive. The storage unit 56 stores various execution programs 60, an extraction learner 61, a classification learner 62, and an identification learner 67, all of which have been trained, and learning data 63, 64, and 68 used for machine learning of the learners 61, 62, and 67. The storage unit 56 also stores the captured image 44 captured by the camera unit 40, a judgment image 65 used to estimate the state of the grinding surface 211, particle size information 66 and dressing information input through the operation of the input unit 54.

[0032] The execution program 60 includes a program for executing machine learning by the learning devices 61, 62, and 67, an estimation program for executing various processes related to estimating the state of the grinding surface 211, and a judgment program for executing various processes related to judging the state of the grinding surface 211.

[0033] <Extractive learning machine> The extraction learning device 61 will now be described. The extraction learning device 61 is a learning device used to generate an extracted image, which will be described later. In this embodiment, a Generative Adversarial Network (GAN), more specifically, an image generation model called "pix2pix", is used as the extraction learning device 61.

[0034] The extraction learning device 61 is machine-trained using learning data 63. The learning data 63 is data for performing machine learning on the extraction learning device 61 so as to acquire the ability to generate an extracted image, which is an image obtained by extracting only the tip portion of the grinding surface 211 of the grinding wheel 21 from an image related to the grinding surface 211 of the grinding wheel 21. As shown in FIG. 5, the learning data 63 is composed of an image set in which a pre-conversion image 63B and a post-conversion image 63A form a pair. The pre-conversion image 63B is an image that is expected to be used as input data for the extraction learning device 61. The post-conversion image 63A is an image that is expected to become output data for the extraction learning device 61.

[0035] In this embodiment, the extraction learning device 61 is set for each type of grindstone 21. In this embodiment, the types of grindstone 21 include general grindstones having general abrasive grains and superabrasive grindstones having superabrasive grains, and are divided by manufacturer and model. Specifically, grindstones 21 of different manufacturers or models are treated as different types of grindstones 21. In this embodiment, one extraction learning device 61 corresponding to abrasive grains of a predetermined particle size (specifically, the average particle size of abrasive grains 212 corresponding to grain size #200) is determined for each type of grindstone 21. In this embodiment, in order to realize such a configuration, grindstones 21 having abrasive grains 212 of a predetermined particle size are adopted as grindstones 21 used to create the learning data 63 from among the multiple types of grindstones 21.

[0036] The learning data 63 is generated for each type of grindstone 21 in the following procedure. First, a large number of captured images of the grinding surface 211 of the grinding wheel 21 are prepared. The captured images include a first image captured at the first time and a second image captured at the second time. The first image is an image corresponding to the basic image 44B, and the second image is an image corresponding to the latest image 44N. The first image and the second image are acquired, for example, by using the camera unit 40. The first image and the second image are converted to grayscale and divided into a plurality of divided images in a manner that cuts the grinding surface 211 in a lattice shape in the circumferential direction and the rotation axis direction of the grinding wheel 21. In this embodiment, these divided images are used as the pre-conversion image 63B.

[0037] Then, for each divided image, an extracted image is created by extracting only the tip portion of the grinding surface 211 (more specifically, the portion of the grinding surface 211 that functions as a cutting edge) from the divided image. In this embodiment, these extracted images are used as the converted image 63A. The following method can be used to create the extracted image. First, the grinding surface 211 of the grinding wheel 21 before imaging is placed against an ink pad to adhere ink to the tip portion of the grinding surface 211. Then, the grinding surface 211 of the grinding wheel 21 is imaged. Then, only the colored portion to which ink is adhered is extracted from this captured image (more specifically, the divided image) to create the extracted image. According to this method, the boundary between the tip portion of the grinding surface 211 and the other portions in the captured image becomes clear, so that the tip portion of the grinding surface 211 can be easily extracted from the divided image. Therefore, the extracted image can be easily created.

[0038] Thereafter, as shown in Fig. 5, a large number of image sets are created, each pair consisting of a pre-conversion image 63B and a post-conversion image 63A of the same location on the grinding surface 211. In this embodiment, these image sets are used as learning data 63 for the extraction learner 61. In detail, the electronic control device 50 (Fig. 2) extracts one of the multiple image sets constituting the learning data 63, and performs machine learning on the extraction learner 61 using the image set as input data. This learning process is repeatedly executed for each image set constituting the learning data 63. Then, the machine learning model constructed through the execution of the machine learning is stored in the storage unit 56 as the trained extraction learner 61.

[0039] When an image relating to the grinding surface 211 of the grinding wheel 21, more specifically, a captured image 44 capturing the grinding surface 211, is input to the thus-trained extraction learning device 61, the learning device 61 outputs an extracted image in which only the tip portion of the grinding surface 211 is extracted from the captured image 44. In this embodiment, a judgment image 65 used to estimate the state of the grinding surface 211 is generated based on the extracted image.

[0040] <Classification Learner> The classification learner 62 will now be described. The classification learner 62 is a learner used to classify the judgment image 65, which will be described later. In this embodiment, a Convolutional Neural Network (CNN) is adopted as the classification learner 62. The classification learner 62 is machine-learned using learning data 64. The learning data 64 is data for performing machine learning on the classification learner 62 so that the classification learner 62 acquires the ability to discriminate the state of the grinding surface 211. In this embodiment, the classification learner 62 corresponds to a first learner, and the learning data 64 corresponds to the first learning data.

[0041] In this embodiment, the classification learner 62 is set for each type of grindstone 21. Also, in this embodiment, one classification learner 62 is set for each type of grindstone 21, corresponding to abrasive grains 212 of a predetermined particle size (specifically, the average particle size of abrasive grains 212 corresponding to grain size #200). In this embodiment, in order to realize such a configuration, grindstones 21 having abrasive grains 212 of a predetermined particle size are adopted as grindstones 21 used to create the learning data 64 from among a plurality of types of grindstones 21.

[0042] Basically, the learning data 64 is generated for each type of grindstone 21 in the following manner. First, an image set is prepared, which is two captured images of the same location on the grinding surface 211 of the grinding wheel 21, the first image being captured at the first time and the second image being captured at the second time. The first image is an image corresponding to the basic image 44B, and the second image is an image corresponding to the latest image 44N. The first and second images are acquired, for example, by using the camera unit 40. The first and second images are converted to grayscale.

[0043] Thereafter, for each of the first and second grayscale images, an extracted image is created by extracting only the tip portion of the grinding surface 211 (more specifically, the portion of the grinding surface 211 that functions as a cutting edge) from the image. Methods for creating an extracted image include a method using the ink pad described above, a method using the converted image 63A that constitutes the learning data 64 for the extraction learning device 61, and a method using the extraction learning device 61.

[0044] A difference image is then generated that indicates the difference between the extracted image corresponding to the first image and the extracted image corresponding to the second image. In this embodiment, this difference image is used as training data 64 for classification learner 62.

[0045] In this embodiment, as the differential images constituting the learning data 64, a large number of images corresponding to three types of grinding surface 211 conditions, namely, "evenly shedding grinding surface", "unbalanced shedding grinding surface", and "chip-welded grinding surface", are prepared.

[0046] The "evenly shedding abrasive surface" is defined as the state of the grinding surface 211 when the abrasive grains 212 are evenly shedding from each part of the grinding surface 211. When generating the learning data 64 corresponding to the evenly shedding abrasive surface, a large number of image sets consisting of the first image and the second image of the grinding wheel 21 having the evenly shedding abrasive surface are prepared, and the above-mentioned difference image is generated based on this image set. In this case, as the difference image, an image showing a part where the tip part of the grinding surface 211 is reduced in the extracted image corresponding to the second image is generated based on the extracted image corresponding to the first image. Then, this difference image is used as the learning data 64 corresponding to the evenly shedding abrasive surface. As an example shown in FIG. 6, the learning data 64 corresponding to the evenly shedding abrasive surface is an image in which the region RT showing the tip part of the grinding surface 211 is evenly distributed over the entire surface of the learning data 64 (more specifically, the above-mentioned difference image).

[0047] The "unbalanced shed grinding surface" is defined as the state of the grinding surface 211 when the portions where the abrasive grains 212 have fallen off are unevenly distributed on the grinding surface 211. When generating the learning data 64 corresponding to the unbalanced shed grinding surface, a large number of image sets consisting of the first image and the second image of the grinding wheel 21 of the unbalanced shed grinding surface are prepared, and the above-mentioned difference image is generated based on this image set. In this case, as the difference image, an image showing a portion where the tip portion of the grinding surface 211 is reduced in the extracted image corresponding to the second image is generated based on the extracted image corresponding to the first image. Then, this difference image is used as the learning data 64 corresponding to the unbalanced shed grinding surface. As an example shown in FIG. 7, the learning data 64 corresponding to the unbalanced shed grinding surface is an image in which the region RT showing the tip portion of the grinding surface 211 is distributed unevenly in a part of the learning data 64 (more specifically, the above-mentioned difference image).

[0048] The "chip-welded grinding surface" is defined as the state of the grinding surface 211 when chips are welded to the grinding surface 211. When generating the learning data 64 corresponding to the chip-welded grinding surface, a large number of image sets consisting of a first image and a second image of the grinding wheel 21 with the chip-welded grinding surface are prepared, and the above-mentioned difference image is generated based on this image set. In this case, as the difference image, an image showing an increased tip portion of the grinding surface 211 in the extracted image corresponding to the second image is generated based on the extracted image corresponding to the first image. Then, this difference image is used as the learning data 64 corresponding to the chip-welded grinding surface. As shown in an example in FIG. 8, the learning data 64 corresponding to the chip-welded grinding surface is an image in which the region RT showing the tip portion of the grinding surface 211 extends over a wide range including a plurality of abrasive grains 212.

[0049] In this embodiment, a learning process in which one of the multiple images constituting the learning data 64 is extracted and the classification learner 62 is made to perform machine learning using the extracted image as input data is repeatedly executed for each image constituting the learning data 64. Then, the machine learning model constructed through the execution of the machine learning is stored in the storage unit 56 as the trained classification learner 62.

