Model generation method, model generation apparatus, and method for determining the lifespan of hot rolling tools.

Abstract

To provide a model generation method, a model generation device and a life determination method of a tool for hot rolling which can generate a model capable of objectively determining a life of a tool for hot rolling in a short time.SOLUTION: A model generation method for generating a model for determining a life of a tool for hot rolling includes: an image acquisition step of acquiring an image obtained by photographing appearance of a tool for hot rolling; a label designation step of imparting at least one label among a plurality of labels predetermined for the appearance in the image; and a generation step of generating a learned model by machine learning with the appearance as an explanatory variable and at least the one label as an objective variable, using the image to which at least the one label is imparted.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a model generation method, a model generation apparatus, and a method for determining the life of a hot rolling tool. In particular, the present disclosure relates to a model generation method, a model generation apparatus, and a method for determining the life of a hot rolling tool for generating a model for determining the life of a hot rolling tool.

Background Art

[0002] For example, there is a plug mill rolling method as one of the methods for manufacturing seamless steel pipes. In this method, a steel slab having a round or square cross section is heated to a predetermined temperature in a heating furnace, pierced with a piercer, for example, and the hole is expanded with an elongator to form a base pipe. The base pipe is rolled while reducing the wall thickness with a plug mill, the inner and outer surfaces are smoothed with a reel, and finished to a target dimension with a sizing mill or a hot stretch reducer. For example, Patent Document 1 discloses a plug configured such that the axis of the supplied plug is spontaneously horizontal on a plug table.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Here, the life of a hot rolling tool such as a plug used in the manufacture of seamless steel pipes is generally determined by reaching the upper limit of the number of uses or by visual inspection by a person. The determination of the life of a hot rolling tool based on the number of uses tends to have a low upper limit set, and in many cases, a hot rolling tool that can actually be used is discarded. In addition, the determination of the life by visual inspection is based on the rule of thumb of the person making the determination, and variations occur in the determination results due to differences in years of experience. In addition, even for the same person, the determination result may change depending on the day, and it is difficult to achieve reproducibility and objective determination.

[0005] Generally, an image analysis method is known in which the shape of a tool in use is measured using a 2D or 3D measuring instrument for quantitative evaluation, and if the error from a reference shape exceeds a threshold, the tool's lifespan is determined to be over. However, steam is generated in hot rolling mills, and the light or infrared rays detected by non-contact measuring instruments are absorbed or their paths are bent, making accurate measurement difficult. Furthermore, when using contact-type measuring instruments, not only is measurement time-consuming, but the high temperature of the tool surface (e.g., above 100°C) makes measurement itself difficult. In addition, hot rolling tools undergo thermal expansion during use, and correction is necessary according to the temperature at the time of measurement. Moreover, if there is uneven heating, the correction value needs to be changed depending on the measurement location, making accurate measurement difficult.

[0006] This disclosure is made in view of the above circumstances and aims to provide a model generation method, a model generation apparatus, and a method for determining the lifespan of hot rolling tools that can generate a model that enables objective determination of the lifespan of hot rolling tools in a short time. [Means for solving the problem]

[0007] (1) A model generation method according to one embodiment of the present disclosure is: A model generation method for generating a model for determining the lifespan of a hot rolling tool, An image acquisition step to acquire an image of the appearance of the hot rolling tool, In the aforementioned image, a label designation step is performed in which at least one label from a predetermined set of labels is assigned to the aforementioned appearance, The process includes a generation step of generating a trained model by machine learning using the image to which the at least one label has been assigned, with the appearance as the explanatory variable and the at least one label as the target variable.

[0008] (2) As one embodiment of the present disclosure, in (1), The acquired image shows the hot rolling tool with a single color or a background of five or fewer colors.

[0009] (3) In one embodiment of the present disclosure, in (1) or (2), The label assignment step involves extracting the portion of the image that shows the appearance before assigning the at least one label.

[0010] (4) In one embodiment of the present disclosure, in any of (1) to (3), The explanatory variables further include at least one of the hot rolling tool in the image, the time used, and the length of the rolled material.

[0011] (5) In one embodiment of the present disclosure, in any of (1) to (4), The aforementioned hot rolling tool is a plug used in the manufacture of seamless steel pipes.

[0012] (6) A model generation apparatus according to one embodiment of the present disclosure is A model generation device for generating a model for determining the lifespan of hot rolling tools, An image acquisition unit that acquires an image of the appearance of the hot rolling tool, In the aforementioned image, a label designation unit assigns at least one label from a predetermined set of labels to the aforementioned appearance, The system includes a generation unit that generates a trained model by machine learning using the image to which the at least one label has been assigned, with the appearance as the explanatory variable and the at least one label as the target variable.

