Continuous casting system and melt level estimation method

Abstract

A continuous casting system (1) is provided with: a camera (21) that captures an image of the inner surface of a crystallizer (5) and the melt surface of molten steel housed in the crystallizer (5) for continuous casting; and an estimation unit (123) that estimates, on the basis of new input data and an estimation model generated by machine learning so as to indicate the relationship between the input data based on moving image data including a plurality of images captured in time series by the camera (21) and the boundary position of the melt surface within the field of view of the camera (21) and the inner surface of the crystallizer (5), the new input data. And estimating the boundary position.

Description

Technical Field

[0001] The present disclosure relates to a continuous casting system and a melt level estimation method. Background Art

[0002] Patent Document 1 discloses a method of capturing an image of the contact point between the inner surface of a mold and the molten metal surface using an optical liquid level gauge, and detecting the position of the contact point between the inner surface of the mold and the molten metal surface in the captured image.

[0003] Prior art literature

[0004] Patent Literature

[0005] Patent Document 1: Japanese Patent Application Laid-Open No. 01-293953 Summary of the Invention

[0006] Problems to be solved by the invention

[0007] The present disclosure provides a continuous casting system that is effective in estimating the boundary position between the melt surface and the inner surface of the mold with high reliability.

[0008] Solutions for solving problems

[0009] A continuous casting system of a technical solution disclosed herein includes: a camera that captures the melt surface of molten steel contained in a continuous casting crystallizer and the inner surface of the crystallizer; and an estimation unit that estimates the boundary position based on an estimation model generated by machine learning in a manner that represents the relationship between input data based on dynamic image data including multiple images captured by the camera in a time series and the boundary position between the melt surface within the camera's field of view and the inner surface of the crystallizer, and new input data.

[0010] In an image, when there is a large brightness difference between the molten metal surface and the inner surface of the mold, the boundary position can be easily identified by comparing the image with a threshold value, thereby dividing the image into two areas corresponding to the molten metal surface and the inner surface of the mold. However, in continuous casting, since the brightness difference is small, there are cases where the boundary position cannot be accurately identified by the above method. For example, continuous casting is sometimes performed by supplying a powdered release agent to the molten steel injected into the mold. Continuous casting performed in this manner will hereinafter be referred to as "molding slag casting." In mold slag casting, the brightness difference between the molten metal surface and the inner surface of the mold is reduced due to the release agent floating on the molten metal surface. In contrast, according to the present device, the boundary position is estimated based on an estimation model generated by machine learning in a manner that represents the relationship between input data based on dynamic image data and the boundary position, and new input data, and based on dynamic image data for estimating the boundary position. Based on dynamic image data, information that cannot be obtained from a single image, such as information on the movement of each part within the image, can be obtained. Therefore, according to this device, even when the brightness difference is small, the boundary position can be estimated with high reliability.

[0011] The system may further include a displacement map generating unit that generates a displacement map based on the moving image data, the displacement map being a representation of a time-series displacement vector in each of a plurality of regions within the camera's field of view; an estimation model generated to represent the relationship between input data including the displacement map and the boundary position; and an estimation unit estimating the boundary position based on the estimation model and new input data including the newly generated displacement map. By converting the time-series displacement trend distribution data into the displacement map and including it in the input data, the boundary position can be easily estimated with high reliability.

[0012] Alternatively, the system may further include a brightness map generator that generates a brightness map representing the brightness of each of the plurality of regions based on the dynamic image data, wherein the estimation model is generated in a manner that represents the relationship between the input data including the brightness map and the boundary position, and the estimation unit estimates the boundary position based on the estimation model and new input data including the newly generated brightness map. When mold slag floats on the melt surface, a certain degree of brightness difference may occur between the melt surface and the inner surface of the mold. Therefore, by further including the brightness map in the input data, the boundary position can be estimated with higher reliability.

[0013] Alternatively, the brightness map generation unit may generate a brightness map for each of the plurality of colors, the estimation model may be generated to represent the relationship between the input data further including the brightness maps for the plurality of colors and the boundary position, and the estimation unit may estimate the boundary position based on the estimation model and new input data further including the newly generated brightness maps for the plurality of colors. Even when the brightness difference between the melt surface and the inner surface of the mold is small, a relatively large difference can be generated between the color of the melt surface and the color of the mold. Therefore, by further including the brightness maps for the plurality of colors in the input data, the boundary position can be estimated with higher reliability.

[0014] Alternatively, the system may further include: a correct data acquisition unit that acquires correct data, which is not based on the estimated model but represents the boundary position identified based on the dynamic image data; an accumulation unit that accumulates learning records obtained by associating the input data with the correct data in a database; and a model generation unit that generates the estimated model through machine learning based on the plurality of learning records accumulated in the database. Since the system performs a series of processes from accumulating the learning records to generating the estimated model based on the accumulated learning data, convenience can be improved.

