Method for determining belt damage of a belt conveyor, method for generating a belt damage determination model, method for generating training data for a belt damage determination model, belt damage determination device, and device for generating a belt damage determination model.

Abstract

To provide a belt damage detection method and a belt damage detection device for a belt conveyor that can accurately determine belt damage even when there are fluctuations in the environment in which the belt conveyor is used. [Solution] The belt damage determination method for a belt conveyor according to the present invention includes a determination step in which a belt damage determination model is generated by machine learning using multiple training data, which determines belt damage by inputting an image of the belt taken while the belt conveyor is in operation. The model inputs a composite image, which is a composite image obtained by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt on an image of the belt taken as an image of the belt taken as an input image, and information about the belt damage on the composite image is output data.

Description

Technical Field

[0001] The present invention relates to a method for determining belt damage of a belt conveyor, a method for generating a belt damage determination model, a method for generating learning data for a belt damage determination model, a belt damage determination device, and a device for generating a belt damage determination model.

Background Art

[0002] Iron ore and coal, which are the main raw materials required in the iron-making process, are transported by ship from overseas and stored in the raw material yard. After the composition is adjusted, they are transported to blast furnace plants, sintering plants, coke plants, etc. At this time, a belt conveyor is used for transporting the raw materials. However, the belt of the belt conveyor is laid over a long distance and is also installed at a high place, so it is difficult for people to see and it is difficult to monitor its condition. However, since the belt conveyor is an important facility for transporting raw materials and the like to the production equipment, if the belt is damaged and raw materials and the like cannot be transported to the production equipment, it is necessary to stop the operation of the production equipment, etc., and the impact is large. Therefore, conventionally, in order to detect the damage formed on the belt of the belt conveyor at an early stage and repair the belt as necessary, techniques for determining belt damage have been proposed.

[0003] For example, Patent Document 1 describes a technique for determining belt damage by irradiating a laser beam onto a belt of a belt conveyor using a line laser and imaging a contour line drawn on the belt conveyor with a camera. Further, Patent Document 2 describes a technique for determining the damage level of a belt by detecting the damage site of the belt based on an image of the belt surface of the belt conveyor and specifying the shape of the damage site of the belt. Further, Patent Document 2 describes specifying the area and depth of the damage site of the belt based on the shading of the image of the belt surface. Further, Patent Document 3 describes a method for determining the deterioration level of the end portion in the width direction of the belt by inputting image data of a predetermined range including the end portion in the width direction of the belt of the belt conveyor into a deterioration level determination model learned by machine learning, and performing abnormal monitoring of the belt.

Prior Art Documents

[0004] [Patent Document 1] Japanese Patent Publication No. 2017-32346 [Patent Document 2] Japanese Patent Publication No. 2018-122995 [Patent Document 3] Japanese Patent Publication No. 2021-17296 [Overview of the project] [Problems that the invention aims to solve]

[0005] The technology described in Patent Document 1 involves irradiating a belt with laser light and capturing the outline of the laser light reflected from the belt's surface using a camera. However, the surface of belt conveyors used outdoors often becomes soiled with the transported objects and is contaminated with rainwater. Therefore, the technology described in Patent Document 1 may not be able to capture a clear image of the reflected light. Furthermore, when using a line laser to capture reflected light, it may not be possible to accurately detect damage depending on the type of damage to the belt.

[0006] On the other hand, the technology described in Patent Document 2 identifies the area and depth of damaged parts of the belt based on the density of the image of the belt surface. However, the density of the image of the belt surface changes depending on the illuminance of sunlight during the day and night, as well as the weather. For this reason, the accuracy of determining the belt damage level may not be stable with the technology described in Patent Document 2. In addition, if dirt from the conveyed object or rainwater adheres to the surface of the belt, these can become disturbances to the determination of the belt damage level, potentially leading to misjudgments of the belt damage level.

[0007] Furthermore, the technology described in Patent Document 3 uses a degradation level determination model generated by machine learning using a convolutional neural network method. According to such a machine learning model, by collecting multiple training data corresponding to changes in sunlight intensity during the day and night and weather conditions, it is possible to eliminate the effects of the aforementioned disturbances to some extent. However, the belt of a conveyor belt is not only damaged at its widthwise ends by contact with adjacent equipment or frames, but the surface of the belt can also be damaged due to the objects being conveyed. In this case, the damage formed in areas other than the widthwise ends of the belt is diverse in form, and the required level of repair usually differs. For this reason, simply determining the degradation level at the widthwise ends of the belt, as in the technology described in Patent Document 3, is not sufficient, and there is room for improvement.

[0008] The present invention has been made to solve the above problems, and its objective is to provide a belt damage detection method and a belt damage detection device for a belt conveyor that can accurately determine belt damage even when there are fluctuations in the environment in which the belt conveyor is used. Another objective of the present invention is to provide a method and a device for generating a belt damage detection model for a belt conveyor that can efficiently generate a belt damage detection model for determining belt damage to a belt conveyor. Another objective of the present invention is to provide a method for generating training data for a belt damage detection model for a belt conveyor that can efficiently generate training data for training a belt damage detection model for determining belt damage to a belt conveyor. [Means for solving the problem]

[0009] The present invention relates to a belt damage determination method for a belt conveyor, which determines damage formed on the belt of a belt conveyor, and includes a determination step of determining belt damage by inputting an image of the belt taken while the belt conveyor is in operation to a belt damage determination model that determines belt damage, which is generated by machine learning using a plurality of training data, with input being an image of the belt taken while the belt conveyor is in operation. The model inputs a composite image, which is obtained by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt on an image of the belt taken as an input, and an image of a damaged area which is an image of a damaged area formed on the belt. The composite image is used as input data, and information about the belt damage on the composite image is used as output data.

[0010] The information relating to the damage to the belt may include information relating to at least one of the type and level of damage to the belt.

[0011] The information regarding the belt damage includes information regarding the area of ​​damage to the belt, and the belt damage determination model is preferably generated by machine learning using a convolutional neural network method.

[0012] The present invention relates to a method for generating a belt damage determination model for a belt conveyor, which generates a belt damage determination model for determining damage formed on the belt of a belt conveyor, and includes a generation step of generating a composite image by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt on an image of the belt, and a damaged area image which is an image of a damaged area formed on the belt, and a learning step of generating the belt damage determination model by machine learning using a plurality of training data, with the composite image as input data and information on belt damage on the composite image as output data.

[0013] The information relating to the damage to the belt may include information relating to at least one of the type and level of damage to the belt.

[0014] The information regarding the belt damage includes information regarding the area of ​​damage to the belt, and the belt damage determination model is preferably generated by machine learning using a convolutional neural network method.