[0013] (7) A method for determining the lifespan of a hot rolling tool according to one embodiment of the present disclosure is: The lifespan of the hot rolling tool is determined using the trained model generated by the model generation device of (6). [Effects of the Invention]

[0014] This disclosure provides a model generation method, a model generation apparatus, and a method for determining the lifespan of a hot rolling tool, which can generate a model that enables objective determination of the lifespan of a hot rolling tool in a short amount of time.

Brief Description of the Drawings

[0015] [Figure 1] FIG. 1 is a block diagram showing a schematic configuration of a model generation device according to an embodiment of the present disclosure. [Figure 2] FIG. 2 is a diagram showing a process of a method for determining the life of a hot rolling tool using a learned model. [Figure 3] FIG. 3 is a diagram for explaining the assignment of labels. [Figure 4] FIG. 4 is a diagram illustrating an image with a single - color background. [Figure 5] FIG. 5 is a diagram for explaining the extraction of a portion showing the appearance. ​​​​​​​​​​​​​​​​Figure 1 is a block diagram showing a schematic configuration of the model generation device 20 according to this embodiment. As shown in Figure 1, the plug life learning system 1 comprises a higher-level computer 10, a model generation device 20, and a storage device 30. The higher-level computer 10 manages the model generation device 20 and the storage device 30. The higher-level computer 10 acquires data such as images stored in the storage device 30 and outputs it to the model generation device 20. The model generation device 20 performs machine learning using the data obtained from the higher-level computer 10 and generates a trained model 31. In this embodiment, the trained model 31 is a model for determining plug life. The storage device 30 is accessible from the higher-level computer 10 and the model generation device 20, and may be, for example, memory or a hard disk drive (HDD). The storage device 30 stores data used in the processing performed by the plug life learning system 1. The storage device 30 stores at least the trained model 31 and images of the plug's appearance.

[0018] The higher-level computer 10 outputs an image of the plug's appearance (hereinafter referred to as the "plug appearance image") to the model generation device 20. In other words, the higher-level computer 10 functions as an input device for the model generation device 20. The plug appearance image is captured by a known imaging device such as a digital camera, video camera, or surveillance camera, and stored in the storage device 30. Here, the imaging device is not limited to equipment used for so-called macro photography, such as a camera. For example, equipment that captures micro-images, such as a digital microscope or optical microscope, may be used as the imaging device. In this case, multiple images may be combined in two or three dimensions to create an image that shows the entire plug. The combined image may then be used as the plug appearance image.

[0019] The model generation device 20 includes an image acquisition unit 21 that performs an image acquisition step, a label assignment unit 22 that performs a label assignment step, and a generation unit 23 that performs a generation step. The image acquisition step is the step of acquiring an image (a plug appearance image in this embodiment) of the appearance of a hot rolling tool. The label assignment step is the step of assigning at least one label from a predetermined number of labels to the appearance of the plug in the image acquired in the image acquisition step. The generation step is the step of generating a trained model 31 by machine learning using the image to which at least one label has been assigned in the label assignment step. As shown by the arrows in the model generation device 20 in Figure 1, in this embodiment, the model generation device 20 performs the image acquisition step, the label assignment step, and the generation step in this order as a model generation method.

[0020] The trained model 31 is a model that determines the lifespan of a plug by inputting an image of the plug's appearance. The trained model 31 is generated before the lifespan is determined by machine learning, using the plug's appearance as an explanatory variable and at least one label as the target variable, as described above. The generated trained model 31 is stored in the memory device 30.

[0021] In this embodiment, the predetermined number of labels includes "OK," which indicates that the plug is not at the end of its lifespan and can be used continuously, and "NG," which indicates that the plug has reached the end of its lifespan. Furthermore, the predetermined number of labels is not limited to two, but may be three or more. For example, if images of the plug's appearance showing many different plug states can be obtained, then three or more labels indicating each state may be defined as the number of labels. Specific examples of three or more labels will be described later.

