Metal material inspection equipment, machine learning model generation method, metal material inspection method, metal material manufacturing equipment

Abstract

To provide a metal material inspection device capable of being easily installed, and determining a correct welding state.SOLUTION: There is provided a metal material inspection device for inspecting a metal material which has a weld part where a base end side of one metal plate as a preceding material and a tip end side of another metal plate as a subsequent material are welded. The metal material inspection device comprises: an image acquiring part for acquiring an image including the weld part in an imaging range; a weld state information generation part for generating weld state information being information related to a weld state of the weld part; and a totaling part for generating weld result data related to the appropriateness of quality of the weld part on the basis of the weld state, by totalizing the weld state information.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to an inspection apparatus for a metal material having a welded portion where the proximal end side of a metal plate of a preceding member and the distal end side of a metal plate of a succeeding member are welded, and the like.

Background Art

[0002] Connection is made by welding the proximal end side of a metal plate of a preceding member and the distal end side of a metal plate of a succeeding member. In such connection by welding of metal plates, confirmation of the welding state is performed. Confirmation of the welding state is performed, for example, visually by an operator. Visual confirmation has a problem that determination errors are likely to occur because the technical level for determination varies depending on the operator. Also, visual confirmation may miss an inappropriate welding state.

[0003] Therefore, confirmation of the welding state is performed by a machine. For example, in Patent Document 1, a light emission image captured during welding and a welding light feature amount are calculated based on welding conditions, and based on the calculated welding light feature amount, determination of the quality of welding is disclosed.

[0004] Also, in Patent Document 2, a quality inspection method for performing overlay welding by irradiating a welded metal object in which a plurality of welded metal members are overlapped with laser light while moving the welded metal object in the welding direction is disclosed.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0006] Patent documents 1 and 2 both use images taken during welding in a welding machine. However, the area around the welding machine is narrow, making it difficult to secure space to install a camera. Furthermore, when a camera is installed around the welding machine, spatter and other contaminants tend to adhere to the surface of the camera lens, resulting in poor maintainability.

[0007] This invention has been made in view of the above-mentioned problems, and aims to provide a metal material inspection device that can be easily installed and that can appropriately determine the welding condition. [Means for solving the problem]

[0008] To solve the above problems, the present invention has the following features. [1] A metal material inspection device for inspecting a metal material having a welded joint formed by welding the base end of one metal plate as a leading material and the front end of another metal plate as a trailing material, An image acquisition unit that acquires an image that includes the welded portion within the imaging range, A welding state information generation unit generates welding state information, which is information relating to the welding state of the welded part, based on the aforementioned image. An aggregation unit that aggregates the welding status information and generates welding result data regarding the suitability of the quality of the welded part based on the welding status, A metal material inspection device having [a specific feature / feature]. [2] The welding state information generation unit has a machine learning model that takes the image as input data and outputs welding state information of the welded part. The metal material inspection apparatus according to [1], wherein the aggregation unit aggregates the welding state information of the welded part output by the machine learning model and generates the welding result data based on the welding state. [3] The machine learning model outputs welding status information of the weld in a manner that is categorized according to the evaluation items of the weld. The metal material inspection apparatus according to [2], wherein the aggregation unit generates the welding result data in a manner divided according to the evaluation items of the welded part. [4] The aforementioned machine learning model, A first machine learning model that classifies the image of the input data into either a containing image containing the weld or a non-containing image not containing the weld, and outputs the result accordingly. The metal material inspection apparatus according to [2] or [3], comprising a second machine learning model that takes the contained image as input data and outputs welding condition information of the welded part. [5] The machine learning model has a third machine learning model that takes the contained image as input data and outputs a coordinate range corresponding to the welded area containing the welded part. The system includes a preprocessing unit that uses the coordinate range output by the third machine learning model to extract the weld-containing region from the containing image output by the first machine learning model and generate a preprocessed image. The metal material inspection apparatus described in [4], wherein the second machine learning model takes the preprocessed image generated by the preprocessing unit as input and outputs welding condition information of the welded part. [6] The preprocessing unit divides the welded portion-containing region into multiple regions in the width direction of the metal material and generates the preprocessed image. The second machine learning model takes the preprocessed images generated for each of the divided regions as input and outputs the welding state information of the welded part. The metal material inspection apparatus according to [5], wherein the aggregation unit generates welding result data regarding the suitability of the quality of the welded parts based on the welding state of the welded parts for each region. [7] A method for generating a machine learning model, wherein an image of a metal material including a weld formed by welding the base end of one metal plate as a leading material and the front end of another metal plate as a trailing material is generated as training data, and an image including the weld in its imaging range is taken as input, and welding state information of the weld is output, A method for generating a machine learning model, wherein the image used as training data comprises a region where 90% or more of the imaging range is occupied by the welded area. [8] The images used as training data are divided according to the evaluation items of the welded part, The method for generating a machine learning model according to [7], wherein the length in the width direction of the metal material of the welded portion in the divided image is defined according to the evaluation item. [9] A method for inspecting a metal material having a welded joint formed by welding the base end of one metal plate as a leading material and the front end of another metal plate as a trailing material, An image acquisition step of acquiring an image that includes the welded portion within the imaging range, An output step in which the aforementioned image is used as input data and the welding status information of the welded part is output from a machine learning model that outputs welding status information of the welded part, The aggregation step involves aggregating the welding status information of the welded part output in the output step, and generating welding result data regarding the suitability of the quality of the welded part based on the welding status, A method for inspecting metal materials, comprising [a specific characteristic].