[0015] The present invention relates to a method for generating training data for a belt damage determination model for a belt conveyor, which generates training data for training a belt damage determination model that outputs information regarding damage to the belt when an image of the belt is input, and includes a generation step of generating a composite image by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt on the image of the belt, and a damaged area image which is an image of a damaged area formed on the belt, and an assignment step of generating training data by associating information regarding damage to the belt with the damaged area image on the composite image.

[0016] The belt damage determination device for a belt conveyor according to the present invention is a belt damage determination device for determining damage formed on the belt of a belt conveyor, and comprises a determination unit that determines the damage to the belt by inputting an image of the belt captured by inputting an image of the belt into a belt damage determination model, wherein the belt damage determination model is a machine learning model generated by machine learning using multiple training data, with input data being a composite image obtained by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt on the image of the belt captured by input data being a composite image obtained by combining a composite image obtained by combining a processed in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt, and an image of a damaged area formed on the belt, and output data being information about the damage to the belt on the composite image.

[0017] The present invention relates to a belt conveyor belt damage determination model generation apparatus, which generates a belt damage determination model for determining damage formed on the belt of a belt conveyor, comprising: an image generation unit that generates a composite image by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt on an image of the belt, and a damaged area image which is an image of a damaged area formed on the belt; and a learning unit that generates the belt damage determination model by machine learning using a plurality of training data, with the composite image as input data and information on belt damage on the composite image as output data. [Effects of the Invention]

[0018] According to the belt damage detection method and belt damage detection device of the present invention, belt damage can be accurately detected even when there are fluctuations in the environment in which the belt conveyor is used. Furthermore, according to the belt damage detection model generation method and generation device of the present invention, a belt damage detection model for detecting damage to a belt conveyor can be efficiently generated. Furthermore, according to the belt damage detection model training data generation method of the present invention, training data for training a belt damage detection model for detecting damage to a belt conveyor can be efficiently generated. [Brief explanation of the drawing]

[0019] [Figure 1] Figure 1 is a schematic diagram showing one example of the configuration of a belt conveyor that is the target of a belt damage determination method and belt damage determination device, which is one embodiment of the present invention. [Figure 2] Figure 2 is a block diagram showing the configuration of a belt damage detection device for a belt conveyor, which is one embodiment of the present invention. [Figure 3] Figure 3 is a block diagram showing the configuration of a belt damage determination model generation device for a belt conveyor, which is one embodiment of the present invention. [Figure 4] Figure 4 shows an example of belt wetting. [Figure 5] Figure 5 is a diagram showing an example of belt contamination. [Figure 6] Figure 6 is a diagram showing an example of image data obtained by imaging damage occurring on a belt and a damage site image corresponding to the damage. [Figure 7] Figure 7 is a schematic diagram for explaining a method of generating a synthetic image. [Figure 8] Figure 8 is a schematic diagram showing representative examples of the types and levels of damage occurring on a belt. [Figure 9] Figure 9 is a diagram showing an example in which a rectangular box is attached to a damage area shown in a synthetic image. [Figure 10] Figure 10 is a diagram showing a structural example of a convolutional neural network that performs object detection. [Figure 11] Figure 11 is a flowchart showing the flow of a method for generating a belt damage determination model of a belt conveyor according to an embodiment of the present invention. [Figure 12] Figure 12 is a flowchart showing the flow of a belt damage determination method of a belt conveyor according to an embodiment of the present invention. [Figure 13] Figure 13 is a schematic diagram showing an example in which data stored in an image server is displayed on a display device. [Figure 14] Figure 14 is a diagram showing an example of damage determined on a belt.

Embodiments for Carrying Out the Invention

[0020] Hereinafter, referring to the drawings, a belt damage determination method, a method for generating a belt damage determination model, a method for generating learning data for a belt damage determination model, a belt damage determination device, and a device for generating a belt damage determination model according to an embodiment of the present invention will be described.

[0021] 〔Belt Conveyor〕 First, referring to FIG. 1, the configuration of a belt conveyor which is a determination target of a belt damage determination method and a belt damage determination device according to an embodiment of the present invention will be described.

[0022] Figure 1 is a schematic diagram showing an example of the configuration of a belt conveyor that is the target of a belt damage determination method and belt damage determination device, which is one embodiment of the present invention. As shown in Figure 1, the belt conveyor 1 is a device that transports raw materials 16 to a chute 17 by rotating a pulley 13 using an electric motor 11 and a reduction gear 12, thereby causing an endless belt 15 to run along a roller 14. The belt conveyor 1 is equipped with a control device 10 that controls the operation of the electric motor 11.

[0023] The control device 10 controls the starting and stopping of the pulley 13. The control device 10 may also control the rotational speed of the pulley 13. In this embodiment, the belt damage detection of the belt conveyor is used to monitor abnormalities in the belt 15 by determining the type and level of damage to the belt surface of the belt 15.

[0024] In this embodiment, a camera 2 is positioned on the belt conveyor 1 to capture an image of the belt surface 15. Since the camera 2 is intended to capture an image of the belt surface, it is preferable that the raw material 16 is positioned so that it does not appear in the image captured by the camera 2. Multiple cameras 2 may be positioned at different locations. In addition, a light fixture 3 is positioned next to the camera 2 to stabilize the illumination of the belt surface 15. Multiple light fixtures 3 may be provided.

[0025] Depending on the placement of camera 2, it is preferable to install a light-shielding plate or enclose the belt surface of belt 15 to create a darkroom in order to avoid direct sunlight from behind camera 2. This is because direct sunlight may shine on the belt surface of belt 15 and cause disturbances.

[0026] In this embodiment, a two-dimensional camera is used as camera 2. It is preferable to use a color camera for camera 2 so that damage to the belt 15 can be determined from the non-uniformity of the color tone of the belt surface of the belt 15. Camera 2 is positioned so as to be able to capture an image of the belt surface of the belt 15. It is preferable to position camera 2 so as to be able to capture an image of the belt surface including the edges at the widthwise ends of the belt 15. It is preferable to position camera 2 so as to be able to capture an image of the belt surface including both edges at the widthwise ends of the belt 15.

[0027] [Belt damage detection device] Next, with reference to Figures 1 and 2, the configuration of a belt damage detection device for a belt conveyor, which is one embodiment of the present invention, will be described.

[0028] Figures 1 and 2 show the connection relationships between the camera 2, control device 10, belt damage detection device 4, display device 5, and image server 6, all located on the belt conveyor 1. The belt damage detection device 4 is connected to the camera 2, control device 10, and display device 5 via communication, either directly or via a network. Preferably, the belt damage detection device 4 is connected to the image server 6, and the image data of the belt 15 captured by the camera 2 is stored on the image server 6.