[0050] An image relating to the grinding surface 211 of the grinding wheel 21 (in this embodiment, a judgment image 65, described later) is input to the classification learner 62, which has been trained by machine learning in this way. As a result, the classification learner 62 performs so-called classification, in which the input image is classified into one of an evenly-shed grinding surface (class 1), an unbalancedly-shed grinding surface (class 2), and a chip-welded grinding surface (class 3).

[0051] <Specific learning device> The specific learning device 67 will now be described. The specific learning device 67 is a learning device used to identify one of the multiple extraction learning devices 61 set for each type of grinding wheel 21 and one of the multiple classification learning devices 62 set for each type of grinding wheel 21.

[0052] In this embodiment, a convolutional neural network (CNN) is adopted as the specific learner 67. The specific learner 67 is machine-learned using learning data 68. The learning data 68 is data for performing machine learning on the specific learner 67 so that the specific learner 67 acquires the ability to identify one of the multiple extraction learners 61 and one of the multiple classification learners 62 as learners corresponding to grinding wheels 21 having a similar basic structure to the grinding wheel 21 to be estimated. In this embodiment, the specific learner 67 corresponds to the second learner, and the learning data 68 corresponds to the second learning data.

[0053] The training data 68 is generated in the following manner. First, for each type of grindstone 21, a large number of captured images of the grinding surface 211 of the grindstone 21 having the abrasive grains 212 of the predetermined particle diameter are prepared. The captured images include a first image captured in the first period and a second image captured in the second period. The first image is an image corresponding to the basic image 44B, and the second image is an image corresponding to the latest image 44N. The first image and the second image are acquired, for example, by using the camera unit 40. Each captured image (the first image and the second image) is converted to grayscale. In this embodiment, the grayscale captured images are used as learning data 68.

[0054] In this embodiment, the electronic control device 50 repeatedly executes a learning process for each captured image constituting the learning data 68, in which one of the captured images constituting the learning data 68 is extracted and the specific learning device 67 is trained using the captured image as input data. Then, the machine learning model constructed through the execution of the machine learning is stored in the storage unit 56 as the trained specific learning device 67. Note that in this embodiment, the process of storing the trained specific learning device 67 in the storage unit 56 corresponds to a second learning device storage process.

[0055] An image relating to the grinding surface 211 of the grinding wheel 21 (in this embodiment, a grayscale resized image 44R, which will be described later) is input to the trained specific learning device 67 that has been trained in this manner. As a result, the specific learning device 67 outputs specific data that identifies one of the multiple extraction learning devices 61 and one of the multiple classification learning devices 62 as learning devices corresponding to grinding wheels 21 having a similar basic structure to the grinding wheel 21 to be estimated.

[0056] <Specific part> The machined surface condition estimating device 30 (more specifically, the electronic control device 50) of this embodiment has a specifying unit 58 as a functional unit.

[0057] When estimating the state of the grinding surface 211 of the grinding wheel 21 to be estimated, the identification unit 58 uses the captured image 44 and the identification learner 67 to identify one of the multiple extraction learners 61 and identify one of the multiple classification learners 62. As a result, one extraction learner 61 and one classification learner 62 corresponding to the grinding wheel 21 having a similar basic structure to the grinding wheel 21 to be estimated are identified from the multiple extraction learners 61 and the multiple classification learners 62 stored in advance. According to this embodiment, it is possible to identify one extraction learner 61 and one classification learner 62 suitable for estimating the state of the grinding wheel 21 to be estimated in this manner. In this embodiment, the extraction learner 61 and the classification learner 62 identified in this manner are used in the determination process for determining the state of the grinding surface 211.

[0058] FIG. 9 shows a procedure for identifying the extraction learner 61 and the classification learner 62. As shown in FIG. 9, first, the camera unit 40 is attached to the automatic grinding device 20 on which the grinding wheel 21 to be estimated is provided, and the camera unit 40 and the operation control unit 25 of the automatic grinding device 20 are connected to the electronic control device 50 (step S10). Then, the electronic control device 50 controls the operation of the camera unit 40 and the operation control unit 25. In detail, illumination is performed by the illumination unit 43. As a result, the imaging target area AR on the grinding surface 211 is illuminated. In addition, the operation control of each of the driving units 26, 27, 28, and 29 of the automatic grinding device 20 is performed. As a result, the grinding wheel 21 is rotated at a speed suitable for imaging the grinding surface 211, and the relative position between the camera unit 40 and the grinding wheel 21 is changed in a predetermined pattern. Furthermore, the operation control of the camera 41 is performed based on the detection signal of the rotation phase sensor. As a result, a captured image 44 of the grinding surface 211 is acquired. In this embodiment, the grinding surface 211 is divided into a plurality of regions by dividing the grinding surface 211 into a lattice shape in the circumferential direction and the rotation axis direction of the grinding wheel 21. Then, the camera 41 captures an image of the grinding surface 211 for each of these regions. As a result, a plurality of captured images 44 are obtained. In this embodiment, an image including a large number of abrasive grains 212 is obtained as each captured image 44. In this embodiment, the process of step S10 corresponds to the acquisition step. In this embodiment, the electronic control device 50 that executes the process of step S10 corresponds to the acquisition unit.

[0059] Then, based on the plurality of captured images 44 acquired in the process of step S10, a resized image 44R is generated by changing the image size of the captured image 44 at a change ratio RG (step S11). Here, the predetermined particle size is set to "G1" and the average particle size in the particle size information is set to "G2". In this case, in the process of step S11, the image size of the captured image 44 is changed at the change ratio RG expressed by the formula "G1 / G2". In this embodiment, the process of step S11 corresponds to a resizing step, and the electronic control device 50 that executes the process of step S11 corresponds to a resizing unit. Then, the plurality of resized images 44R are converted into grayscale (step S12).

[0060] Thereafter, one of the plurality of extraction learners 61 and one of the plurality of classification learners 62 are identified using the grayscale resized image 44R and the trained specific learner 67 (step S13). Specifically, first, the trained specific learner 67 classifies the resized image 44R using the grayscale resized image 44R as input data. Then, based on the classification result by the specific learner 67, the extraction learner 61 and classification learner 62 corresponding to the type of grinding wheel 21 that has the most similar basic structure among the plurality of types of grinding wheels 21 are identified. In this embodiment, the process of step S13 corresponds to the identification step. Thereafter, a determination process is executed to determine the state of the grinding surface 211 using the extraction learner 61 and classification learner 62 identified in the process of step S13 (step S14).

[0061] <Determination process> The determination process will now be described. When the grindstone 21 is used, it is desirable that the condition of the grindstone surface 211 of the grindstone 21 change uniformly over the entire grindstone surface 211 .

[0062] In a general grinding wheel, the condition of the tip portion of the grinding surface 211 changes in such a way that the abrasive grains 212 fall off. As the deterioration of the grinding surface 211 progresses, the deviation in the falling off of the abrasive grains 212 at each portion of the grinding surface 211 increases. For this reason, in a general grinding wheel, when the deviation in the falling off of the abrasive grains 212 increases, it can be said that the condition of the grinding surface 211 has changed in an undesirable manner. On the other hand, in a superabrasive grinding wheel, the condition of the tip portion of the grinding surface 211 changes in such a way that foreign matter such as cuttings adheres to the grinding surface 211. For this reason, when the deviation in the adhesion of foreign matter at each portion of the grinding surface 211 increases, it can be said that the condition of the grinding surface 211 has changed in an undesirable manner.

[0063] In this embodiment, the machined surface condition estimating device 30 is constructed on the assumption that occurrence of deterioration abnormality will be determined in a plurality of types of grinding wheels 21, including general grinding wheels and super abrasive grinding wheels. In this embodiment, occurrence of deterioration abnormality in the grinding wheel 21 to be estimated is determined when the distribution bias of the tip portion of the grinding surface 211 in the judgment image 65 is equal to or greater than a predetermined level.

[0064] Hereinafter, a specific execution procedure of the determination process will be described with reference to Fig. 10. Fig. 10 is a flowchart showing the execution procedure of the determination process. Note that the series of processes shown in this flowchart conceptually show the execution procedure of the determination process, and the actual process is executed by the electronic control unit 50 as a process at a predetermined cycle.

[0065] 10, in this process, first, it is determined whether or not it is the first period immediately after dressing of the grinding wheel 21 (step S20). If it is the first period (step S20: YES), a plurality of grayscale resized images 44R are stored as basic images 44B in the storage unit 56 together with position information on the grinding surface 211 (step S21). In this case, this process is temporarily terminated without executing the processes of steps S22 to S29.

[0066] On the other hand, if it is the second time period that is later than the first time period (step S20: NO), the multiple grayscale resized images 44R are stored in the memory unit 56 as the latest images 44N together with position information on the grinding surface 211 (step S22).

[0067] Thereafter, a process is executed in which the extraction learning device 61 specified by the specific learning device 67 outputs an extracted image (hereinafter, converted basic image) using the basic image 44B as input data (step S23). Also, a process is executed in which the extraction learning device 61 specified by the specific learning device 67 outputs an extracted image (hereinafter, converted latest image) using the latest image 44N as input data. In the process of step S23, in detail, a generation process is repeatedly executed for each of the multiple basic images 44B, such as extracting one of the multiple basic images 44B and causing the extraction learning device 61 to output a converted basic image using the basic image 44B as input data. Also, a generation process is repeatedly executed for each of the multiple latest images 44N, such as extracting one of the multiple latest images 44N and causing the extraction learning device 61 to output a converted latest image using the latest image 44N as input data.

[0068] Thereafter, a difference image (hereinafter, judgment image 65) showing the difference between two images (the converted basic image and the converted latest image) corresponding to the same location on the grinding surface 211 is generated (step S24). The judgment image 65 is generated for each imaging position on the grinding surface 211, i.e., for each of the aforementioned regions. Furthermore, as the judgment image 65, an image showing a portion where the tip portion of the grinding surface 211 has decreased in the converted latest image and an image showing a portion where the tip portion has increased are generated based on the converted basic image. The judgment image 65 generated in the processing of step S24 is stored in the memory unit 56.