[0015] The model generation unit may generate the estimation model through deep learning. Deep learning can model the difficult-to-formulate relationship between input data based on moving image data and boundary positions with high reliability.

[0016] The mold may further include a control unit that controls the flow of molten steel into the mold based on the estimated boundary position so that the melt level approaches a target height. The boundary position estimation result can be effectively used to automatically adjust the melt level.

[0017] Alternatively, the control unit may calculate the height of the melt surface based on the amount of molten steel flowing into the mold, and if the difference between the height of the melt surface corresponding to the boundary position estimated by the estimation unit and the result of the calculation of the melt surface height based on the amount of molten steel flowing in exceeds a predetermined level, the control unit may not perform control of the amount of molten steel flowing in based on the estimated result of the boundary position. This can improve the reliability of control based on the estimated result of the boundary position.

[0018] Another technical solution of the present disclosure includes a method for estimating a molten steel surface, comprising: obtaining dynamic image data comprising a plurality of images captured in a time series using a camera that captures the molten steel surface and the inner surface of the molten steel contained in a continuous casting mold; generating input data based on the dynamic image data; obtaining correct data representing a recognition result of the dynamic image data based on the boundary position between the molten steel surface and the inner surface of the molten steel mold within the camera's field of view; accumulating learning records obtained by associating the input data with the correct data in a database; generating an estimation model representing the relationship between the input data and the boundary position based on the plurality of learning records accumulated in the database; generating new input data based on the newly acquired dynamic image data; and estimating the boundary position based on the estimation model and the new input data.

[0019] Effects of the Invention

[0020] According to the present disclosure, it is possible to provide a continuous casting system that is effective in estimating the boundary position between the molten metal surface and the inner surface of the mold with high reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a diagram schematically illustrating the structure of a continuous casting system.

[0022] Figure 2 This is a block diagram illustrating the functional structure of a control system.

[0023] Figure 3 is a diagram schematically illustrating a displacement map.

[0024] Figure 4 is a diagram schematically illustrating an estimation model.

[0025] Figure 5 This is a block diagram illustrating the hardware configuration of a control system.

[0026] Figure 6 This is a flowchart illustrating the steps of generating an estimation model.

[0027] Figure 7 This is a flowchart illustrating the steps of controlling the melt level. DETAILED DESCRIPTION

[0028] Hereinafter, the embodiments will be described in detail with reference to the accompanying drawings. In the description, the same elements or elements having the same functions are denoted by the same reference numerals, and repeated descriptions will be omitted.

[0029] 〔Continuous Casting System〕

[0030] Figure 1 It is a diagram schematically illustrating the structure of a continuous casting system. Figure 1The continuous casting system 1 shown is a system for producing steel billets by continuous casting. As an example, the continuous casting system 1 produces steel billets by mold flux casting in which a powdered release agent is supplied to molten steel contained in a mold.

[0031] The continuous casting system 1 includes a ladle 2, a nozzle 3, a tundish 4, a mold 5, a vibrating device 11, an electromagnetic stirring device 12, a camera 21, and a control system 100. The ladle 2 holds molten steel produced in a blast furnace. The ladle 2 has a pouring hole 6 and a spout 7 at its bottom. The pouring hole 6 discharges the molten steel downward. The spout 7 opens and closes the pouring hole 6.

[0032] The nozzle 3 is located below the pouring hole 6 and guides the molten steel flowing from the pouring hole 6 downward. The tundish 4 receives the molten steel guided from above by the nozzle 3. The tundish 4 includes a nozzle 8 and a pouring gate 9. The nozzle 8 is located below the tundish 4 and guides the molten steel contained in the tundish 4 downward. The pouring gate 9 opens and closes the flow path of the molten steel in the nozzle 8.

[0033] The mold 5 receives molten steel, which is guided from above by a nozzle 8, and delivers it downward while shaping it. The lower end of the nozzle 8 can be immersed in the molten steel received by the mold 5. The molten steel gradually solidifies as it passes through the mold 5. The molten steel delivered downward through the mold 5 is conveyed by multiple conveyor rollers and cut into billets using a gas cutter or other similar machine.

[0034] The vibrating device 11 vibrates the mold 5 in the vertical direction so that the mold release agent melted on the molten steel surface flows between the inner surface of the mold 5 and the molten steel. For example, the vibrating device 11 vibrates the mold 5 in the vertical direction using power from an electric motor or a hydraulic motor (e.g., a hydraulic motor).

[0035] The electromagnetic stirring device 12 stirs the molten steel in the mold 5 to prevent impurities in the molten steel from being fixed to the solidified portion of the molten steel. For example, the electromagnetic stirring device 12 generates a magnetic field in the mold 5 to flow the molten steel in response to power supply.

