Method and device for creating training data for machine learning to build wear determination model for cutting edge of cutting tool
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2026-01-29
- Publication Date
- 2026-08-13
Smart Images
Figure JP2026003099_13082026_PF_FP_ABST
Abstract
Description
Method and apparatus for creating training data for machine learning to build a cutting tool cutting edge wear determination model.
[0001] The present invention relates to a method and apparatus for creating training data for machine learning to construct a cutting edge wear determination model for cutting tools.
[0002] In factories that mass-produce workpieces by machining them using machine tools, managing the lifespan of the tools used for machining is a crucial factor that directly impacts production costs. Replacing tools with new ones before they reach the end of their lifespan unnecessarily increases costs, and continuing to use tools after they have reached the end of their lifespan can lower the quality of the machined surface, increase cutting resistance, and cause tool breakage, potentially resulting in defective workpieces. For this reason, systems for evaluating the lifespan of cutting tools have been disclosed with the aim of replacing them with new ones just before they reach the end of their lifespan. For example, Patent Document 1 discloses a cutting tool management system comprising: an individual identification database that stores the characteristics of the image of the shank portion of the tool taken by an individual identification camera in association with a tool ID; a wear history database that stores wear history data obtained from images of the cutting edge of the tool taken by 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 from the wear history data.
[0003] Furthermore, in such systems, machine learning, which has made significant progress in recent years, is being utilized to evaluate the remaining lifespan from images of the blade. In systems that utilize machine learning, it is generally known that the more training data there is, the better the estimation accuracy becomes. However, preparing a sufficient number of images for training data is not easy and requires time and cost.
[0004] Therefore, for example, Patent Document 2 discloses an image generation device that generates an image for learning a classifier used for the appearance inspection of an object. In this image generation device, an abnormal image obtained by photographing an object with an abnormal appearance is input, and after masking a portion indicating an abnormality in the abnormal image, it is used as a normal image for learning. Further, Patent Document 3 discloses an image creation device that creates an abnormal image by performing processing such as superimposing or transforming a predetermined figure on a normal image and uses it for machine learning.
[0005] However, the cutting edge of a cutting tool is used while being worn over time, and it is difficult to simply distinguish between normal and abnormal. In addition, there are various types of cutting tools (for example, end mills, face mills, drills, etc.), and furthermore, there are a wide variety of indicators for evaluating wear states such as the amount of wear, the location of wear, and the chipping (so-called chipping) of the cutting edge. If we try to fully cover all of these elements, it is necessary to collect a large amount of image data. Even if a large amount of image data is collected, classifying according to each indicator for evaluating the wear state and evaluating the excess and deficiency for each indicator will result in a large number of man-hours. In addition, a photographed image has a huge amount of information. For example, a grayscale image consisting of 10,000 pixels with 100 columns vertically and horizontally and having a black-and-white binary value has an information amount of 10,000 dimensions, and in a generally used color image, it exceeds several million to tens of millions of pixels. Therefore, it is not practical to directly evaluate a photographed image having such a huge amount of information, and it is necessary to perform some preprocessing before evaluation. In addition, during actual photographing, photographing disturbances such as deviation of the imaging device's field of view, change in ambient brightness, blurring due to lens dirt or mist may occur, so it is necessary to ensure robustness against these photographing disturbances before evaluation.
[0006] Japanese Patent Application Laid-Open No. 2024-155460, Japanese Patent Application Laid-Open No. 2021-93004, International Publication No. 2022 / 065271
[0007] In view of the above circumstances, an object of the present invention is to provide a learning data creation method and a creation device for machine learning for constructing a cutting tool cutting edge wear determination model for accurately and efficiently generating machine learning data.
[0008] One aspect of the present invention is a method for creating training data for machine learning to construct a wear determination model for the cutting edge of a cutting tool using multiple captured images of the cutting edge of the cutting tool, the method comprising: capturing images of the cutting edge and acquiring multiple captured images; compressing the dimensions of the multiple captured images and extracting two-dimensional or three-dimensional features; plotting the two-dimensional or three-dimensional features extracted by dimensionality compression in a feature space; identifying areas of feature data deficiency in the feature space from the density of the plots in the feature space; and generating generated images from the captured images for extracting features so that feature data can be supplemented in the identified deficiency areas.
