Method and device for creating learning data for machine learning to build a cutting tool cutting edge wear judgment model

Abstract

A method and device for creating learning data for machine learning to build a cutting tool cutting edge wear determination model are provided. [Solution] A method for creating learning data for machine learning to construct a wear judgment model for the cutting edge CE of a cutting tool using photographed images PI of the cutting edge CE of the cutting tool, characterized by photographing the cutting edge CE and obtaining multiple photographed images PI, compressing the dimensions of the multiple photographed images PI to extract two-dimensional or three-dimensional features FA, plotting the features FA extracted by the dimension compression in feature spaces FS2 and FS3, identifying blank areas ER of feature FA data in feature spaces FS2 and FS3 from the density of the plots in feature spaces FS2 and FS3, and generating a generated image for extracting feature FA from the photographed images PI so that the identified blank areas ER can be filled with feature FA data.

Description

[Technical Field]

[0001] The present invention relates to a method and device for creating learning data for machine learning to build a cutting edge wear determination model for a cutting tool. [Background technology]

[0002] In factories that use machine tools to mass-produce workpieces, managing the tool lifespan is an important factor directly linked to production costs. Replacing a tool before its lifespan unnecessarily increases costs. Furthermore, continuing to use a tool after its lifespan can degrade the quality of the machined surface, increase cutting resistance, and potentially cause tool breakage, resulting in a defective workpiece. For this reason, systems for evaluating the tool lifespan of cutting tools have been disclosed, with the aim of enabling tool replacement just before the end of the tool's lifespan. For example, Patent Document 1 discloses a cutting tool management system that includes an individual identification database that associates and stores the characteristics of images of the tool's shank taken with an individual identification camera with the tool ID; a wear history database that stores wear history data obtained from images of the tool's cutting edge taken with a cutting edge condition confirmation camera for each tool ID stored in the individual identification database; and a control device that estimates the remaining lifespan of each tool based on the wear history data.

[0003] Furthermore, in these systems, machine learning, which has been developing in recent years, is being used to evaluate remaining life from images of the cutting edge. It is generally known that in systems that use machine learning, the more training data there is, the more accurate the estimation becomes, but preparing a sufficient number of images to use as training data is not easy and requires cost and time.

[0004] For example, Patent Document 2 discloses an image generation device that generates images for training a classifier used in visual inspection of objects. This image generation device takes an abnormal image, a photograph of an object with an abnormal appearance, as input, and uses the abnormal image as a normal image for training after masking the abnormal portion of the abnormal image. Patent Document 3 also discloses an image generation device that creates an abnormal image for use in machine learning by processing a normal image, such as by superimposing a predetermined figure on it or by deforming it.

[0005] However, cutting tool cutting edges wear over time, making it difficult to simply distinguish between normal and abnormal. There are various types of cutting tools (e.g., end mills, face mills, drills, etc.), and a wide variety of indicators for assessing wear, such as wear volume, wear location, and chipping (so-called chipping), are available. Comprehensive analysis of all these factors requires the collection of large amounts of image data. Even if such data were collected, classifying the data into indicators for assessing wear and evaluating the excess or deficiency of each indicator would require significant work. Furthermore, captured images contain a vast amount of information. For example, a grayscale image consisting of 10,000 pixels (100 columns and 100 rows) with black and white binary data would contain 10,000 dimensions of information, while commonly used color images contain millions to tens of millions of pixels. Therefore, directly evaluating captured images containing such a large amount of information is impractical; they must be preprocessed before evaluation. In addition, during actual shooting, disturbances such as shifts in the field of view of the imaging device, changes in the ambient brightness, and blurring due to dirt on the lens or mist may occur, so evaluation must ensure robustness against these disturbances. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Japanese Patent Application Laid-Open No. 2024-155460 [Patent Document 2] Patent Publication No. 2021-93004 [Patent Document 3] International Publication No. 2022 / 065271 Summary of the Invention [Problem to be solved by the invention]

[0007] In view of the above circumstances, the present invention aims to provide a method and device for creating learning data for machine learning to construct a cutting tool cutting edge wear judgment model for generating machine learning data accurately and efficiently. [Means for solving the problem]

[0008] One aspect of the present invention is a method for creating learning data for machine learning to construct a wear judgment model for the cutting edge of a cutting tool using multiple photographed images of the cutting edge of the cutting tool, the learning data creation method including: photographing the cutting edge and obtaining multiple photographed images; dimensionally compressing the multiple photographed images to extract two-dimensional or three-dimensional features; plotting the two-dimensional or three-dimensional features extracted by the dimensional compression in a feature space; identifying areas in the feature space where feature data is lacking based on the density of the plots in the feature space; and generating a generated image for extracting the features from the photographed images so that the identified areas where feature data is lacking can be filled with feature data.

