Method for manufacturing glass in a container, apparatus for manufacturing the same, and method for estimating the temperature of molten glass.

Abstract

The present invention provides a method for manufacturing containerized glass, a manufacturing apparatus for that method, and a method for estimating the temperature of molten glass, which reduces the frequency of misestimation when estimating the temperature of molten glass near the inlet of a flow nozzle by applying machine learning with training data, compared to when a comparative estimation model is used. [Solution] The temperature of the molten glass near the inlet of the flow nozzle is estimated using a machine learning model created based on training data including the following data (a) to (d): (a) training images of molten glass flowing down from the nozzle, (b) the temperature of the molten glass near the inlet of the nozzle in (a), (c) the concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), and (d) temperature information of the molten glass flowing down from the nozzle in (a).

Description

[Technical Field]

[0001] The present invention relates to a method for manufacturing glass in a container, an apparatus for manufacturing the same, and a method for estimating the temperature of molten glass. [Background technology]

[0002] High-level radioactive liquid waste generated from the reprocessing of spent nuclear fuel is disposed of in geological formations as vitrified waste (buried in stable geological formations at depths of 300 meters or more). Vitrified waste is produced by heating radioactive liquid waste and glass raw materials (e.g., borosilicate glass: specific gravity approximately 2.2 to 2.7) charged into a melting furnace to form molten glass, then flowing the formed molten glass from a flow nozzle in the melting furnace into a stainless steel storage container called a canister, and finally solidifying the molten glass contained in the storage container (Patent Document 1). The vitrified waste is stored in a dedicated storage facility for 30 to 50 years, after which it is disposed of in geological formations.

[0003] Radioactive waste liquid contains platinum group metal elements such as ruthenium (Ru, specific gravity approximately 12.4), rhodium (Rh, specific gravity approximately 12.5), and palladium (Pd, specific gravity approximately 12.0). Since the specific gravity of platinum group metal elements is greater than that of the glass raw materials in the solid state, platinum group metal elements tend to deposit at the bottom of the melting furnace (Patent Document 1). At the bottom of the melting furnace, where the concentration of metal elements has increased due to the deposition of platinum group metal elements, the electrical resistance of the molten glass decreases, thus reducing the Joule heating capacity when electricity is applied. Furthermore, as the concentration of metal elements increases, the viscosity of the molten glass increases, so the fluidity of the glass also decreases. In such a situation, there is a risk that the flow nozzle at the bottom of the melting furnace may become blocked.

[0004] Therefore, in the production of glass solidified bodies by the above method, temperature control of the molten glass in the melting furnace, particularly at the bottom of the melting furnace near the inlet of the flow nozzle, is important. Conventionally, as an example of a method for controlling the temperature of molten glass, there has been an attempt to estimate the temperature of the molten glass in the melting furnace based on the shape (e.g., diameter) of the molten glass flowing down from the flow nozzle and the concentration of platinum group metal elements in the molten glass (Non-Patent Literature 1). This method utilizes the correlation between the viscosity of the molten glass and the temperature of the molten glass and the concentration of platinum group metal elements. [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2015-081224 [Non-patent literature]

[0006] [Non-Patent Document 1] Ryoya Koyama et al., "Prediction of the state at the bottom of a glass melting furnace by simulation of molten glass flow," Proceedings of the 53rd Autumn Meeting of the Society of Chemical Engineers, Japan (2022), Presentation No. DB108. [Overview of the project] [Problems that the invention aims to solve]

[0007] In recent years, machine learning with supervised data has been applied and developed in various technological fields. Machine learning with supervised data is a data analysis technique that creates a machine learning model that learns the correlation between data based on supervised data, and then has the machine learning model output predicted or estimated values ​​for unknown data for the input data.

[0008] The inventors believed that machine learning with training data would contribute to improving the efficiency of data analysis and further enhancing measurement accuracy, and attempted to apply machine learning with training data to estimating the temperature of molten glass in a melting furnace. Through diligent research, the inventors created a machine learning model based on training data including (a) training images of molten glass flowing down from a nozzle, (b) the temperature of the molten glass near the nozzle inlet (a), and (c) the concentration of metal elements in the molten glass near the nozzle inlet (a). By inputting analytical images of the flowing molten glass into this machine learning model, the inventors found that the temperature of the molten glass and the concentration of metal elements near the nozzle inlet could be estimated. Hereinafter, the estimation model created by machine learning only the correlation between the above data (a) to (c) will be referred to as the "comparative estimation model."

[0009] However, the inventors have also found that the comparative estimation model can result in large errors in the estimated temperature of the molten glass near the nozzle inlet (e.g., errors exceeding 2%), and that misestimations may occur.

[0010] Therefore, one of the objectives of the present invention is to provide a method and apparatus for manufacturing glass in storage containers that can reduce the frequency of misestimation when estimating the temperature of molten glass near the inlet of the flow nozzle by applying machine learning with training data, compared to when a comparative estimation model is used.

[0011] Furthermore, considering that the cause of nozzle blockage is the deposition of metal having a specific gravity greater than that of the glass raw material, it can be inferred that the scope of application of the above method for estimating the properties of molten glass in a melting furnace is not limited to cases where the metal element is a platinum group metal element, nor is it limited to cases where the metal element-containing raw material is radioactive waste liquid. In other words, the above method for estimating the properties of molten glass in a melting furnace may be more generally applicable to the manufacture of glass containing metal elements that have a specific gravity greater than that of the glass raw material in the solid state.

[0012] Therefore, the object of the present invention is to provide a method and apparatus that can reduce the frequency of misestimation when estimating the temperature of molten glass near the inlet of a flow nozzle by applying machine learning with training data to various glass manufacturing methods and apparatus, not limited to glass in storage containers, compared to when a comparative estimation model is used.

