JP7345867B2 - Abrasive surface determination device, learning device, determination program, and abrasive surface determination method - Google Patents

Description

Application of Article 30, Paragraph 2 of the Patent Act Presented at the 2020 Student Web Lecture of the Kyushu Branch of the Society for Precision Engineering, held online on December 9, 2020 (website publication date: 2020) December 6th to 11th)

The present invention relates to an abrasive surface determination device for determining the abrasive surface of a whetstone, a learning device used in the device, a determination program used in the device, and an abrasive surface determination method for determining the abrasive surface of a whetstone. .

In grinding using a grinding device, the finish quality of the processed surface of a workpiece is influenced by the condition of the processed surface (so-called grinding surface) of a grindstone serving as a grinding tool. The condition of the abrasive surface is mainly determined by the shape and size of the surface portion of the abrasive grains that function as cutting edges, and the amount of protrusion of the abrasive grains. Therefore, various techniques are being researched to measure and evaluate the coordinates, shape, surface area, etc. of individual abrasive grains exposed on the abrasive surface of a whetstone.

In grinding, in which a workpiece is ground down minute by minute using abrasive grains that are dispersed on the abrasive surface of a grindstone, the distribution state of the abrasive grains is transferred to the machined surface of the workpiece, so that the workpiece becomes the desired shape. Finished in shape. Therefore, in order to perform grinding with high precision, it is important to appropriately manage the state of individual abrasive grains and the distribution state of abrasive grains on the abrasive surface (hereinafter referred to as surface texture).

Conventionally, methods have been proposed for evaluating the surface quality of the grinding surface of a grindstone using an electron microscope or a laser (see, for example, Non-Patent Documents 1 to 3).

Sakamoto, Sasaki, Kobayashi, Shimizu, "Identification of specific grinding resistance based on the grinding wheel working surface profile measured on the machine", Proceedings of the 2013 Japan Society for Precision Engineering Spring Conference, pp. 539-540 Kakino, Matsubara, Yamaji, Matsuda, Nakagawa, Hirogaki, Kida, "Research on on-the-machine measurement of grindstone working surface topography (first report)", Journal of the Japan Society for Precision Engineering, vol. 63, NO. 2, pp. 228-232 (1997) Shoji, Zhou, "Study on truing and dressing of diamond grinding wheels (2nd report)", Journal of the Japan Society of Precision Engineering, vol. 55, NO. 12, pp. 2267-2272 (1989)

In an apparatus that uses an electron microscope, the surface properties of the abrasive surface can be precisely observed using the electron microscope. However, since the observation range of an electron microscope is extremely narrow, it takes a long time to observe the surface texture over the entire surface of the abrasive surface using an electron microscope.

On the other hand, with a device that uses a laser to evaluate the surface quality of an abrasive surface, the shape of an abrasive surface can be measured with high precision and relatively easily. However, since the information that can be obtained by measurement using a laser is the unevenness profile of the abrasive surface, it is difficult to determine whether the acquired information is information about the part that is in contact with the abrasive grains, or information about the part that is in contact with the binder that fixes the abrasive grains. difficult to distinguish. This is a factor that hinders improvement in evaluation accuracy when evaluating the surface properties of the abrasive surface. Furthermore, in order to improve the accuracy of measurement using a laser, it is necessary to reduce the spot diameter of the laser. If the spot diameter is made smaller, the measurement range becomes narrower accordingly. Therefore, in this case, it takes a long time to measure the surface quality of the abrasive surface and to evaluate whether the surface is normal.

An abrasive surface determination device for solving the above problems is an abrasive surface determination device that determines the state of the abrasive surface of a whetstone having a structure in which a plurality of abrasive grains are fixed by a binder. an input unit that stores, as a determination image, an image containing information regarding a portion of the grinding surface of the grindstone to be determined that extends in the grinding direction of the grindstone; a storage unit that stores a learning device trained by training data consisting of an image that is an image that has been used and that includes information regarding a portion of the abrasive surface that extends in the grinding direction in the specific usage state; and inputs the judgment image. a difference output unit that outputs, as data, an index value of the difference between the usage state of the abrasive surface corresponding to the judgment image and the specific usage state from the trained learning device stored in the storage unit; Be prepared.

A learning device for solving the above problem is the learned learning device included in the abrasive surface determination device.
A determination program for solving the above problem is a determination program that causes an electronic control device included in the abrasive surface determination device to execute the processing of each unit included in the abrasive surface determination device.

The abrasive surface determination method for solving the above problems is a method for determining the condition of the abrasive surface of a whetstone having a structure in which a plurality of abrasive grains are fixed by a binder, and the abrasive surface of the abrasive surface of the whetstone to be determined is an image that is captured and includes information about a portion of the grinding surface of the grindstone to be judged that extends in the grinding direction of the grindstone, and uses the image as a judgment image to be used as an input section of the grinding surface judgment device. and acquiring an image that is a captured image of the abrasive surface in a specific usage state and that includes information regarding a portion of the abrasive surface in the specific usage state that extends in the grinding direction; a storage step of storing the learning device in a storage unit of the abrasive surface determination device after the learning device is trained using teacher data consisting of images obtained by the training; and a learning step in which the learning device is stored in the storage unit using the determination image as input data. and a difference output step of outputting an index value of a difference between the use state of the abrasive surface corresponding to the determination image and the specific use state from the learned learning device.

In the abrasive surface determination device, learning device, determination program, and abrasive surface determination method described above, an image of the abrasive surface in a specific state of use is prepared, and a trained image that has been trained in advance using the same image as training data is prepared. A learning device will be prepared. Therefore, by using a trained learning device, it is possible to determine in a short time and with high accuracy whether the grinding surface of the target grindstone is in a specific usage state.

According to the present invention, the condition of the abrasive surface can be determined in a short time and with high accuracy.