[10] In the output step, the welding status of the weld is output in a manner that is categorized according to the evaluation items of the weld. The metal material inspection method according to [9], wherein in the aggregation step, the welding result data is generated in a manner that is categorized according to the evaluation items of the welded part.

[11] The system includes a classification step of classifying the image of the input data into either a containing image that includes the welded portion or a non-containing image that does not include the welded portion, and outputting the result. The method for inspecting a metal material according to [9] or

[10] , wherein the output step uses the contained image as input data and outputs welding condition information of the welded part.

[12] A welding part-containing area specifying step of using the contained image as input data and outputting a coordinate range corresponding to a welding part-containing area in which the welding part is included; A pre-processed image generation step of using the coordinate range to extract the welding part-containing area from the contained image and generating a pre-processed image, and having, The inspection method of a metal material according to

[11] , in the output step, using the pre-processed image as input and outputting welding state information of the welding part.

[13] In the pre-processed image generation step, the welding part-containing area is divided into a plurality of areas in the width direction of the metal material and the pre-processed image is generated, In the output step, using the pre-processed image for each of the divided plurality of areas as input and outputting the welding state of the welding part, The inspection method of a metal material according to

[12] , in the aggregation step, based on the welding state of the welding part for each area, generating welding result data regarding the suitability of the quality of the welding part.

[14] A welding apparatus for welding the base end side of a metal plate of a preceding member and the tip end side of a metal plate of a succeeding member; A welding apparatus for welding the base end side of one metal plate as a preceding member and the tip end side of another metal plate as a succeeding member; The inspection apparatus for a metal material according to any one of [1] to [6]; A manufacturing apparatus for a metal material, having.

Advantages of the Invention

[0009] The metal material inspection device of the present invention includes an image acquisition unit that acquires an image including the welded area within its imaging range, and a welding state information generation unit that generates welding state information, which is information regarding the welding state of the welded area, based on the image. Therefore, the metal material inspection device can output welding state information of the welded area without necessarily using an image taken while welding is being performed. Accordingly, with the metal material inspection device, the imaging device can be easily installed without special consideration of installation space or installation location. Furthermore, since the metal material inspection device outputs welding state information of the welded area using the welding state information generation unit, it becomes possible to appropriately determine the welding state of the welded area. [Brief explanation of the drawing]

[0010] [Figure 1] This is an explanatory diagram showing an overview of a metal material manufacturing apparatus. [Figure 2] This is a flowchart illustrating the output processing performed by a machine learning model. [Figure 3] This is an explanatory diagram showing the steps from step S101 (image acquisition step) to step S105 (extraction of the welded area) in Figure 2. [Figure 4] This is an explanatory diagram showing the subdivision steps of step S106 in Figure 2. [Figure 5] This is an explanatory diagram showing how the welding status of a welded joint is output. [Figure 6] This is an explanatory diagram showing another mode in which the welding status of a welded joint is output. [Modes for carrying out the invention]