[0009] One aspect of the present invention is a machine learning data creation device for constructing a wear determination model for the cutting edge of a cutting tool using a plurality of captured images of the cutting edge of a cutting tool, comprising: an image data input unit that photographs the cutting edge and acquires a plurality of captured images; a storage unit that stores the plurality of captured images acquired by the image data input unit; a dimensionality reduction parameter setting unit that sets parameters for dimensionality reduction of the plurality of captured images; a dimensionality reduction unit that compresses the plurality of captured images using the parameters set by the dimensionality reduction parameter setting unit and extracts two-dimensional or three-dimensional feature quantities; 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 deficiency region identification unit that identifies a region of deficiency in feature quantity data in the feature space; and a deficiency image generation unit that generates a generated image for extracting feature quantities from a plurality of captured images stored in the storage unit so that feature quantity data can be supplemented in the identified deficiency region.
[0010] According to one aspect of the present invention, a method for creating training data for machine learning to construct a wear determination model for the cutting edge of a cutting tool, by plotting two-dimensional or three-dimensional features extracted by dimensionality reduction of multiple captured images in a feature space, it is possible to identify areas where feature data is lacking in the feature space. Therefore, it is possible to easily grasp the wear status included in the acquired captured images and the wear status that is not included. Furthermore, generated images for extracting features can be generated from the captured images so that feature data can be supplemented in the identified areas where feature data is lacking. Therefore, generated images can be generated with a focus on areas where feature data is lacking, and training data for machine learning can be created efficiently.
[0011] According to one aspect of the present invention, a machine learning training data creation device for constructing a wear determination model for the cutting edge of a cutting tool, the dimensionality reduction unit compresses the dimensions of multiple captured images, extracts two-dimensional or three-dimensional features, and the display unit plots and displays the two-dimensional or three-dimensional features in a feature space. Furthermore, the deficiency region identification unit can identify regions in the feature space where feature data is lacking. This makes it easy to grasp the wear status included in and not included in the acquired captured images. In addition, the deficiency image generation unit can generate generated images for extracting features from multiple captured images stored in the memory unit so that feature data can be supplemented in the identified deficiency regions. This makes it possible to generate generated images with a focus on deficiency regions, and to efficiently create training data for machine learning.
[0012] Figure 1 shows a block diagram of the learning data creation device according to this embodiment. Figure 2 shows a flowchart of the process of plotting feature quantities in the learning data creation method according to this embodiment in the feature space and identifying missing regions. Figure 3 shows a flowchart of the process of generating a generated image and supplementing the blank region ER with data or increasing the distribution density according to the learning data creation method according to this embodiment. Figure 4 shows a flowchart of the process of creating learning data according to the learning data creation method according to this embodiment. Figure 5 shows the feature space with two-dimensional feature quantities plotted according to this embodiment. Figure 6 shows the feature space with three-dimensional feature quantities plotted according to this embodiment. Figure 7 shows an example of supplementing feature quantity data in the learning data creation method according to this embodiment, where (a) the range is specified so that the blank region ER is included in the supplemented region SR, (b) the range is specified so that the blank region ER is not included in the supplemented region SR, and (c) an example showing that the blank region ER is replaced by the existing region PR due to the addition of feature quantities to the generated image GI. Figure 8 shows an example of clustering of captured images by tool type and the creation of learning data based on clustering according to this embodiment. Figure 9 shows examples of cutting edge damage determination results for a cutting edge wear model trained using single training data created for each type of tool according to this embodiment, and a cutting edge wear model trained using clustered mixed training data.
[0013] The following describes a method and apparatus for creating training data for machine learning according to the embodiment, with reference to the attached drawings. Similar or corresponding elements are denoted by the same reference numerals, and redundant explanations are omitted. The scale of the figures may be changed in some cases to facilitate understanding.