[0069] Thereafter, classification of each judgment image 65 is executed using the multiple judgment images 65 and the trained classification learner 62 identified by the identification learner 67 (step S25). In detail, one of the multiple judgment images 65 is extracted, and the judgment image 65 is used as input data to classify the judgment image 65 by the classification learner 62. This classification process is executed repeatedly for each of the multiple judgment images 65. Note that in this embodiment, the process of step S25 corresponds to an output step, and the electronic control device 50 that executes the process of step S25 corresponds to an output unit.

[0070] Then, based on the classification result for each evaluation image 65 and the position information of each evaluation image 65 on the grinding surface 211, the degree of bias in distribution of the tip portions of the grinding surface 211 in the evaluation images 65 is calculated (step S26).

[0071] In the process of step S26, if the classification learner 62 identified by the specific learner 67 is a learner corresponding to a general grinding wheel, a proportion R1 of the portion classified as an evenly-shed abrasive surface (class 1) in the process of step S25 is calculated for each rotation phase of the grinding wheel 21. Here, if a predetermined number (N2) of judgment images 65 out of a plurality (N1) of judgment images 65 corresponding to the same rotation phase are classified as an evenly-shed abrasive surface (class 1), a value "N2 / N1" is calculated as the proportion R1. In this case, a difference ΔR1 (=R1max-R1min) between the maximum value R1max and the minimum value R1min of the proportion R1 calculated for each rotation phase is calculated as the degree of bias RR1.

[0072] On the other hand, in the process of step S26, if the classification learning device 62 identified by the identification learning device 67 is a learning device corresponding to the superabrasive grindstone, the proportion R1 of the portion classified as the chip-welded grinding surface (class 3) in the process of step S25 is calculated for each rotation phase of the grindstone 21. Here, if a predetermined number (N2) of judgment images 65 out of a plurality (N1) of judgment images 65 corresponding to the same rotation phase are classified as the chip-welded grinding surface (class 3), the value "N2 / N1" is calculated as the proportion R1. In this case, the difference ΔR1 (=R1max-R1min) between the maximum value R1max and the minimum value R1min of the proportion R1 calculated for each rotation phase is calculated as the degree of bias RR1.

[0073] Thereafter, it is determined whether the difference ΔR1 is equal to or greater than a predetermined value J1 (step S27). As the predetermined value J1, a value capable of properly determining the occurrence of deterioration abnormality of the grinding wheel 21 is previously determined for each of the multiple classification learners 62 used in the determination process, specifically, for each type of grinding wheel 21. More specifically, a value (predetermined value J1) capable of properly determining the occurrence of deterioration abnormality of the grinding wheel 21 is determined for each of the multiple classification learners 62 based on the results of various experiments and simulations conducted by the inventors, etc. Then, the predetermined value J1 is previously stored in the storage unit 56 in a manner associated with each of the multiple classification learners 62. In the process of step S27, the predetermined value J1 corresponding to the classification learner 62 identified by the identification learner 67 is read and used.

[0074] Then, if the difference ΔR1 is less than the predetermined value J1 (step S27: NO), it is determined that the distribution bias of the tip portions of the grinding surface 211 in the determination image 65 is less than a predetermined level, and it is determined to be normal (step S28). On the other hand, if the difference ΔR1 is equal to or greater than the predetermined value J1 (step S27: YES), it is determined that the distribution bias of the tip portions of the grinding surface 211 in the determination image 65 is equal to or greater than a predetermined level, and it is determined that the grinding wheel 21 has deteriorated abnormally (step S29). In this case, it is determined that the grinding wheel 21 has deteriorated to such an extent that dressing of the grinding surface 211 is required.

[0075] <Action and effect> The effects of the machined surface condition estimating device 30 of this embodiment will be described below. First, a case where the grinding wheel 21 to be estimated is a general grinding wheel will be described.

[0076] As described above, in the case of the grinding wheel 21, which is a general grinding wheel, if there is a large bias in the falling off of the abrasive grains 212 at various parts of the grinding surface 211, it can be said that the condition of the grinding surface 211 has changed in an undesirable way.

[0077] In this embodiment, an image showing a portion where the tip portion of the grinding surface 211 has been reduced in the latest image 44N (specifically, the latest image after conversion) is generated as the judgment image 65 based on the basic image 44B (specifically, the basic image after conversion). In this judgment image 65, the amount of change in the state of the tip portion of the grinding surface 211, which is a portion that affects the finishing quality of the workpiece, specifically, the amount of abrasive grains that have fallen off from each portion of the grinding surface 211 during the period from the first period to the second period, appears. Therefore, based on the position and size of the tip portion of the grinding surface 211 in the judgment image 65, the amount of abrasive grains that have fallen off from each portion of the grinding surface 211 can be grasped. Therefore, according to this embodiment, it is understood that the distribution of the tip portion of the grinding surface 211 in the judgment image 65 is biased, and therefore the bias of the falling off of the abrasive grains 212 from each portion of the grinding surface 211 is large. In this embodiment, in such a case (step S27: YES in FIG. 10), it can be determined that the deterioration abnormality of the grinding wheel 21 has occurred.

[0078] Next, a case where the grindstone 21 to be estimated is a super abrasive grindstone will be described. As described above, in the case of the superabrasive grinding wheel 21, if there is a large deviation in the amount of foreign matter adhering to various parts of the grinding surface 211, it can be said that the condition of the grinding surface 211 has changed in an undesirable manner.

[0079] In this embodiment, an image showing an increased portion of the tip portion of the grinding surface 211 in the latest image 44N (specifically, the latest image after conversion) is generated as the judgment image 65 based on the basic image 44B (specifically, the basic image after conversion). In this judgment image 65, the amount of change in the state of the tip portion of the grinding surface 211, which is a portion that affects the finishing quality of the workpiece, specifically, the amount of foreign matter attached to each portion of the grinding surface 211 during the period from the first period to the second period, appears. Therefore, based on the position and size of the tip portion of the grinding surface 211 in the judgment image 65, the amount of foreign matter attached to each portion of the grinding surface 211 can be grasped. Therefore, according to this embodiment, it is understood that the uneven distribution of the tip portion of the grinding surface 211 in the judgment image 65 has caused the unevenness of the amount of foreign matter attached to each portion of the grinding surface 211 to increase. In this embodiment, in such a case (step S27: YES in FIG. 10), it can be determined that the deterioration abnormality of the grinding wheel 21 has occurred.

[0080] Furthermore, according to this embodiment, the judgment image 65 can be generated by using two captured images 44 (basic image 44B and latest image 44N), both of which contain a large number of abrasive grains 212, and the learned extraction learning device 61. Therefore, the amount of change in state in each part of the grinding surface 211 can be easily grasped in a short time compared to the case where the amount of change in state in each part of the grinding surface 211 is grasped by analyzing the state of the grinding surface 211 for each finely divided area containing an extremely small number of abrasive grains 212. Therefore, the occurrence of deterioration abnormality in the grinding wheel 21 can be judged in a short time.

[0081] Furthermore, according to this embodiment, by periodically executing the work of attaching the camera unit 40, the process of identifying the learning devices 61, 62, and the judgment process, it is possible to perform grinding processing using the automatic grinding device 20 while checking the condition of the grinding surface 211 at appropriate times.

[0082] Here, if the grindstones 21 are of the same type (more specifically, made by the same manufacturer and model), the structure of the grinding surfaces 211 of those grindstones 21 is approximately the same even if the sizes of the grindstones 21 (external size and particle diameter of the abrasive grains 212) are different.

[0083] In this embodiment, as the learned extraction learning device 61 and classification learning device 62, learning devices shared by multiple grinding wheels 21 of different sizes among the same type of grinding wheel 21, specifically, learning devices corresponding to grinding wheels 21 having abrasive grains 212 of a predetermined particle diameter G1, are stored. Then, in the judgment process, a basic image 44B and a latest image 44N to be used as input data for the extraction learning device 61 are generated based on a captured image 44 of the grinding wheel 21 to be estimated, while changing the image size by a change ratio RG. As a result, the basic image 44B and the latest image 44N are generated with an image size corresponding to the predetermined particle diameter G1, specifically, an image size in which the abrasive grains 212 of the predetermined particle diameter G1 are displayed.

[0084] Therefore, a converted basic image and a converted latest image can be output from an extraction learning device 61 corresponding to a grinding wheel 21 having abrasive grains 212 of the predetermined particle diameter G1 as input data, and the converted basic image and the converted latest image can be output from the extraction learning device 61 corresponding to the grinding wheel 21 having abrasive grains 212 of the predetermined particle diameter G1. Thus, according to this embodiment, even when estimating the state of the grinding surface 211 of grinding wheels 21 of the same type but different sizes, the estimation can be performed using a common extraction learning device 61 without preparing a trained extraction learning device 61 for each size of the grinding wheel 21. Therefore, the time and effort required for setting up the extraction learning device 61 can be reduced.

[0085] In the judgment process, a judgment image 65 to be used as input data for the classification learner 62 is generated based on a captured image 44 of the grinding wheel 21 to be estimated, while changing the image size by a change ratio RG. As a result, an image size corresponding to the predetermined particle diameter G1, more specifically, an image size in which the abrasive grains 212 of the predetermined particle diameter G1 are displayed, is generated as the judgment image 65.

[0086] Therefore, in this embodiment, a judgment image 65 having an image size corresponding to the predetermined particle diameter G1 is used as input data, and the judgment image 65 can be classified by a classification learner 62 corresponding to a grinding wheel 21 having abrasive grains 212 of the predetermined particle diameter G1. Thus, according to this embodiment, even when estimating the state of the grinding surface 211 of a plurality of types of grinding wheels 21, the estimation can be performed using a common classification learner 62 without preparing a trained classification learner 62 for each type of grinding wheel 21. Therefore, the time and effort required for the task of setting the classification learner 62 can be reduced.