[0011] Embodiments of the present invention will be described below with reference to the drawings. Figure 1 shows an overview of a metal material manufacturing apparatus. The metal material is not particularly limited as long as it is a material to be welded. In this embodiment, an example using steel material as an example of a metal material and a steel plate as an example of a metal plate will be described. As shown in Figure 1, the metal material manufacturing apparatus 100 has an inspection table 20 on which to place the steel material 10 as a metal material, and a welding apparatus 30 for welding the steel material 10.

[0012] The steel material 10 is placed, for example, on an inspection table 20 that is formed in a rectangular shape when viewed from above. The steel material 10 has a leading steel plate 11 located at one end of the longitudinal direction D1 of the inspection table 20, a trailing steel plate 12 located at the other end of the longitudinal direction D1, and a welded joint 13 where the leading and trailing steel plates 11 and 12 are connected by welding.

[0013] The welding apparatus 30 is not particularly limited, but various known welding machines such as laser welding machines and arc welding machines can be used. For example, a mash seam welding machine can be used as the welding apparatus 30.

[0014] The steel material manufacturing apparatus 100 has a camera 40 positioned to capture images of the welded portion 13 of the steel material 10. Preferably, the camera 40 is provided so as to be able to capture images across the width direction of the steel material 10. Multiple cameras 40 may be provided. In this case, the imaging areas of the multiple cameras 40 should be combined so that the entire width direction of the steel material 10 is captured. The camera 40 continuously captures images and sequentially transmits the data to the inspection apparatus 50, which will be described later.

[0015] The steel manufacturing apparatus 100 includes a steel inspection apparatus (hereinafter also simply referred to as the inspection apparatus) 50 for the steel material 10. The inspection apparatus 50 has a computer including a central processing unit, a storage device, and volatile memory. The inspection apparatus 50 is communicatively connected to a camera 40. Image data transmitted from the camera 40 is stored in the storage device.

[0016] The inspection device 50 includes an image acquisition unit 51 that acquires images captured by the camera 40, and a machine learning model 52 that takes the images captured by the camera 40 as input and outputs welding state information of the welded part 13. The inspection device 50 includes an aggregation unit 53 that aggregates the welding state information of the welded part 13 output by the machine learning model 52 and generates welding result data regarding the suitability of the quality of the welded part 13 based on the welding state information. The inspection device 50 includes a pre-processing unit 54 that performs image processing on the input data images. The inspection device 50 includes an input unit 55 and a display unit 56.

[0017] In other words, the machine learning model 52 generates welding state information, which is information about the welding state of the welded joint 13, based on the image captured by the camera 40. The machine learning model functions as a welding state information generation unit that generates welding state information. Examples of machine learning models 52 include well-known machine learning models such as neural networks and deep learning. The machine learning model 52 is trained using, for example, a dataset stored in a memory device as training data.

[0018] The machine learning model 52 includes a first machine learning model 52a that classifies the input image data into containing images that contain a weld or non-containing images that do not contain a weld, and outputs the result. The machine learning model 52 includes a second machine learning model 52b that takes the containing images classified by the first machine learning model 52a as input data and outputs welding state information of the weld 13. The machine learning model 52 includes a third machine learning model 52c that takes the containing images as input data and outputs a coordinate range corresponding to the weld containing region that contains the weld 13.

[0019] The first machine learning model 52a is generated using contained and non-contained images as training data. By learning using these images as training data, the first machine learning model 52a can learn the shape features of the weld 13. Therefore, the first machine learning model 52a can recognize the region in the image that matches the features of the weld 13 as the weld 13, and can classify and output the input data image into contained and non-contained images.

[0020] For the training data of the second machine learning model 52b, for example, an image including a welded joint 13 formed by welding the base end of one steel plate 11 as a leading material and the tip end of another steel plate 12 as a trailing material, and data in which the welding state of the welded joint 13 in the image are linked as a set of data can be used.