[0013] In another embodiment, an object of the present invention also includes providing a method for estimating the temperature of molten glass in the above-mentioned manufacturing method and apparatus, which can reduce the frequency of misestimation when estimating the temperature of molten glass near the inlet of the flow nozzle by applying machine learning with training data, compared to when a comparative estimation model is used. [Means for solving the problem]

[0014] The above problem was solved by creating a machine learning model based on training data including the dataset containing the following data (a) to (d). Specifically, the above problem was solved by the invention described in [1] below, and more preferably by the inventions described in [2] and subsequent inventions. [1] The process involves heating metal element-containing raw materials and glass raw materials in a melting furnace to form molten glass. The process involves flowing the formed molten glass down from the flow nozzle of the melting furnace, and A method for manufacturing containerized glass, which includes placing the flowing molten glass into a container, Metal element-containing raw materials include metal elements (hereinafter also referred to as "specific metal elements") that have a specific gravity greater than that of the glass raw material in a solid state. The manufacturing method further includes obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle by the estimation method described below, and controlling the heating based on the obtained estimated value so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature: ·Estimation method Prepare an analytical image obtained by imaging the molten glass flowing down from the flow nozzle of the melting furnace, the analytical image including temperature information of the flowing molten glass. Inputting the analysis image into a pre-trained estimation model, and An estimation method comprising obtaining an estimated value of the temperature of molten glass in the vicinity of the inlet of the flowing-down nozzle from the estimation model into which the analysis image has been input: However, the estimation model is a machine learning model created by machine learning the correlation between data based on teacher data including the following data (a) to (d): (a) A learning image of molten glass flowing down from a nozzle, (b) The temperature of molten glass in the vicinity of the inlet of the nozzle in (a), (c) The concentration of a specific metal element in the molten glass in the vicinity of the inlet of the nozzle in (a), (d) Temperature information of molten glass flowing down from the nozzle in (a). [2] The temperature information included in the analysis image is at least one of the temperature of a predetermined point of the molten glass flowing down in the analysis image and the temperature distribution in the molten glass flowing down in the analysis image, The temperature information of the data (d) is at least one of the temperature of a predetermined point of the molten glass flowing down in the learning image and the temperature distribution in the molten glass flowing down in the learning image, the manufacturing method according to [1]. [3] The metal element-containing raw material contains a specific metal element having a specific gravity of 7 or more in a solid state, the manufacturing method according to [1] or [2]. [4] The metal element-containing raw material contains a platinum group metal element, the manufacturing method according to [3]. [5] The estimation method further includes obtaining an estimated value of the concentration of the specific metal element in the molten glass in the vicinity of the inlet of the flowing-down nozzle in addition to the estimated value of the temperature, the manufacturing method according to any one of [1] to [4]. [6] The metal element-containing raw material is radioactive waste liquid, the manufacturing method according to any one of [1] to [5]. [7] The analysis image is a moving image or a still image, the manufacturing method according to any one of [1] to [6]. [8] A melting furnace equipped with a flow nozzle, which heats a metal element-containing raw material and a glass raw material to form molten glass, A control unit that controls heating by the melting furnace, An imaging unit for imaging the molten glass flowing down from the aforementioned flow nozzle, A manufacturing apparatus for containerized glass, comprising a receiving section for containing the flowing molten glass in a container, Metal element-containing raw materials include metal elements (hereinafter also referred to as "specific metal elements") that have a specific gravity greater than that of the glass raw material in a solid state. The control unit obtains an estimated value of the temperature of the molten glass near the inlet of the flow nozzle using the following estimation method, and controls the heating based on the obtained estimated value so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature, in a manufacturing apparatus: ·Estimation method Prepare an analytical image obtained by imaging the molten glass flowing down from the flow nozzle of the melting furnace with the imaging unit, the analytical image including temperature information of the flowing molten glass. Inputting the aforementioned analysis images into a machine learning-based estimation model, and An estimation method comprising obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle from the estimation model into which the aforementioned analytical image has been input: However, the estimation model described above is a machine learning model created by machine learning the correlation between data based on training data including the following data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the nozzle entrance of (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a). [9] The temperature information included in the analytical image is at least one of the temperature at a predetermined point in the flowing molten glass in the analytical image and the temperature distribution in the flowing molten glass in the analytical image. The manufacturing apparatus according to [8], wherein the temperature information of the data (d) is at least one of the temperature of a predetermined point in the flowing molten glass in the training image and the temperature distribution in the flowing molten glass in the training image.

[10] The manufacturing apparatus according to [8] or [9], wherein the metal element-containing raw material contains a specific metal element having a specific gravity of 7 or more in a solid state.

[11] The manufacturing apparatus according to

[10] , wherein the metal element-containing raw material comprises a platinum group metal element.

[12] The manufacturing apparatus according to any one of [8] to

[11] , further comprising obtaining an estimated value of the concentration of the specific metal element in the molten glass near the inlet of the flow nozzle, in addition to the estimated value of the temperature.

[13] The manufacturing apparatus according to any one of items [8] to

[12] , wherein the metal element-containing raw material is radioactive waste liquid.

[14] The manufacturing apparatus according to any one of items [8] to

[13] , wherein the analytical image is either a video or a still image.

[15] A method for estimating the temperature of molten glass near the inlet of the flow nozzle, used in the method for manufacturing a containerized glass described in [1], Prepare an analytical image obtained by imaging the molten glass flowing down from the flow nozzle of the melting furnace, the analytical image including temperature information of the flowing molten glass. Inputting the aforementioned analysis images into a machine learning-based estimation model, and An estimation method comprising obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle from the estimation model into which the aforementioned analytical image has been input: However, the estimation model described above is a machine learning model created by machine learning the correlation between data based on training data including the following data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the nozzle entrance of (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a).

[16] The temperature information included in the analytical image is at least one of the temperature at a predetermined point in the flowing molten glass in the analytical image and the temperature distribution in the flowing molten glass in the analytical image. The estimation method according to

[15] , wherein the temperature information of the data (d) is at least one of the temperature of a predetermined point in the flowing molten glass in the training image and the temperature distribution in the flowing molten glass in the training image. [Effects of the Invention]

[0015] The method and apparatus for manufacturing containerized glass according to the present invention allows for the application of machine learning with training data to estimate the temperature of molten glass near the nozzle inlet, thereby reducing the frequency of misestimations compared to when a comparative estimation model is used.

[0016] In addition, the method for estimating the temperature of molten glass according to the present invention makes it possible to reduce the frequency of misestimations when estimating the temperature of molten glass near the inlet of the flow nozzle in the above-mentioned manufacturing method and manufacturing apparatus by applying machine learning with training data, compared to when a comparative estimation model is used. [Brief explanation of the drawing]