[0029] Figure 2 is a block diagram showing the configuration of a belt damage detection device for a belt conveyor, which is one embodiment of the present invention. As shown in Figure 2, the belt damage detection device for a belt conveyor (hereinafter abbreviated as "detection device") 4, which is one embodiment of the present invention, includes an acquisition unit 41, a determination unit 42, a specification unit 43, a storage unit 44, and an output unit 45.

[0030] The determination device 4 is composed of a well-known information processing device that includes a processor, memory, and a display. The determination device 4 functions as a determination unit 42 and a identification unit 43 when the processor executes a computer program. The functions of each of these units will be described later.

[0031] The acquisition unit 41 may include a communication interface for acquiring captured images P of the belt surface of the belt 15 captured by the camera 2. The acquisition unit 41 acquires the captured images P of the belt surface of the belt 15 captured by the camera 2 directly or via a network. The acquisition unit 41 may also acquire the captured images P from the image server 6 where the captured images captured by the camera 2 are stored. The acquisition unit 41 may also acquire control variables representing the driving state of the belt 15 from the control device 10, such as data on the time the image of the belt 15 was taken and the speed of the belt 15 at that time.

[0032] Preferably, the captured image P acquired by the acquisition unit 41 captures an area that includes at least one widthwise end of the belt 15. Preferably, the captured image P acquired by the acquisition unit 41 captures an area that includes both widthwise ends of the belt 15. The captured image P that includes the widthwise ends of the belt 15 allows for the determination of damage to the belt surface including the widthwise ends of the belt 15.

[0033] The determination unit 42 determines whether the belt 15 is damaged by inputting the captured image P acquired by the acquisition unit 41 into the belt damage determination model M, which will be described later. The information regarding the damage to the belt 15 includes whether or not the belt 15 is damaged as captured in the captured image P. If the belt 15 is damaged as captured in the captured image P, the information regarding the damage to the belt 15 may include information regarding the type, level, and area of ​​the damage. If the determination unit 42 receives an output indicating that the belt 15 is not damaged as captured in the captured image P, it determines that the belt 15 is not damaged as captured in the captured image P.

[0034] When the determination unit 42 receives an output indicating that the belt 15 captured in the captured image P is damaged, it determines that the belt 15 captured in the captured image P is damaged. When the determination unit 42 determines that the belt 15 captured in the captured image P is damaged, it includes information regarding at least one of the type and level of damage to the belt 15 in the determination result. If the information regarding the damage to the belt 15 includes information regarding the area of ​​the damage, the determination unit 42 may include information regarding the area of ​​the damage in the determination result.

[0035] If the determination unit 42 determines that the belt 15 is damaged, the identification unit 43 identifies information regarding the location of the damage to the belt 15. The information regarding the location of the damage to the belt represents the position of the belt 15 in the conveying direction. The position of the belt 15 in the conveying direction can be represented, for example, by using the distance along the longitudinal direction of the belt 15 from a reference point attached to the belt 15.

[0036] The identification unit 43 may obtain and identify information regarding the location of damage to the belt 15 from the image server 6 which stores the captured image P in which the damage to the belt 15 was determined. Alternatively, the identification unit 43 may detect reference points that have been pre-assigned as markers to the belt 15 from the captured image P, and identify information regarding the location of damage to the belt 15 from the time difference between the time the captured image P of the reference points was taken and the time the captured image P in which the damage to the belt 15 was determined was taken, and from the data on the speed of the belt 15.

[0037] The determination unit 42 and the identification unit 43 may be configured to include at least one processor, such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). The determination unit 42 and the identification unit 43 may each be configured with a different processor. The determination unit 42 and the identification unit 43 may be configured with a single processor. The determination unit 42 or the identification unit 43 may be configured with a single processor or with multiple processors. The processors constituting the determination unit 42 or the identification unit 43 may realize the functions of the determination device 4 by reading and executing a computer program stored in the storage unit 44.

[0038] The storage unit 44 stores the captured image P and its determination result, which have been determined to be damaged by the determination unit 42. Preferably, the storage unit 44 stores the captured image P and its determination result, which have been determined to be damaged by the determination unit 42, in association with information regarding the location of the damage identified by the identification unit 43. The storage unit 44 is, for example, an information recording medium such as a flash memory, hard disk, or memory card that can be updated. In addition to the determination result for determining damage to the belt 15, the storage unit 44 may also store computer programs or data for executing each function of the determination device 4. The storage unit 44 stores the belt damage determination model M for determining damage to the belt 15.

[0039] The output unit 45 outputs the determination result regarding the damage to the belt 15, as determined by the determination unit 42, to the display device 5. Preferably, the output unit 45 outputs the captured image P in which the damage was determined by the determination unit 42, the determination result, and information regarding the location of the damage identified by the identification unit 43 to the display device 5. The output unit 45 may also output the determination result regarding the damage to the belt 15 to the control device 10 or the image server 6.

[0040] The display device 5 is comprised of a liquid crystal display or an organic EL panel, etc. The display device 5 may also be comprised of the display of a terminal device such as a smartphone or tablet. This allows the operator of the belt conveyor 1 to take necessary actions before an operational problem occurs with the belt conveyor 1, based on the captured image P, which includes an image of the damaged part of the belt 15, and the judgment result, displayed on the display device 5. In other words, the operator can formulate a repair plan for the belt 15 according to the judgment result regarding the damage to the belt 15. In addition, the operator can easily identify the repair location of the belt 15 based on the damaged location displayed on the display device 5.

[0041] The determination device 4 is preferably connected to an image server 6 that stores image data of the belt 15 captured by the camera 2. The image server 6 may, for example, collect image data of one full rotation of the belt 15 captured by the camera 2 as captured image P and store the collected captured image P. The image server 6 may preferably store the captured image P of the belt 15 captured by the camera 2 in association with the position information of the belt 15. By accumulating the captured image P of one full rotation of the belt 15 along with the determination result from the determination unit 42, the image server 6 can detect changes in damage to the belt 15 over time. For example, it may detect changes in the damage level at a specific position on the endless belt 15 over time and output information on the changes in damage to the belt 15 over time from the image server 6 to the display device 5. This allows the operator to take necessary actions before operational problems occur with the belt conveyor 1.

[0042] [Belt damage assessment model generation device] Next, with reference to Figures 3 to 10, the configuration of a belt damage determination model generation device for a belt conveyor, which is one embodiment of the present invention, will be described.

[0043] Figure 3 is a block diagram showing the configuration of a belt damage detection model generation device for a belt conveyor, which is one embodiment of the present invention. As shown in Figure 3, the belt damage detection model generation device for a belt conveyor (hereinafter abbreviated as "generation device") 7, which is one embodiment of the present invention, comprises an acquisition unit 71, an image generation unit 72, a learning data generation unit 73, a database unit 74, and a learning unit 75.