[0036] The camera 21 captures the surface of the molten steel contained in the mold 5 and the inner surface of the mold 5. For example, the camera 21 is a visible light camera capable of capturing color images. It captures the surface of the molten steel and the inner surface of the mold 5 from obliquely above, passing between the tundish 4 and the mold 5. For example, the camera 21 includes an imaging element such as a CCD (Charge Coupled Device) or a CMOS (Complementary Metal Oxide Semiconductor) image sensor. Image data of the molten steel surface and the inner surface of the mold 5 are generated based on the electrical signals generated by the imaging element in response to incident light. The image data, for example, includes pixel values ​​for each of a plurality of pixels arranged in a matrix. A pixel value represents brightness. For each of the plurality of pixels, the image data may also include pixel values ​​for a plurality of colors. For example, the plurality of colors may be RGB (Red, Green, and Blue).

[0037] The control system 100 identifies the boundary position between the melt surface and the inner surface of the mold 5 within the image captured by the camera 21 (e.g., the aforementioned image data). Based on the identification result of the boundary position, the control system 100 controls the gate 9 to bring the melt surface level closer to the target height. If the brightness difference between the melt surface and the inner surface of the mold 5 (the difference between the brightness of the melt surface and the brightness of the inner surface of the mold 5) within the image is large, the boundary position can be easily identified by comparing the image with a threshold value, thereby dividing the image into a region corresponding to the melt surface and a region corresponding to the inner surface of the mold 5.

[0038] However, in continuous casting, the aforementioned brightness difference is relatively small, making it difficult to accurately identify the boundary using this method. For example, in mold flux casting, the release agent floating on the melt surface tends to reduce the brightness difference. Currently, mold flux casting is the mainstream method in continuous casting, and therefore, the small brightness difference often makes it difficult to accurately identify the boundary.

[0039] Therefore, the control system 100 is configured to: estimate the boundary position based on an estimation model generated through machine learning to represent the relationship between input data based on dynamic image data, including multiple images captured in a time series by the camera 21, and the boundary position; and to adjust the melt surface to a target height based on the boundary position estimation result. Using dynamic image data, information such as motion information of each part within the image, which cannot be obtained from a single image, can be obtained. Therefore, the control system 100, which estimates the boundary position based on an estimation model generated through machine learning to represent the relationship between input data based on dynamic image data and the boundary position, and the new input data, can estimate the boundary position with high reliability. This allows the melt surface to be maintained near the target height with high reliability, thereby improving the quality of the aforementioned billet.

[0040] Figure 2 is a block diagram illustrating the functional structure of the control system 100. Figure 2 As shown, the control system 100 has a recording unit 111, a dynamic image data storage unit 112, a displacement map generation unit 113, a brightness map generation unit 114, an input data generation unit 115, a correct data acquisition unit 116, an accumulation unit 117 and a database 118 as functional components (hereinafter referred to as "functional blocks").

[0041] The recording unit 111 acquires multiple images captured in a time series by the camera 21. For example, the camera 21 is caused to repeatedly capture images at a predetermined cycle, and the recording unit 111 acquires multiple images obtained through this repeated capturing over a predetermined period of time. The recording unit 111 stores the acquired multiple images in a time series in the motion picture data storage unit 112. Thus, the motion picture data containing the multiple images in a time series is at least temporarily stored in the motion picture data storage unit 112.

[0042] Dynamic image data is data that enables visual recognition of the displacement of elements contained in an image by displaying a plurality of images in sequence in a time series. The reason for visually recognizing the displacement of elements contained in an image through dynamic image data is that elements contained in a previous image are also contained in a subsequent image at a position different from that in the previous image. Hereinafter, such common elements contained in both the previous image and the subsequent image at mutually different positions will be referred to as "common displacement elements". In the case where the common displacement elements are not contained in the previous image and the subsequent image because the above-mentioned prescribed period is too long, the multiple images obtained by repeatedly shooting at the prescribed period do not belong to dynamic image data. The prescribed period for obtaining multiple images belonging to dynamic image data is, for example, less than 0.1 seconds, less than 0.06 seconds, or less than 0.03 seconds.

[0043] The displacement map generation unit 113 generates a displacement map based on the moving image data. The displacement map is data that represents a temporal displacement vector for each of the multiple regions within the field of view of the camera 21. For example, the displacement map generation unit 113 uses one image included in the moving image data as a reference image and the temporally preceding or succeeding image as a comparison image, and generates the displacement map by comparing the reference image with the comparison image. For example, the displacement map generation unit 113 generates the displacement map by performing the following processing on each of the multiple regions.

[0044] Process 1) In the reference image, common displacement elements are identified within the region.

[0045] Process 2) Calculate the displacement vector of the region based on the difference between the position of the common displacement element in the reference image and the position of the common displacement element in the comparison image.

[0046] The displacement map generation unit 113 may also generate a displacement map using multiple images included in the dynamic image data as reference images. For example, the displacement map generation unit 113 may perform the above-described processing on each of the multiple reference images to calculate multiple displacement vectors for each of the multiple regions, and then average the multiple calculated displacement vectors to calculate a displacement vector for each of the multiple regions.