[0017] [Figure 1] Figure 1 is a schematic diagram of one embodiment of a manufacturing apparatus for containerized glass according to the present invention. [Figure 2] Figure 2 is a simulation image that mimics molten glass flowing down from a nozzle (region A in Figure 1). [Figure 3]Figure 3 is a graph showing the temperature distribution of flowing molten glass. (a) shows the change in temperature distribution due to differences in platinum group metal element (PGM) concentration (0% or 3.5%), and (b) shows the change in temperature distribution due to differences in the temperature of the molten glass (PGM concentration 3.5%) when it flows into the nozzle inlet (1400K or 1500K). [Figure 4] Figure 4 is a conceptual diagram of the dataset that makes up the training data. [Figure 5] Figure 5 shows the analysis framework in computational fluid dynamics simulations. [Figure 6] Figure 6 is a graph showing the temperature dependence of the viscosity of molten glass at each PGM concentration. [Figure 7] Figure 7 is a simulation image showing the flow motion (0-10 seconds) of molten glass (PGM concentration 0%, temperature 1300K) through a nozzle. (a) is a cross-sectional view passing through the center of the nozzle, and (b) is a front view from outside the nozzle. [Figure 8] Figure 8 is a simulation image showing the flow motion (0-10 seconds) of molten glass (PGM concentration 0%, temperature 1400K) through a nozzle. (a) is a cross-sectional view passing through the center of the nozzle, and (b) is a front view from outside the nozzle. [Figure 9] Figure 9 is a simulation image showing the flow motion (0-10 seconds) of molten glass (PGM concentration 0%, temperature 1500K) through a nozzle. (a) is a cross-sectional view passing through the center of the nozzle, and (b) is a front view from outside the nozzle. [Figure 10] Figure 10 is a simulation image showing the flow motion (0-10 seconds) of molten glass (PGM concentration 3.5%, temperature 1300K) through a nozzle. (a) is a cross-sectional view passing through the center of the nozzle, and (b) is a front view from outside the nozzle. [Figure 11] Figure 11 is a simulation image showing the flow motion (0-10 seconds) of molten glass (PGM concentration 3.5%, temperature 1400K) through a nozzle. (a) is a cross-sectional view passing through the center of the nozzle, and (b) is a front view from outside the nozzle. [Figure 12]Figure 12 is a simulation image showing the flow motion (0-10 seconds) of molten glass (PGM concentration 3.5%, temperature 1500K) through a nozzle. (a) is a cross-sectional view passing through the center of the nozzle, and (b) is a front view from outside the nozzle. [Figure 13] Figure 13 is a graph showing the interface position of molten glass at various temperatures as it flows down from the nozzle. (a) shows the case when the PGM concentration is 0%, and (b) shows the case when the PGM concentration is 3.5%. [Figure 14] Figure 14 is a graph showing the relationship between the temperature of the molten glass flowing down from the nozzle and the jet diameter. (a) shows the relationship at 5 cm below the nozzle exit, and (b) shows the relationship at 10 cm below the nozzle exit. [Figure 15] Figure 15 is a conceptual diagram of reference training data. [Figure 16] Figure 16 is a graph showing the results of temperature estimation using a machine learning model (reference) created based on the reference training data shown in Figure 15. The PGM concentration of the analysis image input to the machine learning model (reference) was set to the same value (3.5%) as that of the training data (reference) shown in Figure 15. [Figure 17] Figure 17 is a conceptual diagram of the training data (comparison) used to create the comparative estimation model. [Figure 18] Figure 18 is a graph showing the estimation results using a comparative estimation model created based on training data (comparison). (a) shows the estimation results for temperature, and (b) shows the estimation results for PGM concentration. [Figure 19] Figure 19 is a conceptual diagram of the training data used to create the estimation model according to an embodiment of the present invention. [Figure 20] Figure 20 is a graph showing the estimation results using an estimation model created based on training data according to the present invention. (a) shows the estimation result for temperature, and (b) shows the estimation result for PGM concentration. [Modes for carrying out the invention]

[0018] [Method and apparatus for manufacturing glass containers] The present invention's method for manufacturing glass containers is: The process involves heating metal element-containing raw materials and glass raw materials in a melting furnace to form molten glass. The process involves flowing the formed molten glass down from the flow nozzle of the melting furnace, and This includes collecting the flowing molten glass into a container, Here, the metal element-containing raw material includes a metal element that has a specific gravity greater than that of the glass raw material in its solid state (hereinafter also referred to as "specific metal element"), The manufacturing method further includes obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle using the estimation method described below, and controlling the heating based on the obtained estimated value so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature. ·Estimation method Prepare an analytical image obtained by imaging molten glass flowing down from a flow nozzle in a melting furnace, which includes temperature information of the flowing molten glass. Inputting images for analysis into a machine learning-based estimation model, and An estimation method that includes obtaining an estimated value of the temperature of molten glass near the inlet of the flow nozzle from an estimation model into which analytical images have been input: However, the estimation model is a machine learning model created by machine learning the correlations between data based on training data including the following data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the nozzle entrance of (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a).

[0019] The method for manufacturing containerized glass of the present invention can be carried out using the manufacturing apparatus for containerized glass of the present invention.

[0020] The present invention provides a manufacturing apparatus for containerized glass. A melting furnace equipped with a flow nozzle, which heats a metal element-containing raw material and a glass raw material to form molten glass, A control unit that controls heating by a melting furnace, An imaging unit that images the molten glass flowing down from the flow nozzle, It includes a container for containing the flowing molten glass, Here, the metal element-containing raw material includes a metal element that has a specific gravity greater than that of the glass raw material in its solid state. The control unit obtains an estimated value of the temperature of the molten glass near the inlet of the flow nozzle using the estimation method described above, and controls the heating based on the obtained estimated value so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature.

[0021] The following explanation will focus particularly on the case where the metal element-containing raw material is radioactive liquid waste.

[0022] Figure 1 is a schematic diagram of one embodiment of a manufacturing apparatus for containerized glass according to the present invention. The manufacturing apparatus 1 includes a melting furnace 10 equipped with a heating section (main electrodes 11a and 11b and a bottom electrode 11c) and a flow nozzle 12, a control unit 20, an imaging unit 21, and a storage unit 30. The melting furnace 10 is protected by casing material such as refractory bricks 14 and an insulating board, and further has a raw material loading inlet 15 and an exhaust port 16. The manufacturing apparatus 1 can be used, for example, to manufacture glass solidified bodies for geological treatment of radioactive liquid waste.

[0023] Molten glass 5 is a mixed molten material of metal element-containing raw material 2 and glass raw material 3. The molten glass is used to produce a vitrified body, and the vitrified body is produced by cooling the molten glass.

[0024] Metal element-containing raw material 2 is a raw material containing a metal element (specific metal element) that has a specific gravity greater than that of the glass raw material in its solid state, and is, for example, radioactive waste liquid discharged from a nuclear facility. Alternatively, when the present invention is applied to the manufacture of general glass products, the metal element-containing raw material can be a metal material used as an additive to the glass raw material.

[0025] Radioactive waste liquid may contain, for example, platinum group metals such as ruthenium (Ru, specific gravity approximately 12.4), rhodium (Rh, specific gravity approximately 12.5), and palladium (Pd, specific gravity approximately 12.0); minor actinides such as neptunium (Np, specific gravity approximately 20.5), americium (Am, specific gravity approximately 13.7), and curium (Cm, specific gravity approximately 13.5); and lanthanides such as lanthanum (La, specific gravity approximately 6.2), cerium (Ce, specific gravity approximately 6.8), and neodymium (Nd, specific gravity approximately 7.0). In addition, radioactive waste liquid generally contains nitric acid solution in addition to the metal components.