[0044] The generation device 7 is composed of a well-known information processing device that includes a processor, memory, and a display. The generation device 7 functions as an image generation unit 72, a training data generation unit 73, and a training unit 75 when the processor executes a computer program. The functions of each of these units will be described later. The generation device 7 may be provided with the determination device 4 and may be configured as a separate information processing device. The generation device 7 may send the generated belt damage determination model M to the determination device 4 via a network.

[0045] The acquisition unit 71 may include a communication interface for acquiring captured images P of the belt surface of the belt 15, which are captured by the camera 2. The acquisition unit 71 has the same functions as the acquisition unit 41 of the determination device 4.

[0046] The image generation unit 72 generates a composite image PS by combining a processed image PI, which has a pattern simulating liquid wetting and / or dirt on the belt 15 applied to the region of the captured image P acquired by the acquisition unit 71 that corresponds to the belt 15, and a damaged area image PD, which is an image of the damaged area formed on the belt 15.

[0047] The processed image PI is an image in which a pattern simulating liquid wetting of the belt 15 of the belt conveyor 1, and / or dirt caused by dust, dirt, etc. from the conveyed object, is added to the area corresponding to the belt 15 in the captured image P. Liquid wetting of the belt 15 refers to a state in which the belt 15 is wet when the conveyed object contains liquid such as slurry, or when the belt 15 is wet with cooling water, rainwater, etc.

[0048] Figure 4 is a photograph showing an example of wetting of belt 15. The area R1 shown in Figure 4 contains a pattern caused by wetting of belt 15. The pattern caused by wetting of belt 15 is, for example, a flow-like pattern extending in the direction of conveyance on the surface of belt 15, and is darker in color than belt 15 itself. The pattern caused by wetting occurs sparsely in the widthwise center of belt 15 where the conveyed object is placed, and extends in the direction of conveyance near the widthwise ends of belt 15. The pattern simulating wetting of belt 15 is applied to have the characteristics of such a pattern caused by wetting.

[0049] On the other hand, the soiling of the belt 15 occurs when residue from the conveyed object adheres to the surface of the belt 15. Figure 5 is a photograph showing an example of soiling of the belt 15. The pattern caused by the soiling of the belt 15 is visible in region R2 shown in Figure 5. The pattern caused by the soiling of the belt 15 is a shape composed of numerous spot-like patterns, and as a whole, it shows a shape that extends in the conveying direction of the belt 15. The pattern caused by the soiling of the belt 15 often appears with a higher brightness than the belt 15 itself, mainly in the area excluding the widthwise ends of the belt 15. In addition, the color may correspond to the conveyed object, mainly in the area excluding the widthwise ends of the belt 15. The pattern that simulates the soiling of the belt 15 is applied so as to have the characteristics of such a pattern caused by soiling.

[0050] The processed image PI is generated by applying an image processing technique to the region of the captured image P corresponding to the belt 15, which adds a pattern simulating liquid wetting and / or dirt on the belt 15. The image processing technique may include at least one of the following: brightness correction, color correction, and blurring. Brightness correction simulates a change in brightness of a portion of the belt 15's surface, resulting in a difference in brightness between the clean and unclean areas of the belt. Color correction simulates a change in color tone of a portion of the belt 15's surface, resulting in a difference in color tone between the clean and unclean areas of the belt. Blurring blurs the boundaries where differences in brightness or color tone occur on the belt 15's surface. For example, by blurring the boundaries where differences in brightness or color tone occur on the belt 15 through brightness correction or color correction, patterns caused by liquid wetting and dirt on the belt 15 can be simulated.

[0051] The damaged area image PD is an image of the damaged area formed on the belt 15. The damaged area image PD is generated by first capturing images of the damage occurring on the belt 15 and then demarcating the damaged area. Figure 6 schematically shows an example of image data of damage occurring on the belt 15 and the corresponding damaged area image PD. Figure 6(a) shows image data (a-1) of a penetration defect, which is a type of damage occurring on the belt 15, and the corresponding damaged area image PD (a-2). A penetration defect is a point-like or circular defect that occurs on the belt 15, and is a state in which a hole has occurred in the belt 15, as well as damage in the stage before a hole occurs. In other words, a penetration defect includes a minor gouge that occurs on the belt 15. The damaged area image PD corresponding to a penetration defect is generated by demarcating the area in which the penetration defect has occurred.

[0052] Figure 6(b) shows image data (b-1) of a transverse tear, which is damage to the belt 15, and the corresponding damaged area image PD (b-2). A transverse tear is a crack-like defect that occurs from the widthwise end of the belt 15 toward the widthwise center. A transverse tear is damage that occurs only at the widthwise end of the belt 15 and progresses toward the widthwise center of the belt. The damaged area image PD corresponding to the transverse tear is generated by demarcating the area of ​​the crack.

[0053] Figure 6(c) shows image data (c-1) of a longitudinal tear, which is damage occurring in the belt 15, and the corresponding damaged area image PD (c-2). A longitudinal tear is a crack-like defect that occurs in the conveying direction of the belt 15. A longitudinal tear is damage that extends in the conveying direction of the belt 15. The damaged area image PD corresponding to the longitudinal tear is generated by demarcating the area of ​​the crack that extends in the conveying direction of the belt 15.

[0054] Damage area image PD can be generated by partitioning the damaged area of ​​belt 15, taking advantage of the characteristic that the damaged area of ​​belt 15 has lower brightness compared to the color tone of belt 15 itself. Damage area image PD can be generated by partitioning the damaged area using an image processing method that utilizes the difference in brightness of the image. Alternatively, damage area image PD can be generated by partitioning the damaged area using a well-known neural network that performs image segmentation.

[0055] As described above, the damaged area image PD is generated using an image of the damage that occurred on the belt 15. However, the damaged area image PD is not limited to this and may include images obtained by resizing, rotating, etc., of an image that demarcates the damaged area. It may also include images obtained by changing the brightness or color tone of an image that demarcates the damaged area. This makes it possible to generate damaged area image PDs that correspond to various forms of damage that occur on the belt 15. However, for damage such as edge tears, transverse tears, and longitudinal tears, since the damage occurs at a specific angle with respect to the conveying direction of the belt 15, it is preferable not to apply rotation processing with a rotation angle of 30° or more to these.

[0056] The composite image PS is an image created by combining the processed image PI and the damaged area image PD. The patterns formed on the belt 15 of the conveyor belt 1 are numerous due to liquid wetting and dirt. Furthermore, the type, level, and location of damage occurring on the belt 15 can also be formed in numerous patterns. Therefore, a large amount of training data is required to generate the belt damage judgment model M, and collecting this training data requires time and effort. In contrast, by generating a composite image PS, which is created by combining the processed image PI and the damaged area image PD, and using this as training data, the belt damage judgment model M can be generated efficiently.