[0014] Figure 1 shows a block diagram of a machine learning data creation device 10 (hereinafter referred to as the "training data creation device 10") for constructing a cutting edge wear determination model for a cutting tool TL (see Figure 5). The training data creation device 10 is a device that is incorporated into or connected to a cutting tool TL management system (not shown) that has a function to estimate the remaining lifespan of a cutting tool TL, as described in Patent Document 1. The training data creation device 10 is a device that creates training data for constructing a cutting edge wear determination model for a cutting tool TL using a machine learning system (not shown) when such a cutting tool TL management system estimates the remaining lifespan of the cutting tool TL. The training data creation device 10 consists of a computer incorporated into the NC device of a machine tool, a personal computer prepared separately from the machine tool, or a more high-performance workstation and 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, and these computing devices can be appropriately selected according to the scale of the training data to be created. Furthermore, the learning data creation device 10 can be configured by combining multiple of these computing devices. For example, a personal computer can be used for the dimensionality compression 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 dimensionality compression unit 18, the deficient region identification unit 24, the deficient image generation unit 20, and the clustering unit 26, which require large-scale calculations.
[0015] Therefore, the learning data creation device 10 uses the captured image PI (see Figure 5) of the cutting tool TL to create learning data for machine learning according to the following steps, and outputs the learning data for use in machine learning. (Step a) Identify areas in the feature space where data for feature quantity FA is insufficient. (Step b) If there are areas where data for feature quantity FA is insufficient, generate images to supplement the data for feature quantity FA. (Step c) Plot the feature quantities FA (see Figure 5) of all types of cutting tools TL stored in the training data creation device 10 in the same feature spaces FS2 and FS3 (see Figures 5 and 6), and cluster them according to the type of cutting tool TL. (Step d-1) If clustering can be performed according to the type of cutting tool TL, create mixed training data containing mixed types of cutting tool TL. (Step d-2) If clustering cannot be performed according to the type of cutting tool TL, create single training data for each type of cutting tool TL. (Step e) To ensure robustness against imaging disturbances, apply image processing to the created mixed training data or single training data. (Step f) Output the created training data for executing machine learning.
[0016] The learning data creation device 10 has an image data input unit 12 to execute step a. The image data input unit 12 is configured to capture images of the cutting edge CE of the cutting tool TL used for machining by a camera (not shown) located in the cutting chamber, tool magazine, or tool storage section of a cutting tool TL management system or a machine tool managed by the management system, and to acquire multiple captured images PI. The image data input unit 12 can also acquire captured images PI that have already been captured from a storage medium on the cloud or management system side. The learning data creation device 10 also has a storage unit 14 for storing the multiple captured images PI thus acquired by the image data input unit 12.
[0017] Furthermore, the learning data creation device 10 includes a dimensionality reduction parameter setting unit 16 and a dimensionality reduction unit 18 for dimensionality reduction of multiple captured image PIs. The dimensionality reduction unit 18 uses parameters for dimensionality reduction of multiple captured image PIs set by the operator in the dimensionality reduction parameter setting unit 16 to dimensionally reduce the multiple captured image PIs and extract two-dimensional or three-dimensional feature quantities FA. The dimensionality reduction parameter setting unit 16 is configured so that the operator can set dimensionality reduction methods such as t-SNE (t-distribution type stochastic neighbor embedding), PCA (principal component analysis), and UMAP (homogeneous manifold approximation and projection) to be executed by the dimensionality reduction unit 18, as well as various parameters used in these dimensionality reduction methods. Note that the dimensionality reduction methods shown herein are examples, and the dimensionality reduction methods to which the present invention can be applied are not limited to these.
[0018] Furthermore, the learning data creation device 10 has a display unit 22 for plotting and displaying the two-dimensional or three-dimensional feature quantities FA extracted by the dimensionality compression unit 18 in the feature spaces FS2 and FS3 in order to execute step b using the feature quantities FA extracted by the dimensionality compression unit 18. At this time, the learning data creation device 10 sets the size of the feature spaces FS2 and FS3 (i.e., the scale of the two-axis or three-axis representing the feature spaces FS2 and FS3) to an appropriate value according to the distribution of the feature quantities FA. The learning data creation device 10 is also configured so that the operator can manually adjust the size of the feature spaces FS2 and FS3. The display unit 22 has a monitor, display panel, etc. (not shown) and is configured so that the distribution of feature quantities FA in the feature spaces FS2 and FS3 can be visually confirmed, and the operator can zoom in and out as needed. Furthermore, 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 they can grasp the three-dimensional extent of the feature space FS3. Alternatively, it may be a stereoscopic monitor, configured so that the operator can view the three-dimensional feature space FS3 in three dimensions.