[0009] One aspect of the present invention is a learning data creation device for machine learning that uses multiple photographed images of the cutting edge of a cutting tool to construct a wear determination model for the cutting edge of a cutting tool, the learning data creation device comprising: an image data input unit that photographs the cutting edge and acquires the multiple photographed images; a memory unit that stores the multiple photographed images acquired by the image data input unit; a dimension compression parameter setting unit that sets parameters for dimensionally compressing the multiple photographed images; a dimension compression unit that compresses the multiple photographed images using the parameters set by the dimension compression parameter setting unit and extracts two-dimensional or three-dimensional features; a display unit that plots the two-dimensional or three-dimensional features extracted by the dimension compression unit in a feature space and displays them; a missing area identification unit that identifies areas in the feature space where feature data is missing; and a missing image generation unit that generates generated images for extracting features from the multiple photographed images stored in the memory unit so that the identified missing areas can be filled with feature data. [Effects of the Invention]

[0010] According to one aspect of the present invention, a method for creating learning data for machine learning to construct a wear determination model for a cutting edge of a cutting tool can identify areas in the feature space where feature data is lacking by plotting two-dimensional or three-dimensional feature values ​​extracted by dimensionality compression of multiple captured images in a feature space. This makes it possible to easily grasp the wear states included and excluded from the captured images. Furthermore, generated images for extracting feature values ​​can be generated from the captured images so that the identified areas where feature data is lacking can be filled. This allows generated images to be generated with a focus on the areas where feature data is lacking, thereby efficiently creating learning data for machine learning.

[0011] According to one aspect of the present invention, a machine learning learning data creation device for constructing a wear determination model for a cutting tool cutting edge can use a dimensional compression unit to compress the dimensions of multiple captured images, extract two-dimensional or three-dimensional feature values, and display the two-dimensional or three-dimensional feature values ​​in a feature space. Furthermore, a missing area identification unit can identify areas in the feature space where feature data is missing. This allows for easy understanding of the wear states included and not included in the acquired captured images. Furthermore, a missing image generation unit can generate generated images for extracting feature values ​​from multiple captured images stored in a storage unit so that the identified missing areas can be filled with feature data. This allows for the generation of generated images that focus on missing areas, thereby efficiently creating learning data for machine learning. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 shows a block diagram of a training data creation device according to this embodiment. [Figure 2] FIG. 2 shows a flowchart of the process of plotting feature quantities in a feature space and identifying missing regions in the training data creation method according to this embodiment. [Figure 3] FIG. 3 shows a flowchart of the steps of generating a generated image in the learning data generating method according to this embodiment, and filling blank areas ER with data or increasing the distribution density. [Figure 4] FIG. 4 is a flowchart showing the steps of creating training data in the training data creation method according to this embodiment. [Figure 5] FIG. 5 shows a feature space in which two-dimensional feature amounts according to this embodiment are plotted. [Figure 6] FIG. 6 shows a feature space in which three-dimensional feature amounts according to this embodiment are plotted. [Figure 7]Figure 7 shows an example of feature data supplementation in the learning data creation method of this embodiment, where (a) is an example of specifying a range so that the blank area ER is included within the supplement area SR, (b) is an example of specifying a range so that the blank area ER is not included within the supplement area SR, and (c) is an example showing that the blank area ER has been replaced with the existing area PR due to the addition of features of the generated image GI. [Figure 8] FIG. 8 shows an example of clustering of photographed images for each tool type and creation of learning data based on the clustering according to this embodiment. [Figure 9] Figure 9 shows an example of the results of determining cutting edge damage using a cutting edge wear model trained using single learning data created for each type of tool in this embodiment, and a cutting edge wear model trained using mixed learning data that has been clustered. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, a method and an apparatus for generating training data for machine learning according to an embodiment will be described with reference to the accompanying drawings. Similar or corresponding elements are designated by the same reference numerals, and redundant explanations will be omitted. For ease of understanding, the scale of the drawings may be changed.