[0057] Figure 7 is a schematic diagram illustrating the method for generating a composite image PS. As shown in Figure 7(a), the captured image P taken by camera 2 includes a region BR corresponding to belt 15 and a background BG for belt 15. In contrast, as shown in Figure 7(b), the processed image PI has patterns A simulating liquid wetting of belt 15 and patterns B simulating dirt applied to the region BR corresponding to belt 15 on the captured image P. On the other hand, as shown in Figure 7(c), the damaged area image PD is generated using an image of damage to belt 15 (such as a penetration or longitudinal tear), and can also include images that have been resized or rotated after demarcating the damaged area. Then, as shown in Figure 7(d), the composite image PS can be generated by combining the processed image PI shown in Figure 7(b) and the damaged area image PD shown in Figure 7(c). It is preferable that the composite image PS includes damaged area images PD corresponding to multiple types and levels of damage. By labeling each damaged area image PD, a highly accurate belt damage detection model M can be generated with a small amount of training data.

[0058] Since the captured image P includes the background BG of the belt 15, it is possible to perform a trimming process to remove the background BG from the captured image P beforehand, generate a processed image PI using the captured image P from which the background BG has been removed, and then generate a composite image PS by combining the damaged area image PD. In this case, multiple background BG images are generated by performing a brightness correction process on the trimmed background BG. Then, an image is generated by recombining these background BG images with the composite image PS, and this can be used as training data for generating the belt damage detection model M. By performing a brightness correction process on the background BG, a highly accurate belt damage detection model M can be generated even in environments with differences in illumination, such as day and night.

[0059] Returning to Figure 3, the training data generation unit 73 generates training data for generating a belt damage determination model M that determines damage formed on the belt 15. The training data generation unit 73 associates information about damage, including information about at least one of the type and level of damage, with the damaged areas of the belt 15 shown in the composite image PS generated by the image generation unit 72.

[0060] Information regarding the type of damage is information such as letters, symbols, codes, or numerical values ​​that identify the type of damage associated with the damaged area image PD. Since the damaged area image PD used when generating the composite image PS is generated from images of damaged areas formed on the belt 15, the type of damage is predetermined, and it is sufficient to associate it with information that identifies the type of damage. Alternatively, the information regarding the type of damage may be associated based on the shape of the damaged area image PD.

[0061] Information regarding the level of damage is information such as letters, symbols, codes, and numbers that identify the level of damage associated with the damaged area image PD. The level of damage can be determined based on the size of the damaged area shown in the composite image PS. The level of damage may also be determined based on the brightness of the damaged area shown in the composite image PS. Generally, the higher the level of damage, i.e., the more severe the damage, the larger and lower the brightness of the damaged area shown in the composite image PS, so the level of damage may be determined based on these factors.

[0062] Figure 8 is a schematic diagram showing typical examples of the types and levels of damage that occur on the belt 15. As shown in Figure 8, examples of the types of damage that occur on the belt 15 include through-holes, edge tears, transverse tears, longitudinal tears, and abrasion. Through-holes are point-like or circular scratches that occur on the belt 15, and include damage that occurs before a hole is formed in the belt 15. The level of through-holes can be determined by the size or brightness of the damaged area shown in the composite image PS. Edge tears are damage where the widthwise edge of the belt 15 is not straight but irregular and uneven, and a tear can be observed at the widthwise edge of the belt 15. The level of edge tears can be determined by the size of the tear in the damaged area shown in the composite image PS.

[0063] Transverse tears are crack-like defects that occur from the widthwise end of the belt 15 toward the widthwise center. Transverse tears can be identified by the direction in which the damage occurs. The level of transverse tears can be identified by the length of the crack that occurs from the widthwise end of the belt 15 toward the widthwise center. Longitudinal tears are crack-like defects that occur in the conveying direction of the belt 15. Longitudinal tears can be identified by the direction in which the damage occurs and the linear shape. The level of longitudinal tears can be identified by the length or thickness of the line on the belt 15. Abrasion is a band-like damage along the conveying direction of the belt 15 that occurs due to the wear of the surface of the belt 15. Abrasion is characterized by a wider damage width and longer length in the longitudinal direction than longitudinal tears. Also, because abrasion exposes the canvas inside the belt 15, it appears as a band with lower brightness than longitudinal tears. The level of abrasion can be identified by the thickness or brightness of the band on the belt 15.

[0064] The training data generation unit 73 generates training data by associating information about the type and level of damage with the damaged areas of the belt 15 shown in the composite image PS. The level of damage should be labeled in 2 to 5 stages according to the severity of the damage. Depending on the level of damage, the operator of the belt conveyor 1 can take appropriate action before operational problems occur.

[0065] The training data generation unit 73 preferably generates training data that associates information about the area of ​​damage with the type or level of damage of the damaged area displayed on the composite image PS. Information about the area of ​​damage refers to information that represents the area of ​​the damaged area shown on the composite image PS. The area of ​​damage can be generated by enclosing the damaged area shown on the composite image PS with a rectangular box (bounding box). Enclosing the damaged area shown on the composite image PS with a bounding box creates an annotation (label) for the area of ​​damage of the damaged area shown on the composite image PS. Furthermore, enclosing the damaged area shown on the composite image PS with a bounding box allows for visual recognition of the area where damage has occurred on the composite image PS.

[0066] Figure 9 shows an example where a rectangular box is placed around the damaged area in the composite image PS. As shown in Figure 9, the rectangular box is placed around the damaged area, corresponding to the type of damage. Alternatively, to place a rectangular box around the damaged area in the composite image PS, a general-purpose image recognition neural network may be used to detect the damaged area and generate a rectangular box surrounding the detected area.

[0067] Returning to Figure 3, the database unit 74 is composed of, for example, an information recording medium such as an updatable flash memory, hard disk, or memory card. The number of datasets stored in the database unit 74 is at least 10, preferably 30, and more preferably 100. The number of damaged areas shown in the composite image PS, which serves as training data, is at least 30, preferably 100, and more preferably 500.

[0068] The training data stored in the database unit 74 may include other images, as well as the composite image PS generated by the image generation unit 72. For example, if the captured image P taken by the camera 2 includes damaged areas of the belt 15, information regarding the damage to the belt 15 as seen in the captured image P may be included in the training data. Using an image P that shows patterns due to liquid wetting and / or dirt, and includes damaged areas of the belt 15, improves the accuracy of the belt damage determination model M. Furthermore, the training data stored in the database unit 74 may include a dataset that associates information indicating that the belt 15 is undamaged as seen in the captured image P, when the captured image P does not include damaged areas of the belt 15. Using images with and without belt damage as training data improves the accuracy of the belt damage determination model M in determining the presence or absence of damage.