1 is a schematic diagram showing a schematic configuration of an automatic grinding device to which an abrasive surface determination device of an embodiment is applied. FIG. 3 is a cross-sectional view showing the whetstone cover in an open state and the structure around it. FIG. 1 is a block diagram showing a schematic configuration of an abrasive surface determination device. 5 is a flowchart showing the procedure for executing imaging processing. A schematic diagram showing images forming a judgment image and teacher data. An explanatory diagram for explaining a learning mode of a learning device. An explanatory diagram for explaining a generation error output from a learning device. 5 is a flowchart illustrating the execution procedure of determination processing. FIG. 3 is a front view showing display contents of a display device. (a) to (c) Explanatory diagrams showing changes in usage conditions of the abrasive surface. FIG. 7 is an explanatory diagram for explaining the generation error ΔP1 output from the learning device 51A. A time chart showing changes in the generation error ΔP1. FIG. 5 is an explanatory diagram for explaining the generation error ΔP2 output from the learning device 51B. A time chart showing changes in the generation error ΔP2. FIG. 7 is an explanatory diagram for explaining a learning mode of a learning device according to an abrasive surface determination device according to another embodiment. An explanatory diagram for explaining input data of the learning device. FIG. 7 is an explanatory diagram for explaining a learning mode of a learning device in an abrasive surface determination device according to another embodiment. FIG. 7 is a perspective view showing an imaging unit used in an abrasive surface determination device according to another embodiment.

An embodiment of an abrasive surface determination device, a learning device, a determination program, and an abrasive surface determination method will be described below.
As shown in FIG. 1, the abrasive surface determination device 30 of this embodiment is applied to a numerical control (NC) type automatic grinding device 20.

The automatic grinding device 20 has a substantially disc-shaped grindstone 21 as a grinding tool. The whetstone 21 has a structure in which a large number of abrasive grains are dispersed and fixed to the outer peripheral surface of a base material using a binder. The grindstone 21 is rotatably supported. A rotation drive unit 22 having, for example, a servo motor is connected to the grindstone 21 . The automatic grinding device 20 is capable of relatively moving the grindstone 21 and the workpiece. In the automatic grinding device 20, the upper surface of the workpiece is ground by the grinding surface 23 of the grindstone 21 by moving the grindstone 21 and the workpiece relative to each other while rotating the grindstone 21 and bringing the grindstone 21 into contact with the workpiece. Ru.

The automatic grinding device 20 has a dresser 24 for dressing the grindstone 21. In the automatic grinding device 20, the dressing of the grindstone 21 is performed by the dresser 24 at an arbitrary timing. This dressing keeps the grinding surface 23 of the grindstone 21 in good condition.

As shown in FIGS. 1 and 2, the automatic grinding device 20 has a grindstone cover 25. As shown in FIGS. The grindstone cover 25 has a square box shape with an open bottom. The whetstone 21 is housed inside the whetstone cover 25 with only the lower part of the whetstone 21 exposed to the outside.

An opening 27 that communicates between the inside and outside of the whetstone cover 25 is provided in an upper wall portion 26 that forms the upper wall of the whetstone cover 25 . The grindstone cover 25 is provided with a plate-shaped shutter section 28 that opens and closes the opening 27, and an actuator (hereinafter referred to as a shutter operation section 29) that moves the shutter section 28 back and forth. By moving the shutter section 28 in one direction through the operation control of the shutter operating section 29, the opening 27 of the grindstone cover 25 is closed by the shutter section 28 (the state shown in FIG. 1). On the other hand, by moving the shutter section 28 in the other direction through the operation control of the shutter operation section 29, the shutter section 28 is brought into the open state (the state shown in FIG. 2) where it is retracted from the opening 27 of the grindstone cover 25. In this embodiment, the opening 27 of the grindstone cover 25 can be switched between a closed state and an open state by controlling the operation of the shutter operating section 29.

A camera 31 serving as an imaging section is attached to the grindstone cover 25. The camera 31 is attached in such a manner that it can image the grinding surface 23 of the grindstone 21 housed inside the grindstone cover 25 through the opening 27 of the grindstone cover 25 . As the camera 31, an area scan camera is adopted. In this embodiment, the camera 31 captures an image of the grinding surface 23 of the grindstone 21 at a predetermined magnification, and the captured image (more specifically, image data) is output to the electronic control device 40, which will be described later. There is. In this embodiment, the camera 31 is integrally provided with an illuminator (not shown) for illuminating the grinding surface 23 of the grindstone 21 when the camera 31 takes an image.

The camera 31 is attached to the grindstone cover 25 via an X-axis drive section 32, a Y-axis drive section 33, and a Z-axis drive section 34. These X-axis drive section 32, Y-axis drive section 33, and Z-axis drive section 34 have a slide mechanism and an actuator (a servo motor in this embodiment) that operates the slide mechanism. In this embodiment, the camera 31 can be moved relative to the grindstone cover 25 in the X direction (the left-right direction in FIG. 1) through the operation control of the X-axis drive unit 32. Furthermore, by controlling the operation of the Y-axis drive unit 33, it is possible to move the camera 31 relative to the grindstone cover 25 in the Y direction (vertical direction in FIG. 1). Furthermore, by controlling the operation of the Z-axis drive unit 34, it is possible to move the camera 31 relative to the grindstone cover 25 in the Z direction (the depth direction of the paper in FIG. 1).

The automatic grinding device 20 has an electronic control device 40. As shown in FIG. 3, the electronic control device 40 includes a CPU 41, a ROM 42, a RAM 43, and a storage section 44 that stores various programs and data. The storage unit 44 is constituted by a nonvolatile memory such as a hard disk drive or solid state drive that can be written to and read from at any time. The electronic control device 40 is configured to perform various calculation processes based on various programs and data. Based on the calculation results, the electronic control unit 40 performs numerical control for the grinding process, various controls for dressing the grindstone 21, and drives each drive unit 22, 32 for determining the usage state of the grinding surface 23 of the grindstone 21. , 33, 34 and the camera 31. In this embodiment, a control section including a CPU 41, a ROM 42, and a RAM 43 corresponds to a difference output section and an image output section.

The automatic grinding device 20 has a display device 35. This display device 35 is connected to an electronic control device 40. The display device 35 is adapted to display the determination result of the usage state of the grinding surface 23 of the grindstone 21.

Here, the abrasive surface determination device 30 of the present embodiment uses an image of the abrasive surface 23 of the whetstone 21 (judgment image 50) as input data, and uses the two trained learning devices 51A and 51B to determine the abrasive surface 23. This is a device that determines usage status. Specifically, one learning device 51A is used to determine whether or not the following [determination condition A] is satisfied, and the other learning device 51B is used to determine whether the following [determination condition B] is satisfied. It is determined whether or not.