[0019] Furthermore, the learning data creation device 10 has a deficiency region identification unit 24 for identifying data blank regions ER based on the distribution state between feature quantities FA in the feature spaces FS2 and FS3 displayed by the display unit 22. The deficiency region identification unit 24 identifies the region inside a circle or sphere with a threshold TV of the radius centered on the feature quantities FA plotted in the two-dimensional or three-dimensional feature spaces FS2 and FS3 as the data existence region PR, and identifies the region outside the existence region PR as the blank region ER. The deficiency region identification unit 24 also has a track bar 22a (see Figure 7) for the operator to change the threshold TV on the display unit 22, and is configured so that when the operator changes the threshold TV on the display unit 22, the deficiency region identification unit 24 recalculates the existence region PR and the blank region ER and displays them on the display unit 22. Alternatively, instead of the track bar 22a, there may be a numerical input unit (not shown) to input the threshold TV to be changed as a numerical value.
[0020] Furthermore, the learning data creation device 10 has a missing image generation unit 20 for supplementing the identified blank region ER with data of feature quantities FA. The missing image generation unit 20 generates a generated image GI using a GAN (Generative Adversarial Network) from a plurality of captured images PI stored in the storage unit 14 in order to extract feature quantities FA that can be supplemented in the blank region ER. As shown in Figure 7, the operator can specify a supplement region SR in the feature space displayed on the display unit 22. In the example shown in Figure 7, a method of specifying a rectangular area is shown using mouse operation, but the method is not limited to this, and the area may be specified using an input device such as a touch panel or keyboard, and the area may be a shape other than a rectangle, such as a circle or an ellipse.
[0021] When the operator specifies a supplement area SR, the missing image generation unit 20 determines whether or not a blank area ER is included in the supplement area SR. If a blank area ER is included in the supplement area SR, the missing image generation unit 20 determines that it is necessary to supplement feature quantities FA of different wear conditions and generates generated images GI of different wear conditions by referring to captured images PI of cutting edges of different cutting tool types. In the example of generating generated images GI shown here, StarGAN and CycleGAN are used among the GANs. When the missing image generation unit 20 generates the generated images GI to be supplemented, the dimensionality compression unit 18 compresses the dimensions of the newly generated generated images GI to extract feature quantities FA, and the display unit 22 plots the feature quantities FA, including the newly extracted feature quantities FA, in the feature spaces FS2 and FS3. As a result, the learning data creation device 10 is configured to draw a circle (in the case of 2D) or a sphere (in the case of 3D) with the feature quantity FA of the generated image GI as the center and the radius as the threshold TV, and to update the display of the existing region PR. This allows the operator to visually confirm whether or not the blank region ER has been replaced by the existing region PR.
[0022] Furthermore, if the supplementary region SR does not include a blank region ER, that is, if the entire supplementary region SR is the existing region PR, the missing image generation unit 20 determines that it is necessary to further increase the distribution density of feature quantities FA within the supplementary region SR, and generates a generated image GI by referring to an image PI of the same tool type as the image PI in which the feature quantities FA exist within the supplementary region SR, thereby increasing the distribution density of feature quantities FA within the supplementary region SR. In the example of generating the generated image GI shown here, SAGAN and DCGAN are used among the GANs. When the missing image generation unit 20 generates the generated image GI to be supplemented, the dimensionality compression unit 18 compresses the dimensions of the newly generated generated image GI to extract the feature quantities FA, and the display unit 22 of the learning data creation device 10 is configured to plot the feature quantities FA, including the newly extracted feature quantities FA, in the feature spaces FS2 and FS3. This allows the operator to visually confirm that the distribution density of feature quantities FA within the supplementary region SR, which is located within the existing region PR, has increased.