[0069] The learning unit 75 uses the training data stored in the database unit 74 to generate a belt damage determination model M that determines damage formed on the belt 15 by machine learning using multiple training data sets, with the input being a composite image PS and the output being information about the damage of the damaged area displayed on the composite image PS. The machine learning used to generate the belt damage determination model M is not particularly limited as long as it is capable of practically sufficient belt damage determination. For example, machine learning may be performed using commonly used neural networks (including deep learning and convolutional neural networks), decision tree learning, random forests, support vector regression, etc. Furthermore, machine learning may be performed using an ensemble model that combines multiple models. In addition, machine learning may be performed using classification models such as k-nearest neighbors or logistic regression.

[0070] The learning unit 75 preferably uses machine learning with a convolutional neural network (CNN) method to generate the belt damage detection model M. The convolutional neural network comprises convolutional layers and pooling layers, and compresses the information contained in the input composite image PS while maintaining the features contained in the composite image PS. Then, by converting it into one-dimensional information using a connecting layer, it can be associated with information about the damage to the belt 15. This allows for the efficient extraction of features contained in the input captured image P as the belt damage detection model M.

[0071] The learning unit 75 preferably uses machine learning with a convolutional neural network method, outputting information about the damaged area as information about the damage to the belt 15 on the synthesized image PS in order to generate a belt damage determination model M. In particular, it is preferable to apply machine learning that includes a convolutional neural network that performs object detection.

[0072] Object detection is a technique that, in response to an image input, detects the bounding box of each object region in the image and performs class identification of that object. In other words, it detects bounding boxes that enclose the region of interest (ROI) of the objects contained in the image as rectangular areas, and performs class identification of the objects in the region of interest. Specifically, the learning unit 75 detects bounding boxes that enclose the damaged area of ​​the belt 15 as the region of interest (ROI) on the synthesized image PS, and constructs a belt damage judgment model M using a convolutional neural network that identifies "information about damage" as the object class of the detected region of interest. As a result, the belt damage judgment model M outputs an image in which the damaged area of ​​the belt 15 is enclosed by a bounding box, so the damaged area of ​​the belt 15 is output on the image.

[0073] For object detection, several convolutional neural networks can be applied, including R-CNN (Regions with CNN features), Fast R-CNN, and Faster R-CNN. R-CNN extracts multiple regions from an input image that appear to be objects, and then performs image classification on these extracted regions to detect objects. Fast R-CNN has a similar neural network structure to R-CNN for extracting regions of interest, but it also uses a neural network to identify objects within those regions. Faster R-CNN, on the other hand, improves processing speed by sharing the neural network used to generate feature maps from the input image with the neural network used to extract regions of interest.

[0074] Figure 10 shows an example of the structure of a convolutional neural network that performs object detection. In the convolutional neural network shown in Figure 10, the synthesized image PS that constitutes the training data is input to the input layer L1. If the synthesized image PS is a color image, the 2D image data is converted into RGB channel-specific image data (data obtained by converting the brightness value of the image into numerical information from 0 to 255), and the 3-channel image data is input to the input layer L1. However, the synthesized image PS may be subjected to data compression processing to compress the number of pixels in the horizontal and vertical directions before being input to the input layer L1. For example, the synthesized image PS input to the input layer L1 may be compressed to 1064 × 1064 image data. Alternatively, for example, the synthesized image PS input to the input layer L1 may be compressed to 224 × 224 image data. These processes may be performed, for example, by providing a pre-processing unit.

[0075] The composite image PS input to the input layer L1 is converted into a feature map by the feature map generation unit 81, in which the features of the composite image PS are compressed. The feature map generation unit 81 can use a convolutional neural network consisting of convolutional layers and pooling layers. This allows for compression of image data while maintaining the features of the composite image PS. In the feature map generation unit 81, the image data of the input layer L1 is compressed to approximately 1 / 76 to 1 / 16 of its original size. The feature map for the composite image PS generated by the feature map generation unit 81 is then sent to the damage information identification unit 83 and also to the region proposal unit 82.

[0076] The region proposal unit 82 is a convolutional neural network called a Region Proposal Network, which identifies candidate regions of objects in the input image using anchor boxes. An anchor box is a region specification determined by the position of a point on the feature map and its size in the horizontal and vertical directions. For each point on the feature map, the region proposal unit 82 generates anchor boxes of different sizes and aspect ratios. The region proposal unit 82 then calculates a score using the softmax function based on the region information from the generated anchor boxes and whether an object (e.g., "damaged area") is present within the anchor box or if it is part of the background. Depending on the calculated score, positive and negative example labels are assigned, and it is identified whether an object (represented here as "damaged area") is contained within the anchor box or if it is part of the background. In this way, the region proposal unit 82 identifies the region surrounding an object in the input image as a region of interest.

[0077] Furthermore, the convolutional neural network that performs object detection shown in Figure 10 uses the feature map generated by the feature map generation unit 81 and the region of interest proposed by the region proposal unit 82 based on that feature map to identify objects (information about damage) included in the region of interest using the damage information identification unit 83. The neural network constituting the damage information identification unit 83 includes a convolutional neural network (ROI pooling) that converts the feature map generated by the region proposal unit 82 into a fixed-size feature map from among the feature maps generated by the feature map generation unit 81. As a result, the size of the feature map included in the region of interest is compressed to about 7x7. The damage information identification unit 83 also applies a fully connected layer to the feature map compressed by the ROI pooling layer to generate a one-dimensional feature vector.

[0078] Furthermore, the damage information identification unit 83 takes a one-dimensional feature vector as input and outputs the classification of objects included in the region of interest to the output layer L2. In this case, the output layer L2 of the convolutional neural network shown in Figure 10 outputs information about the damage to the belt 15, specifically information about the type and level of damage. However, since the region of interest generated by the region proposal unit 82 is combined with the composite image PS, the damaged area of ​​the belt 15 is enclosed by a bounding box on the composite image PS and output, so the damaged area of ​​the belt 15 can be identified from the acquired image. Here, a bounding box refers to a region enclosed by the smallest rectangle around the damaged area detected by the convolutional neural network performing object detection, and an anchor box refers to a region on the image that is a candidate for a bounding box.