[Judgment Condition A] The usage state of the abrasive surface 23 to be determined is equivalent to the usage state of the abrasive surface 23 immediately after dressing has been properly performed (specific usage state A).
[Judgment Condition B] The usage state of the abrasive surface 23 to be determined is equivalent to the usage state of the abrasive surface 23 at a timing suitable for starting execution of dressing (specific usage state B).

The storage unit 44 of the electronic control device 40 has a storage area for storing various execution programs 52 and a storage area for storing teacher data 53A and 53B used for machine learning of the learning devices 51A and 51B. The storage unit 44 has a storage area that stores learned learning devices 51A and 51B (more specifically, learning data created through execution of machine learning). The storage unit 44 also stores a storage area serving as an input unit for storing an image (judgment image 50) taken of the abrasive surface 23 to be determined, and determination result data 54 indicating the result of determining the state of the abrasive surface 23. It has a storage area.

The execution program 52 includes a program for causing the learning devices 51A and 51B to execute machine learning and to generate learning data that is data indicating the execution results of the machine learning. The execution program 52 also includes a determination program for executing a process (determination process) related to determining the usage state of the grinding surface 23 of the grindstone 21 and generating determination result data 54 indicating the execution result of the determination process. There is.

The teacher data 53A and 53B are data for performing machine learning on the learning devices 51A and 51B so as to acquire the ability to determine the usage state of the abrasive surface 23.
In this embodiment, in order to obtain the teacher data 53A, 53B and the determination image 50, an imaging process is performed in which the camera 31 images the grinding surface 23 of the grindstone 21.

Hereinafter, the execution procedure of the imaging process will be described with reference to FIG. 4. Note that FIG. 4 conceptually shows the execution procedure of the imaging process. A series of processes shown in the flowchart of FIG. 4 are executed by the electronic control device 40.

As shown in FIG. 4, in this process, first, the shutter section 28 is opened by controlling the operation of the shutter operation section 29 (step S11).
Thereafter, the camera 31 captures an image of the abrasive surface 23 while rotating the grindstone 21 in a manner suitable for capturing an image of the abrasive surface 23 (step S12). In this embodiment, each pattern that enables the camera 31 to efficiently image the entire surface of the grinding surface 23 is set in advance as the rotational drive pattern of the grindstone 21, the position control pattern of the camera 31, and the imaging pattern of the camera 31. It is determined and stored in the electronic control unit 40. As a method for storing each of the above-mentioned patterns in the storage unit 44, the following (method A) and (method B) are assumed. (Method A) After the user of the automatic grinding device 20 attaches the grindstone 21 that will actually be used, the automatic grinding device 20 is operated to find each of the above-mentioned patterns suitable for the same grindstone 21 and stored in the electronic control device 40. . (Method B) Based on the results of various experiments and simulations, the manufacturer of the automatic grinding device 20 has determined in advance each of the above patterns that make it possible to efficiently image the entire surface of the grinding surface 23 of the grinding wheel 21. . The data related to these patterns are then stored in the automatic grinding device 20 before shipment or in the automatic grinding device 20 installed in the user's factory.

In the process of step S12, the operation control of the rotation drive unit 22 is executed based on the rotation drive pattern of the grindstone 21, and the X-axis drive unit 32, the Y-axis drive unit 33, Operation control of the Z-axis drive section 34 is executed. Further, based on the imaging pattern of the camera 31, the camera 31 performs imaging of the abrasive surface 23. Then, the electronic control device 40 captures image data of the abrasive surface 23 captured by the camera 31 and stores it in the storage unit 44 . Specifically, the camera 31 images the abrasive surface 23 in multiple times. Then, by combining the plurality of captured images arranged vertically and horizontally in a predetermined order, a single image showing the entire abrasive surface 23 (judgment image 50 or teacher data 53A, 53B) is generated. The electronic control device 40 stores the image generated in this manner in the storage unit 44. Although not shown, a unique pattern created by the abrasive grains and the binder appears in this image.

When the imaging of the abrasive surface 23 by the camera 31 is completed, the shutter section 28 is closed by controlling the operation of the shutter operation section 29 (step S13).
The imaging process shown in FIG. 4 is executed on the condition that the grinding process of the workpiece by the grindstone 21 and the dressing of the grindstone 21 are not executed. Note that when performing the imaging process, prior to the execution, an operation of removing cutting oil from the surface of the abrasive surface 23 by blowing air onto the abrasive surface 23 to be imaged is performed as necessary. Specifically, the imaging process is executed as follows.

When acquiring the teacher data 53A used for machine learning of the learning device 51A, the grindstone 21 [A] in a state where optimal dressing has been performed (specific usage state A) is prepared, and the same grindstone 21 [A] is prepared. [A] is attached to the automatic grinding device 20. Then, in this state, the electronic control device 40 executes the imaging process. The image data acquired through this imaging process is stored in the storage unit 44 as teacher data 53A used for machine learning by the learning device 51A. In the present embodiment, image data regarding a plurality of grindstones 21[A] are acquired, and the teacher data 53A is composed of the image data. In this embodiment, the grindstone 21 [A] is one that has been determined by an operator such as a skilled worker to be in a state where optimal dressing has been performed. As the grindstone 21 [A], it is also possible to use a grindstone whose grinding surface 23 has been analyzed by an analysis device and which has been determined to be in a state where optimal dressing has been performed.

When acquiring the teacher data 53B used for machine learning of the learning device 51B, the grindstone 21[B] in the state most suitable for starting the dressing execution (specific usage state B) is prepared, and the grindstone 21[B] is 21 [B] is attached to the automatic grinding device 20. Then, in this state, the electronic control device 40 executes the imaging process. The image data acquired through this imaging process is stored in the storage unit 44 as teacher data 53B used for machine learning by the learning device 51B. In the present embodiment, image data regarding a plurality of grindstones 21[B] are acquired, and the teacher data 53B is composed of the image data. In this embodiment, the grindstone 21 [B] that is determined by an operator such as a skilled worker to be in the most suitable state for starting execution of dressing is employed. As the grindstone 21 [B], it is also possible to use a grindstone whose grinding surface 23 is analyzed by an analysis device and which is determined to be in the most suitable state for starting the dressing.

During the period in which dressing is performed by the automatic grinding device 20, the dressing is stopped each time the total dressing period reaches a predetermined period (a predetermined time, a period in which the grindstone 21 rotates a predetermined number of times, etc.). Imaging processing is performed on the condition that The image data acquired through this imaging process is stored in the storage unit 44 as a determination image 50.