[0014] FIG. 1 shows a block diagram of a machine learning learning data creation device 10 (hereinafter referred to as the learning data creation device 10) for constructing a cutting edge wear determination model for a cutting tool TL (see FIG. 5). The learning data creation device 10 is a device incorporated into or connected to a cutting tool TL management system (not shown) having a function for estimating the remaining life of the cutting tool TL, as described in, for example, Patent Document 1. The learning data creation device 10 creates learning data for constructing a cutting edge wear determination model for the cutting tool TL using a machine learning system (not shown) when the cutting tool TL management system estimates the remaining life of the cutting tool TL. The learning data creation device 10 is composed of a computer incorporated into the NC device of a machine tool, a personal computer provided separately from the machine tool, a higher-performance workstation or mainframe, a dedicated computer consisting of dedicated integrated circuits such as ASICs and FPGAs, or a cloud computing system that performs large-scale calculations via the Internet. These computing devices can be selected appropriately depending on the scale of the learning data to be created. The learning data creation device 10 can also be configured by combining multiple of these computing devices. For example, a personal computer can be used for the dimension reduction parameter setting unit 16 and the disturbance addition parameter setting unit 30, which will be described in detail later, while a cloud computing system can be used for the dimension reduction unit 18, the missing region identification unit 24, the missing image generation unit 20, and the clustering unit 26, which require large-scale calculations.

[0015] For this reason, the learning data creation device 10 uses the captured image PI of the cutting tool TL (see FIG. 5) to create learning data for machine learning in accordance with the following steps, and outputs the learning data for use in machine learning. (Step a) Identifying areas in the feature space where feature FA data is insufficient (Step b) If there is an area where the feature FA data is insufficient, generate an image to supplement the feature FA data. (Step c) The feature quantities FA (see FIG. 5) of all types of cutting tools TL stored in the learning data creation device 10 are plotted in the same feature spaces FS2 and FS3 (see FIGS. 5 and 6), and clustered for each type of cutting tool TL. (Step d-1) If clustering is possible for each type of cutting tool TL, mixed learning data containing a mixture of types of cutting tools TL is created. (Step d-2) If clustering cannot be performed for each type of cutting tool TL, single learning data is created for each type of cutting tool TL. (Step e) In order to ensure robustness against photographic disturbances, image processing is performed on the created mixed learning data or single learning data. (Step f) Output the created learning data for machine learning

[0016] The learning data creation device 10 includes an image data input unit 12 for executing step a. The image data input unit 12 is configured to capture images of the cutting edge CE of the cutting tool TL being used for machining using a camera (all of which are not shown) located in a management system for the cutting tool TL, a machining chamber of a machine tool managed by the management system, a tool magazine, or a tool storage unit of a tool transport system that transports tools to multiple machine tools, to acquire multiple captured images PI. The image data input unit 12 can also acquire previously captured images PI from the cloud or a storage medium on the management system side. The learning data creation device 10 also includes a storage unit 14 for storing the multiple captured images PI thus acquired by the image data input unit 12.

[0017] Furthermore, the training data creation device 10 includes a dimension compression parameter setting unit 16 and a dimension compression unit 18 for dimensionally compressing the plurality of captured images PI. The dimension compression unit 18 compresses the dimensions of the plurality of captured images PI using parameters for dimensionally compressing the plurality of captured images PI set by an operator in the dimension compression parameter setting unit 16, and extracts two-dimensional or three-dimensional feature quantities FA. The dimension compression parameter setting unit 16 is configured to allow an operator to set dimension compression methods, such as t-SNE (t-distribution type stochastic neighbor embedding), PCA (principal component analysis), and UMAP (uniform manifold approximation and projection), to be executed by the dimension compression unit 18, as well as various parameters used in these dimension compression methods. Note that the dimension compression methods described in this specification are merely examples, and the dimension compression methods to which the present invention can be applied are not limited to these.

[0018] The training data creation device 10 also includes a display unit 22 for plotting and displaying the two-dimensional or three-dimensional feature values ​​FA extracted by the dimensionality reduction unit 18 in feature spaces FS2 and FS3 in order to perform step b using the feature values ​​FA extracted by the dimensionality reduction unit 18. At this time, the training data creation device 10 sets the size of the feature spaces FS2 and FS3 (i.e., the scale of the two or three axes representing the feature spaces FS2 and FS3) to an appropriate value depending on the distribution of the feature values ​​FA. The training data creation device 10 is also configured to allow an operator to manually adjust the size of the feature spaces FS2 and FS3. The display unit 22 includes a monitor, a display panel, or the like (not shown), and is configured to visually confirm the distribution of the feature values ​​FA in the feature spaces FS2 and FS3, allowing the operator to zoom in and out as needed. The display unit 22 is a general-purpose monitor, and when displaying the three-dimensional feature space FS3, the operator can freely rotate the display of the feature space FS3 so that the three-dimensional extent of the feature space FS3 can be grasped. Alternatively, a stereoscopic monitor may be used, allowing the operator to stereoscopically confirm the three-dimensional feature space FS3.