[0026] The higher the specific gravity of a metal element in its solid state, the more easily it is deposited in molten glass, leading to a higher concentration at the bottom of the furnace. In the bottom of the furnace, where the concentration of metal elements increases due to deposition, the electrical resistance of the molten glass decreases, thus reducing its Joule heating capacity. Furthermore, as the concentration of metal elements increases, the viscosity of the molten glass also increases, reducing its fluidity. Thus, high concentrations of metal elements reduce the fluidity of molten glass. This invention, which enables the estimation and control of the temperature of the molten glass at the bottom of the furnace, is particularly useful when the metal elements in the metal element-containing raw material 2 are easily deposited in the molten glass.

[0027] The present invention can be used when the metal element-containing raw material 2 preferably contains a specific metal element with a specific gravity of 7 or more in the solid state, more preferably contains a specific metal element with a specific gravity of 9 or more, and even more preferably contains a specific metal element with a specific gravity of 11 or more. The upper limit of the specific gravity of the specific metal element contained in the metal element-containing raw material in the solid state is not particularly limited, but for example it is 22 or less, and may be 21 or less, or 20 or less. In particular, platinum group metal elements tend to form flocs (aggregates) and deposit in molten glass. The present invention can be used particularly preferably when the metal element-containing raw material 2 contains platinum group metal elements.

[0028] The glass raw material 3 is not particularly limited, and various glass materials can be used depending on the application and purpose. In applications of radioactive waste liquid treatment, the glass raw material is generally borosilicate glass (specific gravity approximately 2.2 to 2.7), and iron phosphate glass, magnesium phosphate glass, or PbO-B2O3-ZnO glass may also be used. The form of the glass raw material is not particularly limited, and can be, for example, in block, bead, fiber, or powder form.

[0029] The method of charging the metal element-containing raw material 2 and the glass raw material 3 into the melting furnace is not particularly limited. These raw materials may be charged into the melting furnace individually or simultaneously. Alternatively, these raw materials may be charged into the melting furnace in the form of a mixture obtained by impregnating the glass raw material 3 with liquid metal element-containing raw material 2. The proportion of metal element-containing raw material 2 to the total amount of raw materials is, for example, about 15 to 30% by mass, and the proportion of glass raw material 3 is, for example, about 70 to 85% by mass.

[0030] The shape of the melting furnace 10 is not particularly limited, but the lower part of the melting furnace usually has a tapered shape that narrows towards the bottom. The capacity of the melting furnace is appropriately selected according to the purpose of manufacture and the desired scale. The metal element-containing raw material 2 and the glass raw material 3 are charged into the melting furnace from the raw material charging port 15.

[0031] In one embodiment of the present invention, the heating section is composed of main electrodes 11a and 11b and a bottom electrode 11c. The main electrodes 11a and 11b are arranged facing each other on the inner wall near the midpoint in the vertical direction of the melting furnace 10. The raw material in the melting furnace is melted by Joule heating by energizing the area between the main electrodes. A bottom electrode 11c is located at the bottom of the melting furnace 10, and Joule heating by energizing is also performed between the bottom electrode 11c and the main electrodes 11a and 11b. The heating section may further include an indirect heating section (not shown) that indirectly heats the molten glass by heating the unfilled space at the top of the melting furnace and / or heat-resistant bricks, for example.

[0032] The flow nozzle 12 is located at the bottom of the melting furnace 10, but is not limited to this location. The flow nozzle 12 has, for example, an inner diameter of 2 to 3 cm and a length of 15 to 30 cm. The manufacturing apparatus of the present invention may include a heating coil (not shown) surrounding the flow nozzle 12.

[0033] The control unit 20 is a computer system (not shown) equipped with, for example, a control device such as a CPU or GPU, a storage device such as internal or external storage, and a display. The control unit 20 obtains an estimated value of the temperature of the molten glass near the inlet of the flow nozzle by an estimation method described in detail below, and controls heating by the melting furnace based on the obtained estimate so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature. Heating control is performed by controlling the heating section (for example, the main electrodes 11a and 11b and the bottom electrode 11c). The set temperature is appropriately selected so that the temperature of the molten glass near the inlet of the flow nozzle is, for example, about 1300 to 1500K. If the heating section includes an indirect heating section, the control unit 20 may be configured to control the indirect heating section.

[0034] The imaging unit 21 includes a furnace monitoring camera. The imaging unit 21 is positioned near the outlet of the flow nozzle 12 and generates an image of the molten glass 6 flowing down from the flow nozzle 12. Figure 2 is a simulated image of the molten glass flowing down from the nozzle (region A in Figure 1). As shown in Figure 2, the imaging unit 21 is positioned to capture an image of the flow state of the molten glass 6 near the outlet of the flow nozzle 12. The imaging unit 21 transmits the generated image data to the control unit 20. The control unit 20 receives the image data from the imaging unit 21 and uses the image data as an analysis image in the estimation method described above.

[0035] The storage section 30 may include a stainless steel storage container 31 called a canister and a trolley 32. The storage container 31 and trolley 32 are not particularly limited. Molten glass 5 in the melting furnace 10 is collected in the storage container 31 through the flow nozzle 12, resulting in glass in storage containers. The trolley 32 is used to transport the glass in storage containers. Alternatively, when the present invention is applied to the manufacture of general glass products, the container may be, for example, a mold for glass products or a container for temporarily receiving molten glass, and is not limited to these.

[0036] The manufacturing method of the present invention is carried out, for example, using the above-described manufacturing apparatus, and includes heating metal element-containing raw materials and glass raw materials charged into a melting furnace to form molten glass, allowing the formed molten glass to flow down from a flow nozzle of the melting furnace, and storing the flowed molten glass in a container.

[0037] When the present invention is applied to the treatment of radioactive liquid waste, the glass in the storage container is in a molten state immediately after manufacture, but solidifies through natural cooling or active cooling, becoming a vitrified body. The glass in the storage container is stored for approximately 30 to 50 years in a dedicated storage facility, for example, and then disposed of in a geological site.

[0038] Furthermore, the manufacturing method of the present invention includes obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle by an estimation method described in detail below, and controlling the heating by the melting furnace based on the obtained estimated value so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature. The set temperature is appropriately selected so that the temperature of the molten glass near the inlet of the flow nozzle is, for example, around 1300 to 1500 K.

[0039] [Method for estimating the temperature of molten glass] The following describes the method for estimating the temperature of molten glass according to the present invention.