During the period in which the automatic grinding device 20 executes the grinding process of the workpiece, the grinding process is performed every time the total execution period of the grinding process reaches a predetermined period (a predetermined time, a period in which machining of a predetermined number of workpieces is completed, etc.). Imaging processing is executed on the condition that execution of is stopped. The image data acquired through this imaging process is stored in the storage unit 44 as a determination image 50.

In this embodiment, a generative adversarial network is employed as the two learning devices 51A and 51B. In this embodiment, machine learning of the learning devices 51A and 51B is performed based on the teacher data 53A and 53B stored in the storage unit 44. Specifically, a learning process in which one of the plurality of image data forming the teacher data 53A is extracted and the same image data is used as input data to perform machine learning on the learning device 51A is repeated for each image data forming the teacher data 53A. executed. Further, a learning process of extracting one of the plurality of image data constituting the teacher data 53B and causing the learning device 51B to perform machine learning using the same image data as input data is repeatedly executed for each image data constituting the teacher data 53B. be done. Learning data created through the execution of machine learning is then stored in the storage unit 44. In this embodiment, an execution program 52 is constructed and stored in the storage unit 44 so that a series of processes related to machine learning by the learning devices 51A and 51B are automatically executed. In this embodiment, the process of executing machine learning of the learning devices 51A and 51B in this manner corresponds to a storage process.

As conceptually shown in FIG. 6, the machine learning of each learning device 51A, 51B is performed using the same input data and generated image data generated by the learning devices 51A, 51B based on the same input data. The process is executed in such a manner that data representing the image (image data A) is generated. Then, the learned learning devices 51A and 51B output values corresponding to the difference between the input data and the generated image data as generation errors ΔP1 and ΔP2.

As a result, when the learning devices 51A and 51B are properly trained, when the determination process for determining the usage state of the target grindstone 21 (specifically, its grinding surface 23) is executed, the learning device 51A , 51B output the following values (generation errors ΔP1, ΔP2).

In the determination process using the learning device 51A, if the grindstone 21 to be determined is in the state immediately after optimal dressing (specific usage state A), the input data (determination image 50) and the learning The generated image data generated by the device 51A is substantially the same. Therefore, in this case, the generation error ΔP1 output from the learning device 51A becomes a very small value (for example, "0").

On the other hand, for example, if the dressing execution time is insufficient or if the grinding process by the grindstone 21 is executed after the dressing is completed, the actual usage state of the grindstone 21 to be determined will not be the specific usage state A. In this case, as shown in FIG. 7, the input data (judgment image 50) and the generated image data do not match, so the generation error ΔP1 output from the learning device 51A is the difference between the input data and the generated image data. It becomes a "positive value" according to the difference. Moreover, in this case, the larger the difference between the actual usage state of the grindstone 21 and the specific usage state A, the larger the difference between the input data and the generated image data. The error ΔP1 also becomes a large value.

Therefore, in the determination process using the learning device 51A, if the generation error ΔP1 is a value smaller than the predetermined level (determination value J1), the actual usage state of the grinding surface 23 of the whetstone 21 to be determined is "specific. It can be determined that the usage state is "A".

In the determination process using the learning device 51B, if the grindstone 21 to be determined is in the state most suitable for starting dressing (specific usage state B), the input data and the learning device 51B generate the The generated image data becomes substantially the same (see FIG. 6). Therefore, in this case, the generation error ΔP2 output from the learning device 51B becomes a very small value (for example, "0").

On the other hand, for example, if the usage time of the grindstone 21 after dressing is too short or the same usage time is too long, the actual usage state of the grindstone 21 to be determined will not be the specific usage state B. In this case, as shown in FIG. 7, the input data (judgment image 50) and the generated image data do not match, so the generation error ΔP2 output from the learning device 51B is the difference between the input data and the generated image data. It becomes a "positive value" according to the difference. Moreover, in this case, the larger the difference between the actual usage state of the grindstone 21 and the specific usage state B, the larger the difference between the input data and the generated image data. The error ΔP2 also becomes a large value.

Therefore, in the determination process using the learning device 51B, if the generation error ΔP2 is a value smaller than the predetermined level (determination value J2), the actual usage state of the grinding surface 23 of the whetstone 21 to be determined is "specific. It can be determined that the usage state is "B".

Hereinafter, the determination process for determining the usage state of the grinding surface 23 of the grindstone 21 to be determined will be specifically described. FIG. 8 shows the execution procedure of the determination process. A series of processes shown in the flowchart of FIG. 8 are executed by the electronic control device 40 as processes at predetermined intervals.

As shown in FIG. 8, in this process, first, an image (judgment image 50) of the grinding surface 23 of the grindstone 21 attached to the automatic grinding device 20 is acquired (step S21). Specifically, the imaging process (FIG. 4) is executed, and the image data acquired by the imaging process is stored in the storage unit 44 as the determination image 50. In this embodiment, the process in step S21 corresponds to an input process.

Thereafter, using the determination image 50 stored in the storage unit 44 as input data, the generation error ΔP1 is output from the trained learning device 51A stored in the storage unit 44 (step S22). This generation error ΔP1 is stored in the storage unit 44 as determination result data 54. Further, using the determination image 50 stored in the storage unit 44 as input data, the generation error ΔP2 is output from the trained learning device 51B stored in the storage unit 44 (step S23). This generation error ΔP2 is stored in the storage unit 44 as determination result data 54. In this embodiment, the process of step S22 and the process of step S23 correspond to the difference output process.

Then, based on the generation errors ΔP1 and ΔP2, it is determined whether the usage state of the grinding surface 23 of the grindstone 21 to be determined is in a specific usage state A or B (step S24).
Specifically, the generation error ΔP1 is compared with a predetermined determination value J1. If the generation error ΔP1 is less than or equal to the determination value J1, it is determined that the abrasive surface 23 is in a specific usage state A. On the other hand, if the generation error ΔP1 is less than the determination value J1, it is determined that the abrasive surface 23 is not in the specific usage state A. Furthermore, the generation error ΔP2 is compared with a predetermined determination value J2. If the generation error ΔP2 is less than or equal to the determination value J2, it is determined that the abrasive surface 23 is in a specific usage state B. On the other hand, if the generation error ΔP2 is less than the determination value J2, it is determined that the abrasive surface 23 is not in the specific usage state B. The results of these determinations are stored in the storage unit 44 as determination result data 54.