[0019] The training data creation device 10 further includes a missing region identifying unit 24 for identifying a data blank region ER based on the distribution state between feature values ​​FA in the feature spaces FS2 and FS3 displayed by the display unit 22. The missing region identifying unit 24 identifies the region inside a circle or sphere whose center is the feature value FA plotted in the two-dimensional or three-dimensional feature space FS2 or FS3 and whose radius is a threshold value TV as a data presence region PR, and identifies the region outside the presence region PR as a blank region ER. The missing region identifying unit 24 also includes a track bar 22a (see FIG. 7 ) that allows an operator to change the threshold value TV on the display unit 22. When the operator changes the threshold value TV on the display unit 22, the missing region identifying unit 24 recalculates the presence region PR and blank region ER and displays them on the display unit 22. Alternatively, a numerical input unit (not shown) may be provided instead of the track bar 22a, allowing the operator to input the changed threshold value TV as a numerical value.

[0020] The learning data creation device 10 also includes a missing image generation unit 20 for filling the identified blank area ER with feature FA data. The missing image generation unit 20 generates a generated image GI using a generative adversarial network (GAN) from multiple captured images PI stored in the storage unit 14 to extract feature FA that can be filled in the blank area ER. As shown in FIG. 7 , the missing image generation unit 20 allows an operator to specify a fill area SR to which data is to be filled in the feature space displayed on the display unit 22. While the example shown in FIG. 7 illustrates a method for specifying a rectangular range using a mouse, the method is not limited to this, and the range may also be specified using an input device such as a touch panel or keyboard. The range may also be a shape other than a rectangle, such as a circle or an ellipse.

[0021] When the operator specifies a fill region SR, the missing image generation unit 20 determines whether the fill region SR includes a blank region ER. If the fill region SR includes a blank region ER, the missing image generation unit 20 determines that feature values ​​FA of different wear states need to be filled and generates a generated image GI of different wear states by referencing photographed images PI of cutting edges of different cutting tool types. In the example shown here, the generated image GI is generated using StarGAN and CycleGAN, among GANs. When the missing image generation unit 20 generates a generated image GI to be filled, the dimensionality reduction unit 18 reduces the dimension of the newly generated generated image GI to extract feature values ​​FA. The display unit 22 plots the feature values ​​FA, including the newly extracted feature values ​​FA, in feature spaces FS2 and FS3. The training data creation device 10 is configured to draw a circle (in the case of two dimensions) or a sphere (in the case of three dimensions) centered on the feature value FA of the generated image GI and with a radius equal to a threshold value TV, and update the display of the existence region PR. This allows the operator to visually confirm whether the blank region ER has been replaced with the existence region PR.

[0022] Furthermore, if the fill area SR does not include a blank area ER, i.e., if the entire fill area SR is an existing area PR, the missing image generation unit 20 determines that the distribution density of the feature FA within the fill area SR needs to be further increased. The missing image generation unit 20 generates a generated image GI by referencing the captured image PI of the same tool type as the captured image PI in which the feature FA exists within the fill area SR, thereby increasing the distribution density of the feature FA within the fill area SR. In the example shown here, the generated image GI is generated using SAGAN and DCGAN, among GANs. Once the missing image generation unit 20 generates the generated image GI to be filled, the dimensionality reduction unit 18 dimensionally reduces the newly generated generated image GI to extract the feature FA. The display unit 22 of the training data creation device 10 is configured to plot the feature FA, including the newly extracted feature FA, in the feature spaces FS2 and FS3. This allows the operator to visually confirm that the distribution density of the feature FA within the fill area SR within the existing area PR has increased.