[0021] The welding condition of the welded joint 13 may be linked in a manner corresponding to the evaluation items of the welded joint 13. For example, the evaluation items of the welded joint 13 can be those listed in Table 1.

[0022] [Table 1]

[0023] For example, in the case of item No. 1 in Table 1, images showing a hole within X mm of the edge (which is the end of the steel material 10 in the width direction D2) and images showing no hole within X mm of the edge are used as training data.

[0024] By using these images as training data to train the second machine learning model 52b, the second machine learning model 52b can learn the features of edge shapes. Therefore, the second machine learning model 52b can recognize areas in an image that match the edge features as edges, and recognize areas within X mm of the edge. In addition, the second machine learning model 52b can learn the features of hole shapes. As a result, the second machine learning model 52b can recognize areas in images captured by the camera 40 that match the features of hole shapes. The second machine learning model 52b can output whether there is a hole within X mm of the edge or whether there is no hole within X mm of the edge as the welding status of the welded part 13.

[0025] It is preferable that the images used as training data for the second machine learning model 52b have a welded area occupying 90% or more of the imaging range. By using such images as training data, the shape characteristics of the welded area 13 can be precisely learned, and the accuracy of the data output from the second machine learning model 52b can be improved.

[0026] The images used as training data should have a defined length in the width direction of the steel material 10, according to the evaluation items of the welded joint 13. For example, item No. 4 in Table 1 is an item that evaluates whether the bead is formed straight, that is, whether the bead is formed linearly. In the case of such an evaluation item, the evaluation accuracy can be improved if the shape of the bead is formed to be longer. For this reason, the images used as training data for evaluation item No. 4 in Table 1 should be registered with a length D2 in the width direction of the steel material 10 that is longer than that of other evaluation items. By registering images in this way, machine learning can be performed with training data that is appropriate for the evaluation item.

[0027] The third machine learning model 52c is generated using an image containing the weld 13 and the coordinate range of the image associated with the weld 13 as training data. The images used as training data for the third machine learning model 52c should have the same coordinate system as the images used as input data. For example, the images used as training data can be those taken with coordinate systems that match those of the images used as input data, determined by shooting conditions such as the field of view and lens magnification, and the number of pixels in the images. Alternatively, a correction coefficient for matching the coordinate systems of the training data and input data images can be determined in advance, and these correction coefficients can be used to match them. In this way, even if the coordinate systems of the images used as training data and the images used as input data are misaligned, the coordinate systems of both can be matched without any problems as long as the object being photographed is within the field of view.

[0028] The aggregation unit 53 aggregates the welding condition information of the welded joint 13 output by the second machine learning model 52b. For example, if the second machine learning model 52b outputs welding condition information indicating that there is a hole within X mm from the edge for evaluation item No. 1 in Table 1, the aggregation unit 53 aggregates this as an output indicating that the quality of the welded joint 13 is unacceptable. Based on the aggregated welding condition information, the aggregation unit 53 generates welding result data regarding the suitability of the quality of the welded joint 13.

[0029] The preprocessing unit 54 is implemented, for example, by reading software stored in a storage device. The preprocessing unit 54 uses the coordinate range output by the third machine learning model 52c to extract the weld-containing region from the containing image output by the first machine learning model 52a and generates a preprocessed image.

[0030] The preprocessing unit 54 can generate a preprocessed image by dividing the weld-containing region into multiple regions in the width direction D2 of the steel material 10 and subdividing it. For example, the preprocessing unit 54 can generate a preprocessed image by dividing the weld 13 into multiple regions in a manner corresponding to the evaluation items.

[0031] Furthermore, the preprocessing unit 54 may divide the weld-containing region of image 1 into multiple regions in the width direction D2 of the steel material 10 and generate preprocessed images. Each of these preprocessed images of the divided regions may be used as training data. In this way, it is possible to create multiple training data from a single image. These training data, as described above, may be stored in a memory device, for example.