[0087] In this embodiment, when estimating the condition of the grinding surface 211 of the grinding wheel 21, the captured image 44 and the identification learner 67 are used to identify one extraction learner 61 and one classification learner 62 from a plurality of extraction learners 61 and a plurality of classification learners 62 that are stored in advance.

[0088] As a result, when an extraction learner 61 and a classification learner 62 corresponding to a grindstone 21 of the same type as the grindstone 21 to be estimated are set, it is possible to identify those extraction learners 61 and classification learners 62. Moreover, when an extraction learner 61 and a classification learner 62 corresponding to a grindstone 21 of the same type as the grindstone 21 to be estimated are not set, it is possible to identify an extraction learner 61 and a classification learner 62 corresponding to a grindstone 21 having a similar basic structure to the grindstone 21 to be estimated. According to this embodiment, in this way, it is possible to identify an extraction learner 61 and a classification learner 62 suitable for estimating the state of the grindstone 21 to be estimated from a plurality of extraction learners 61 and a plurality of classification learners 62 stored in advance.

[0089] Then, by using the identified extraction learning device 61 and the captured image 44 (specifically, the base image 44B and the latest image 44N), it is possible to obtain a converted base image and a converted latest image used to generate the judgment image 65. In addition, by using the identified classification learning device 62 and the judgment image 65, it is possible to classify the judgment image 65.

[0090] According to this embodiment, rather than storing a separate extraction learner 61 and classification learner 62 for each type of grinding wheel 21, it is possible to estimate the state of the grinding wheel 21 to be estimated by using one of a plurality of pre-stored extraction learners 61 and classification learners 62. This makes it possible to reduce the number of extraction learners 61 and classification learners 62 that are pre-stored, thereby shortening the time and reducing the effort required for the work of setting the extraction learner 61 and the work of setting the classification learner 62.

[0091] Furthermore, even if an extraction learner 61 and a classification learner 62 corresponding to a grindstone 21 of the same type as the grindstone 21 to be estimated have not yet been set, it is possible to estimate the state of the grindstone 21 by reusing an existing extraction learner 61 and classification learner 62 corresponding to a grindstone 21 having a similar basic structure. Therefore, in this case, it is possible to estimate the state of the grindstone 21 without newly determining an extraction learner 61 and a classification learner 62 corresponding to the type of grindstone 21 to be estimated.

[0092] <Action and effect> According to this embodiment, the following advantageous effects can be obtained. (1-1) The machined surface state estimation device 30 includes a storage unit 56, an acquisition unit, a specification unit 58, and an output unit. The storage unit 56 stores a plurality of classification learners 62, each of which is determined for a plurality of types of grinding wheels 21, and a trained specification learner 67 that has been trained to specify one of the plurality of classification learners 62. The acquisition unit acquires a resized image 44R, which is an image related to the grinding surface 211 of the grinding wheel 21 to be estimated. The specification unit 58 specifies one of the plurality of classification learners 62 using the resized image 44R and the specification learner 67. The output unit uses the resized image 44R (specifically, the judgment image 65) and the classification learner 62 specified by the specification unit 58 to output machined surface information (specifically, the classification result of the judgment image 65) indicating the state of the grinding surface 211 of the grinding wheel 21 to be estimated. According to this embodiment, it is possible to reduce the time and effort required for the task of setting the trained classification learner 62.

[0093] (1-2) Images constituting the learning data 63, 64, 68 (pre-conversion image 63B, post-conversion image 63A, difference image, captured image) are converted into grayscale images. In addition, images used as input data for the learning devices 61, 62, 67 (captured image 44, judgment image 65, resized image 44R) are converted into grayscale images.

[0094] Grindstones 21 come in a variety of colors (red, blue, gray, etc.). Therefore, if the grinding surfaces 211 of multiple types of grindstones 21 are simply imaged to obtain captured images 44, the captured images 44 will include grinding surfaces 211 of different colors. In this embodiment, when estimating the state of the grindstone 21 to be estimated, an extraction learner 61 or a classification learner 62 corresponding to different types of grindstones 21, specifically grindstones 21 having a similar basic structure, may be used.

[0095] In this embodiment, in such a case, differences in color of the grinding surface 211 in the captured image 44 may be a factor that causes a decrease in the accuracy of generation of the extracted image by the extraction learning device 61, or a factor that causes a decrease in the classification accuracy of the judgment image 65 by the classification learning device 62.

[0096] According to this embodiment, images of the same color, specifically grayscale images, can be used as images constituting the learning data 63, 64, 68 and images used as input data for the learners 61, 62, 67. This makes it possible to suppress the influence of differences in color of the grinding surface 211 represented in each image, specifically, the influence on the generation accuracy of the extracted image by the extraction learner 61 and the classification accuracy of the judgment image 65 by the classification learner 62. This makes it possible to suppress a decrease in the generation accuracy of the extracted image by the extraction learner 61 and a decrease in the classification accuracy of the judgment image 65 by the classification learner 62.

[0097] (1-3) Each of the learning data 63, 64, 68 is composed of an image related to the grinding surface 211 of the grinding wheel 21 having abrasive grains of a predetermined grain diameter. The memory unit 56 stores particle diameter information related to the particle diameter of the abrasive grains in the grinding wheel 21 to be estimated. The machined surface condition estimation device 30 has a resizing unit that changes the image size of the captured image 44 by a change ratio RG. A resized image 44R, which is an image after resizing by the resizing unit, is used as input data for the extraction learner 61, input data for the classification learner 62, and input data for the identification learner 67.

[0098] According to this embodiment, even when estimating the state of the grinding surface 211 of grinding wheels 21 of the same type but different sizes, the estimation can be performed using a common extraction learner 61 without preparing a trained extraction learner 61 for each size of the grinding wheel 21. Therefore, it is possible to reduce the time and effort required for setting the extraction learner 61. Furthermore, even when estimating the state of the grinding surface 211 of grinding wheels 21 of the same type but different sizes, the estimation can be performed using a common classification learner 62 without preparing a trained classification learner 62 for each size of the grinding wheel 21. Therefore, it is possible to reduce the time and effort required for setting the classification learner 62.

[0099] (1-4) The machined surface condition estimation device 30 includes a camera 41 that captures an image of the grinding surface 211 from a direction perpendicular to the grinding surface 211, and an illumination unit 43 that illuminates an imaging target area AR on the grinding surface 211. The illumination unit 43 has a first irradiating unit 45 and a second irradiating unit 46. The first irradiating unit 45 irradiates the imaging target area AR with first illumination light 45L having a first emission color from a direction perpendicular to the grinding surface 211. The second irradiating unit 46 irradiates the imaging target area AR with second illumination light 46L having a second emission color different from the first emission color from a direction along the grinding surface 211.

[0100] According to this embodiment, an image clearly showing the top surface of the grinding wheel 21, which is a part that affects the finishing quality of the workpiece, can be obtained as the captured image 44 captured by the camera 41. Therefore, the captured image 44 can be used to accurately estimate the condition of the grinding surface 211 of the grinding wheel 21, more specifically, the tip portion of the abrasive grains 212.

[0101] Second embodiment Hereinafter, a second embodiment of a machined surface condition estimating device, an estimating program, and a machined surface condition estimating method will be described with reference to Figs. 11 to 13, focusing on differences from the first embodiment.

[0102] The present embodiment differs from the first embodiment in the extraction learning device, classification learning device, and specific learning device. The extraction learning device, classification learning device, and specific learning device of the present embodiment will be described below. Note that, in the following, the same components as those of the first embodiment will be denoted by the same reference numerals, and detailed descriptions thereof will be omitted.

[0103] <Extractive learning machine> First, the extraction learning device 71 of this embodiment will be described. 11, in this embodiment, the extraction learning device 71 is set for each type of grindstone 21, and also for each particle size (specifically, grain size defined in the standard) of the abrasive grains 212 in the same type of grindstone 21. In this embodiment, in order to realize such a configuration, a plurality of grindstones 21 of different types and particle sizes are adopted as the grindstones 21 used to create the learning data 73.

[0104] In this embodiment, the learning data 73 for the extraction learning device 71 is generated for each of a plurality of grinding wheels 21 of different types and particle sizes as follows. Based on an image of the grinding surface 211 of the grinding wheel 21, a large number of image sets are created, each pair consisting of a pre-conversion image 63B and a post-conversion image 63A of the same location on the grinding surface 211. These image sets are then used as the learning data 73 for the extraction learning device 71.

[0105] In this embodiment, a learning process is repeatedly performed for each image set constituting the learning data 73, in which one of the multiple image sets constituting the learning data 73 is extracted and the extraction learner 71 is trained by machine learning using the image set as input data. The learning process is performed by the electronic control device 50 for each of the multiple grinding wheels 21 having different types and particle sizes, more specifically, for each of the multiple extraction learners 71. Then, the machine learning model constructed through the execution of the machine learning is stored in the storage unit 56 as the trained extraction learner 71.

[0106] When an image of the grinding surface 211 of the grinding wheel 21, more specifically, an image 44 capturing the grinding surface 211, is input to the trained extraction learning device 71, which has been machine-learned in this manner, the learning device 71 outputs an extracted image in which only the tip portion of the grinding surface 211 is extracted from the image 44.

[0107] <Classification Learner> Next, classification learner 72 of this embodiment will be described. In this embodiment, the classification learner 72 is set for each type of grindstone 21, and also for each particle size (specifically, grain size defined in the standard) of the abrasive grains in the same type of grindstone 21. In this embodiment, in order to realize such a configuration, many grindstones 21 of different types and particle sizes are used as the grindstones 21 used to create the learning data 74.