[0079] As for the training method of the convolutional neural network that performs object detection as shown in Figure 10, known methods may be used. For example, for the convolutional neural network of the region proposal unit 82, training is performed using a loss function that includes both the probability of identifying whether an object (damaged area) is contained within the anchor box or if it is part of the background, and the prediction probability of the region information of the anchor box, as described above. Then, the neural network constituting the trained region proposal unit 82 is fixed, and the structure of the entire neural network is learned using the loss function between the input layer L1 and the output layer L2. Furthermore, by alternately performing the training of the region proposal unit 82 and the overall training between the input layer L1 and the output layer L2, the judgment accuracy of the belt damage judgment model M is improved.

[0080] The belt damage detection model M may be updated to a new model by retraining, for example, every six months or every year. By updating the belt damage detection model M based on the latest data, a belt damage detection model that reflects changes in the operating conditions of the belt conveyor 1 can be generated.

[0081] [Method for generating a belt damage assessment model and method for generating training data] Next, with reference to Figure 11, a method for generating a belt damage detection model and a method for generating training data, which are embodiments of one of the present inventions, will be described.

[0082] Figure 11 is a flowchart showing the flow of a method for generating a belt damage detection model for a belt conveyor, which is one embodiment of the present invention. The flowchart shown in Figure 11 includes the flow of a method for generating training data for a belt damage detection model for a belt conveyor, which is one embodiment of the present invention. The flowchart shown in Figure 11 can be started at any time while the belt conveyor 1 is in operation.

[0083] As shown in Figure 11, in a method for generating a belt damage determination model for a belt conveyor, which is one embodiment of the present invention, first, the acquisition unit 71 of the generation device 7 acquires an image P of the belt surface of the belt 15 while the belt conveyor 1 is in operation (step S11). Next, the image generation unit 72 of the generation device 7 generates a composite image PS by combining a processed image PI, which has a pattern simulating liquid wetting and / or dirt on the belt 15 applied to the region of the image P corresponding to the belt 15, and a damaged area image PD, which is an image of a damaged area formed on the belt 15 (step S12).

[0084] Next, the training data generation unit 73 of the generation device 7 generates a dataset in which information about the damage is added to the damaged areas of the belt 15 shown in the composite image PS (step S13). Preferably, the training data generation unit 73 adds information about at least one of the type and level of damage to the belt 15 as information about the damage. Preferably, the training data generation unit 73 also adds information to identify the damaged area by dividing the damaged areas of the belt 15 on the composite image PS with rectangular boxes. By repeating the above steps S11 to S13 multiple times, training data for training the belt damage judgment model M can be generated.

[0085] Next, the learning unit 75 of the generation device 7 generates a belt damage determination model M for determining damage to the belt 15 by machine learning, using the training data generated by repeating the processes in steps S11 to S13 multiple times, with the synthesized image PS as input and information regarding damage to the belt 15 on the synthesized image PS as output (step S14). Preferably, the learning unit 75 performs machine learning so that the information regarding damage to the belt 15 includes information regarding at least one of the type and level of damage. Furthermore, preferably, the learning unit 75 performs machine learning so that the information regarding damage to the belt 15 includes a bounding box that demarcates the damaged area of ​​the belt 15.

[0086] [Method for determining belt damage] Next, with reference to Figure 12, a method for determining belt damage in a belt conveyor, which is one embodiment of the present invention, will be described.

[0087] Figure 12 is a flowchart showing the flow of a method for determining belt damage to a belt conveyor, which is one embodiment of the present invention. The flowchart shown in Figure 12 starts when an execution command for the belt damage determination process of the belt conveyor (hereinafter abbreviated as the determination process) is input to the determination device 4, and the determination process proceeds to step S21.

[0088] In step S21, the determination device 4 determines whether the belt damage determination model M has already been generated. If the determination result is that the belt damage determination model M has already been generated (step S21: Yes), the determination device 4 proceeds to step S23. On the other hand, if the belt damage determination model M has not been generated (step S21: No), the determination device 4 proceeds to step S22.

[0089] In step S22, the generation device 7 generates the belt damage determination model M. This completes the process in step S22, and the determination process proceeds to step S23.

[0090] In step S23, the acquisition unit 41 of the determination device 4 acquires an image P of the belt surface of the belt 15, which is captured by the camera 2 while the belt conveyor 1 is in operation. With this, the process of step S23 is completed, and the determination process proceeds to step S24.

[0091] In step S24, first, the determination unit 42 of the determination device 4 inputs the captured image P acquired in step S23 to the belt damage determination model M. Next, the determination unit 42 determines whether or not the belt 15 captured in the image P is damaged, based on the information about belt damage output by the belt damage determination model M. If it determines that the belt 15 is damaged, the determination unit 42 includes information about at least one of the type and level of damage in the determination result. Furthermore, if it determines that the belt 15 is damaged, it is preferable for the determination unit 42 to include information about the damaged area demarcated by the bounding box in the determination result. Also, if it determines that the belt 15 is damaged, it is preferable for the determination unit 42 to include information about the location of the damage identified by the identification unit 43 in the determination result. This makes it easy to identify the repair location of the belt 15. With this, the process of step S24 is completed, and the determination process proceeds to step S25.

[0092] In step S25, the determination device 4 determines whether it has received a signal from the control device 10 to stop the operation of the belt conveyor 1, or whether it has received a command from an external signal to stop the damage determination of the belt 15. If the determination result is that a command to stop the operation of the belt conveyor 1 or stop the damage determination has been received (step S25: Yes), the determination device 4 terminates the series of determination processes. On the other hand, if a command to stop the operation of the belt conveyor 1 or stop the damage determination has not been received (step S25: No), the determination device 4 returns to the process in step S23. [Examples]

[0093] As an example, damage to the belt 15 of the belt conveyor 1 shown in Figure 1 was determined. In this example, a belt damage determination model M was generated by machine learning using multiple training datasets, with a synthesized image PS as input and the type of damage to the belt 15 on the synthesized image PS as output. Faster R-CNN, a convolutional neural network that performs object detection, was used as the machine learning method to generate the belt damage determination model M. Specifically, an image P of the belt surface of the belt 15 was acquired, and a processed image PI was generated by adding patterns simulating liquid wetting and dirt on the belt 15 to the region corresponding to the belt 15 on the image P. The image P used was one in which the damaged part of the belt 15 was not captured. For the processed image PI, 200 images with patterns simulating liquid wetting on the belt 15 and 40 images with patterns simulating dirt were generated.

[0094] For the damaged area images (PD), images of damage that had occurred on belt 15 in the past were accumulated, so damaged area images were generated from these. As damaged area images, 10 images of through-holes, 100 images of edge tears, and 10 images of longitudinal tears were used. For the through-hole images, the through-hole portion was resized, rotated, and brightness corrected to generate 600 damaged area images (PD) representing the type of damage as a through-hole. For the edge tears, the edge tear portion was resized and brightness corrected to generate 600 damaged area images (PD) representing the type of damage as an edge tear. For the longitudinal tears, the longitudinal tear portion was resized and brightness corrected to generate 600 damaged area images (PD) representing the type of damage as a longitudinal tear. Finally, 6000 composite images (PS) were generated by combining a total of 240 processed images (PI) and a total of 1800 damaged area images (PD). Furthermore, for each damaged area on the composite image PS, the damaged area was enclosed by a rectangular box, and training data was generated by associating information about the type of damage (penetrating cut, edge tear, longitudinal tear) with the damaged area within the anchor box.