[0032] The second machine learning model 52b takes the preprocessed image generated by the preprocessing unit 54 as input and outputs welding state information of the welded part 13. This makes it possible to improve the output accuracy of the third machine learning model 52c.

[0033] The input unit 55 is an input device that accepts input operations from an operator. Examples of input devices include keyboards, mice, and touch panels.

[0034] The display unit 56 is a display that shows the output data of the machine learning model 52. The display unit 56 may also be equipped with a speaker that outputs audio data.

[0035] Figure 2 shows the process of inspecting steel materials using the inspection device 50. The process by the inspection device 50 is started, for example, when the steel material manufacturing device 100 is turned on. As shown in Figure 2, the image acquisition unit 51 acquires the image captured by the camera 40 by referring to the storage device (step S101).

[0036] When the image acquisition step of step S101 is executed, the first machine learning model 52a processes the image as input. The first machine learning model 52a classifies the input image into either a containing image that includes the weld 13 or a non-containing image that does not include the weld 13 and outputs the result (step S102).

[0037] The inspection device 50 determines whether the image output in step S102 is a contained image (step S103).

[0038] In the determination in step S103, if the image is not a containing image, that is, if the output from the first machine learning model 52a is a non-containing image (step S103: NO), the inspection device 50 terminates the process.

[0039] In the determination in step S103, if the output from the first machine learning model 52a is a containing image (step S103: YES), the third machine learning model 52c identifies the weld containing region that contains the weld 13 in the image (step S104).

[0040] Specifically, in the step of identifying the welded area in step S104, the third machine learning model 52c takes the welded area image output by the first machine learning model 52a as input and outputs a coordinate range corresponding to the welded area in the said welded area image.

[0041] When the step of identifying the welded area in step S104 is performed, the preprocessing unit 54 extracts the welded area from the containing image using the coordinate range output by the third machine learning model 52c and generates a first preprocessed image (step S105).

[0042] When the preprocessing image generation step of step S105 is executed, the preprocessing unit 54 generates a second preprocessing image by dividing the weld-containing region of the first preprocessing image generated in the preprocessing image generation step of step S105 into multiple regions in the width direction D2 of the steel material 10 (step S106).

[0043] When the subdivision step of step S106 is executed, the output step is executed (step S107) by taking each of the second preprocessed images generated in step S106 as input and outputting welding state information of the welded part 13 from the second machine learning model 52b.

[0044] When the output step of step S107 is executed, the aggregation unit 53 performs an aggregation step (step S108) to aggregate the suitability of the quality of the welded part 13 based on the welding status of the welded part 13 output in the output step.

[0045] More specifically, the aggregation unit 53 aggregates the suitability of the quality of the welded joint 13 for each second preprocessed image. Based on the aggregation results, the aggregation unit 53 generates welding result data regarding the suitability of the quality of the welded joint 13, for example, determining whether the overall quality of the welded joint 13 is appropriate. The aggregation unit 53 displays the welding result data generated in the aggregation step of step S108 on, for example, the display unit 56.

[0046] Furthermore, in the output step of step S107, the aggregation unit 53 may generate welding result data in a manner that is categorized according to the multiple evaluation items listed in Table 1.

[0047] For example, a second machine learning model 52b may be provided for each item listed in Table 1, and the aggregation unit 53 may aggregate the suitability of the quality of the welded joint 13 output from each second machine learning model 52b.

[0048] In this case, during the aggregation step of step S108, it is preferable to display the welding result data to be displayed on the display unit 56 on the screen of the display unit 56 in a selectable manner. Alternatively, during the aggregation step, the welding status information of the welded parts 13 corresponding to these evaluation items may be displayed on the screen of the display unit 56 based on the individual outputs, and the welding result data may be displayed on the screen of the display unit 56.

[0049] Figure 3 shows the steps from step S101 (image acquisition step) to step S105 (weld joint area identification step) in Figure 2. As shown in Figure 3(a), in step S101, an image captured by the camera 40 is acquired. The image shows the preceding and succeeding steel plates 11 and 12, the weld joint 13 of the steel material 10, and the inspection table 20 on which the steel material 10 is placed.