[0023] As shown in Figure 1, the training data creation device 10 has a clustering unit 26 for classifying the feature quantities FA by type of cutting tool TL so that, after generating the generated image GI, it can execute steps c and d-1 or d-2 once it confirms that no blank areas ER occur in the feature spaces FS2 and FS3 where the feature quantities FA are plotted, and that the distribution density of the feature quantities FA has increased sufficiently. The clustering unit 26 is configured to cluster the feature quantities FA plotted in the feature spaces FS2 and FS3 by type of cutting tool TL, such as a drill, radius end mill, or ball end mill. For clustering, algorithms such as K-means, DBSCAN, mean shift, Gaussian mixture model, and agglomerative clustering can be used. The clustering unit 26 clusters the feature quantities FA using the algorithm and settings selected by the operator, and displays symbols of different colors or shapes on the feature quantities FA for each tool type on the display unit 22. As shown in the example in Figure 8, a triangular symbol is superimposed for drills, a rhombus symbol for radius end mills, and a circular symbol for ball end mills. When the operator confirms this display and determines that the feature quantities FA have been correctly clustered for each type of cutting tool TL (see upper left of Figure 8), that is, when the distribution of feature quantities FA for each type of cutting tool TL can be spatially separated without intersecting in the feature spaces FS2 and FS3, the clustering unit 26 mixes the types of cutting tool 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 lower left of Figure 8). Furthermore, if the operator checks this display and determines that the feature quantity FA has not been clustered for each type of cutting tool TL (see upper right of Figure 8), that is, if the distribution of feature quantity 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 training data for each type of cutting tool TL from the captured image PI and generated image GI for each type of cutting tool TL (see lower right of Figure 8).In the example shown in Figure 8, a two-dimensional feature space FS2 is displayed, but it goes without saying that clustering can be performed in a three-dimensional feature space FS3. Figure 9 shows the results of determining chipping of the cutting edge for three types of cutting tools TL: ball end mills, drills, and radius end mills, using a cutting edge wear model trained with single training data and a cutting edge wear model trained with mixed training data that has been classified and mixed to distinguish between these types. As shown in Figure 9, in this embodiment, the inventor has experimentally found that a cutting edge wear determination model trained with mixed training data, which is a mixture of cutting tool TL types and classified to distinguish between each tool type, has higher determination accuracy than a model trained with single training data, and therefore it is preferable to use mixed training data. Accordingly, while operators can train with single training data, it is preferable to train with mixed training data obtained by appropriately changing the algorithm and settings used for clustering.
[0024] The learning data creation device 10, once learning data is created in the clustering unit 26, executes step e and has a disturbance addition unit 28 to add images to the learning data in which the captured image PI and generated image GI constituting the learning data have been subjected to image processing that simulates disturbances at the time of image capture, so as to ensure robustness against shooting disturbances in machine learning. This makes it possible to generate images in which disturbances have been added to the captured image PI and generated image GI. The learning data creation device 10 also has a disturbance addition parameter setting unit 30, in which the operator can set disturbance addition parameters that determine the type of disturbance and the degree of disturbance, such as image rotation, brightness change, and noise addition. The learning data creation device 10 further has a learning data output unit 32 to execute step f, and can output the package of the thus generated captured image PI, generated image GI, and disturbance-added image to the machine learning system as learning data.
[0025] The method for creating training data according to this embodiment will be explained in detail below using the flowcharts shown in Figures 2 to 4 and Figures 5 to 8.
[0026] Figure 2 shows a flowchart of the process of plotting feature quantities FA in feature spaces FS2 and FS3 and identifying blank regions ER. The learning data creation device 10 moves to step S10 to start creating learning data, and then moves to step S20. When the learning data creation device 10 moves to step S20, it activates the image data input unit 12 to acquire multiple captured images PI and stores the captured images PI in the storage unit 14. Once the captured images PI are stored, the learning data creation device 10 moves to step S30, and the dimension to be compressed is set in the dimension compression parameter setting unit 16. Here, if compression is to be to two dimensions, it moves to step S40, and if compression is to be to three dimensions, it moves to step S50.