[0022] In this embodiment, the higher-level computer 10 also functions as a determination device that performs life determination on the target plug. That is, in this embodiment, the higher-level computer 10 executes a life determination method for hot rolling tools. After the model generation device 20 generates a trained model 31 and the generated trained model 31 is stored in the storage device 30, the higher-level computer 10 determines the life of the target plug according to the flow shown in Figure 2, for example. Figure 2 shows the processing of a plug life determination method using a trained model 31 that determines "OK" or "NG". The higher-level computer 10 acquires an image of the plug appearance of the plug to be determined and the trained model 31 (step S1). The higher-level computer 10 inputs the acquired plug appearance image into the trained model 31 and performs the determination (step S2). If the output of the trained model 31 is "OK", the higher-level computer 10 indicates to the operator performing the hot rolling work that it is possible to continue using the tool (step S3). If the output of the trained model 31 is "NG", the higher-level computer 10 indicates to the operator performing the hot rolling operation that it is unusable (step S4). In this case, the plug in question will be discarded, and a message prompting the operator to replace the plug may be displayed.

[0023] Here, in machine learning using the model generation device 20, it is necessary to have a sufficient number of plug appearance images to be used as training data. After the labeling step, the model generation device 20 may perform image processing aimed at increasing the number of plug appearance images (N-increasing the number of samples). Such image processing may be performed by the labeling unit 22 or by another functional block (e.g., the image processing unit). For image processing aimed at increasing N, general techniques used in image learning may be used. Examples include brightness change, color tone change, rotation, contrast change, enlargement, reduction, blurring, noise addition, cropping, vertical flipping, and horizontal flipping. By using one or a combination of these techniques, it is possible to increase N and prepare a sufficient number of plug appearance images for generating a trained model 31, even if the number of images acquired by the image acquisition unit 21 is small. However, it is preferable to avoid image processing that generates images that are not actually possible. For example, if the color tone cannot change in an actual image, increasing the number of plug appearance images with color tone change image processing may result in the generation of a trained model 31 with reduced judgment accuracy.

[0024] Figure 3 is a diagram illustrating the labeling process. Figure 3 shows two images of the appearance 300 of different plugs. The plug in the upper image has sufficient remaining life and is usable, so it is given the OK label 301. The plug in the lower image has adhesion of the rolled material to the tip and is at the end of its lifespan, so it is given the NG label 302. Here, general criteria may be used to determine the end of its lifespan. For example, plugs that have irregularities on the outer surface of the plug transferred to the inner surface of the rolled material, or that cannot roll the rolled material to the end, may be considered to have reached the end of their lifespan.

[0025] Here, the label designation unit 22 assigns further subdivided labels, generating a trained model 31 that allows for consideration of life-saving measures based on the judgment results. For example, instead of a uniform "NG" label, the label designation unit 22 can differentiate between, for example, NG at the tip and NG at the body, thereby subdividing the reasons for NG. In this case, if the judgment result is NG at the tip, it is thought to be deformation of the plug due to heat input at the plug tip, so the operator can consider components that increase high-temperature strength. Also, if the judgment result is NG at the body, it is presumed to be caused by peeling of the oxide scale formed on the body, so the operator can consider optimizing the heat treatment for oxide scale formation or the rolling method. In this way, subdividing the labels and generating the corresponding trained model 31 enables more detailed consideration of countermeasures.

[0026] As described above, plug exterior images can be captured using, for example, a surveillance camera. For instance, if an operator observes the plug on a surveillance camera screen, the image can be stored in the storage device 30 and acquired as a plug exterior image during the image acquisition process. In this case, it is preferable that the background of the acquired plug exterior image is a single color or five or fewer colors. By limiting the background color, the outline of the plug can be made clearer. Furthermore, it is preferable that all acquired images have a predetermined field of view and resolution. That is, it is preferable that the input image used for determining the lifespan of the target plug also has the same field of view and resolution as the plug exterior image used in machine learning.

[0027] Figure 4 shows an example of an image with a solid-color background. The more clearly the plug can be distinguished from the background, the better the accuracy of the generated trained model 31 in its detection. According to several experiments, when the background has more than five colors, the amount of information increases, and the accuracy of the trained model 31 decreases. Also, if the background contains factory wiring or other similar elements, the amount of information increases, and the accuracy of the trained model 31 decreases. It is preferable that the background be solid-colored. It is even preferable that the solid-colored background color is significantly different from the color of the plug (for example, a complementary color).

[0028] Furthermore, in the labeling step, before assigning a label, a process may be performed to extract the portion of the image that shows the appearance 300 of the plug. That is, the plug portion may be cut out from the image, or only the plug may be trimmed. Since unnecessary background information is removed, the judgment accuracy of the trained model 31 can be improved. For example, an image in which only the plug is extracted, as shown in Figure 5, is preferable as a plug appearance image because disturbances other than the plug can be removed. Alternatively, an image in which only the part to be judged (a part of the plug) is extracted may be used.