[0023] As shown in FIG. 1, after generating the generated image GI, the learning data creation device 10 confirms that no blank areas ER exist in the feature spaces FS2 and FS3 in which the feature values ​​FA are plotted and that the distribution density of the feature values ​​FA has increased sufficiently. The learning data creation device 10 then performs step c and step d-1 or d-2. The clustering unit 26 classifies the feature values ​​FA by type of cutting tool TL. The clustering unit 26 is configured to cluster the feature values ​​FA plotted in the feature spaces FS2 and FS3 by type of cutting tool TL, such as drills, radius end mills, and ball end mills. Clustering algorithms, such as K-means, DBSCAN, mean shift, Gaussian mixture model, and agglomerative clustering, are used. The clustering unit 26 clusters the feature values ​​FA using an algorithm and setting values ​​selected by the operator, and displays symbols of different colors or shapes for each tool type superimposed on the feature values ​​FA on the display unit 22. As shown in the example of FIG. 8, a triangle symbol is superimposed on a drill, a diamond symbol is superimposed on a radius end mill, and a circle symbol is superimposed on a ball end mill. If the operator checks this display and determines that the feature values ​​FA have been correctly clustered for each type of cutting tool TL (see the upper left of FIG. 8), that is, if the distributions of the feature values ​​FA for each type of cutting tool TL can be spatially separated without crossing each other in the feature spaces FS2 and FS3, the clustering unit 26 mixes the types of cutting tools TL and creates mixed learning data from a set of captured images PI and generated images GI that have been classified so that each tool type can be identified (see the lower left of FIG. 8). On the other hand, if the operator checks this display and determines that the feature values ​​FA have not been clustered for each type of cutting tool TL (see the upper right of FIG. 8), that is, if the distributions of the feature values ​​FA for each type of cutting tool TL cannot be spatially separated in the feature spaces FS2 and FS3, the clustering unit 26 creates single learning data for each type of cutting tool TL from the captured images PI and generated images GI for each type of cutting tool TL (see the lower right of FIG. 8).Although the example shown in FIG. 8 displays a two-dimensional feature space FS2, clustering can be performed in a three-dimensional feature space FS3. FIG. 9 shows the results of determining chipping of cutting edges for three types of cutting tools TL (ball end mills, drills, and radius end mills) using a cutting edge wear model trained using single-type learning data and a cutting edge wear model trained using mixed learning data that is classified and mixed to enable identification of these cutting tools. As shown in FIG. 9 , in this embodiment, the inventors experimentally found that the cutting edge wear determination model trained using mixed learning data that combines the types of cutting tools TL and classifies each tool type to enable identification has higher determination accuracy than when trained using single-type learning data, and therefore, the use of mixed learning data is preferable. Therefore, although an operator can perform training using single-type learning data, it is preferable to perform training using mixed learning data obtained by appropriately changing the algorithm and setting values ​​used for clustering.

[0024] When the clustering unit 26 creates the training data, the training data creation device 10 executes step e. The training data creation device 10 includes a disturbance addition unit 28 for adding images obtained by performing image processing on the captured images PI and the generated images GI that constitute the training data, simulating disturbances that occurred during image capture, to the training data to ensure robustness against image capture disturbances in machine learning. This allows for the generation of images with disturbances added to the captured images PI and the generated images GI. The training data creation device 10 also includes a disturbance addition parameter setting unit 30, which allows an operator to set disturbance addition parameters that determine the type and degree of disturbance, such as image rotation, brightness change, or noise addition. The training data creation device 10 further includes a training data output unit 32 for executing step f. The training data creation device 10 can output the package of captured images PI, generated images GI, and images with disturbances added as training data to the machine learning system.

[0025] The learning data creation method according to this embodiment will be described in detail below with reference to the flowcharts shown in FIGS. 2 to 4 and FIGS. 5 to 8.

[0026] 2 shows a flowchart of the process of plotting feature quantities FA in feature spaces FS2 and FS3 and identifying blank areas ER. The training data creation device 10 proceeds to step S10 to start creating training data, and then proceeds to step S20. When proceeding to step S20, the training data creation device 10 activates the image data input unit 12 to acquire multiple captured images PI and stores the captured images PI in the memory unit 14. After storing the captured images PI, the training data creation device 10 proceeds to step S30, where the dimension compression parameter setting unit 16 sets the dimension to be compressed. Here, if compression to two dimensions is desired, the process proceeds to step S40, and if compression to three dimensions is desired, the process proceeds to step S50.