[0040] The estimation method of the present invention is Prepare an analytical image obtained by imaging molten glass flowing down from a flow nozzle in a melting furnace, which includes temperature information of the flowing molten glass. Inputting images for analysis into a machine learning-based estimation model, and This includes obtaining an estimated value of the molten glass temperature near the inlet of the flow nozzle from an estimation model that takes analytical images as input. However, the estimation model is a machine learning model created by machine learning the correlations between data based on training data including the following data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the nozzle entrance of (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a).

[0041] The estimation method of the present invention includes preparing an analytical image obtained by imaging molten glass flowing down from a flow nozzle of a melting furnace, the analytical image containing temperature information of the flowing molten glass. The preparation of the analytical image is carried out, for example, by imaging the flowing molten glass 6 with an imaging unit 21 to generate image data, transmitting the image data to a control unit 20, and storing it in the control unit 20. The control unit 20 uses the stored image data as an analytical image in the estimation method described above.

[0042] The images used for analysis can be either video or still images, and still images can be a single still image or two or more consecutive still images. If the images used for analysis are a single still image, one estimate can be obtained for this still image. If the images used for analysis are consecutive still images, an estimate can be obtained for each still image. If the images used for analysis are video, an estimate can be obtained for each frame that makes up the video, or for each of the frames extracted at predetermined intervals. If multiple estimates are obtained using video or consecutive still images as the images used for analysis, the average of these estimates can be used as the output estimate. Alternatively, if video or consecutive still images are used as the images used for analysis, the time change of the estimate can be determined by arranging the multiple estimates in a time series. The time change of the estimate is thought to contribute to improving the predictability of temperature control.

[0043] The analytical image includes temperature information of the flowing molten glass. The analytical image may include temperature information as the file name and / or as metadata of the image data. In another embodiment, the analytical image may include temperature information by representing the temperature distribution by the magnitude of the raw or normalized pixel values ​​(e.g., RGB values) of the image data. The analytical image may be a color image or a grayscale image.

[0044] The temperature information included in the analytical image is preferably at least one of the temperature at a predetermined point in the flowing molten glass in the analytical image and the temperature distribution in the flowing molten glass in the analytical image. The predetermined point in the flowing molten glass is a point located at a predetermined distance (e.g., 0 cm, 1 cm, 2 cm, 3 cm, 4 cm, 5 cm, 6 cm, 9 cm, or 12 cm) below the outlet of the flowing nozzle, and is not particularly limited.

[0045] The temperature and temperature distribution of molten glass flowing down a given point (particularly near the nozzle exit) are correlated with the temperature of the molten glass and the concentration of specific metal elements near the nozzle inlet. In this invention, this correlation is utilized in machine learning. Specifically, it is as follows:

[0046] Figure 3 is a graph showing the temperature distribution of flowing molten glass. (a) shows the change in temperature distribution due to differences in platinum group metal element (PGM) concentration (0% or 3.5%), and (b) shows the change in temperature distribution due to differences in the temperature of the molten glass (PGM concentration 3.5%) when it flows into the nozzle inlet (1400K or 1500K).

[0047] Figure 3(a) shows that the surface temperature of the molten glass decreases near the nozzle outlet (0 cm) compared to the temperature when it flows into the nozzle inlet (1500 K). When the PGM concentration is 0%, the temperature drop is approximately 125 K, and when the PGM concentration is 3.5%, the temperature drop is approximately 200 K, showing a difference in the degree of temperature drop between the two. This is because, while the temperature of the molten glass is high, the nozzle temperature is low, so heat was transferred from the molten glass to the nozzle. Furthermore, it is thought that heat transfer was suppressed in the molten glass with a PGM concentration of 0% because it has lower viscosity and higher flow velocity and flow rate compared to the molten glass with a PGM concentration of 3.5%. As a result, the temperature drop at the nozzle outlet of the molten glass with a PGM concentration of 0% was smaller than that of the molten glass with a PGM concentration of 3.5%.

[0048] Figure 3(b) also shows that the surface temperature of the molten glass decreases near the nozzle outlet (0 cm) compared to the temperature at which it enters the nozzle inlet (1400K or 1500K). When the temperature at entry is 1400K, the temperature decrease is approximately 150K, and when the temperature at entry is 1500K, the temperature decrease is approximately 200K, showing a difference in the degree of temperature decrease between the two. Although the temperature decrease is smaller than in the 1500K case, it can be seen that even in the 1400K case, the surface temperature of the molten glass decreases near the nozzle outlet.

[0049] As shown in Figure 3, the temperature distribution on the surface of the molten glass differs depending on the temperature of the molten glass and the concentration of specific metal elements when it flows into the nozzle inlet. In other words, temperature information such as the temperature and temperature distribution at a predetermined point (especially near the nozzle outlet) of the flowing molten glass has a correlation with the temperature of the molten glass and the concentration of specific metal elements near the nozzle inlet. Therefore, in this invention, the temperature information of the flowing molten glass is included in the analysis image and the model is subjected to machine learning to utilize the above correlation. As a result, the temperature of the molten glass near the nozzle inlet can be accurately estimated based on the temperature distribution information in addition to the shape of the flowing molten glass.

[0050] In the present invention, when the temperature information is the temperature distribution in the flowing molten glass, it is preferable to normalize the pixel values ​​of the image data. This allows for efficient machine learning. In the present invention, it is preferable that the temperature information is the temperature distribution in the flowing molten glass. This is because, even after normalization, the temperature distribution makes it easier to extract the correlation between the temperature of the molten glass near the nozzle inlet and the concentration of specific metal elements using machine learning.

[0051] The estimation method of the present invention includes inputting an analysis image into a machine learning-trained estimation model, and obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle from the estimation model into which the analysis image has been input. The estimation model is a machine learning model created by machine learning the correlation between data based on training data including the above data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the nozzle entrance of (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a).

[0052] Figure 4 is a conceptual diagram of the datasets that make up the training data. The training data includes, for example, multiple datasets from data 1 to n. The number of datasets n is, for example, 10 to 100, preferably 30 to 90, and may also be 50 to 80. Each dataset includes the above-mentioned data (a) to (d), and may contain only data (a) to (d). Data (a) to data (d) will be described below.

[0053] Data (a) Data (a) is an image showing molten glass with known properties, providing the visual characteristics of the flowing molten glass to the estimation model. The training image is, for example, a simulation image obtained from a computational fluid dynamics simulation. In the case of the simulation image, the properties of the molten glass (density ρ, viscosity μ, thermal conductivity k, and specific heat capacity C) are taken into consideration, taking into account conditions such as the type of specific metal element contained in the metal element-containing raw material, the type of glass raw material, and the mixing ratio of these raw materials. p It is appropriate to select such images. Alternatively, the training images may be real images obtained by pre-images of molten glass flowing down from a flow nozzle in an actual melting furnace.