[0050] As shown in Figure 3(b), in step S104, the weld-containing region 14 is identified. The identification of the weld-containing region 14 may be performed by the third machine learning model 52c as described above.

[0051] As shown in Figure 3(c), in step S105, the welded area 14 is extracted. After the welded area 14 is extracted, a threshold can be set using the brightness difference between the steel material 10 and the background, the brightness distribution, etc., so that the background is removed from the extracted image.

[0052] Figure 4 shows an embodiment of the subdivision step S106 in Figure 2. As shown in Figure 4(a), the pre-processing unit 54 subdivides the weld-containing region 14 into multiple regions in the width direction D2 of the steel material 10. In the example shown in Figure 4(a), the pre-processing unit 54 subdivides the weld-containing region 14 into seven regions 14a to 14g.

[0053] The subdivision of the weld-containing region 14 by the pre-processing unit 54 is not limited to this configuration and can be performed arbitrarily depending on the implementation. For example, when evaluating items such as the straightness of the bead as shown in evaluation item No. 4 of Table 1 above, the region may be divided into two regions, 14h and 14i, that extend along the width direction D2 of the steel material 10, as shown in Figure 4(b). By dividing the weld-containing region 14 in this shape, it is possible to extract defects according to their characteristics and improve the accuracy of the judgment.

[0054] Furthermore, the regions in which the welded area 14 is subdivided by the preprocessing unit 54 may be pre-defined. For example, the regions in which the welded area 14 is subdivided may be set to correspond to the evaluation items listed in Table 1 above.

[0055] As described above, the steel material inspection device 50 of the present invention includes an image acquisition unit 51 that acquires an image that includes the welded portion 13 in its imaging range, and a machine learning model 52 that uses the image as input data to output welding state information of the welded portion 13. Therefore, the steel material inspection device 50 can output welding state information of the welded portion 13 without necessarily using an image taken when welding is being performed. Accordingly, with the steel material inspection device 50, the camera 40 can be easily installed without special consideration of installation space or installation location. Furthermore, since the steel material inspection device 50 outputs welding state information of the welded portion 13 using the machine learning model 52, it is possible to appropriately determine the welding state of the welded portion 13.

[0056] In particular, the steel material inspection device 50 of the present invention can also use images of the welded joint 13 after welding is completed, which allows for easy installation of the camera 40 and reduces the maintenance burden.

[0057] Conventional image processing methods for determining the welding condition of a welded joint 13 make it difficult to achieve the same level of accuracy as visual inspection by an operator. The steel inspection device 50 of the present invention uses a machine learning model 52 that takes the above-mentioned image as input, making it possible to achieve the same level of accuracy as visual inspection by an operator. Furthermore, visual inspection by an operator may result in judgment errors due to the operator's skill level or the possibility of overlooking defects. The steel inspection device 50 of the present invention outputs the welding condition of the welded joint 13 using the machine learning model 52, thus suppressing the occurrence of such judgment errors.

[0058] Furthermore, the welded joints 13 of the steel material 10 on a continuous line tend to be formed with a long width. As a result, the captured image contains many areas other than the welded joint 13. If this image is used as the input image for the machine learning model 52, the machine learning model 52 may capture features of areas other than the welded joint 13 and output welding condition information for the welded joint 13. Also, since steel material 10 of various sizes is transported on a continuous line, the welded joint 13 is not necessarily shown in a fixed position in the captured image. Therefore, in order to improve the output accuracy of the welding condition of the welded joint 13 by the machine learning model 52, it is desirable that the welded joint 13 be identified from the captured image or extracted from the captured image. The machine learning model 52 of the steel material inspection device 50 of the present invention identifies and extracts the welded joint 13 from the image. This makes it possible to further improve the output accuracy of the welding condition of the welded joint 13 by the machine learning model 52.