[0047] Figure 3 is a diagram schematically illustrating a displacement map. Figure 3 In the displacement map 200 shown, a displacement vector 211 is shown in each of a plurality of regions 210. Each region 210 is a region obtained by dividing the field of view of the camera 21 into a matrix. Each region 210 has a plurality of pixels arranged in a matrix.

[0048] The displacement vector 211 in the time series indicates the direction and amount of displacement of the common displacement element in the image. Figure 3In the figure, the dotted line represents the boundary between the molten steel surface and the inner surface of the mold 5. In region 210A corresponding to the molten steel surface and region 210B corresponding to the inner surface of the mold 5, at least one of the direction and magnitude of the displacement vector 211 is likely to differ. For example, in the figure, the displacement vector 211 in region 210B corresponding to the inner surface of the mold 5 tends to be in the vertical direction, corresponding to the vibration direction of the vibrating device 11. In contrast, the displacement vector 211 in region 210A corresponding to the molten steel surface tends to be in the direction of flow by the electromagnetic stirring device 12. As such, due to the difference in the inclination of the displacement vector 211 in region 210A and region 210B, the boundary between the molten steel surface and the inner surface of the mold 5 is likely to appear in the displacement map.

[0049] return Figure 2 The brightness map generation unit 114 generates a brightness map based on the dynamic image data. The brightness map represents the brightness of each of the plurality of regions. For example, the brightness map generation unit 114 includes the brightness value of each of the plurality of regions 210. The brightness map generation unit 114 may also calculate the brightness value of each of the plurality of regions 210 based on the brightness values ​​of the plurality of pixels. For example, the brightness map generation unit 114 may calculate the average of the brightness values ​​of the plurality of pixels as the brightness value of each of the plurality of regions 210.

[0050] The brightness map generation unit 114 may calculate the brightness value of each of the plurality of regions 210 based on the brightness values ​​in the plurality of images of the moving image data. For example, the brightness map generation unit 114 may calculate the average of the brightness values ​​in the plurality of images as the brightness value of each of the plurality of regions 210.

[0051] The brightness map generation unit 114 may also generate a brightness map for each of the plurality of colors described above. For example, the brightness map generation unit 114 calculates the brightness value of R (Red) for each of the plurality of regions 210 to generate an R brightness map. For example, the brightness map generation unit 114 calculates the brightness value of R based on the brightness values ​​of the pixels for R included in each of the plurality of regions 210. Similarly, the brightness map generation unit 114 calculates the brightness value of G (Green) for each of the plurality of regions 210 to generate a G brightness map, and calculates the brightness value of B (Blue) for each of the plurality of regions 210 to generate a B brightness map.

[0052] The input data generation unit 115 generates input data based on the dynamic image data. For example, the input data generation unit 115 generates input data that includes at least the displacement map generated by the displacement map generation unit 113. As an example, the input data generation unit 115 generates input data that includes the displacement map generated by the displacement map generation unit 113 and the brightness map generated by the brightness map generation unit 114. If the brightness map generation unit 114 generates brightness maps for each of a plurality of colors, the input data generation unit 115 generates input data that includes the displacement map generated by the displacement map generation unit 113 and the brightness map generated by the brightness map generation unit 114 for each of the plurality of colors.

[0053] The correct data acquisition unit 116 acquires correct data. The correct data is not based on the above-mentioned estimation model but represents the boundary position identified based on the dynamic image data. For example, the correct data acquisition unit 116 may also acquire data representing the boundary position identified by the user based on the dynamic image data based on input to the user interface 196 (described later). The accumulation unit 117 accumulates the learning records obtained by associating the input data generated by the input data generation unit 115 with the correct data acquired by the correct data acquisition unit 116 in the database 118. For example, the accumulation unit 117 generates learning records by associating input data based on the same dynamic image data with the correct data, and accumulates them in the database 118.

[0054] The control system 100 further includes a model generation unit 121, a model storage unit 122, an estimation unit 123, and a control unit 124 as functional blocks. The model generation unit 121 generates the estimation model by machine learning based on a plurality of learning records accumulated in the database 118 and stores the generated estimation model in the model storage unit 122.

[0055] When the input data generated by the input data generating unit 115 includes a displacement map, the model generating unit 121 generates an estimation model so as to express the relationship between the input data including the displacement map and the boundary position.

[0056] When the input data generated by the input data generating unit 115 further includes a luminance map, the model generating unit 121 generates an estimation model so as to express the relationship between the input data further including the luminance map and the boundary position.

[0057] When the input data generated by the input data generating unit 115 further includes luminance maps of multiple colors, the model generating unit 121 generates an estimation model to express the relationship between the input data further including luminance maps of multiple colors and the boundary position.