[0054] The line of sight in the training images is preferably in a plane perpendicular to the flow direction. To increase the amount of data, it is also preferable to rotate the line of sight within the above plane to create training images. Increasing the amount of data in the training data can improve the generalization ability of the machine learning model. Some of the training images may be images taken with a line of sight that is not perpendicular to the flow direction (i.e., images taken at an angle to the flow direction). By including images taken from various directions in the training images, accurate estimation becomes possible even for analysis images with a line of sight that is not perpendicular to the flow direction.

[0055] Data (b) Data (b) is the temperature of the molten glass near the nozzle inlet of (a). If the flow nozzle is located at the bottom of the melting furnace, this corresponds to the temperature of the molten glass at the bottom of the melting furnace. Data (b) can be the average temperature near the nozzle inlet or the temperature at any location near the nozzle inlet. The shape of the flowing molten glass (especially the diameter D) has a correlation with the temperature of the molten glass near the nozzle inlet (see Figure 14 below), and this correlation is used in machine learning. In this invention, "near the nozzle inlet" refers to a region in the melting furnace where the temperature of the molten glass tends to decrease or where metal elements tend to deposit, for example, within 10 cm or 5 cm from the nozzle inlet, and includes the position of the nozzle inlet.

[0056] If the training image is a simulation image, data(b) can be the temperature of the molten glass near the inlet of the model nozzle, set or calculated in a computational fluid dynamics simulation. If the training image is a real-world image, data(b) can be the temperature of the molten glass near the inlet of the flow nozzle, determined in advance by another temperature measurement method.

[0057] Data (c) Data (c) represents the concentration of a specific metal element in the molten glass near the nozzle inlet of (a). If the flow nozzle is located at the bottom of the melting furnace, this corresponds to the concentration of the specific metal element in the molten glass at the bottom of the melting furnace. Data (c) can be the average concentration of the specific metal element near the nozzle inlet or the concentration of the specific metal element at any location near the nozzle inlet. The shape of the flowing molten glass (especially the diameter D) has a correlation with the concentration of the specific metal element in the molten glass near the nozzle inlet (see Figure 14 below), and this correlation is used for machine learning.

[0058] The specific metal elements in data(c) are determined according to the metal element-containing raw material charged into the melting furnace. The types of specific metal elements in data(c) do not necessarily have to match the specific metal elements contained in the molten glass in the analytical image. However, from the viewpoint of improving the accuracy of estimation by machine learning, it is preferable, and more preferable, for the types of specific metal elements in data(c) to match, at least partially, the specific metal elements contained in the molten glass in the analytical image. When the present invention is applied to the treatment of radioactive waste liquid, the specific metal element in data (c) is preferably at least one metal element selected from platinum group metals, minor actinides and lanthanides, more preferably at least one metal element selected from ruthenium (Ru, specific gravity about 12.4), rhodium (Rh, specific gravity about 12.5), palladium (Pd, specific gravity about 12.0), neptunium (Np, specific gravity about 20.5), americium (Am, specific gravity about 13.7), curium (Cm, specific gravity about 13.5), lanthanum (La, specific gravity about 6.2), cerium (Ce, specific gravity about 6.8) and neodymium (Nd, specific gravity about 7.0), and even more preferably at least one metal element selected from ruthenium, rhodium and palladium. In particular, the specific metallic elements in data (c) are preferably three platinum group metallic elements: ruthenium, rhodium, and palladium, or preferably two: ruthenium and palladium.

[0059] If the training image is a simulation image, data(c) may be the concentration of a specific metal element set or calculated in a computational fluid dynamics simulation for the molten glass near the inlet of the model nozzle. If the training image is a real-world image, data(c) may be the concentration of a specific metal element for the molten glass near the inlet of the flow nozzle, determined in advance by another concentration measurement method.

[0060] Data (d) Data (d) is temperature information of the molten glass flowing down from the nozzle in (a). Preferably, the temperature information in data (d) is at least one of the temperature at a predetermined point in the flowing molten glass in the training image and the temperature distribution in the flowing molten glass in the training image. As described in relation to the temperature information contained in the analysis image, the temperature and temperature distribution at a predetermined point in the flowing molten glass (especially near the nozzle exit) have a correlation with the temperature of the molten glass near the nozzle inlet and the concentration of a specific metal element. In this invention, this correlation is used in machine learning, and the model is trained to perform machine learning. The temperature and temperature distribution at the predetermined point are the same as described in relation to the temperature information contained in the analysis image.

[0061] If the training image is a simulation image, data(d) may be the temperature set or calculated in a computational fluid dynamics simulation for molten glass flowing from a model nozzle, or the color values ​​and / or color distribution of the simulation image. If the training image is a real-world image, data(d) may be the temperature obtained in advance by another temperature measurement method (e.g., a thermographic camera), or the color values ​​and / or color distribution of the real-world image.

[0062] In this invention, by including data (d) in the training data for creating the estimation model, the frequency of misestimation can be reduced when estimating the temperature of molten glass near the nozzle inlet compared to using a comparative estimation model (a machine learning model created based on training data consisting of multiple datasets of only data (a) to (c)). This is because temperature information such as the temperature and temperature distribution at a predetermined point in the flowing molten glass has a correlation with the temperature of the molten glass near the nozzle inlet and the concentration of specific metal elements, and is useful information for estimating the temperature of molten glass near the nozzle inlet.

[0063] In addition to the estimated value of the temperature, the estimation method of the present invention can further include obtaining an estimated value of the concentration of a specific metal element in the molten glass near the inlet of the flowing-down nozzle. Based on the estimated value of the concentration of the specific metal element, it becomes easier to predict the timing of charging the additional raw material.

Example

[0064] The present invention will be further specifically described below with reference to examples. However, the scope of the present invention is not limited to the specific examples shown below.

[0065] 1. Simulation of molten glass flowing from a nozzle 1-1.Analysis conditions FIG. 5 shows an analysis system in numerical fluid dynamics simulation. The calculation region is a rectangular parallelepiped space with a width and depth L = 7 cm and a height H = 35 cm. On the upper surface of the rectangular parallelepiped space, a circular tube simulating a nozzle with an inner diameter d = 3 cm, an outer diameter D = 6 cm, and a length h = 20 cm was installed. The circular tube was filled with molten glass at a temperature of 1200 K at time 0 seconds.

[0066] The density ρ of the molten glass in the calculation region g , the thermal conductivity k g and the specific heat capacity C p、g , the density ρ of the nozzle n , the thermal conductivity k n and the specific heat capacity C p、n , and the density ρ of air a , the viscosity μ a , the thermal conductivity k a and the specific heat capacity C p、a were set as shown in Table 1 below. It has been confirmed that these physical property values do not affect the flow and heat transfer of the glass within the temperature range of the simulation.