[0027] Once the dimensions to be compressed are set, the learning data creation device 10 moves to step S60, where the dimensionality compression parameter setting unit 16 sets the details of the dimensionality compression method, that is, which dimensionality compression method to use individually or in combination, and sets the parameters required for each dimensionality compression method. Here, if dimensionality compression is performed using a single dimensionality compression method, the device moves to step S70. If dimensionality compression is performed using a combination of multiple dimensionality compression methods, such as compressing up to 40 dimensions using PCA and then further compressing down to 2 dimensions using UMAP, the device moves to step S80. Once dimensionality compression is complete, the device moves to step S90, where the learning data creation device 10 activates the display unit 22 to plot and display the 2D or 3D feature quantities FA extracted by the dimensionality compression unit 18 in the feature spaces FS2 and FS3, as shown in Figures 5 and 6. When the feature vector FA is plotted and displayed in the feature spaces FS2 and FS3, the process moves to step S100, where the learning data creation device 10 activates the missing region identification unit 24 and identifies the data existence region PR and the blank region ER based on the threshold TV set by the operator, and displays them on the feature spaces FS2 and FS3.
[0028] Figure 5 shows an example of the distribution of two-dimensional feature quantities FA plotted in a two-dimensional feature space FS2. The values of feature quantities FA on the vertical and horizontal axes of the two-dimensional feature space FS2 are displayed as dimensionless numbers. Here, the feature space FS2 displayed on the display unit 22 is plotted with only point clouds of feature quantities FA, and the system is configured so that when the operator moves the cursor on the display unit 22 to the position of the point cloud of feature quantities FA, the corresponding captured image PI is displayed. The display unit 22 can display the captured image PI that is the basis for all feature quantities FA, and the operator can switch the display state as appropriate according to the amount of information to be displayed. In addition, the missing area identification unit 24 identifies a blank area ER based on the threshold TV set by the operator sliding the track bar 22a, and the identified blank area ER is displayed in the feature space FS2 so that the operator can see it.
[0029] Figure 6 shows an example of the distribution of three-dimensional feature quantities FA plotted in a three-dimensional feature space FS3. The values of feature quantities FA on the vertical and horizontal axes of the three-dimensional feature space FS3 are displayed as dimensionless numbers. The point cloud of feature quantities FA is displayed with different patterns and display colors for each type of cutting tool TL. In addition, the display unit 22 shown in Figure 6 is configured so that when the operator moves the cursor on the display unit 22 to the position of the point cloud of feature quantities FA, the corresponding captured image PI is displayed. Furthermore, the operator can switch from displaying the point cloud to displaying the captured image PI, which is the source of all feature quantities FA. When displaying the captured image PI instead of the point cloud, the color and type of the frame of the captured image PI can be changed for each type of cutting tool TL.
[0030] Figure 3 shows a flowchart of the process for generating the generated image. The learning data creation device 10 displays the two-dimensional or three-dimensional feature quantities FA, the existing region PR, and the missing region identification unit 24 in the feature spaces FS2 and FS3, and then proceeds to step S110. When the process moves to step S110, the operator specifies the supplement region SR in the feature space where they want to supplement the data.
[0031] When the supplementary area SR is specified, the learning data creation device 10 proceeds to step S120. The learning data creation device 10 activates the deficiency area identification unit 24 to determine whether or not the blank area ER is included in the supplementary area SR. If the blank area ER is included in the supplementary area SR, the device proceeds to step S130; otherwise, it proceeds to step S140.
[0032] If the deficiency area identification unit 24 determines that a blank area ER is included in the supplement area SR and proceeds to step S130, the deficiency image generation unit 20 determines that it is necessary to supplement feature quantities FA of different wear states, reads captured images PI of cutting edges of different cutting tool types from the storage unit 14 to increase the wear states, and proceeds to step S150. On the other hand, if the deficiency area identification unit 24 determines that a blank area ER is not included in the supplement area SR and proceeds to step S140, the deficiency image generation unit 20 determines that it is necessary to further increase the distribution density of feature quantities FA in the supplement area SR, reads captured images PI of the same tool type from the storage unit 14 to increase the distribution density, and proceeds to step S150.