Thereafter, the determination result data 54 is output to the display device 35, and the result of the determination is displayed on the display device 35 (step S25).
As shown in FIG. 9, the following content is displayed on the display device 35. The display device 35 displays changes in the production error ΔP1 and changes in the production error ΔP2. Further, in the area AR1 of the display device 35, a character indicating that the optimal dressing is being executed (specific usage state A) when it is determined that the generation error ΔP1 is less than or equal to the determination value J1. Information (in detail, dressing OK) is displayed. Furthermore, in the area AR2 of the display device 35, when it is determined that the generation error ΔP2 is less than or equal to the determination value J2, it is shown that the state is the most suitable for starting the dressing execution (specific usage state B). Text information (more specifically, dress timing) is displayed.

Hereinafter, the effects of the abrasive surface determination device 30 of this embodiment will be explained.
FIG. 10 shows an example of how the usage state of the grinding surface 23 changes when the grinding process is performed using the grindstone 21. As shown in FIG. Note that FIG. 10 shows a part of the surface of the abrasive surface 23 in an enlarged manner.

As shown in FIG. 10(a), immediately after proper dressing, the outermost peripheral surface of the abrasive surface 23 is composed only of the tip surfaces of the abrasive grains 70. As shown in FIG. 10(b), after the grinding process by the grinding wheel 21 has been executed to some extent, the usage condition of the grinding surface 23 changes and a portion 71 other than the abrasive grains 70 (mainly , a so-called bond tail formed by the binder appears. These portions 71 other than the abrasive grains 70 tend to extend in the grinding direction from the rear portion of the abrasive grains 70 in the grinding direction. As shown in FIG. 10(c), as the total execution time of the grinding process by the grinding wheel 21 becomes longer, as the usage state of the grinding surface 23 progresses, the portion 71 other than the abrasive grains 70 on the outermost peripheral surface of the grinding wheel 21 increases. However, the abrasive grains 70 arranged in the grinding direction are extended in a manner that connects them. In this way, the degree of change in the usage state of the abrasive surface 23 appears in the portion of the outermost peripheral surface of the abrasive surface 23 that extends in the grinding direction.

In this embodiment, an image of the entire abrasive surface 23 is used as the image data that constitutes the determination image 50 and the teacher data 53A, 53B. Therefore, the image data constituting the judgment image 50 and the teacher data 53A, 53B includes information regarding the portion of the outermost peripheral surface of the abrasive surface 23 that extends in the grinding direction as information representing the usage state of the abrasive surface 23. I can say that it is. Therefore, in this embodiment, when the learning devices 51A and 51B perform machine learning, "the portion extending in the grinding direction on the outermost surface of the abrasive surface 23" can be extracted as a feature amount, and based on this feature amount, the learning devices 51A and 51B 51B machine learning can be advanced.

Then, the generation error ΔP1 output from the trained learning device 51A has the following value.
As shown in "state A1" in FIG. 11, when the abrasive surface 23 is in the state immediately after optimal dressing has been performed (specific usage state A), the state generated by the input data and the learning device 51A is The generated image data will be almost the same as the generated image data. Therefore, in this case, the generation error ΔP1 becomes a very small value. In this case, the generation error ΔP1 becomes the smallest value.

As shown in "state B1" in FIG. 11, when the cutting process by the grinding wheel 21 has been performed and the usage state of the grinding surface 23 has changed, a state in which a difference has occurred between the input data and the generated image data Therefore, the generation error ΔP1 becomes a "positive value".

As shown in "state C1" in FIG. 11, when the total execution time of the grinding process by the grinding wheel 21 becomes longer and the usage state of the grinding surface 23 is changing, the difference between the input data and the generated image data is large, the generation error ΔP1 becomes a large "positive value."

From this, the generation error ΔP1 output from the trained learning device 51A basically changes as follows according to the execution of dressing.
As shown in FIG. 12, the generation error ΔP1 output from the learning device 51A gradually becomes smaller after the execution of dressing is started (time t11). The generation error ΔP1 becomes the smallest at the timing (time t12) when the abrasive surface 23 is in a state where optimal dressing has been performed (specific usage state A). At this time, the generation error ΔP1 becomes less than the determination value J1. Note that when cutting is performed using the grindstone 21 after the dressing is completed, the production error ΔP1 gradually increases in value as the cutting process is performed.

In this embodiment, as shown in FIG. 9, the transition of such generation error ΔP1 is displayed on the display device 35. Therefore, by visually observing the generation error ΔP1 displayed on the display device 35, it is possible to grasp how close the actual usage state of the abrasive surface 23 is to the specific usage state A. Moreover, in this embodiment, when it is determined that the generation error ΔP1 is equal to or less than the determination value J1, text information indicating that the optimal dressing has been executed is displayed in the area AR1 of the display device 35. Is displayed. By visually observing this display, it is possible to easily understand that the optimal dressing has been performed.

On the other hand, the generation error ΔP2 output from the trained learning device 51B has the following value.
As shown in "state A2" in FIG. 13, when the total execution time of the grinding process by the grinding wheel 21 is relatively short and the usage state of the grinding surface 23 has not changed much, the state is generated using the input data and the learning device 51B. There is a difference between the generated image data and the generated image data. In this case, the generation error ΔP2 becomes a "positive value."

As shown in "state B2" in FIG. 13, when the abrasive surface 23 is in the most suitable state for starting the dressing (specific usage state B), the input data and the generated image data are approximately the same. Therefore, the generation error ΔP2 becomes a very small value. In this case, the generation error ΔP1 becomes the smallest value.

As shown in "state C2" in FIG. 13, when the total execution time of the grinding process by the grinding wheel 21 is relatively long and the usage state of the grinding surface 23 is changing, the input data and the generated image data are Since the difference is large, the generation error ΔP2 becomes a large "positive value."

From this, the generation error ΔP2 output from the trained learning device 51B basically changes as follows according to the execution of the grinding process by the grindstone 21.
As shown in FIG. 14, the generation error ΔP2 output from the learning device 51B gradually becomes smaller after the completion of dressing and the execution of cutting by the grindstone 21 is started (before time t21). Then, the abrasive surface 23 becomes the smallest at the timing when it becomes the most suitable state for starting the dressing (specific usage state B) (time t21). Note that when the execution of dressing is started, the generation error ΔP2 gradually increases as the dressing is executed (after time t21).