[0058] For example, the model generation unit 121 generates an estimation model that represents the relationship between the input data and the boundary position through a multi-stage input-output relationship. The input-output relationship can also be represented by a case division process (Japanese: 場合分け処理) or a probabilistic case division process (Japanese: 確率的場合分け処理) that divides the input into cases and generates outputs corresponding to the case division results. As an example, the model generation unit 121 can also generate an estimation model through deep learning. Figure 4 FIG. is a diagram schematically illustrating an estimation model generated by deep learning. Figure 4 The illustrated estimation model 300 is a neural network and includes an input layer 311, one or more intermediate layers 313, and an output layer 312. The input layer 311 outputs an input vector to the next intermediate layer 313. The intermediate layer 313 transforms the input from the previous layer through an activation function and outputs it to the next layer. The output layer 312 transforms the input from the intermediate layer 313 farthest from the input layer 311 through an activation function and outputs the transformation result as an output vector. In the estimation model 300, the activation functions in each intermediate layer 313 and the activation function in the output layer 312 are an example of the above multi-stage input-output relationship. According to the process of transforming the input through the activation function, the transformation result changes corresponding to the input. That is, since the transformation result changes corresponding to the input situation, the transformation based on the activation function is an example of the above case division process or probabilistic case division process.

[0059] Return Figure 2 , the estimation unit 123 estimates the boundary position based on the estimation model stored in the model storage unit 122 and the new input data. The new input data is the input data newly generated by the input data generation unit 115 after the estimation model generated by the model generation unit 121 is stored in the model storage unit 122. For example, the new input data is the input data newly generated by the input data generation unit 115 based on the displacement map newly generated by the displacement map generation unit 113 and the luminance map newly generated by the luminance map generation unit 114 after the estimation model generated by the model generation unit 121 is stored in the model storage unit 122. For example, the estimation unit 123 inputs the new input data into the estimation model stored in the model storage unit 122 and estimates the boundary position based on the output of the estimation model corresponding to the input of the new input data. <0—000141>

[0060] When the new input data generated by the input data generation unit 115 includes the displacement map newly generated by the displacement map generation unit 113, the estimation unit I23 estimates the boundary position based on the estimation model and the new input data including the newly generated displacement map.

[0061] When the new input data generated by the input data generating unit 115 further includes the luminance map newly generated by the luminance map generating unit 114 , the estimating unit 123 estimates the boundary position based on the estimation model and the new input data further including the newly generated luminance map.

[0062] When the input data generated by the input data generating unit 115 further includes the luminance maps of multiple colors newly generated by the luminance map generating unit 114 , the estimating unit 123 estimates the boundary position based on the estimation model and the new input data including the newly generated luminance maps of multiple colors.

[0063] The control unit 124 controls the flow rate of molten steel into the mold 5 to bring the melt level closer to the target height based on the boundary position estimated by the estimation unit 123. This allows the boundary position estimation result to be effectively used for automatic adjustment of the melt level.

[0064] For example, the control unit 124 controls the opening of the gate 9 to bring the molten steel surface closer to the target height. Since the amount of molten steel flowing into the mold 5 is affected by the opening of the gate 9, controlling the opening of the gate 9 is one example of controlling the amount of molten steel flowing into the mold 5. The control unit 124 can also control the gate 9 so that the boundary position within the image approaches a target position corresponding to the target height.

[0065] The control unit 124 may also calculate the melt level based on the correspondence between the boundary position and the melt level and the boundary position recognition result, and control the gate 9 so that the calculated melt level approaches the target height. The correspondence between the boundary position and the melt level is prepared in advance through actual machine testing or simulation.

[0066] The control unit 124 may further calculate the molten steel level based on the amount of molten steel flowing into the mold 5, in addition to the calculation result of the molten steel level based on the boundary position identification result (hereinafter referred to as the "first calculation result"). For example, the control unit 124 may calculate the molten steel level by dividing the difference between the cumulative amount of molten steel flowing in from the initial state before the molten steel enters the mold 5 and the cumulative amount of molten steel flowing out from the initial state by the cross-sectional area within the mold 5. Hereinafter, the calculation result of the molten steel level based on the amount of molten steel flowing in will be referred to as the "second calculation result." The control unit 124 may also suspend control of the molten steel flow based on the first calculation result if the difference between the first calculation result and the second calculation result exceeds a predetermined level (e.g., a predetermined threshold value). Silencing control of the molten steel flow based on the first calculation result includes continuing control of the molten steel flow based on the previous first calculation result.

[0067] Figure 5 1 is a block diagram illustrating the hardware structure of the control system 100. Figure 5 As shown, the control system 100 has a circuit 190 , which includes a processor 191 , a memory 192 , a storage 193 , a graphics circuit 194 , a control circuit 195 , and a user interface 196 .

[0068] Memory 193 is composed of one or more nonvolatile memory devices, such as flash memory or a hard disk. Memory 193 stores a program for causing control system 100 to: estimate the boundary position based on the estimation model and new input data; and adjust the melt level to a target height based on the boundary position estimation result. For example, memory 193 stores a program for configuring the aforementioned functional blocks into control system 100.