[0027] After setting the dimensions to be reduced, the training data creation device 10 proceeds to step S60, where the dimension reduction parameter setting unit 16 sets the details of the dimension reduction method, i.e., which dimension reduction method to use alone or in combination, and sets the parameters required for each dimension reduction method. If a single dimension reduction method is used for dimension reduction, the device proceeds to step S70. If a combination of dimension reduction methods, such as reducing dimensions up to 40 dimensions using PCA and then further reducing dimensions down to 2 dimensions using UMAP, the device proceeds to step S80. Once the dimension reduction is complete, the device proceeds to step S90, where the training data creation device 10 activates the display unit 22 to plot and display the two-dimensional or three-dimensional feature values ​​FA extracted by the dimension reduction unit 18 in feature spaces FS2 and FS3, as shown in FIGS. 5 and 6. After plotting and displaying the feature FA in the feature spaces FS2 and FS3, the process proceeds to step S100, where the learning data creation device 10 activates the missing area identification unit 24 to identify the data presence areas PR and blank areas ER based on the threshold TV set by the operator, and displays them in the feature spaces FS2 and FS3.

[0028] FIG. 5 shows an example of the distribution of two-dimensional feature quantities FA plotted in a two-dimensional feature space FS2. The values ​​of the feature quantities FA on the vertical and horizontal axes of the two-dimensional feature space FS2 are displayed as dimensionless numbers. Here, the feature quantities FA are plotted in the feature space FS2 displayed on the display unit 22 only as point clouds. When the operator moves the cursor on the display unit 22 to the position of the point cloud of the feature quantities FA, the corresponding captured image PI is displayed. The display unit 22 can display the captured image PI that is the source of all feature quantities FA, and the operator can switch the display state as appropriate depending on the amount of information to be displayed. Furthermore, the missing area identifying unit 24 identifies a blank area ER based on a threshold value TV set by the operator by sliding the track bar 22a. The identified blank area ER is displayed in the feature space FS2 so that it can be visually recognized by the operator.

[0029] FIG. 6 shows an example of the distribution of three-dimensional feature quantities FA plotted in a three-dimensional feature space FS3. The values ​​of the feature quantities FA on the vertical and horizontal axes of the three-dimensional feature space FS3 are displayed as dimensionless numbers. The point clouds of the feature quantities FA are displayed with different point cloud patterns and display colors for each type of cutting tool TL. The display unit 22 shown in FIG. 6 plots the feature quantities FA only as point clouds in the feature space FS3, and is configured to display the corresponding captured image PI when the operator moves the cursor on the display unit 22 to the position of the point cloud of the feature quantities FA. The operator can also switch from displaying the point clouds to displaying the captured image PI, which is the source of all the feature quantities FA. When displaying the captured image PI instead of the point clouds, the color and type of the border of the captured image PI can be changed for each type of cutting tool TL.

[0030] 3 shows a flowchart of the process for generating a generated image. After the learning data creation device 10 displays the two-dimensional or three-dimensional feature amount FA, the presence region PR, and the missing region identification unit 24 in the feature spaces FS2 and FS3, the process proceeds to step S110. In step S110, the operator specifies a supplementary region SR in the feature space where data is to be supplemented.

[0031] When the supplementary region SR is specified, the training data creation device 10 proceeds to step S120. The training data creation device 10 activates the missing region identification unit 24 to determine whether or not a blank region ER is included in the supplementary region SR. If the result shows that a blank region ER is included in the supplementary region SR, the process proceeds to step S130; if not, the process proceeds to step S140.

[0032] If the missing region identifying unit 24 determines that a blank region ER is included in the supplement region SR and proceeds to step S130, the missing image generating unit 20 determines that feature amounts FA of different wear states need to be added, reads from the storage unit 14 captured images PI of cutting edges of different cutting tool types to increase the wear states, and proceeds to step S150. On the other hand, if the missing region identifying unit 24 determines that a blank region ER is not included in the supplement region SR and proceeds to step S140, the missing image generating unit 20 determines that the distribution density of feature amounts FA in the supplement region SR needs to be further increased, reads from the storage unit 14 captured images PI of the same tool type to increase the distribution density, and proceeds to step S150.

[0033] In step S150, the operator selects one or more captured images PI required for image generation from the captured images PI read by the missing image generation unit 20 from the memory unit 14. Then, in step S160, the operator sets image generation conditions such as the number of images to be generated. Then, in step S170, the missing image generation unit 20 generates a generated image GI based on the already selected captured images PI and the set image generation conditions. Once the generated image GI has been generated, the process proceeds to step S180, where the dimensional compression unit 18 performs dimensional compression on the generated image GI, and then, in step S190, the display unit 22 displays the generated image GI.