In this embodiment, as shown in FIG. 9, the transition of such generation error ΔP2 is displayed on the display device 35. Therefore, by visually observing the generation error ΔP2 displayed on the display device 35, it is possible to grasp how close the actual usage state of the abrasive surface 23 is to the specific usage state B. Moreover, in this embodiment, when it is determined that the generation error ΔP2 is equal to or less than the determination value J2, text information indicating that the state is most suitable for starting the dressing is displayed in the area AR2 of the display device 35. is displayed. By visually observing this display, it can be easily understood that the most suitable state for starting the dressing has been reached.

In this embodiment, since the determination result is displayed on the display device 35 in this manner, the operator can easily grasp the usage status of the abrasive surface 23 at any given time by visually observing the displayed content on the display device 35. be able to.

According to this embodiment, the effects described below can be obtained.
(1) Prepare images taken of the abrasive surface 23 in specific usage states A and B, and prepare in advance trained learning devices 51A and 51B that have learned the same images as teacher data 53A and 53B. I made it. Therefore, by using the trained learning devices 51A and 51B, it is possible to determine with high accuracy whether the grinding surface 23 of the grindstone 21 to be determined is in the specific usage state A or B. Moreover, compared to a device that evaluates the usage state of the abrasive surface 23 by closely observing or measuring the abrasive surface 23, the abrasive surface 23 of the whetstone 21 to be determined is in a specific usage state A or B. It is possible to determine in a short time what is happening.

(2) Obtain an image of the grinding surface 23 of the grinding wheel 21 [A] in a state where optimal dressing has been performed (specific usage state A), and use the image as training data 53A for machine learning of the learning device 51A. I decided to use it for. Therefore, by using the trained learning device 51A, it is possible to determine in a short time and with high accuracy that the dressing of the grinding surface 23 of the grindstone 21 to be determined has been appropriately performed.

(3) Obtain an image of the grinding surface 23 of the grinding wheel 21 [B] in the state most suitable for starting execution of dressing (specific usage state B), and use the image as training data 53B to machine the learning device 51B. I decided to use it for learning. Therefore, by using the trained learning device 51B, it is possible to determine in a short time and with high accuracy whether the grinding surface 23 of the target grindstone 21 is in a state suitable for starting the dressing operation. .

(4) With a simple configuration that uses a single image of the entire abrasive surface 23, it is possible to determine that the abrasive surface 23 is in a specific state of use.
Here, deterioration of the grinding wheel 21 (specifically, the grinding surface 23) does not appear over the entire area of the grinding surface 23, but occurs in a region (specific region) that has a characteristic structure. Based on this point, it is conceivable to have a learning device learn the deterioration mode of the abrasive surface 23 in such a specific area, and then use the learning device to determine the usage state of the abrasive surface 23. Thereby, the usage status of each part of the abrasive surface 23 can be determined in detail. However, in this case, during the process of using the grindstone 21, it is necessary to sequentially observe the same area (specific area) of the grinding surface 23 and grasp the state of change. Further, a complicated task of cutting out an image of a specific region from an image of the abrasive surface 23 as teacher data for learning the learning device is also required.

In this embodiment, since it is only necessary to prepare a single image showing the entire abrasive surface 23 when determining the usage state of the abrasive surface 23, such complicated work is unnecessary, and the abrasive surface 23 can be sharpened with very simple work. The state of use of the surface 23 can be determined.

(5) Machine learning of the learning devices 51A and 51B is performed in such a manner that input data and generated image data generated by the learning devices 51A and 51B based on the same input data are data representing the same image. . Then, values corresponding to the difference between the input data and the generated image data are outputted from the learning devices 51A and 51B as generation errors ΔP1 and ΔP2. Therefore, based on the generation errors ΔP1 and ΔP2 output from the learning devices 51A and 51B, it is possible to determine how close the actual usage state of the grinding surface 23 of the grindstone 21 to be determined is to the specific usage states A and B. can be grasped.

(6) A camera 31 is attached to the automatic grinding device 20 to take an image of the grinding surface 23 of the grindstone 21. The image captured by this camera 31 is stored in the storage unit 44 of the electronic control device 40 as the determination image 50. Therefore, the work of taking an image of the abrasive surface 23 and the work of storing the taken image of the abrasive surface 23 in the electronic control device 40 can be performed without having to attach or detach the camera 31 to the automatic grinding device 20 in accordance with the work. This can be easily done with the grindstone 21 still attached to the wheel. In addition, the control related to the judgment of the usage state of the grinding surface 23 by the abrasive surface determination device 30, the control related to the grinding process of the workpiece by the automatic grinding device 20, and the control related to the dressing of the grindstone 21 by the automatic grinding device 20 are combined. can be implemented as a series of automated controls. Thereby, it becomes possible to perform grinding of a workpiece by the automatic grinding device 20 and dressing of the grindstone 21 while checking the actual usage state of the grinding surface 23 at an appropriate timing.

(7) By controlling the operation of the shutter operation unit 29, the opening 27 of the grindstone cover 25 can be switched between a closed state and an open state. Therefore, when the camera 31 does not capture an image of the grinding surface 23, the shutter section 28 can be closed to create a closed state in which the camera 31 and the grindstone 21 are partitioned off. Thereby, it is possible to suppress adhesion of foreign substances (chips, oil, etc.) to the camera 31. When the camera 31 takes an image of the abrasive surface 23, the camera 31 can take an image of the abrasive surface 23 by opening the shutter section 28 to create an open state in which there is no partition between the camera 31 and the grindstone 21. can be put into a state.

Note that the above embodiment can be modified and implemented as follows. The above embodiment and the following modification examples can be implemented in combination with each other within a technically consistent range.

- Instead of using the image data obtained by the imaging process (FIG. 4) as is as the teacher data 53A, 53B and the judgment image 50, the same image data may be Fourier transformed (for example, fast Fourier transform [FFT]) and then It may also be used in

In this configuration, as conceptually shown in FIG. 15, image data C2 obtained by Fourier-transforming image data C1 acquired by imaging processing is used as input data in machine learning of each learning device 51A, 51B. Then, each learning device 51A, 51B uses the generated image data generated by the learning devices 51A, 51B based on this input data and the same input data as data indicating the same image (image data C2). Machine learning is performed. In addition, in the determination process for determining the usage state of the target grindstone 21, as shown in FIG. Data D2 is used. In the determination process, the learning devices 51A and 51B output a value corresponding to the difference between the input data (image data D2) and the generated image data D3 as a generation error.