[0067] For the properties of molten glass, we referred to the following: Toru Sugawara et al., "High-temperature heat capacity and density of simulated high-level glass waste," Journal of Nuclear Materials, Vol. 454, 2014, pp. 298-307; MBRemizov et al., "Thermal and electrical conductivity of molten aluminophosphate and borosilicate glass containing mimics of highly active waste by SNF treatment," Glass Physics and Chemistry, Vol. 44, 2018, pp. 557-563; and J. Puig et al., "Rheological properties of reactor glass melts containing platinum group elements," Procedia Materials Science, Vol. 7, 2014, pp. 156-162. The concentration of platinum group metal elements (PGM concentration) in the molten glass was set to 0% by mass or 3.5% by mass. The assumed glass material (PGM concentration 0 mass%) is a borosilicate glass containing 45-46 mass% SiO2, 13-14 mass% B2O3, 4-5 mass% Al2O3, 10-11 mass% Na2O, 3-4 mass% CaO, 2-3 mass% ZnO, 3-4 mass% ZrO2, and 6-7 mass% rare earth oxides (see the paper by J. Puig et al. above). The molten glass with a PGM concentration of 3.5 mass% contains 2.3 mass% RuO2 and 1.2 mass% Pd (see the paper by J. Puig et al. above). Hereafter, the concentrations of metal elements are based on mass.

[0068] The viscosity of molten glass is highly dependent on temperature and is also affected by the PGM concentration. Therefore, the viscosity of molten glass was defined by the following function, depending on whether the PGM concentration is 0% or 3.5% (see the paper by J. Puig et al. mentioned above). The graph of this function is shown in Figure 6. μ g, 0 = 0.31 exp [ exp {- 0.0030 ( T -273.15 ) + 4.5}] μ g, 3.5 = 0.60 exp [ exp {- 0.0031 ( T - 273.15 ) + 4.5}]

[0069] For the nozzle characteristics, we used the properties of Inconel 690, a nickel-based alloy commonly used as a material for downstream nozzles. For the air properties, we used typical values.

[0070] [Table 1]

[0071] Initial air temperature T 0、a The initial temperature of the nozzle is 300K. 0、n The temperature was set to 1200K. The variable in this simulation is the temperature of the molten glass flowing into the nozzle (in other words, the molten glass at the bottom of the furnace), which was varied within the range of 1300 to 1500K.

[0072] At the top of the nozzle, there is a constant pressure (1.0 × 10⁻¹⁰) equivalent to the hydrostatic pressure at the bottom of the melting furnace. 4 A pressure of Pa was applied, and molten glass was introduced into the nozzle inlet and allowed to flow out into the air from the nozzle outlet. Adiabatic and no-slip conditions were applied as boundary conditions for the upper part of the calculation domain (excluding the nozzle). For the other parts of the calculation domain, an open condition was applied as the boundary condition, and the entire nozzle was constantly heated at 1173K. No-slip conditions were applied to the inner surface, outer surface, and edges of the nozzle.

[0073] The governing equations used in the calculations are the Navier-Stokes equations, the continuity equation, the advection-diffusion equation for the VOF function, and the heat advection-diffusion equation.

[0074] These governing equations were approximated using the finite volume method and solved using OpenFOAM (Open source Field Operation And Manipulation). OpenFOAM is a C++ toolbox for numerical analysis and pre- and post-processing of continuous fluid dynamics, including computational fluid dynamics (see OpenFOAM Foundation, https: / / openfoam.org / , Open CAE Society, Numerical Analysis of Heat Transfer and Flow with OpenFOAM, Morikita Publishing (2016)). The VOF method used in this example also utilizes an algorithm already implemented in OpenFOAM. The computational domain was uniformly divided into a structured grid (rectangular mesh), with a mesh count of 384,000 (32 × 40 × 300).

[0075] 1-2. Flow motion of molten glass Figures 7-12 are simulation images showing the flow motion (0-10 seconds) of molten glass (PGM concentration 0% or 3.5%, temperature 1300K, 1400K, or 1500K) through a nozzle. In each of Figures 7-12, (a) is a cross-sectional view passing through the center of the nozzle, and (b) is a front view from outside the nozzle. Figures 7-12 also show the temperature distribution of the molten glass from t=0s to t=10s, when the molten glass flows in from the nozzle inlet.

[0076] At t=0s, the nozzle is filled with molten glass at 1200K. As time progresses, the 1200K molten glass flows down from the nozzle exit as a jet, and then molten glass at 1300-1500K supplied from the top of the nozzle flows down from the nozzle exit as a jet. In all cases, the flowing molten glass (also simply called the "jet") gradually widens, and by t=10s, it reaches a near-steady state.

[0077] Figure 13 is a graph showing the interface position of molten glass at various temperatures as it flows down from the nozzle. (a) shows the case when the PGM concentration is 0%, and (b) shows the case when the PGM concentration is 3.5%. From Figure 13, it can be seen that, regardless of the PGM concentration, the interface position moves outward as the temperature of the molten glass increases. This is due to Hagen-Poiseuille's law. That is, as the temperature of the molten glass increases, the viscosity of the molten glass decreases, which increases the flow rate of the molten glass and increases the jet diameter (the diameter D of the flowing molten glass).

[0078] Figure 14 is a graph showing the relationship between the temperature of the molten glass flowing down from the nozzle and the jet diameter. (a) shows the relationship at 5 cm below the nozzle exit, and (b) shows the relationship at 10 cm below the nozzle exit. In Figure 14, in addition to the temperatures of 1300K, 1400K, and 1500K, the calculation results for other temperatures (1325K, 1350K, 1375K, 1425K, 1450K, 1475K) are also shown.

[0079] Figure 14 also shows that the jet diameter increases as the temperature of the molten glass increases. Furthermore, it was found that there is a linear correlation between the temperature of the molten glass and the jet diameter, and that the rate of increase in jet diameter is greater as the PGM concentration decreases. These results suggest that there is a linear correlation between the temperature of the molten glass at the bottom of the furnace and the jet diameter for each PGM concentration, and that it may be possible to estimate the temperature of the molten glass at the bottom of the furnace based on the measured jet shape.

[0080] 2. Estimation of the state of the molten furnace bottom using machine learning We attempted to apply machine learning with training data to the above temperature estimation. Specifically, we created a machine learning model based on training data consisting of a dataset including training images created by computational fluid dynamics simulations. We then input analytical images obtained by imaging flowing molten glass into this machine learning model and investigated whether it is possible to estimate the temperature of the molten glass at the bottom of the melting furnace by having the machine learning model output the state, physical properties, or characteristics of the molten glass.