[0108] In this embodiment, the learning data 74 for the classification learner 72 is generated for each of the multiple grinding wheels 21 of different types and particle sizes as follows. A difference image showing the difference between the extracted image corresponding to the first image and the extracted image corresponding to the second image is generated based on a captured image of the grinding surface 211 of the grinding wheel 21. In this embodiment, this difference image is used as the learning data 74 for the classification learner 72. In this embodiment, a large number of images corresponding to each of three types of grinding surface 211 conditions, namely, the "evenly shed grinding surface", the "unbalanced shed grinding surface", and the "chip-welded grinding surface", are prepared as the difference images constituting the learning data 74.

[0109] In this embodiment, a learning process is repeatedly performed for each image constituting the learning data 74, in which one of the multiple images constituting the learning data 74 is extracted and the classification learner 72 is trained using the extracted image as input data. The learning process is performed by the electronic control device 50 for each of the multiple grinding wheels 21 having different types and particle sizes, more specifically, for each of the multiple extraction learners 71. Then, the machine learning model constructed through the execution of the machine learning is stored in the storage unit 56 as the trained classification learner 72. Note that in this embodiment, the process of storing the trained classification learner 72 in the storage unit 56 corresponds to a first learner storage process.

[0110] An image relating to the grinding surface 211 of the grinding wheel 21 (in this embodiment, a judgment image 75, described later) is input to the classification learner 72, which has been trained by machine learning in this way. As a result, the classification learner 72 performs so-called classification, in which the input image is classified into one of an evenly-shed grinding surface (class 1), an unbalancedly-shed grinding surface (class 2), and a chip-welded grinding surface (class 3).

[0111] <Specific learning device> Next, the specific learner 77 of this embodiment will be described. In this embodiment, the specific learner 77 is used to identify one of the multiple extraction learners 71 set for each of the multiple grinding wheels 21 of different types and particle sizes, and one of the multiple classification learners 72 set for each of the multiple grinding wheels 21 of different types and particle sizes. The specific learner 77 is machine-learned using the learning data 78. Through this machine learning, the specific learner 77 acquires the ability to identify one of the multiple extraction learners 71 and one of the multiple classification learners 72 as learners corresponding to grinding wheels 21 having the same structure as or a similar structure to the grinding wheel 21 to be estimated. In this embodiment, in order to realize such a configuration, a large number of grinding wheels 21 of different types and particle sizes are adopted as the grinding wheels 21 used to create the learning data 78.

[0112] In this embodiment, the learning data 78 is generated as follows. A large number of captured images of the grinding surfaces 211 are prepared for each of a plurality of grinding wheels 21 of different types and particle sizes. Each captured image (more specifically, the first image and the second image) is converted to grayscale. In this embodiment, the grayscale captured images are used as the learning data 78.

[0113] In the present embodiment, the electronic control device 50 repeatedly executes a learning process for each captured image constituting the learning data 78, in which one of the captured images constituting the learning data 78 is extracted and the specific learning device 77 is trained using the captured image as input data. Then, the machine learning model constructed through the execution of the machine learning is stored in the storage unit 56 as the trained specific learning device 77. Note that in the present embodiment, the process of storing the trained specific learning device 77 in the storage unit 56 corresponds to a second learning device storage process.

[0114] An image (grayscale captured image 44 in this embodiment) relating to the grinding surface 211 of the grinding wheel 21 is input to the trained specific learning device 77 that has been trained in this manner. As a result, the specific learning device 77 outputs specific data that identifies one of the multiple extraction learning devices 71 and one of the multiple classification learning devices 72 as learning devices corresponding to grinding wheels 21 having the same structure as or a similar structure to the grinding wheel 21 to be estimated.

[0115] The procedure for identifying extraction learner 71 and classification learner 72 using identified learner 77 will be described below. FIG. 12 shows a procedure for identifying the extraction learner 71 and the classification learner 72.

[0116] 12, first, a camera unit 40 is attached to the automatic grinding device 20 on which the grinding wheel 21 to be estimated is provided, and the camera unit 40 and the operation control unit 25 of the automatic grinding device 20 are connected to an electronic control device 50 (step S10). The electronic control device 50 then controls the operation of the camera unit 40 and the operation control unit 25. As a result, a plurality of captured images 44 are obtained. In this embodiment, an image including a large number of abrasive grains 212 is obtained as each captured image 44.

[0117] Thereafter, the multiple captured images 44 acquired in the process of step S10 are converted into grayscale (step S31). Thereafter, one of the plurality of extraction learners 71 and one of the plurality of classification learners 72 are identified using the grayscale captured images 44 and the trained identification learner 77 (step S32). Specifically, first, the plurality of grayscale captured images 44 are used as input data, and the trained identification learner 77 classifies the captured images 44. Then, based on the classification result by the identification learner 77, the extraction learner 71 and the classification learner 72 corresponding to the grinding wheel 21 having the most similar basic structure among the plurality of grinding wheels 21 having different types and particle sizes are identified. Note that in this embodiment, the process of step S32 corresponds to the identification step.

[0118] Thereafter, a determination process is executed to determine the state of the grinding surface 211 using the extraction learning device 71 and classification learning device 72 identified in the process of step S32 (step S33). <Determination process> Hereinafter, a specific execution procedure of the determination process of this embodiment will be described with reference to Fig. 13. Fig. 13 is a flowchart showing the execution procedure of the determination process. Note that the series of processes shown in this flowchart conceptually show the execution procedure of the determination process, and the actual process is executed by the electronic control unit 50 as a process at a predetermined cycle.

[0119] 13, in this process, first, it is determined whether or not it is the first period immediately after dressing of the grinding wheel 21 (step S40). If it is the first period (step S40: YES), a plurality of grayscale captured images 44 are stored in the storage unit 56 as basic images 44B together with position information on the grinding surface 211 (step S41). In this case, this process is temporarily terminated without executing the processes of steps S42 to S49.

[0120] On the other hand, if it is the second time period that is later than the first time period (step S40: NO), the multiple grayscale captured images 44 are stored in the memory unit 56 as the latest image 44N together with position information on the grinding surface 211 (step S42).

[0121] Thereafter, a process is executed in which the extraction learning device 71 specified by the specified learning device 77 outputs an extracted image (hereinafter, converted basic image) using the basic image 44B as input data (step S43). Also, a process is executed in which the extraction learning device 71 specified by the specified learning device 77 outputs an extracted image (hereinafter, converted latest image) using the latest image 44N as input data. In the process of step S43, in detail, a generation process is repeatedly executed for each of the multiple basic images 44B, such as extracting one of the multiple basic images 44B and causing the extraction learning device 71 to output the converted basic image using the basic image 44B as input data. Also, a generation process is repeatedly executed for each of the multiple latest images 44N, such as extracting one of the multiple latest images 44N and causing the extraction learning device 71 to output the converted latest image using the latest image 44N as input data.

[0122] Thereafter, a difference image (hereinafter, judgment image 75) showing the difference between two images (the converted basic image and the converted latest image) corresponding to the same location on the grinding surface 211 is generated (step S44). The judgment image 75 is generated for each imaging position on the grinding surface 211, i.e., for each of the aforementioned regions. Furthermore, as the judgment image 75, an image showing a portion where the tip portion of the grinding surface 211 has decreased in the converted latest image and an image showing a portion where the tip portion has increased are generated based on the converted basic image. The judgment image 75 generated in the processing of step S44 is stored in the memory unit 56.

[0123] Thereafter, classification of each judgment image 75 is executed using the multiple judgment images 75 and the trained classification learner 72 identified by the identification learner 77 (step S45). In detail, one of the multiple judgment images 75 is extracted, and the judgment image 75 is used as input data to classify the judgment image 75 by the classification learner 72. This classification process is executed repeatedly for each of the multiple judgment images 75. Note that in this embodiment, the process of step S45 corresponds to an output step, and the electronic control device 50 that executes the process of step S45 corresponds to an output unit.

[0124] Then, based on the classification result for each evaluation image 75 and the position information on the grinding surface 211, the degree of bias in distribution of the tip portion of the grinding surface 211 in the evaluation image 75 is calculated (step S46).

[0125] In the process of step S46, if the classification learner 72 identified by the specific learner 77 is a learner corresponding to a general grinding wheel, the proportion R2 of the portion classified as an evenly-shed abrasive surface (class 1) in the process of step S45 is calculated for each rotation phase of the grinding wheel 21. Here, if a predetermined number (N2) of judgment images 75 out of a plurality (N1) of judgment images 75 corresponding to the same rotation phase are classified as an evenly-shed abrasive surface (class 1), the value "N2 / N1" is calculated as the proportion R2. In this case, the difference ΔR2 (=R2max-R2min) between the maximum value R2max and the minimum value R2min of the proportion R2 calculated for each rotation phase is calculated as the degree of bias RR2.

[0126] On the other hand, in the process of step S46, if the classification learning device 72 identified by the identification learning device 77 is a learning device corresponding to the superabrasive grindstone, the proportion R2 of the portion classified as the chip-welded grinding surface (class 3) in the process of step S45 is calculated for each rotation phase of the grindstone 21. Here, if a predetermined number (N2) of judgment images 75 out of a plurality (N1) of judgment images 75 corresponding to the same rotation phase are classified as the chip-welded grinding surface (class 3), the value "N2 / N1" is calculated as the proportion R2. In this case, the difference ΔR2 (=R2max-R2min) between the maximum value R2max and the minimum value R2min of the proportion R2 calculated for each rotation phase is calculated as the degree of bias RR2.

[0127] Thereafter, it is determined whether the difference ΔR2 is equal to or greater than a predetermined value J2 (step S47). As the predetermined value J2, a value capable of properly determining the occurrence of deterioration abnormality of the grinding wheel 21 is previously determined for each of the multiple classification learners 72 used in the determination process (specifically, for each of the multiple grinding wheels 21 having different types and particle sizes). In detail, a value (predetermined value J2) capable of properly determining the occurrence of deterioration abnormality of the grinding wheel 21 is obtained for each of the multiple classification learners 72 used in the determination process based on the results of various experiments and simulations conducted by the inventors, etc. Then, the predetermined value J2 is stored in advance in the storage unit 56 in a manner associated with the multiple classification learners 72 used in the determination process. In the process of step S47, the predetermined value J2 corresponding to the classification learner 72 identified by the identification learner 77 is read and used.