By Fourier transforming the image data acquired through the imaging process, the features of the surface of the abrasive surface 23 including the feature amount (the portion extending in the grinding direction on the outermost circumferential surface of the abrasive surface 23) can be expressed in the concept of frequency. become. From this, it can be said that the features (including feature amounts) of the abrasive surface 23 appear in the image data after Fourier transformation.

According to the above configuration, by using such Fourier-transformed image data, the usage state of the abrasive surface 23 can be determined while taking into account the size and distribution of the feature amount on the abrasive surface 23. Therefore, it is possible to determine with high accuracy whether the grinding surface 23 of the grindstone 21 to be determined is in the specific usage state A or B.

Also, with the above configuration, it is possible to successively observe the same area of the grinding surface 23 in the process of using the grinding wheel 21 to understand the changes in its state, and to take an image of the grinding surface 23 as teacher data for learning the learning device. This eliminates the need to cut out a specific area from an image. Therefore, the usage state of the abrasive surface 23 can be determined with a simple operation.

In addition, in the image data after Fourier transformation, the type of grinding wheel, more specifically, the difference in the size of the abrasive grains and the difference in the distribution mode of the abrasive grains on the abrasive surface also appear in the concept of frequency. According to the above configuration, by understanding this difference, it is also possible to estimate the type of grindstone to be determined.

- Instead of using image data as the generated data generated by the learning devices 51A and 51B based on input data, a feature vector corresponding to the state of use of the abrasive surface can be used.

In this configuration, machine learning of each learning device 51A, 51B is executed as follows. As conceptually shown in FIG. 17, a feature vector F corresponding to a specific usage state is generated by learning devices 51A and 51B based on image data E (input data) obtained by capturing an image of the abrasive surface 23 in a specific usage state. The machine learning of each learning device 51A, 51B is performed in this manner. Note that as the feature vector F, a quantity representing the brightness (or brightness) distribution of the image is generated. In the above configuration, in the determination process, the difference between the feature vector corresponding to the input data (specifically, the usage state of the grindstone 21 to be determined) and the feature vector corresponding to the specific usage state is transmitted from the learning devices 51A and 51B. The value corresponding to is output as the generation error.

Here, if the number of times (for example, once) that the workpiece is processed by the grinding wheel 21 is small, the change in the usage state between the grinding surface 23 before processing and the grinding surface 23 after processing appears to be very small. . On the other hand, by generating feature vectors using the learning devices 51A and 51B as in the above configuration, the usage state of the abrasive surface 23 at any given time can be expressed numerically. Therefore, the difference between the usage state of the abrasive surface 23 before machining and the usage state of the abrasive surface 23 after machining, which is difficult to visually tell, can be expressed numerically.

According to the above configuration, a learning device of a type that generates a feature vector based on such input data is used in the determination process for determining the usage state of the grindstone 21 to be determined. Therefore, based on the generation error, which is a numerical value indicating the difference between the feature vector corresponding to the abrasive surface 23 in a specific usage state and the feature vector corresponding to the usage state of the determination target whetstone 21, the determination target whetstone 21 is It is possible to accurately determine whether a specific usage state has been reached. Furthermore, based on the generation error, which is a numerical value indicating the difference, it is also possible to grasp with high precision how much the usage state of the grindstone 21 to be determined differs from a specific usage state.

- The content displayed on the display device 35 can be changed arbitrarily. For example, the configuration for displaying character information indicating specific usage states A and B may be omitted. In addition, it is also possible to display only text information indicating that the device is in a specific usage state A or B.

- The judgment result of the state of the abrasive surface 23 is not limited to being displayed on the display device 35, but a warning buzzer may be sounded or audio information may be emitted from a speaker in accordance with the judgment result.

- The configuration for displaying the determination result of the state of the abrasive surface 23 on the display device 35 may be omitted.
- Storing in the electronic control device 40 an image captured by a camera installed in a device (such as a dedicated imaging device or other automatic grinding device) other than the automatic grinding device 20 in which the grindstone 21 to be determined is installed; It may also be used as teacher data 53A, 53B.

- The learning devices 51A and 51B are not limited to employing a generative adversarial network , but may also employ an autoencoder or the like.

- The learning mode in which the learning devices 51A and 51B perform machine learning is not limited to the learning mode in which input data and generated image data are data indicating the same image, but can be arbitrarily changed. In short, any learning mode can be used as long as the index value of the difference between the usage state of the abrasive surface 23 corresponding to the judgment image 50 and the specific usage states A and B is output from the trained learning devices 51A and 51B. This learning mode can be adopted.

- One of the two learning devices 51A and 51B can be omitted. According to the abrasive surface determination device having the learning device 51A, it is possible to determine that the abrasive wheel 21 to be determined is in a state where optimal dressing has been performed (specific usage state A). Moreover, according to the abrasive surface determination device having the learning device 51B, it is possible to determine that the abrasive stone 21 to be determined is in a state most suitable for starting execution of dressing (specific usage state B).

- It is also possible to use a learning device that is trained using teacher data consisting of images of the abrasive surface in a specific usage state C, which is different from the specific usage states A and B. As the specific usage state C, for example, a state just before optimal dressing is performed can be adopted. According to this configuration, the abrasive surface determination device can recognize or announce that the dressing of the target whetstone is almost ready for completion. In addition, it is also possible to adopt, as the specific usage state C, a state slightly before the most suitable state for starting the dressing. According to the same configuration, the abrasive surface determination device can recognize and announce that the dressing execution timing for the abrasive wheel to be determined is approaching.

- The work for storing the learned learning devices 51A and 51B in the electronic control device 40 can be performed in any manner such as (work A) to (work C) below.
(Work A) Using the automatic grinding device 20 provided with the grindstone 21 to be determined, the teacher data 53A, 53B are acquired, and the learning devices 51A, 51B are subjected to machine learning using the teacher data 53A, 53B.

(Work B) The manufacturer of the automatic grinding device 20 or the manufacturer of the whetstone 21 is asked to provide the teacher data 53A, 53B corresponding to the whetstone 21 actually used. Then, by storing the teacher data 53A, 53B in the electronic control device 40 of the abrasive surface determination device 30 used by the user, the learning devices 51A, 51B are caused to perform machine learning.