[0033] When the process moves to step S150, the operator selects one or more captured images PI necessary for image generation from the captured images PI read by the missing image generation unit 20 from the storage unit 14. Next, the process moves to step S160, where the operator sets image creation conditions such as the number of images to be generated. Next, the process moves to step S170, where the missing image generation unit 20 generates a generated image GI based on the already selected captured images PI and the set image creation conditions. Once the generated image GI is generated, the process moves to step S180, where the dimensionality compression unit 18 performs dimensionality 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 moves 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 area ER has been replaced by the existing area PR, or that the distribution density of feature quantities FA in the supplementary area SR within the existing area PR has increased. Once the operator confirms that the blank area ER has been replaced by the existing area PR, or that the distribution density of feature quantities FA in the supplementary area SR within the existing area PR has increased, the process moves to step S210 (see Figure 4).
[0035] Figure 4 shows a flowchart of the process for creating training data. When the process moves to step S210, the display unit 22 displays the feature quantities FA of the captured image PI and generated image GI to be clustered in the feature spaces FS2 and FS3. After displaying the feature quantities FA, the process moves to step S220, where the feature quantities FA are clustered using the algorithm and settings selected by the operator. As shown in Figure 8, symbols of different colors or shapes are superimposed on the feature quantities FA for each type of tool and displayed on the display unit 22. Next, the process moves to step S230, and if the operator determines that the feature quantities FA can be clustered for each type of cutting tool TL (see upper left of Figure 8), the process moves to step S240, where the clustering unit 26 mixes the types of cutting tools TL and creates mixed training data (see lower left of Figure 8) from a set of captured image PI and generated image GI that have been classified to identify each type of tool. Furthermore, if the operator determines that the feature quantities FA cannot be clustered for each type of cutting tool TL (see upper right of Figure 8), the process proceeds to step S250. The clustering unit 26 determines that the distribution of feature quantities FA on the feature spaces FS2 and FS3 is appropriate for each type of cutting tool TL, and creates single training data for each type of cutting tool TL from the captured image PI and generated image GI for each type of cutting tool TL (see lower right of Figure 8).
[0036] Once the training data is created, the process moves to step S260, where the disturbance application unit 28 applies image processing (such as image rotation, brightness change, and noise addition) to the captured image PI and generated image GI that constitute the training data, thereby simulating disturbances during image capture. After adding the disturbance-treated image, the process moves to step S270, where a package of the captured image PI, generated image GI, and disturbance-treated image is generated as training data and output to the machine learning system, thus ending the process.
[0037] The above embodiment includes a step in which an operator intervenes to make a decision, but is not limited to this. If the necessary computing resources can be secured, 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 supplementary area SR, and determines whether the blank area ER has been updated to the existing area PR, or whether the distribution density of feature quantities FA in the existing area PR has increased sufficiently. Here, if the dimensionality compression unit 18, the missing image generation unit 20, and the missing area identification unit 24 have sufficient computing power, the generated image GI can be created from the automatically selected captured image PI until all blank areas ER are replaced by existing areas PR by pre-setting the desired distribution density of feature quantities FA relative to the threshold TV. Furthermore, the generated image GI can be automatically generated until the distribution density in the existing area PR reaches a pre-set value. Furthermore, in process c and process d-1 or process d-2, the clustering unit 26 may automatically perform clustering by pre-setting a clustering algorithm and settings based on the operator's empirical rules and past records. In this way, the operator's judgment may be automatically processed by a computer.
[0038] According to the learning data creation device 10 and method of this embodiment, the dimensionality compression unit 18 compresses the dimensions of multiple captured images PI to extract two-dimensional or three-dimensional feature quantities FA, and the display unit 22 plots the two-dimensional or three-dimensional feature quantities FA in feature spaces FS2 and FS3 for display. Furthermore, the deficiency region identification unit 24 can identify blank regions ER of feature quantity FA data in feature spaces FS2 and FS3. This makes it easy to grasp the wear status included in and not included in the acquired captured images PI. In addition, the deficiency image generation unit 20 can generate generated image GI for extracting feature quantities FA from multiple captured images PI stored in the storage unit 14 so that feature quantity FA data can be supplemented in the identified blank regions ER. This makes it possible to generate generated image GI with a focus on blank regions ER, and to efficiently create learning 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 the generated image GI to be supplemented so as to supplement the blank region ER with feature quantity FA data based on the determination result, and generate the generated image GI from the captured image PI. This makes it possible to efficiently generate the generated image GI for supplementing the blank region ER with feature quantity FA.