[0128] Then, if the difference ΔR2 is less than the predetermined value J2 (step S47: NO), it is determined that the distribution bias of the tip portions of the grinding surface 211 in the determination image 75 is less than a predetermined level, and it is determined to be normal (step S48). On the other hand, if the difference ΔR2 is equal to or greater than the predetermined value J2 (step S47: YES), it is determined that the distribution bias of the tip portions of the grinding surface 211 in the determination image 75 is equal to or greater than a predetermined level, and it is determined that the grinding wheel 21 has deteriorated abnormally (step S49). In this case, it is determined that the grinding wheel 21 has deteriorated to such an extent that dressing of the grinding surface 211 is required.

[0129] <Action and effect> The effects of the machined surface state estimating device of this embodiment will be described below. In this embodiment, when estimating the condition of the grinding surface 211 of the grinding wheel 21, the captured image 44 and the identification learner 77 are used to identify one extraction learner 71 and one classification learner 72 from a plurality of extraction learners 71 and a plurality of classification learners 72 that are stored in advance.

[0130] As a result, when an extraction learner 71 and a classification learner 72 corresponding to a grindstone 21 of the same type and the same particle diameter as the grindstone 21 to be estimated are set, it is possible to identify the extraction learner 71 and the classification learner 72. Furthermore, when an extraction learner 71 and a classification learner 72 corresponding to a grindstone 21 of the same type and the same particle diameter as the grindstone 21 to be estimated are not set, it is possible to identify an extraction learner 71 and a classification learner 72 corresponding to a grindstone 21 having a similar basic structure to the grindstone 21 to be estimated. According to this embodiment, in this way, it is possible to identify an extraction learner 71 and a classification learner 72 suitable for estimating the state of the grindstone 21 to be estimated from a plurality of extraction learners 71 and a plurality of classification learners 72 stored in advance.

[0131] Then, by using the identified extraction learning device 71 and the captured image 44 (specifically, the base image 44B and the latest image 44N), it is possible to obtain a converted base image and a converted latest image used to generate the judgment image 75. In addition, by using the identified classification learning device 72 and the judgment image 75, it is possible to classify the judgment image 75.

[0132] According to this embodiment, instead of storing separate learning devices 71, 72 for all types and particle sizes of grinding wheels 21, it is possible to use one of a plurality of pre-stored learning devices 71, 72 to estimate the state of the grinding wheel 21 to be estimated. This makes it possible to reduce the number of extraction learning devices 71 and classification learning devices 72 that are pre-stored, thereby shortening the time and reducing the effort required for setting the extraction learning device 71 and the classification learning device 72.

[0133] Furthermore, even if the learning devices 71, 72 corresponding to grindstones 21 of the same type and particle size as the grindstone 21 to be estimated have not been set, it is possible to use the existing learning devices 71, 72 corresponding to grindstones 21 having a similar basic structure to estimate the state of the grindstone 21. Therefore, in this case, it is possible to estimate the state of the grindstone 21 without newly setting the extraction learning device 71 or classification learning device 72 corresponding to the type of grindstone 21 to be estimated.

[0134] <Effects> According to this embodiment, in addition to the effects described in (1-2) and (1-4) above, the effect described in (2-1) below can be obtained.

[0135] (2-1) The time and effort required for setting up the trained extraction learning device 71 and the trained classification learning device 72 can be reduced. (Example of change) The above-described embodiments may be modified as follows: The above-described embodiments and the following modifications may be combined with each other to the extent that no technical contradiction occurs.

[0136] In the first embodiment, the particle diameter information on the particle diameter of the grindstone 21 to be estimated may be "average particle diameter", "grain size of the abrasive grains 212", or "manufacturer and model of the grindstone 21". In short, any information that can identify the "particle diameter of the grindstone 21" may be used as the particle diameter information 66. In this configuration, the relationship between the particle diameter information 66 and the particle diameter of the abrasive grains 212 may be stored in advance in the storage unit 56. Then, the "particle diameter of the grindstone 21" is identified from the relationship based on the particle diameter information 66, and the identified particle diameter is used in the process of generating the resized image 44R (step S11 in FIG. 9).

[0137] In the second embodiment, the classification learner 72 may be determined to correspond to abrasive grains of a predetermined particle size for each type of grindstone 21. In this configuration, the judgment image 75 to be used as input data for the classification learner 72 may be generated by changing the image size at the change ratio RG based on the captured image 44 capturing the grindstone 21 to be estimated.

[0138] In each embodiment, the following procedure can be adopted as the procedure for generating the judgment image 65, 75. First, a difference image showing the difference between the base image 44B and the latest image 44N is generated. Then, an extraction image is generated by using the difference image and a trained extraction learning device that has been trained to output an extraction image in which the tip portion of the grinding surface 211 is extracted from the difference image. This extraction image is set as the judgment image 65, 75.

[0139] In each embodiment, the process of converting the multiple latest images 44N and the multiple basic images 44B to grayscale separately (the process of step S12 in FIG. 9 and the process of step S31 in FIG. 13) can be omitted when generating the judgment images 65, 75. In this case, the process of converting the first image and the second image to grayscale can be omitted when creating the learning data 63, 73 of the extraction learners 61, 71. Also, the process of converting the first image and the second image to grayscale can be omitted when creating the learning data 64, 74 of the classification learners 62, 72.

[0140] When creating the learning data 63, 73 for the extraction learning device 61, 71, the procedure of dividing the first image and the second image into a plurality of divided images can be omitted. In this case, a large number of image sets, each pair consisting of a captured image 44 of the grinding surface 211 and an extracted image obtained by extracting the tip portion of the grinding surface 211 from the captured image 44, may be prepared, and these image sets may be used as the learning data 63, 73. As in the machined surface state estimation device of each embodiment, by dividing the first image and the second image into a plurality of divided images and then using them to create the learning data 63, 73, it becomes possible to create the learning data 63, 73 using an electronic control device 50 with a relatively low processing capacity.

[0141] The manner in which the judgment process of each embodiment (FIGS. 10 and 12) is executed can be changed arbitrarily, provided that it is possible to judge the occurrence of deterioration abnormality in the grinding wheel 21 to be estimated when the distribution bias of the tip portion of the grinding surface 211 in the judgment images 65, 75 is equal to or greater than a predetermined level.

[0142] The determination process can be performed, for example, as follows. As shown in Fig. 14, in the judgment process, a plurality of converted judgment images are generated by converting each of the judgment images 65, 75 into a bitmap image (step S50). Here, when the extraction learning device 61, 71 identified by the specific learning device 67, 77 is a learning device corresponding to a general grinding wheel, an image showing a portion in which the tip part of the grinding surface 211 has decreased in the latest converted image based on the converted base image is used as the judgment image 65, 75. On the other hand, when the extraction learning device 61, 71 identified by the specific learning device 67, 77 is a learning device corresponding to an ultra-abrasive grinding wheel, an image showing a portion in which the tip part of the grinding surface 211 has increased in the latest converted image based on the converted base image is used as the judgment image 65, 75.

[0143] Thereafter, an integrated value PP3 is calculated by integrating the number of pixels (hereinafter, specific pixels) indicating the tip portion of the grinding surface 211 for each rotation phase of the grindstone 21 based on the multiple conversion determination images (step S51).

[0144] Then, based on the integrated value PP3 calculated for each rotation phase, a degree of bias in distribution of the tip portions of the grinding surface 211 in the conversion determination image (more specifically, a single image showing the entire grinding surface 211 made up of a plurality of conversion determination images) is calculated (step S52). In the process of step S52, the difference ΔPP3 (=PP3max-PP3ave) between the average value PP3ave and the maximum value PP3max of the integrated value PP3 calculated for each rotation phase is calculated as the degree of bias RR3.

[0145] Then, it is determined whether the difference ΔPP3 is equal to or greater than a predetermined value J3 (step S53). As the predetermined value J3, a value capable of properly determining the occurrence of a deterioration abnormality of the grinding wheel 21 is determined in advance for each of the multiple extraction learning devices 61, 71 used in the determination process. In detail, a value (predetermined value J3) capable of properly determining the occurrence of a deterioration abnormality of the grinding wheel 21 is obtained for each of the multiple extraction learning devices 61, 71 used in the determination process based on the results of various experiments and simulations performed by the inventors, etc. Then, the predetermined value J3 is stored in advance in the storage unit 56 in a manner associated with the multiple extraction learning devices 61, 71 used in the determination process. In the process of step S53, the predetermined value J3 corresponding to the extraction learning device 61, 71 identified by the specific learning device 67, 77 is read and used.

[0146] If the difference ΔPP3 is less than a predetermined value J3 (step S53: NO), it is determined that the distribution bias of the tip portions of the grinding surface 211 in the conversion assessment image is less than a predetermined level, and it is determined to be normal (step S54). On the other hand, if the difference ΔPP3 is equal to or greater than the predetermined value J3 (step S53: YES), it is determined that the distribution bias of the tip portions of the grinding surface 211 in the conversion assessment image is greater than a predetermined level, and it is determined that a deterioration abnormality has occurred in the grinding wheel 21 (step S55).

[0147] In the above configuration, when a learning machine corresponding to a general grinding wheel is specified as the extraction learning machine 61, 71, if an abrasive grain 212 falls off from the grinding wheel 21 to be estimated, this fact will appear in the display of a pixel (the specific pixel) corresponding to the part where the abrasive grain 212 has fallen off in the conversion judgment image. Therefore, in this case, the state in which the abrasive grain 212 has fallen off in each part of the grinding surface 211 can be grasped based on the position and range of the specific pixel in the conversion judgment image.