[0069] Memory 192 is composed of one or more volatile memory devices, such as random access memory. Memory 192 temporarily stores programs loaded from memory 193. Processor 191 is composed of one or more computing devices, such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit). Processor 191 executes the programs loaded into memory 192, thereby configuring the control system 100 with the aforementioned functional blocks. Calculation results from processor 191 are temporarily stored in memory 192.

[0070] The graphic circuit 194 controls the camera 21 according to a request from the processor 191. The control circuit 195 operates the gate 9 according to a request from the processor 191. The user interface 196 includes a display device and an input device. The display device displays information to the user. Examples of display devices include liquid crystal displays or organic EL (Electro-Luminescence) displays. The input device obtains user input. Examples of input devices include keyboards and mice. The input device can also be integrated with the display device as a touch panel, for example. The hardware structure shown above is just an example and can be changed appropriately. For example, the circuit 190 can also be divided into a plurality of circuits that can communicate with each other.

[0071] [Control Step]

[0072] Next, the control steps of the control system 100 for the gate 9 are illustrated. As an example of a melt level estimation method, this control step includes a melt level estimation step of the control system 100. The melt level estimation step includes: acquiring dynamic image data; generating input data based on the dynamic image data; acquiring correct data representing the recognition result of the dynamic image data based on the boundary position; accumulating learning records obtained by associating the input data with the correct data in the database 118; generating an estimation model representing the relationship between the input data and the boundary position based on the multiple learning records accumulated in the database 118; generating new input data based on newly acquired dynamic image data; and estimating the boundary position based on the estimation model and the new input data. The control steps are illustrated below, divided into the estimation model generation step and the melt level height control step.

[0073] (Generative step of the estimation model)

[0074] like Figure 6 As shown, the control system 100 first executes steps S01, S02, S03, and S04. In step S01, the recording unit 111 stores dynamic image data, including multiple images captured in a time series by the camera 21, in the dynamic image data storage unit 112. In step S02, the displacement map generation unit 113 generates the displacement map based on the dynamic image data stored in the dynamic image data storage unit 112. In step S03, the brightness map generation unit 114 generates the brightness map based on the dynamic image data stored in the dynamic image data storage unit 112. In step S04, the input data generation unit 115 generates input data including the displacement map generated by the displacement map generation unit 113 and the brightness map generated by the brightness map generation unit 114.

[0075] Next, the control system 100 executes steps S05, S06, and S07. In step S05, the correct data acquisition unit 116 acquires the correct data. In step S06, the accumulation unit 117 generates a learning record by associating the input data generated by the input data generation unit 115 with the correct data acquired by the correct data acquisition unit 116, and accumulates the learning record in the database 118. In step S07, the model generation unit 121 checks whether the number of learning records accumulated in the database 118 exceeds a predetermined threshold.

[0076] If, in step S07, it is determined that the number of learning records does not exceed the threshold, the control system 100 returns the process to step S01. Thereafter, the acquisition of dynamic image data, generation of input data, acquisition of correct data, and accumulation of learning records are repeated until the number of learning records exceeds the threshold.

[0077] If, in step S07, the number of learning records exceeds the threshold, the control system 100 proceeds to step S08. In step S08, the model generation unit 121 generates the aforementioned estimation model through machine learning based on the learning records accumulated in the database 118, and stores the model in the model storage unit 122. This completes the estimation model generation step.

[0078] Furthermore, step S03 may be executed before step S02. The execution period of step S02 and the execution period of step S03 may also partially overlap. In addition, the fact that step S05 is executed after step S04 is merely an example, and step S05 may be executed at any time after step S01 and before step S06.

[0079] (Steps for controlling the melt level)

[0080] After the generation step of the estimation model above, Figure 7 As shown in the steps. Figure 7 As shown, the control system 100 first executes steps S11, S12, S13, and S14. In step S11, after generating the estimation model, the recording unit 111 stores the moving image data including a plurality of images captured in a time series by the camera 21 in the moving image data storage unit 112. Hereinafter, the moving image data stored by the recording unit 111 in the moving image data storage unit 112 in step S11 will be referred to as "new moving image data."

[0081] In step S12, the displacement map generator 113 generates a new displacement map based on the new moving image data. In step S13, the brightness map generator 114 generates a new brightness map based on the new moving image data. In step S14, the input data generator 115 generates new input data including the displacement map generated by the displacement map generator 113 and the brightness map generated by the brightness map generator 114.

[0082] Next, the control system 100 executes steps S15 and S16. In step S15, the estimation unit 123 estimates the boundary position based on the estimation model and the new input data. In step S16, the control unit 124 determines whether the melt surface height exceeds the target height based on the estimated boundary position. If, in step S16, the melt surface height is determined not to exceed the target height, the control system 100 executes step S17. In step S17, the control unit 124 determines whether the melt surface height is less than the target height based on the estimated boundary position. If, in step S17, the melt surface height is determined not to be less than the target height, the control system 100 returns the process to step S11.