[0059] The machine learning model 52 should output welding status information for each of the multiple regions divided in the width direction D2 of the steel material 10 within the weld-containing region 14, and then output overall welding status information for the weld 13 based on the welding status of the weld 13 in each region. For example, even if there is an improperly welded area in a part of the weld 13, if the other regions are in an appropriate welding state, the overall welding state may be output as appropriate. Therefore, the machine learning model 52 should output welding status information indicating that the overall welding state is improper, for example, if there is an improperly welded area in one region, regardless of the results of the other regions. In this way, the machine learning model 52 can improve the output accuracy by outputting overall welding status information for the weld 13 based on the welding status of the weld 13 in each region.

[0060] Images of improperly welded areas are difficult to obtain because the occurrence rate of such improper conditions is low. Therefore, if one training data point is used for each welded area 13, collecting the training data may require an enormous amount of time and effort.

[0061] It is advisable to subdivide the welded area 14 in image 1 and use each area as training data. In this way, multiple training data can be obtained from image 1. This allows for an increase in the training frequency of the machine learning model 52, thereby improving the output accuracy.

[0062] In the above-described embodiment, the preprocessing image generation step in step S105 and the subdivision step in step S106 are processes that can be arbitrarily performed depending on the mode of implementation. That is, the machine learning model 52 may output welding state information of the welded part 13 by taking, for example, a contained image as input, instead of using the preprocessing images generated in steps S104 and S105 as input.

[0063] Furthermore, an example in which a machine learning model 52 is used as a welding state information generation unit to generate welding state information was described. The welding state information generation unit may be configured without using the machine learning model 52. The welding state information generation unit may be configured, for example, by extracting a specific region using pixel differences, pixel gradients, etc., and by generating welding state information using a pre-established reference image and a pattern matching process on the region (not shown). Even if the welding state information generation unit is configured in this way, the same effects and advantages as in the above embodiment can be obtained. [Examples]

[0064] Cameras were installed around the welding equipment, and the welding condition of the welded area was verified using the images captured by the cameras. The evaluation item for the welding condition was surface roughness, item No. 5 in Table 1 above.

[0065] A machine learning model was created using images categorized into evaluation items No. 1 to 6 as shown in Table 1 above as training data. Furthermore, the training data used consisted of images in which the steel material was subdivided in the width direction. Specifically, as shown in Figure 5, images in which the welded area 14 was subdivided into four regions 14j to 14m in the width direction of the steel material were used as training data.

[0066] The output of the machine learning model indicated that the welding condition of the welds in all four regions (14j-14m) was inadequate. Furthermore, the welding result data generated by the aggregation unit also indicated that the welding condition of the welds was inadequate. Upon further visual inspection of these regions, it was confirmed that the welding condition of the welds in all four regions (14j-14m) was indeed inadequate.

[0067] Next, for the four regions 14j to 14m shown in Figure 6, the welding condition was evaluated using the hole No. 2 in Table 1 above as the evaluation item. The output of the machine learning model indicated that the welding condition of the welds in regions 14j and 14k was inadequate. Furthermore, the welding result data generated by the aggregation unit also indicated that the welding condition of the welds was inadequate. Upon visual inspection of these regions, it was confirmed that the welding condition of the welds in regions 14j and 14k was inadequate.

[0068] As described above, it has been confirmed that the steel material inspection device of the present invention can output welding condition information that determines the welding condition of a welded joint with a high level of accuracy equivalent to that of visual inspection by a skilled operator. [Explanation of Symbols]

[0069] 100 Steel manufacturing equipment 10 Steel material 11 Steel plate (preceding material) 12 Steel plate (trailing material) 13. Welded section 30 Welding equipment 50 Inspection equipment 51 Image acquisition unit 52 Machine Learning Models 52a First Machine Learning Model 52b Second Machine Learning Model 52c Third Machine Learning Model 53. Aggregation Department 54 Pre-processing section