[0040] Furthermore, according to the learning 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 feature quantities FA in the existing region PR, it can generate generated image GI using captured image PI of the cutting edge CE of the same type of cutting tool TL in the feature spaces FS2 and FS3. Therefore, by generating generated image GI that mimics the captured image PI, the number of images can be increased, and the distribution density of feature quantities FA in the feature spaces FS2 and FS3 can be increased. As a result, generated image GI can be generated with a focus on the blank region ER, and learning data for machine learning can be created efficiently.
[0041] Furthermore, according to the learning data creation device 10 and creation method of this embodiment, when learning data is created in the clustering unit 26, the disturbance application unit 28 can add images to the learning data that have been processed to simulate disturbances at the time of image capture, such as captured images PI and generated images GI that constitute the learning data. Therefore, in machine learning, robustness against shooting disturbances that occur during imaging, such as shifts in the field of view of the imaging device, changes in ambient brightness, and blurring due to lens dirt or mist, can be ensured.
[0042] As described above, the machine learning data creation device 10 and method for constructing a cutting edge wear determination model for cutting tools according to this embodiment can generate machine learning data accurately and efficiently.
[0043] Although embodiments of the learning data creation device 10 and the creation method have been described above, the present invention is not limited to the above embodiments. In addition to the above, it is expected that those skilled in the art will understand that various modifications of the above embodiments are possible.
[0044] 10 Learning data creation device 12 Image data input unit 14 Storage unit 16 Dimensional compression parameter setting unit 18 Dimensional compression unit 20 Missing image generation unit 22 Display unit 24 Missing region identification unit 28 Disturbance application unit 30 Disturbance application parameter setting unit CE Cutting edge FA Feature quantity FS2 Feature space (2D) FS3 Feature space (3D) GI Generated image PI Captured image ER Blank region PR Existing region SR Replenishment region TL Cutting tool
Claims
1. A method for creating training data for machine learning to construct a wear determination model for the cutting edge of a cutting tool using multiple captured images of the cutting edge of the cutting tool, the method comprising: capturing the cutting edge and acquiring multiple captured images; compressing the dimensions of the multiple captured images and extracting two-dimensional or three-dimensional features; plotting the two-dimensional or three-dimensional features extracted by the dimensionality compression in a feature space; identifying a region of feature data deficiency in the feature space from the density of the plots in the feature space; and generating a generated image from the captured images for extracting the features so that the feature data can be supplemented in the identified region of deficiency.
2. The method for creating training data according to claim 1, further comprising generating a disturbance-added image from the captured image and the generated image by applying image processing that simulates disturbances at the time of image capture.
3. The method for creating learning data according to claim 1, further comprising: determining that the feature quantities of different wear conditions should be supplemented with respect to the identified deficiency region, and generating the generated image that mimics the wear condition using the captured images relating to different types of cutting tools.
4. If the distribution of the feature quantities for each type of cutting tool can be spatially separated in the two-dimensional or three-dimensional feature space, the method for creating training data according to claim 1 further includes mixing the types of cutting tools and creating mixed training data classified so that each type of cutting tool can be identified.
5. A machine learning data creation device for constructing a wear determination model for the cutting edge of a cutting tool using multiple captured images of the cutting edge of the cutting tool, comprising: an image data input unit that photographs the cutting edge and acquires multiple captured images; a storage unit that stores the multiple captured images acquired by the image data input unit; a dimensionality reduction parameter setting unit that sets parameters for dimensionality reduction of the multiple captured images; a dimensionality reduction unit that dimensionally reduces the multiple captured images using the parameters set by the dimensionality reduction parameter setting unit and extracts two-dimensional or three-dimensional feature quantities; a display unit that plots the two-dimensional or three-dimensional feature quantities extracted by the dimensionality reduction unit in a feature space and displays them; a deficiency region identification unit that identifies a region of missing feature quantity data in the feature space; and a deficiency image generation unit that generates a generated image for extracting the feature quantities from the multiple captured images stored in the storage unit so that the feature quantity data can be supplemented in the identified deficiency region.