[0148] In the above configuration, as deterioration of the grindstone 21 progresses, the distribution of the portions (i.e., specific pixels) where the abrasive grains 212 have fallen off across the entire grinding surface 211 becomes more biased. In this case, the variation in the integrated value PP3 calculated for each rotation phase becomes larger, as shown in an example in Fig. 15. In Fig. 15, the solid line indicates the integrated value PP3, and the dashed line indicates the average value PP3ave of the integrated value PP3.

[0149] According to the above configuration, it can be seen that the difference ΔPP3 between the average value PP3ave and the maximum value PP3max of the integrated value PP3 is large, which indicates that the amount of the abrasive grains 212 that have fallen off is largely biased at the tip portion of the grinding surface 211. In this case (step S53 in FIG. 14: YES), the above configuration can determine that the grindstone 21 has deteriorated abnormally.

[0150] On the other hand, in the above configuration, when a learning machine corresponding to an ultra-abrasive grinding wheel is specified as the extraction learning machine 61, 71, if a foreign matter adheres to the grinding surface 211 of the grinding wheel 21 to be estimated, this fact appears in the display of a pixel (hereinafter, a specific pixel) corresponding to the part where the foreign matter adheres in the conversion judgment image. Therefore, in this case, the adhesion state of the foreign matter on each part of the grinding surface 211 can be grasped based on the position and range of the specific pixel in the conversion judgment image.

[0151] In the above configuration, as deterioration of the grinding wheel 21 progresses, the distribution of foreign matter adhering portions (i.e., specific pixels) becomes more biased across the entire grinding surface 211. With the above configuration, it can be seen that the difference ΔPP3 between the average value PP3ave and the maximum value PP3max of the integrated value PP3 is large, which indicates that the bias in the amount of foreign matter adhering to the tip portion of the grinding surface 211 is large. With the above configuration, in such a case (step S53 in FIG. 14: YES), it can be determined that deterioration abnormality has occurred in the grinding wheel 21.

[0152] In the above configuration, the extraction learning devices 61 and 71 correspond to the first learning device that has already been trained. Also, in the above configuration, the extracted image output from the extraction learning devices 61 and 71 corresponds to machining surface information that indicates the state of the machining surface of the tool to be estimated.

[0153] In the judgment process of the first embodiment, in addition to the process of making the classification learner 62 output the classification result using the judgment image 65 as input data, a process of making the classification learner 62 output the classification result using the captured image 44 whose image size has been changed by the change ratio RG as input data may be executed. In addition, in the judgment process of the second embodiment, a process of making the classification learner 72 output the classification result using the captured image 44 as input data may be executed. In this case, the learning data 73 and the extraction learner 71 can be omitted in each embodiment. In each embodiment, the information output from the classification learners 62 and 72 is not limited to the classification result, and any information can be adopted as long as it is grinding surface information indicating the state of the grinding surface 211 of the grinding wheel 21 to be estimated. As such grinding surface information, for example, information indicating that the grinding surface 211 of the grinding wheel 21 to be estimated is in a good state, or a value indicating the degree of deterioration of the grinding surface 211 of the grinding wheel 21 to be estimated, can be adopted.

[0154] As the judgment process in each embodiment, a process for judging the state of the grinding surface 211 can be executed using a fractal analysis technique. The following method can be used to create the learning data 64, 74 for the classification learners 62, 72. First, a difference image showing the difference between the first image and the second image is generated. Then, the difference image is converted to grayscale. Then, an extraction learner that has been trained to output an extraction image in which the tip portion of the grinding surface 211 is extracted from the difference image and the grayscale difference image is used to generate the extraction image. This extraction image is used as the learning data 64, 74 for the classification learners 62, 72.

[0155] As the difference image constituting the judgment image 65, 75, a difference image showing the difference between the converted basic image and the latest converted image can be used, or a difference image showing the difference between the extracted image corresponding to the Nth previous image (for example, the image one time before) and the latest converted image can be used. In the above configuration, the timing when the grinding surface 211 was imaged N times before corresponds to the first period.

[0156] According to the above configuration, it is understood that there is a large bias in the amount of change in the state of the tip portion of the grinding surface 211 from the first period to the second period, because there is a bias in the distribution of the tip portion of the grinding surface 211 in the judgment images 65 and 75. In such a case, the above configuration makes it possible to judge that the grinding wheel 21 has deteriorated abnormally.

[0157] The structure of the lighting unit 43 can be changed as desired. For example, instead of irradiating a part of the second lamp 462 with the second illumination light 46L from a direction that forms "5 degrees" with respect to the grinding surface 211, the second lamp 462 may be irradiated with the second illumination light 46L from a direction that forms "10 degrees" with respect to the grinding surface 211. Alternatively, the second lamp 462 may be irradiated with the second illumination light 46L from a direction that forms "0 degrees" with respect to the grinding surface 211, that is, from a direction parallel to the tangent plane of the grinding surface 211. In order to clarify the boundary between the top surface of the grinding wheel 21 and the portion other than the top surface in the captured image 44, it is preferable that the angle between the optical axis of the part of the second lamp 462 and the tangent plane in the imaging target area AR of the grinding surface 211 is set to a predetermined angle range (0 degrees≦angle≦10 degrees).

[0158] The second irradiating section 46 may be configured solely by the second lamp 462 that irradiates the imaging target area AR of the grinding surface 211 with the second illumination light 46L from a direction along the grinding surface 211. The second irradiating section 46 can be configured solely by a second lamp 462 that irradiates the imaging target area AR of the grinding surface 211 with the second illumination light 46L from a direction parallel to the tangent plane of the grinding surface 211.

[0159] One of the first irradiating section 45 and the second irradiating section 46 may be omitted. The first emission color of the first illumination light 45L in the first irradiating unit 45 and the second emission color of the second illumination light 46L in the second irradiating unit 46 can be changed arbitrarily. In order to clarify the boundary between the top surface of the grinding wheel 21 and a portion other than the top surface in the captured image 44, it is preferable to determine colors different from the first emission color and the second emission color.

[0160] The camera unit 40 may be disposed to the side of the grindstone 21 or above the grindstone 21. In this case, the camera 41 may be disposed so as to capture an image of the grindstone 211 from a direction perpendicular to the grindstone 211. The illumination unit 43 may be disposed so as to illuminate the side or upper end of the grindstone 211, i.e., the area imaged by the camera 41.

[0161] The camera unit 40 may be integrally provided in the automatic grinding device 20 in a manner that makes it impossible to attach or detach it. According to this configuration, the operation of taking an image of the grinding surface 211 and the operation of storing the captured image 44 in the electronic control device 50 can be easily performed with the grinding wheel 21 attached to the automatic grinding device 20 without attaching or detaching the camera unit 40 to or from the automatic grinding device 20. In addition, it is possible to execute a series of combined automatic controls, including control of estimation and judgment of the state of the grinding surface 211, control of grinding of the workpiece by the automatic grinding device 20, and control of dressing of the grinding wheel 21 by the automatic grinding device 20. Therefore, by periodically executing the process of generating the judgment images 65, 75 and the judgment process, the grinding of the workpiece by the automatic grinding device 20 and dressing of the grinding wheel 21 can be executed while checking the state of the grinding surface 211 at an appropriate timing.

[0162] When the above-mentioned configuration is adopted, it is preferable to provide a shutter section that opens and closes the space between the camera unit 40 and the grindstone 21. According to this configuration, when performing grinding processing by the automatic grinding device 20, the shutter section can be closed to create a closed state in which the space between the camera unit 40 and the grindstone 21 is partitioned. This makes it possible to prevent foreign matter (such as cutting chips and oil) from adhering to the camera unit 40. On the other hand, when capturing an image of the grinding surface 211 by the camera unit 40, the shutter section can be opened to create an open state in which the space between the camera unit 40 and the grindstone 21 is not partitioned. This makes it possible to capture an image of the grinding surface 211 by the camera unit 40.

[0163] The task of storing the trained learners (specifically, the extraction learners 61, 71, the classification learners 62, 72, and the specific learners 67, 77) in the electronic control device 50 can be performed in any manner, such as the following (Task A) to (Task C).

[0164] (Task A) Using the automatic grinding device 20 on which the grinding wheel 21 to be estimated is mounted, learning data (specifically, learning data 63, 73, learning data 64, 74, learning data 68, 78) is obtained, and the learning device is trained by machine learning using the learning data.

[0165] (Work B) The manufacturer of the automatic grinding device 20 or the manufacturer of the grinding wheel 21 provides learning data corresponding to the grinding wheel 21 that is actually used. Then, the learning data is stored in the electronic control device 50 of the machined surface condition estimation device 30 used by the user, thereby allowing the learning device to perform machine learning.

[0166] (Operation C) When manufacturing the machined surface condition estimating device 30, the trained learning device is stored in the electronic control device 50. By utilizing the existing learning devices stored in the memory unit 56 and an image of the grinding surface 211 of the grinding wheel 21 to be estimated, a new extraction learning device 61, 71 and classification learning devices 62, 72 corresponding to the grinding wheel 21 to be estimated can be set using a transfer learning technique.

[0167] In this configuration, the extraction learners 61, 71 and classification learners 62, 72 identified by the specific learners 67, 77 can be used as existing learners. This makes it possible to execute transfer learning using a learner corresponding to the grinding wheel 21 that has the most similar basic structure to the grinding wheel 21 to be estimated, and therefore makes it possible to efficiently set the extraction learners 61, 71 and classification learners 62, 72 corresponding to the grinding wheel 21 to be estimated.

[0168] In the above configuration, an image acquired or generated by the machined surface state estimation device 30 can be used as an image related to the grinding surface 211 of the grinding wheel 21 to be estimated. This makes it possible to execute transfer learning by reusing the image (such as the captured image 44) used in estimating the state of the grinding surface 211, and therefore makes it possible to efficiently set the extraction learners 61, 71 and the classification learners 62, 72 corresponding to the grinding wheel 21 to be estimated.

[0169] According to the above configuration, when the extraction learners 61, 71 and classification learners 62, 72 corresponding to the grinding wheel 21 to be estimated have not yet been set, the extraction learners 61, 71 and classification learners 62, 72 can be newly set in a short time and with little effort.