[0083] If, in step S16, it is determined that the melt surface height exceeds the target height, the control system 100 executes step S21. In step S21, the control unit 124 reduces the opening of the gate 9 to reduce the amount of molten steel flowing into the mold 5. For example, the control unit 124 reduces the opening of the gate 9 by a predetermined reduction amount. Alternatively, the control unit 124 calculates the reduction amount by performing a proportional operation, a proportional-integral operation, or a proportional-integral-differential operation on the deviation between the melt surface height and the target height, and reduces the opening of the gate 9 by the calculated reduction amount.

[0084] If, in step S17, it is determined that the melt surface height is less than the target height, the control system 100 executes step S22. In step S22, the control unit 124 increases the opening of the gate 9 to increase the amount of molten steel flowing into the mold 5. For example, the control unit 124 increases the opening of the gate 9 by a predetermined amount. Alternatively, the control unit 124 calculates the amount of increase by performing a proportional operation, a proportional-integral operation, or a proportional-integral-differential operation on the deviation between the target height and the melt surface height, and increases the opening of the gate 9 by the calculated amount of increase.

[0085] 〔Summarize〕

[0086] The above-described exemplary embodiments include the following configurations.

[0087] (1) A continuous casting system 1, comprising: a camera 21 for photographing the melt surface of molten steel contained in a continuous casting mold 5 and the inner surface of the mold 5; and an estimation unit 123 for estimating a boundary position based on an estimation model generated by machine learning in a manner representing a relationship between input data based on dynamic image data including a plurality of images photographed by the camera 21 in a time series and a boundary position between the melt surface within the field of view of the camera 21 and the inner surface of the mold 5, and new input data.

[0088] This device estimates the boundary position based on an estimation model generated through machine learning to represent the relationship between input data based on dynamic image data and the boundary position, new input data, and the dynamic image data used to estimate the boundary position. Using dynamic image data, it is possible to obtain information that cannot be obtained from a single image, such as information about the movement of each part within the image. Therefore, this device can estimate the boundary position with high reliability even when the brightness difference is small.

[0089] (2) The continuous casting system 1 according to (1), wherein the continuous casting system 1 further includes a displacement map generating unit 113, which generates a displacement map based on dynamic image data, and the displacement map is formed by representing the displacement vector on the time series in each of multiple areas in the field of view of the camera 21, and the estimation model is generated in a manner that represents the relationship between the input data containing the displacement map and the intersection position, and the estimation unit 123 estimates the intersection position based on the estimation model and the new input data containing the newly generated displacement map.

[0090] By converting the distribution data of the tendency of displacement in time series into a displacement map and including it in the input data, the boundary position can be easily estimated with high reliability.

[0091] (3) The continuous casting system 1 according to (2), wherein the continuous casting system 1 further includes a brightness map generating unit 114, which generates a brightness map representing the brightness of each of a plurality of areas based on dynamic image data, the estimation model is generated in a manner representing the relationship between input data also including the brightness map and the boundary position, and the estimation unit 123 estimates the boundary position based on the estimation model and new input data also including the newly generated brightness map.

[0092] Even when mold powder floats on the melt surface, a certain degree of brightness difference may occur between the melt surface and the inner surface of the mold 5. Therefore, by further including a brightness map in the input data, the boundary position can be estimated with higher reliability.

[0093] (4) A continuous casting system 1 according to (3), wherein the brightness map generating unit 114 generates a brightness map for each of a plurality of colors, the estimation model is generated in a manner representing the relationship between input data that also includes the brightness map of the plurality of colors and the boundary position, and the estimation unit 123 estimates the boundary position based on the estimation model and new input data that also includes the newly generated brightness map of the plurality of colors.

[0094] Even when the brightness difference between the melt surface and the inner surface of the mold 5 is small, a relatively large difference can be generated between the color of the melt surface and the color of the mold 5. Therefore, by further including brightness maps of multiple colors in the input data, the boundary position can be estimated with higher reliability.

[0095] (5) A continuous casting system 1 according to any one of (1) to (4), wherein the continuous casting system 1 further includes: a correct data acquisition unit 116, which acquires correct data, which is not based on an estimated model but represents a boundary position identified based on dynamic image data; an accumulation unit 117, which accumulates learning records obtained by associating input data and correct data in a database 118; and a model generation unit 121, which generates an estimated model by machine learning based on multiple learning records accumulated in the database 118.

[0096] Since the device performs a series of processes from accumulating learning records to generating an estimation model based on the accumulated learning data, convenience can be improved.

[0097] (6) The continuous casting system 1 according to (5), wherein the model generation unit 121 generates the estimation model by deep learning.

[0098] Deep learning can model the difficult-to-formulate relationship between input data based on dynamic image data and boundary positions with high reliability.

[0099] (7) A continuous casting system 1 according to any one of (1) to (6), wherein the continuous casting system 1 also includes a control unit 124, which controls the inflow of molten steel into the crystallizer 5 based on the estimated boundary position so that the melt surface approaches the target height.

[0100] The estimation result of the boundary position can be effectively used for automatic adjustment of the height of the melt surface.