[0034] When the display unit 22 displays the generated image GI, the process proceeds to step S200, and the learning data creation device 10 repeats steps S110 to S190 to generate the generated image GI until the operator confirms that the blank region ER has been replaced with the existing region PR or that the distribution density of the feature FA in the supplementary region SR within the existing region PR has increased. When the operator confirms that the blank region ER has been replaced with the existing region PR or that the distribution density of the feature FA in the supplementary region SR within the existing region PR has increased, the process proceeds to step S210 (see FIG. 4).

[0035] 4 shows a flowchart of the process of creating learning data. Proceeding to step S210, the display unit 22 displays the feature amounts FA of the captured images PI and the generated images GI to be clustered in the feature spaces FS2 and FS3. After displaying the feature amounts FA, the process proceeds to step S220, where clustering of the feature amounts FA is performed using an algorithm and setting values ​​selected by the operator. As shown in FIG. 8, symbols of different colors or shapes corresponding to the tool types of the feature amounts FA are displayed on the display unit 22, superimposed on the feature amounts FA. Next, proceeding to step S230, if the operator determines that the feature amounts FA can be clustered for each type of cutting tool TL (see the upper left of FIG. 8), the process proceeds to step S240, where the clustering unit 26 mixes the types of cutting tools TL and creates mixed learning data (see the lower left of FIG. 8) from a group of captured images PI and generated images GI that have been classified so that each tool type can be identified. Furthermore, if the operator determines that the feature values ​​FA cannot be clustered for each type of cutting tool TL (see the top right of Figure 8), the process proceeds to step S250, where the clustering unit 26 determines that the distribution of the feature values ​​FA on the feature spaces FS2 and FS3 is appropriate for each type of cutting tool TL, and creates single learning data for each type of cutting tool TL (see the bottom right of Figure 8) from the captured images PI and generated images GI for each type of cutting tool TL.

[0036] Once the learning data has been created, the process proceeds to step S260, where the disturbance adding unit 28 adds to the learning data images obtained by performing image processing (image rotation, brightness change, noise addition, etc.) on the captured images PI and generated images GI that constitute the learning data to simulate disturbances that occurred during image capture. Once the disturbance-added images have been added, the process proceeds to step S270, where a package of the captured images PI, generated images GI, and images with the disturbance added is generated as learning data and output to the machine learning system, thereby completing the process.

[0037] While the above embodiment includes a step in which an operator intervenes to make a decision, this is not limiting. If the necessary computational resources are available, some or all of the steps involving operator intervention may be automated. In step b of the above embodiment, the operator checks the display unit 22 and manually changes the threshold TV to select the range of the supplementary region SR, and determines whether the blank region ER has been updated to the existing region PR or whether the distribution density of the feature amount FA in the existing region PR has sufficiently increased. Here, if the dimensionality reduction unit 18, the missing image generation unit 20, and the missing region identification unit 24 have sufficient computational capabilities, a desired distribution density of the feature amount FA relative to the threshold TV can be preset, allowing the generated image GI to be automatically generated from the selected captured image PI until all of the blank region ER is replaced with the existing region PR. Furthermore, the generated image GI can be automatically generated until the distribution density in the existing region PR reaches a preset value. Furthermore, in steps c, d-1, or d-2, the clustering unit 26 may perform clustering automatically by presetting an algorithm and setting values ​​that enable clustering based on the operator's experience or past records. In this way, the operator's decision may be automatically processed by a computer.

[0038] According to the training data creation device 10 and creation method of this embodiment, the dimensionality reduction unit 18 reduces the dimensions of multiple captured images PI to extract two-dimensional or three-dimensional feature values ​​FA, and the display unit 22 plots and displays the two-dimensional or three-dimensional feature values ​​FA in feature spaces FS2 and FS3. Furthermore, the missing area identification unit 24 identifies blank areas ER of feature value FA data in feature spaces FS2 and FS3. This allows for easy identification of the damage states included and not included in the acquired captured images PI. Furthermore, the missing image generation unit 20 generates generated images GI for extracting feature values ​​FA from multiple captured images PI stored in the storage unit 14 so that the identified blank areas ER can be filled with feature value FA data. This allows for the generation of generated images GI to focus on the blank areas ER, thereby efficiently creating training data for machine learning.

[0039] Furthermore, according to the learning data creation device 10 and creation method of this embodiment, the missing image generation unit 20 can identify a generated image GI to be filled in so as to fill in the blank area ER with feature amount FA data based on the determination result, and generate a generated image GI from the captured image PI. This makes it possible to efficiently generate a generated image GI for filling in the blank area ER with feature amount FA.