[0170] The extraction learners 61 and 71 can be of a type of generative adversarial network other than "pix2pix", such as "cycleGAN" or "CGAN". In addition, the extraction learners 61 and 71 can be of an image generation model other than generative adversarial network (GAN), such as a variational autoencoder (VAE). In short, any image generation model that can generate an extracted image, which is an image in which only the tip portion of the grinding surface 211 of the grinding wheel 21 is extracted from an image related to the grinding surface 211, can be used as the extraction learners 61 and 71.

[0171] An object detection model called YOLO (You Only Look Once) can be used as the classification learners 62 and 72. In addition, a general forward propagation type neural network having a multi-layer structure, a support vector machine, or the like can be used as the classification learners 62 and 72. In short, any object detection model that can determine (estimate) the state of the grinding surface 211 can be used as the classification learners 62 and 72.

[0172] A general forward propagation neural network having a multi-layer structure, a support vector machine, etc. can be used as the specific learning devices 67, 77. In short, any object detection model that can distinguish (identify) the type of grinding wheel 21 can be used as the specific learning devices 67, 77.

[0173] The machined surface condition estimating device, the estimation program, and the machined surface condition estimating method according to the above-described embodiments can also be applied to a grinding wheel having a structure in which a large number of abrasive grains are fixed to the bottom surface of a cylindrical base material by a binder.

[0174] The machined surface condition estimating device, the estimation program, and the machined surface condition estimating method according to the above-described embodiments can be applied to a grinding wheel, as well as to any other tool having a structure in which abrasive grains are fixed, such as a wire tool in a wire saw.

[0175] <Additional Notes> Each of the above embodiments includes the configurations described in the following supplementary notes. [Appendix 1] A machined surface condition estimation device that estimates a state of a machined surface of a tool, comprising: a first learning device memory unit that stores a plurality of first learning devices that are respectively determined for a plurality of types of tools, the first learning devices being trained to use an image of the machined surface of the tool as first learning data and to output machined surface information indicating the state of the machined surface of the tool; a second learning device memory unit that stores a trained second learning device that uses an image of the machined surface of the tool as second learning data and has been trained to identify one of the plurality of first learning devices; an acquisition unit that acquires an estimated image that is an image of the machined surface of the tool to be estimated; an identification unit that uses the estimated image and the second learning device to identify one of the plurality of first learning devices; and an output unit that uses the estimated image and the first learning device identified by the identification unit to output the machined surface information indicating the state of the machined surface of the tool to be estimated.

[0176] [Appendix 2] The machined surface condition estimating device according to [Appendix 1], wherein the images constituting the first learning data, the images constituting the second learning data, and the estimated image are grayscale images.

[0177] [Appendix 3] The machined surface condition estimating device according to [Appendix 1] or [Appendix 2], wherein the tool is a grinding wheel for grinding. [Appendix 4] The machined surface condition estimating device described in [Appendix 3], wherein the first learning device is defined for each type of grinding wheel and for each particle diameter of abrasive grains in the same type of grinding wheel.

[0178] [Appendix 5] The machined surface condition estimating device described in [Appendix 3], wherein each of the first learning data and the second learning data is composed of an image related to the machined surface of the grinding wheel having abrasive grains of a predetermined particle diameter, and the machined surface condition estimating device comprises: an information storage unit that stores particle diameter information related to the particle diameter of the abrasive grains in the grinding wheel to be estimated; and a resizing unit that changes an image size of the estimated image by a change ratio expressed by the formula "G1 / G2" where the predetermined particle diameter is "G1" and the particle diameter in the particle diameter information is "G2", and the identifying unit and the output unit use the estimated image after resizing by the resizing unit.

[0179] [Appendix 6] The machining surface state estimating device includes an imaging unit that images the machining surface from a direction perpendicular to the machining surface (more specifically, a tangent plane), and an illumination unit that illuminates an area on the machining surface to be imaged by the imaging unit, wherein the illumination unit includes a first irradiation unit that irradiates the imaging target area with first illumination light having a first emission color from a direction perpendicular to the machining surface (more specifically, a tangent plane), and a second irradiation unit that irradiates the imaging target area with second illumination light having a second emission color different from the first emission color from a direction along the machining surface (more specifically, the tangent plane). The machining surface state estimating device described in any one of [Appendix 1] to [Appendix 5].

[0180] [Appendix 7] [Appendix 1] to [Appendix 6]. An estimation program that causes an electronic control device provided in the machining surface condition estimation device to execute the processing of the first learning device memory unit, the processing of the second learning device memory unit, the processing of the acquisition unit, the processing of the identification unit, and the processing of the output unit.

[0181] [Appendix 8] A machined surface condition estimating method for estimating a state of a machined surface of a tool, comprising: a first learning device storing step for storing a plurality of first learning devices, each of which is a trained first learning device that has been trained to use an image of the machined surface of the tool as first learning data and to output information indicating the state of the machined surface of the tool, and which is determined separately for a plurality of types of tools; a second learning device storing step for storing a trained second learning device that has been trained to use an image of the machined surface of the tool as second learning data and to identify one of the plurality of first learning devices; an acquisition step for acquiring an estimated image, which is an image of the machined surface of the tool to be estimated; an identification step for identifying one of the plurality of first learning devices using the estimated image and the second learning device; and an output step for outputting machined surface information indicating the state of the machined surface of the tool to be estimated, using the estimated image and the first learning device identified in the identification step. [Explanation of symbols]

[0182] 20...Automatic grinding device 21...Whetstone 211...Abrasive surface 212…Abrasive grain 22...Whetstone support part 23… Foundation 24…Table 25...Operation control unit 26…X-axis drive unit 27…Y-axis drive unit 28…Z-axis drive unit 29…Grinding wheel drive unit 30... Machined surface condition estimation device 40…Camera unit 41...Camera 43…Lighting department 44…Image 44R…Resized image 44B...Basic image 44N...Latest images 45...First irradiation section 451…First ramp 452...Reflective translucent member 45L…First lighting 46…Second irradiation section 461…Lighting dome 462…2nd Ramp 46L…Second illumination light 50...Electronic control device 51…CPU 52...ROM 53…RAM 54...Input section 55...Display section 56...Storage section 58…Specific part 60...Executing program 61,71…Extraction learning 62,72…Classification learner 63.73…Learning data 63B…Image before conversion 63A…Converted image 64,74…Learning data 65,75…Judgment image 66…Particle size information 67,77…Specific learning device 68,78…Learning data

Claims

1. A machined surface state estimation device that estimates a state of a machined surface of a tool, comprising: a first learning device storage unit that stores a plurality of first learning devices that are respectively determined for a plurality of types of tools, the first learning device being a trained first learning device that has been trained to use an image relating to a machining surface of the tool as first learning data and to output machining surface information that indicates a state of the machining surface of the tool; a second learning device storage unit that stores a trained second learning device that is trained to identify one of the plurality of first learning devices and that uses an image related to a machining surface of the tool as second learning data; An acquisition unit that acquires an estimated image which is an image relating to a machining surface of a tool that is an estimation target; an identification unit that identifies one of the first learning devices by using the estimated image and the second learning device; an output unit that uses the estimated image and the first learning device identified by the identification unit to output the machining surface information indicating a state of the machining surface of the tool to be estimated; A machined surface state estimation device comprising:

2. the images constituting the first learning data, the images constituting the second learning data, and the estimated image are grayscale images; The machined surface state estimating device according to claim 1 .

3. The tool is a grinding wheel. The machined surface state estimating device according to claim 1 or 2.

4. The first learning unit is determined for each type of grindstone and for each particle size of abrasive grains in the same type of grindstone. The machined surface state estimating device according to claim 3 .

5. Each of the first learning data and the second learning data is composed of an image relating to a processing surface of the grindstone having abrasive grains of a predetermined particle size, The machined surface state estimating device comprises: an information storage unit that stores particle size information regarding a particle size of the abrasive grains in the grindstone to be estimated; a resizing unit that changes an image size of the estimated image at a change ratio represented by a formula "G1 / G2" where "G1" is the predetermined particle size and "G2" is the particle size in the particle size information, The determination unit and the output unit use the estimated image after resizing by the resizing unit. The machined surface state estimating device according to claim 3 .

6. the machining surface state estimating device includes an imaging unit that images the machining surface from a direction perpendicular to the machining surface (specifically, a tangential plane), and an illumination unit that illuminates an imaging target area on the machining surface by the imaging unit, The illumination unit includes a first illumination unit that irradiates the imaging target area with first illumination light having a first emission color from a direction perpendicular to the processing surface (specifically, a tangent plane), and a second illumination unit that irradiates the imaging target area with second illumination light having a second emission color different from the first emission color from a direction along the processing surface (specifically, a tangent plane). The machined surface state estimating device according to claim 1 or 2.

7. 3. An estimation program that causes an electronic control device provided in the machining surface condition estimation device to execute the processing of the first learning device storage unit, the processing of the second learning device storage unit, the processing of the acquisition unit, the processing of the identification unit, and the processing of the output unit, which are provided in the machining surface condition estimation device according to claim 1 or 2.

8. A machined surface condition estimation method for estimating a condition of a machined surface of a tool, comprising: a first learning device storage step of storing a plurality of first learning devices, each of which is a trained first learning device that has been trained to use an image relating to a machining surface of the tool as first learning data and to output information indicating a state of the machining surface of the tool, the first learning device being determined separately for each of a plurality of types of tools; a second learning device storage step of storing a second learning device that has been trained to identify one of the plurality of first learning devices and that uses an image relating to a machining surface of the tool as second learning data; An acquisition step of acquiring an estimated image which is an image relating to a machining surface of a tool to be estimated; identifying one of the first learning devices using the estimated image and the second learning device; an output step of outputting machining surface information indicating a state of the machining surface of the tool to be estimated, using the estimated image and the first learning device identified in the identification step; The machining surface state estimation method includes:

Images