(Work C) When manufacturing the automatic grinding device 20, the learned learning devices 51A and 51B are stored in the electronic control device 40.
- The camera 31 may be attached to a part of the automatic grinding device 20 other than the grindstone cover 25.

- As an example shown in FIG. 18, the imaging unit 80 having the camera 81 may be configured separately from the automatic grinding device 20. In the same configuration, when determining the usage state of the abrasive surface 23, the imaging unit 80 is temporarily attached to the automatic grinding device 20, and the imaging unit 80 images the abrasive surface 23, so that the whetstone 21 to be determined is sharpened. An image (judgment image 50) of the surface 23 can be obtained. According to the above configuration, when the abrasive surface 23 is imaged by the imaging unit 80, the various driving parts that constitute the automatic grinding device 20 are used to capture the image of the grinding wheel 21 (more specifically, the abrasive surface 23) and the imaging unit 80. It is possible to control the relative position. Note that various drive units include a moving device that moves the grindstone 21, a moving device that moves a table to which a workpiece is fixed, and the like.

- As the judgment image 50, it is not limited to using a single image showing the entire abrasive surface 23, but it is also possible to use a single image showing a part (for example, half) of the abrasive surface 23, or to use a single image showing the entire abrasive surface 23. It is possible to use a plurality of images obtained by dividing the displayed image into a plurality of parts.

- The abrasive surface determination device according to the above embodiment can also be applied to a whetstone having a structure in which a large number of abrasive grains are fixed to the bottom surface of a cylindrical base material with a binder.
- The abrasive surface determination device according to the above embodiment can be configured as a separate device from the automatic grinding device. With the same configuration, when determining the usage state of the abrasive surface, the whetstone to be determined is removed from the automatic grinding device and attached to the abrasive surface determination device. Thereafter, the usage state of the abrasive surface is determined by the abrasive surface determination device.

20... Automatic grinding device 21... Grinding wheel 23... Grinding surface 28... Shutter section 29... Shutter operation section 30... Grinding surface determination device 31... Camera 32... X-axis drive section 33... Y-axis drive section 34... Z-axis drive section 40... Electronic control device 44...Storage unit 50...Judgment image 51A, 51B...Learning device 52...Execution program 53A, 53B...Teacher data 54...Judgment result data 70...Abrasive grain 71...Part 80...Imaging unit 81...Camera

Claims (11)

In an abrasive surface determination device that determines the condition of an abrasive surface in a whetstone having a structure in which a plurality of abrasive grains are fixed by a binder,
an input unit that stores, as a determination image, an image that is a captured image of a grinding surface of a grindstone that is a determination target and that includes information regarding a portion of the grinding surface of the determination target that extends in the grinding direction of the whetstone;
A learning device is stored that is trained using teacher data that is an image of an abrasive surface in a specific state of use and that includes information regarding a portion of the abrasive surface in the specific state of use that extends in the grinding direction. storage section and
Using the determination image as input data, a learned learning device stored in the storage unit outputs an index value of the difference between the usage state of the abrasive surface corresponding to the determination image and the specific usage state. An output section ;
An abrasive surface determination device, wherein the image constituting the determination image and the teacher data is a single image showing the entire abrasive surface . The abrasive surface determination device according to claim 1, wherein the specific usage state is a state of the abrasive surface immediately after dressing of the abrasive surface is properly performed. The abrasive surface determination device according to claim 1, wherein the specific usage state is a state of the abrasive surface at a timing suitable for starting execution of dressing of the abrasive surface. The learning device is trained in such a manner that an image input to the learning device and an image generated by the learning device based on the image are the same,
The abrasive according to any one of claims 1 to 3 , wherein the difference output unit outputs an index value of a difference between the input image and the generated image as the index value of the difference. Surface determination device. The judgment image is an image obtained by Fourier transforming an image of the grinding surface of the grindstone to be judged,
The abrasive surface determination device according to any one of claims 1 to 4, wherein the teacher data is constituted by an image obtained by Fourier transforming an image taken of the abrasive surface in the specific state of use. The learning device is trained in a manner to generate a feature vector corresponding to the specific usage state based on images input to the learning device,
The difference output unit outputs, as the difference index value, a difference between a feature vector generated by the trained learning device using the determination image as input data and a feature vector corresponding to the specific usage state. The abrasive surface determination device according to any one of Items 1 to 3 . The abrasive surface determination device includes:
an imaging unit that is provided in a grinding device that uses the grindstone as a grinding tool and captures an image of the grinding surface of the grindstone of the grinding device;
The abrasive surface determination device according to any one of claims 1 to 6 , further comprising an image output section that outputs an image of the abrasive surface captured by the imaging section to the input section as the determination image. The grinding device includes:
a shutter portion extending along the grinding surface of the whetstone in a manner that covers the grinding surface;
The shutter section is operated to open and close, and when the abrasive surface is imaged, the abrasive surface and the imaging section are in an open state where the distance between the abrasive surface and the imaging section is not partitioned by the shutter section, and when the abrasive surface is not imaged, the shutter section is opened and closed. The abrasive surface determination device according to claim 7 , further comprising: a shutter operating section that sets the abrasive surface and the imaging section to a closed state, where the shutter section partitions the space between the abrasive surface and the imaging section. The learned learning device included in the abrasive surface determination device according to any one of claims 1 to 8 . A determination program that causes an electronic control unit included in the abrasive surface determination device to execute the processing of each unit included in the abrasive surface determination device according to any one of claims 1 to 8 . An abrasive surface determination method for determining the condition of an abrasive surface in a whetstone having a structure in which a plurality of abrasive grains are fixed by a binder, the method comprising:
Obtaining an image of the grinding surface of the grinding wheel to be determined, which includes information regarding a portion of the grinding surface of the grinding wheel to be determined that extends in the grinding direction of the grinding wheel, and using the same image as a determination image. an input step for storing in the input section of the abrasive surface determination device;
Obtaining an image of an abrasive surface in a specific state of use and including information regarding a portion of the abrasive surface in the specific state of use extending in the grinding direction, and using training data consisting of the obtained image. a storage step of storing the learning device in a storage unit of the abrasive surface determination device after learning the learning device;
Using the determination image as input data, a learned learning device stored in the storage unit outputs an index value of the difference between the usage state of the abrasive surface corresponding to the determination image and the specific usage state. Equipped with an output process ,
An abrasive surface determination method , wherein the image constituting the determination image and the teacher data is a single image showing the entire abrasive surface .