[0095] 3,000 images were generated by combining 200 images without any patterns simulating liquid wetting or dirt on the belt 15 with the damaged area image PD, and these were also included in the training data. Using the training data generated in this way, a belt damage determination model M was created that outputs the type and area of ​​belt damage when an image of the belt is input.

[0096] Next, the generated belt damage determination model M was stored in the storage unit 44 of the belt damage determination device 4 shown in Figure 2, and the damage to the belt 15 was determined while the belt conveyor 1 was in operation. In this embodiment, the image server 6 was connected to the determination device 4, and the captured image P in which the determination that the belt 15 was damaged was made, the determination result, and the information of the damage location identified by the identification unit 43 were stored in the image server 6.

[0097] Figure 13 is a schematic diagram showing an example of displaying data stored in the image server 6 on the display device 5. As shown in Figure 13, the display device 5 displays image A, which represents the damaged area of ​​belt 15 as determined on the most recent date A, and the determination result. In addition, a damage map is displayed showing the longitudinal position of belt 15 based on the damage location information for the damaged area of ​​belt 15 determined on the most recent date A. Furthermore, a damage map generated based on determination results from past dates and image B of the damaged area are displayed side by side. This made it possible to visually recognize the changes in the damaged area at the same location on belt 15 over time.

[0098] Figure 14 shows examples of damage detected on belt 15 after operating belt conveyor 1 for one month. Figure 14(a) shows an example of detecting a penetration crack and edge tear on belt 15. Figure 14(b) shows an example of detecting a longitudinal crack on belt 15. As can be seen from Figures 14(a) and (b), the damaged area on belt 15 is displayed by a bounding box, allowing for visual recognition of the location of damage on belt 15 and determination of the corresponding type of damage. This makes it possible to understand the need for repairs to belt 15.

[0099] Although embodiments applying the invention made by the present inventors have been described above, the present invention is not limited by the descriptions and drawings that constitute part of the disclosure of the present invention in this embodiment. That is, all other embodiments, examples, and operational techniques made by those skilled in the art based on this embodiment are included in the scope of the present invention. [Explanation of Symbols]

[0100] 1 Belt conveyor 2 cameras 3 Lighting 4. Belt conveyor belt damage detection device, detection device 5 Display device 6 Image Server 7. Belt conveyor belt damage detection model generation device, generation device 10 Control device 11 Electric motor 12 Reducer 13 Pulley 14 rollers 15 belts 16 Raw materials 17 Shots 41 Acquisition Department 42 Judgment section 43 Specific part 44 Storage Unit 45 Output section 71 Acquisition Department 72 Image generation unit 73. Training Data Generation Unit 74 Database Department 75 Learning Department 81 Feature Map Generation Unit 82 Area Proposal Department 83 Damage information identification department BG background L1 Input Layer L2 output layer M-belt damage assessment model P-image PD injury site images PI processed image PS composite image

Claims

1. A method for determining damage to the belt of a belt conveyor, which determines damage formed on the belt of the belt conveyor, The belt damage determination model, which determines damage to the belt, is generated by machine learning using multiple training data sets, with input data from an image of the belt taken while the belt conveyor is in operation. This model uses a composite image, which is formed by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt in an image of the belt taken from an image of the belt, and a damaged area image, which is an image of a damaged area formed on the belt, as input data, and information about the belt damage on the composite image as output data. The model includes a determination step to determine damage to the belt by inputting an image of the belt taken from the belt taken from the belt during operation. Method for determining belt damage on a conveyor belt.

2. The method for determining belt damage to a belt conveyor according to claim 1, wherein the information relating to the damage to the belt includes information relating to at least one of the type and level of damage to the belt.

3. The method for determining belt damage to a belt conveyor according to claim 1 or 2, wherein the information relating to the belt damage includes information relating to the area of ​​damage to the belt, and the belt damage determination model is generated by machine learning using a convolutional neural network method.

4. A method for generating a belt damage determination model for a belt conveyor, which generates a belt damage determination model for determining damage formed on the belt of a belt conveyor, A generation step to generate a composite image by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt in an image of the belt, and a damaged area image which is an image of a damaged area formed on the belt. A learning step in which a belt damage determination model is generated by machine learning using multiple training data, with the composite image as input data and information regarding belt damage on the composite image as output data; A method for generating a belt damage detection model for a belt conveyor, including [the specified details].

5. The method for generating a belt damage determination model for a belt conveyor according to claim 4, wherein the information relating to the belt damage includes information relating to at least one of the type and level of the belt damage.

6. The method for generating a belt damage determination model for a belt conveyor according to claim 4 or 5, wherein the information relating to the belt damage includes information relating to the area of ​​damage to the belt, and the belt damage determination model is generated by machine learning using a convolutional neural network method.

7. A method for generating training data for a belt damage determination model of a belt conveyor, wherein when an image of the belt is input, training data is generated for training a belt damage determination model that outputs information about damage to the belt, A generation step to generate a composite image by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt in an image of the belt, and a damaged area image which is an image of a damaged area formed on the belt. The steps include: generating training data by associating information about the damage to the belt with the damaged area image on the composite image; A method for generating training data for a belt damage detection model of a belt conveyor.

8. A belt conveyor belt damage determination device for determining damage formed on the belt of a belt conveyor, The system includes a determination unit that determines damage to the belt by inputting the captured image of the belt into a belt damage determination model. The belt damage determination model is a machine learning model generated by machine learning using multiple training datasets, with input data being a composite image obtained by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to the region corresponding to the belt in an image of the belt, and a damaged area image which is an image of a damaged area formed on the belt, and output data being information about the belt damage on the composite image. A device for detecting damage to the belt of a conveyor belt.

9. A belt damage determination model generation apparatus for a belt conveyor generates a belt damage determination model for determining damage formed on the belt of a belt conveyor, An image generation unit generates a composite image by combining a processed image in which a pattern simulating liquid wetting and / or dirt on the belt is applied to a region corresponding to the belt in an image of the belt, and a damaged area image which is an image of a damaged area formed on the belt. A learning unit generates a belt damage determination model by machine learning using multiple training data sets, with the aforementioned composite image as input data and information regarding belt damage on the composite image as output data. A device for generating a belt damage detection model for a belt conveyor.

Images