[0029] The algorithm of the generation unit 23 is not limited to any particular one. The generation unit 23 can use any algorithm commonly used in image learning. For example, GoogLeNet, which is used in image classification, may be used.

[0030] In conventional visual inspection methods for determining the lifespan of plugs, it is possible for operators to record the factors determining the lifespan as data, but this is not practical as it increases the workload and could reduce productivity. In this embodiment, when the host computer 10 performs a lifespan determination on a target plug, it can record the determination result along with the factors determining the lifespan in a storage device 30 or the like. This makes it possible to utilize past determination results and lifespan determination factors in plug development and other processes without increasing the burden on the operator.

[0031] Here, the model generation device 20 may be composed of a different computer from, for example, the host computer 10. The configuration of the computer is not particularly limited and may include, for example, memory, a CPU (processing unit), a communication unit connected to a network, a display device, and an input device. Here, when one or more programs stored in memory are read by the CPU of the computer, the CPU may be made to function as an image acquisition unit 21, a label designation unit 22, and a generation unit 23. The same applies when the host computer 10 functions as a determination device that performs life determination on the plug to be judged. In other words, in the host computer 10, when a program is read by the CPU, the CPU may be made to function as a processing unit that executes each of the processes shown in Figure 2.

[0032] As described above, the model generation method and model generation apparatus 20 according to this embodiment can generate a model that enables objective determination of the lifespan of hot rolling tools such as plugs in a short amount of time. Unlike conventional visual inspection methods, the method for determining the lifespan of hot rolling tools using the generated trained model 31 has reproducibility in the lifespan determination results and can be determined in a short amount of time without requiring special measuring equipment.

[0033] (Examples) The effects of this disclosure will be described in detail below based on examples, but this disclosure is not limited to the contents of the examples.

[0034] In this example, a life determination model for internal tools (piercer plugs) used in the hot piercing rolling (piercer rolling) process in the manufacturing of seamless steel pipes was constructed.

[0035] In this embodiment, the plug appearance images were taken after hot rolling, by removing only the piercer plug and photographing it against a single-color background. IBM's "Visual Insights" was used as the tool to obtain the plug appearance images. GoogLeNet was used as the image classification algorithm in the generation unit 23. The labeling unit 22 used two types of labels: "OK" for plugs that can be reused and "NG" for plugs that cannot be used. A total of 100 plug appearance images were prepared for use in machine learning: 50 images labeled "OK" and 50 images labeled "NG".

[0036] The accuracy of the trained model 31 generated by the higher-level computer 10 was validated using an input image different from the plug appearance image used for machine learning. Figures 6 and 7 show examples of the judgment results, respectively.

[0037] In the example shown in Figure 6, an image of the plug's appearance at the end of its lifespan was used as the input image. The field of view of the input image is the same as that of the plug's appearance image used in machine learning. As shown in Figure 6, the judgment result was "NG" with 99.9% confidence, correctly determining that it had reached the end of its lifespan.

[0038] In the example in Figure 7, an input image with a significantly different field of view from the plug appearance image used in machine learning was used. As shown in Figure 7, the judgment result was "Uncategorized" (not classified). By classifying images that have not been used in training as "Uncategorized," it is possible to accumulate images with judgment results different from "OK" and "NG." By retraining using the accumulated "Uncategorized" images, it is possible to further improve the accuracy of the lifespan judgment model.

[0039] In the examples shown in Figures 6 and 7, the judgment results were displayed within a few seconds, demonstrating sufficient practicality for determining lifespan. Furthermore, the judgment was repeated 10 times using verification images, and the results were the same each time, confirming high reproducibility.

[0040] To further improve the accuracy of the lifespan determination model, it is effective to use time, which indicates the model's usage status, as an explanatory variable in addition to image data (plug appearance). For example, the explanatory variables may further include at least one of the following: the number of times the plug (hot rolling tool) in the image has been used, the usage time (rolling time), and the length of the rolled material (rolling distance). The usage time is, for example, the cumulative time the plug has been used, and like the number of uses, it indicates the usage status of the plug. The length of the rolled material is, for example, the cumulative length of the rolled material rolled using the plug, and like the number of uses, it indicates the usage status of the plug. By including other elements that indicate the usage status of the plug, in addition to appearance, as explanatory variables in machine learning, the reproducibility of the determination results can be further improved. Furthermore, as described above, the labeling unit 22 assigns further subdivided labels to generate the trained model 31, making it possible to consider each lifespan countermeasure based on the determination results. For example, the lifespan determination model may be generated such that the average value up to the end of life is used as a boundary, and the determination results based on the plug appearance image are further divided into cases using a decision tree.