Claims

1. A metal material inspection device for inspecting a metal material having a welded joint formed by welding the base end of one metal plate as a leading material and the front end of another metal plate as a trailing material, An image acquisition unit that acquires an image that includes the welded portion within the imaging range, A welding state information generation unit generates welding state information, which is information relating to the welding state of the welded part according to the evaluation item, based on the aforementioned image. An aggregation unit that aggregates the suitability of the quality of the welded part according to the evaluation items of the welding state information, and generates welding result data regarding the suitability of the quality of the welded part based on the welding state information, It has, The welding state information generation unit has a machine learning model that takes the image as input data and outputs welding state information of the welded part. The aggregation unit aggregates the welding status information of the welded part output by the machine learning model, and generates the welding result data based on the welding status information. The aforementioned machine learning model, A first machine learning model that classifies the image of the input data into either a containing image containing the weld or a non-containing image not containing the weld, and outputs the result accordingly. A second machine learning model takes the aforementioned contained image as input data and outputs welding condition information of the welded part, The third machine learning model takes the aforementioned contained image as input data and outputs a coordinate range corresponding to the welded area containing the welded portion. The system includes a preprocessing unit that uses the coordinate range output by the third machine learning model to extract the weld-containing region from the containing image output by the first machine learning model and generate a preprocessed image. The second machine learning model is a metal material inspection device that takes the preprocessed image generated by the preprocessing unit as input and outputs welding condition information of the welded part.

2. The machine learning model outputs welding status information of the weld in a manner that is categorized according to the evaluation items of the weld. The metal material inspection apparatus according to claim 1, wherein the aggregation unit generates the welding result data in a manner that is divided according to the evaluation items of the welded part.

3. The preprocessing unit divides the welded portion-containing region into multiple regions in the width direction of the metal material and generates the preprocessed image. The second machine learning model takes the preprocessed images generated for each of the divided regions as input and outputs the welding state information of the welded part. The metal material inspection apparatus according to claim 1, wherein the aggregation unit generates welding result data regarding the suitability of the quality of the welded parts based on the welding state of the welded parts for each region.

4. A method for generating a machine learning model according to claim 1, A method for generating a machine learning model, wherein the image used as training data comprises a welded area where 90% or more of the imaging range is occupied by the welded area.

5. The images used as training data are divided according to the evaluation items of the welded part, The method for generating a machine learning model according to claim 4, wherein the length in the width direction of the metal material of the welded portion in the partitioned image is defined according to the evaluation item.

6. A method for inspecting a metal material having a welded joint formed by welding the base end of one metal plate as a leading material and the front end of another metal plate as a trailing material, An image acquisition step of acquiring an image that includes the welded portion within the imaging range, An output step in which a machine learning model, which outputs welding status information of the welded part according to an evaluation item, takes the aforementioned image as input data and outputs the welding status information of the welded part according to the evaluation item, The system includes an aggregation step which aggregates the suitability of the quality of the weld according to the evaluation items of the welding state information of the weld output in the output step, and generates welding result data regarding the suitability of the quality of the weld based on the welding state, The system includes a classification step of classifying the image of the input data into either a containing image that includes the welded portion or a non-containing image that does not include the welded portion, and outputting the result. The output step takes the contained image as input data and outputs welding status information of the welded part. A step to identify a welded area, using the contained image as input data and outputting a coordinate range corresponding to the welded area containing the welded part, The process includes a preprocessing image generation step of generating a preprocessed image by extracting the welded portion containing region from the containing image using the aforementioned coordinate range, A method for inspecting metal materials, wherein in the output step, the preprocessed image is taken as input and welding condition information of the welded part is output.

7. In the output step, the welding status of the weld is output in a manner that is categorized according to the evaluation items of the weld. The method for inspecting a metal material according to claim 6, wherein in the aggregation step, the welding result data is generated in a manner that is categorized according to the evaluation items of the welded part.

8. In the preprocessing image generation step, the welded portion containing region is divided into multiple regions in the width direction of the metal material, and the preprocessing image is generated. In the output step, the pre-processed images for each of the divided regions are input, and the welding state of the welded portion is output. The method for inspecting a metal material according to claim 6, wherein in the aggregation step, welding result data relating to the suitability of the quality of the welded part is generated based on the welding state of the welded part for each region.

9. A welding device for welding the base end of a leading metal plate to the leading end of a trailing metal plate, A welding apparatus for welding the base end of one metal plate as the leading material and the front end of another metal plate as the trailing material, The metal material inspection apparatus according to claim 1, A metal material manufacturing apparatus having the following features.

Images