[0101] (8) The continuous casting system according to (7), wherein the control unit 124 calculates the height of the molten steel surface based on the amount of molten steel flowing into the mold 5, and when the difference between the height of the molten steel surface corresponding to the boundary position estimated by the estimation unit 123 and the result of the calculation of the height of the molten steel surface based on the amount of molten steel flowing in exceeds a predetermined level, the control of the amount of molten steel flowing in based on the estimated result of the boundary position is not performed. The reliability of the control based on the estimated result of the boundary position can be improved.

[0102] (8) A method for estimating a molten steel surface, wherein the method for estimating a molten steel surface comprises: obtaining dynamic image data including a plurality of images captured in a time series by using a camera 21 for capturing the molten steel surface of the molten steel contained in the continuous casting mold 5 and the inner surface of the mold 5; generating input data based on the dynamic image data; obtaining correct data representing a recognition result of the dynamic image data of the boundary position between the molten steel surface and the inner surface of the mold 5 within the field of view of the camera 21; accumulating learning records obtained by associating the input data with the correct data in a database 118; generating an estimation model representing the relationship between the input data and the boundary position based on the plurality of learning records accumulated in the database 118; generating new input data based on the newly acquired dynamic image data; and estimating the boundary position based on the estimation model and the new input data.

[0103] Description of Reference Numerals

[0104] 5. Crystallizer; 21. Camera; 113. Displacement map generation unit; 114. Brightness map generation unit; 116. Correct data acquisition unit; 117. Accumulation unit; 118. Database; 121. Model generation unit; 123. Estimation unit; 124. Control unit.

Claims

1. A continuous casting system, wherein: The continuous casting system includes: a camera that captures images of a molten steel surface and an inner surface of a continuous casting mold; a displacement map generating unit for generating a displacement map based on moving image data including a plurality of images captured in time series by the camera, the displacement map being a displacement map in which a displacement vector in time series is expressed for each of a plurality of regions in the field of view of the camera; and an estimating unit that estimates the boundary position based on an estimation model generated by machine learning in a manner that represents the relationship between the input data including the displacement map and the boundary position between the melt surface and the inner surface of the mold within the field of view of the camera, and new input data including the displacement map newly generated based on the newly captured dynamic image data.

2. The continuous casting system according to claim 1, wherein: The continuous casting system further includes a brightness map generating unit that generates a brightness map indicating the brightness of each of the plurality of regions based on the dynamic image data. The estimation model is generated in a manner that represents the relationship between the input data including the brightness map and the boundary position, The estimation unit estimates the boundary position based on the estimation model and the new input data further including the luminance map newly generated based on the newly captured moving image data.

3. The continuous casting system according to claim 2, wherein: The brightness map generating unit generates the brightness map for each of a plurality of colors. The estimation model is generated in a manner that represents the relationship between the input data including a brightness map of a plurality of colors and the boundary position. The estimation unit estimates the boundary position based on the estimation model and the new input data further including the luminance maps of the plurality of colors newly generated based on the newly captured moving image data.

4. The continuous casting system according to any one of claims 1 to 3, wherein: The continuous casting system also includes: a correct data acquisition unit configured to acquire correct data that is not based on the estimation model but represents the boundary position identified based on the dynamic image data; an accumulation unit that accumulates, in a database, a learning record obtained by associating the input data with the correct data; and A model generating unit generates the estimation model by machine learning based on a plurality of learning records accumulated in the database.

5. The continuous casting system according to claim 4, wherein: The model generation unit generates the estimation model through deep learning.

6. The continuous casting system according to any one of claims 1 to 3, wherein: The continuous casting system further includes a control unit configured to control an inflow rate of the molten steel into the mold based on the estimated boundary position so as to bring the molten steel level close to a target height.

7. The continuous casting system according to claim 6, wherein: The control unit calculates the height of the melt surface based on the inflow of the molten steel into the crystallizer. When the difference between the height of the melt surface corresponding to the boundary position estimated by the estimation unit and the calculation result of the height of the melt surface based on the inflow of the molten steel exceeds a specified level, the inflow of the molten steel based on the estimation result of the boundary position is not controlled.

8. A melt level estimation method, wherein: The melt level estimation method includes: Using a camera that captures images of a molten steel surface and an inner surface of a continuous casting mold, dynamic image data including a plurality of images captured in time series is acquired; generating, based on the dynamic image data, a displacement map in which a time-series displacement vector is expressed in each of a plurality of areas in the field of view of the camera; acquiring correct data representing a recognition result of the dynamic image data based on the boundary position between the melt surface and the inner surface of the mold within the field of view of the camera; accumulating learning records obtained by associating the input data including the displacement map with the correct data in a database; generating an estimation model representing a relationship between the input data and the boundary position based on the plurality of learning records accumulated in the database; newly generating the displacement map based on the newly acquired dynamic image data; and The boundary position is estimated based on the estimation model and input data including the newly generated displacement map.

Images