[0040] Furthermore, according to the training data creation device 10 and creation method of this embodiment, if the missing image generation unit 20 determines that it is necessary to further increase the distribution density of the feature FA in the presence region PR, it can generate a generated image GI using the captured images PI of the cutting edge CE of the same type of cutting tool TL in the feature spaces FS2 and FS3. Therefore, by generating a generated image GI that imitates the captured images PI, it is possible to increase the number of images and increase the distribution density of the feature FA in the feature spaces FS2 and FS3. This allows the generation of generated images GI to focus on the blank region ER, thereby efficiently creating training data for machine learning.

[0041] Furthermore, according to the training data creation device 10 and creation method of this embodiment, once the clustering unit 26 creates training data, the disturbance adding unit 28 can add to the training data images obtained by performing image processing on the captured images PI and the generated images GI that constitute the training data to simulate disturbances that occurred during image capture. This makes it possible to ensure robustness in machine learning against imaging disturbances such as shifts in the field of view of the imaging device that occur during imaging, changes in ambient brightness, and blurring due to lens dirt or mist.

[0042] As described above, the learning data creation device 10 and creation method for machine learning for constructing a cutting tool cutting edge wear judgment model according to this embodiment can generate machine learning data accurately and efficiently.

[0043] Although the embodiments of the training data creation device 10 and creation method have been described above, the present invention is not limited to the above-described embodiments. In addition to the above, it is believed that a person skilled in the art would understand that various modifications of the above-described embodiments are possible. [Explanation of symbols]

[0044] 10. Learning data creation device 12 Image data input section 14 Storage section 16 Dimensional reduction parameter setting section 18 Dimensional compression section 20 Missing image generation unit 22 Display section 24 Deficiency Area Identification Department 28 Disturbance applying section 30 Disturbance parameter setting section CE cutting edge FA feature quantity FS2 Feature Space (2D) FS3 Feature Space (3D) GI generated image PI image ER blank area PR presence area SR replenishment area TL cutting tools

Claims

1. A method for creating learning data for machine learning to construct a wear determination model for a cutting edge of a cutting tool using a plurality of photographed images of the cutting edge of the cutting tool, comprising: Photographing the cutting edge and acquiring a plurality of photographed images; dimensionally compressing the plurality of captured images and extracting two-dimensional or three-dimensional feature quantities; Plotting the two-dimensional or three-dimensional feature amount extracted by the dimensionality reduction in a feature space; Identifying a data-deficient area of ​​the feature amount in the feature space based on the density of the plots in the feature space; generating a generated image for extracting the feature amount from the captured image so that the identified missing area can be filled with the feature amount data; A learning data creation method comprising:

2. The learning data creation method according to claim 1 , further comprising: generating a disturbance-added image from the captured image and the generated image by performing image processing that simulates a disturbance that occurs when the image is captured.

3. Regarding the identified shortage area, determining that the feature amounts of different wear states are to be supplemented, and generating the generated image that imitates the wear state using the captured images of the cutting tools of different types; The training data creation method according to claim 1 , further comprising:

4. When the distribution of the feature amounts for each type of cutting tool can be spatially separated in the two-dimensional or three-dimensional feature space, the types of cutting tools are mixed, and mixed learning data is created that is classified so that each type of cutting tool can be identified. The training data creation method according to claim 1 , further comprising:

5. 1. A learning data creation device for machine learning that uses a plurality of photographed images of a cutting edge of a cutting tool to construct a wear determination model for the cutting edge of the cutting tool, comprising: an image data input unit that photographs the cutting edge and acquires a plurality of photographed images; a storage unit that stores the plurality of captured images acquired by the image data input unit; a dimension compression parameter setting unit that sets parameters for dimension compression of the plurality of captured images; a dimension compression unit that performs dimension compression on the plurality of captured images using the parameters set by the dimension compression parameter setting unit and extracts two-dimensional or three-dimensional feature amounts; a display unit that plots and displays the two-dimensional or three-dimensional feature quantities extracted by the dimensionality reduction unit in a feature space; a missing area identifying unit that identifies an area in the feature space where data of the feature amount is missing; a missing image generating unit that generates a generated image for extracting the feature amount from the plurality of captured images stored in the storage unit so that the identified missing area can be filled with data of the feature amount; A learning data creation device comprising:

Images