[0041] Here, to increase the amount of training data, the number of plug appearance images can be increased by N as described above, but the accuracy of the lifetime determination model may decrease depending on the method of N increase. Table 1 illustrates the correspondence between the image processing applied for N increase, the number of training data, and the evaluation results of the lifetime determination model's accuracy (model accuracy). In case 1, the lifetime determination model was generated by machine learning using only the original image. In cases 2 to 5, several image processing techniques were applied (the applied image processing is indicated as Yes in Table 1), and N increase was performed. Then, the lifetime determination model was generated using the increased amount of training data due to N increase. In cases 3 to 5, where image processing (rotation or inversion) that causes changes that are difficult to anticipate in actual images (original captured images) was applied, the model accuracy decreased. It is thought that the model accuracy decreased because the training data contained unnatural image data. According to the example in Table 1, it is preferable to perform N increase using image processing that corresponds to changes that can be expected in actual images (within the range of blurring, contrast change, cropping, and noise addition) and generate the lifetime determination model. In example 2, this increase in N improves model accuracy by 3%. Here, applying color correction does not significantly affect model accuracy. This is presumably because, for example, brightness and color tone can change in real images, such as between sunny and rainy days, so the plug exterior image after color correction does not appear unnatural. Therefore, it is possible to include color correction in image processing to address changes that can be expected in real images. However, in the example in Table 1, the model accuracy decreases slightly by including color correction.

[0042] [Table 1]

[0043] While embodiments of this disclosure have been described based on the drawings and examples, it should be noted that those skilled in the art will find it easy to make various modifications or alterations based on this disclosure. Therefore, it should be noted that these modifications or alterations are included within the scope of this disclosure. For example, the functions included in each component or step can be rearranged in a logically consistent manner, and multiple components or steps (processes) can be combined into one or divided. Embodiments relating to this disclosure can also be realized as storage media recording programs executed by a processor in the device. It should be understood that these are also included within the scope of this disclosure. [Explanation of Symbols]

[0044] 1. Learning System 10 Higher-level computers 20 Model Generators 21 Image acquisition unit 22 Label specification section 23 Generation part 30 Storage device 31 Pre-trained models 300 Exterior 301 OK Label 302 NG label

Claims

1. A model generation method for generating a model for determining the lifespan of a hot rolling tool, An image acquisition step to acquire an image of the appearance of the hot rolling tool, In the aforementioned image, a label designation step is performed in which at least one label from a predetermined set of labels is assigned to the aforementioned appearance, The process includes a generation step of generating a trained model by machine learning using the image to which the at least one label has been assigned, with the appearance as the explanatory variable and the at least one label as the target variable, The hot rolling tool is a plug used in the manufacture of seamless steel pipes, The aforementioned multiple labels are indicators of whether the plug can be used continuously or not, A model generation method for indicating that the plug is unusable for continued use, the indicators include at least one indicator indicating that the plug is unusable for continued use due to the tip and another indicator indicating that the plug is unusable for continued use due to the body.

2. The model generation method according to claim 1, wherein the acquired image has a background of one color or five or fewer colors for the hot rolling tool.

3. The model generation method according to claim 1 or 2, wherein the label designation step involves extracting a portion of the image that shows the appearance before assigning the at least one label.

4. The model generation method according to claim 1 or 2, wherein the explanatory variables further include at least one of the number of times the hot rolling tool in the image is used, the usage time, and the length of the rolled material.

5. A model generation device for generating a model for determining the lifespan of hot rolling tools, An image acquisition unit that acquires an image of the appearance of the hot rolling tool, In the aforementioned image, a label designation unit assigns at least one label from a predetermined set of labels to the aforementioned appearance, The system comprises: a generation unit that generates a trained model by machine learning using the image to which the at least one label has been assigned, with the appearance as the explanatory variable and the at least one label as the target variable; The hot rolling tool is a plug used in the manufacture of seamless steel pipes, The aforementioned multiple labels are indicators of whether the plug can be used continuously or not, A model generating device in which indicators showing that the plug cannot be used again include at least an indicator showing that the plug cannot be used again due to the tip and an indicator showing that the plug cannot be used again due to the body.

6. A method for determining the lifespan of a hot rolling tool, comprising determining the lifespan of the hot rolling tool using a trained model generated by the model generation device described in claim 5.

Images