[0081] 2-1. Overview of the Dataset The dataset for the machine learning model consists of image data obtained from computational fluid dynamics simulations performed under different temperature and PGM concentration conditions. The filename of each image data file contains embedded values ​​for temperature and PGM concentration, which are used as labels in the machine learning process.

[0082] 2-2. Data Preprocessing The image data is saved as JPEG image files. First, this image data is loaded into the Keras library in Python. Next, each image data is resized to 100x100 pixels, the RGB values ​​of each image data are normalized, and each image data is converted into a NumPy array format. By scaling the RGB values ​​from the range of 0 to 255 to the range of 0 to 1, machine learning can be performed efficiently.

[0083] The temperature and PGM concentration values ​​embedded in the image filenames were extracted and used as labels in machine learning. Specifically, the integer values ​​included in part of the filenames were normalized using the following formula. Temperature normalization: (Temperature part of filename - 1300) / 200 Normalization of density: (density portion of filename) / 3.5

[0084] In addition, to increase the amount of training data, image transformations including random rotations up to 360 degrees were applied to each image data to improve the generalization performance of the machine learning model. Image transformations were applied up to three times to each image data.

[0085] 2-3. Model Construction In machine learning, we used a convolutional neural network (CNN). CNNs are an optimal network architecture for extracting spatial features from images and are excellent at extracting local features. Furthermore, stacking multiple convolutional layers increases the level of abstraction of the extracted features, which helps in the final regression task prediction.

[0086] The CNN architecture was designed as follows: A CNN consists of an input layer, a convolutional layer, a pooling layer, a dropout layer, a fully connected layer, and an output layer. • Input layer: The image is 100x100 pixels in size and is input as an RGB color image (3 channels). • Convolutional layer: Convolutional layers are responsible for extracting features from images, performing convolution operations on the image using a small matrix called a "kernel." As a result of this operation, a "feature map" is generated that captures local features within the image. The first convolutional layer used 128 filters, the next layer used 256 filters, and the last layer used 512 filters, gradually extracting more features from the image. In this example, a 3x3 size kernel was used. ReLU was used as the activation function to introduce nonlinearity. • Pooling layer: After the convolutional layer, MaxPooling2D was applied to downsample features and reduce computational cost. This allows for a reduction in the amount of input data to the next layer while preserving important features, thereby improving training efficiency and inference speed. • Dropout layer: To prevent overfitting, a dropout layer was inserted after each layer. The dropout rate was set to 0.25, randomly disabling neurons to improve the model's generalization ability. ·Fully connected layer: After extracting features using convolutional and pooling layers, the features extracted in each layer were flattened to one dimension to obtain two fully connected layers (512 units each). The final regression prediction is performed through the fully connected layers. Output layer: Ultimately, two output layers were implemented to predict temperature and PGM concentration. However, a single output layer for temperature prediction would also suffice. A linear activation function was applied to each output layer, and predictions were made in a format suitable for the regression problem.

[0087] The Adam optimizer was used to compile the model. The learning rate was set to 0.001. The coefficient for calculating the first moment of the gradient (mean of the gradient) was set to 0.9, and the coefficient for calculating the second moment of the gradient (variance of the gradient) was set to 0.999. For the loss function, the mean squared error (MSE) was used for both temperature and PGM concentration.

[0088] 2-4. Machine Learning Models The batch size for the training data in machine learning was set to 16. To prevent overfitting, an EarlyStopping callback was used to stop machine learning if the validation loss did not improve for 10 consecutive epochs.

[0089] 2-5. Prediction and Evaluation Using the machine learning model described above, we predicted or estimated the temperature and PGM concentration of the molten glass shown in the new image (analysis image).

[0090] 3. Estimation results using machine learning First, we verified whether the temperature of the molten glass could be estimated when the PGM concentration in the training data and the images used for analysis were identical.

[0091] Figure 15 is a conceptual diagram of the reference training data. This training data includes a dataset consisting of three data points: (a) training images of molten glass flowing from the nozzle, (b) the temperature of the molten glass near the nozzle inlet (in 25K intervals within the range of 1300-1500K), and (c) the PGM concentration in the molten glass near the nozzle inlet (3.5%). The training images of the molten glass were given a unified color tone based on the actual color of the molten glass, but the relationship between color tone and temperature was not provided. Then, following the model construction method described above, a machine learning model (reference) was created based on the reference training data.

[0092] The analysis image input to the machine learning model is a simulated image of molten glass (PGM concentration 3.5%) flowing from a nozzle. The simulated image of the molten glass was given a unified color tone based on the actual color of the molten glass, but the relationship between color tone and temperature was not provided. The above analysis image was then input into the machine learning model (reference), and the machine learning model (reference) was made to output an estimated value of the temperature of the molten glass.

[0093] Figure 16 is a graph showing the results of temperature estimation using a machine learning model (reference) created based on the reference training data shown in Figure 15. It was found that when the PGM concentration in the training data and the analysis image are the same, the temperature of the molten glass can be estimated with high accuracy, with an error of less than 2% (absolute value).

[0094] However, since the PGM concentration in molten glass is a physical property that can actually fluctuate over time depending on the deposition of metal elements and the charging of raw materials into the melting furnace, it is desirable to improve the generalization ability of machine learning so that it can be applied to analytical images where the PGM concentration is unknown.

[0095] <Comparative Example> In the comparative example, to enhance the generalization ability of machine learning, training data containing datasets with different PGM concentrations in molten glass was used. Figure 17 is a conceptual diagram of the comparative training data. This training data includes a dataset consisting of three data points: (a) training images of molten glass flowing from the nozzle, (b) the temperature of the molten glass near the nozzle inlet (in 25K intervals within the range of 1300-1500K), and (c) the PGM concentration in the molten glass near the nozzle inlet (0% or 3.5%). The training images of the molten glass were given a unified color tone based on the actual color of the molten glass, but the relationship between color tone and temperature was not provided. Then, a comparative estimation model was created based on the comparative training data according to the model construction method described above.

[0096] The analytical images input to the comparative estimation model are simulated images of molten glass flowing from a nozzle. The simulated images of molten glass were given a unified color tone based on the actual color of the molten glass, but the relationship between color tone and temperature was not provided. Multiple analytical images with PGM concentrations of 0% and 3.5% were prepared, but when inputting them into the comparative estimation model, the PGM concentration was treated as unknown. The above analytical images were then input into the comparative estimation model, and the model was made to output estimated values ​​for the temperature of the molten glass at the bottom of the melting furnace and the PGM concentration.

[0097] Figure 18 is a graph showing the estimation results using a comparative estimation model created based on comparative training data. (a) shows the estimated temperature, and (b) shows the estimated PGM concentration. Looking at Figure 18(a), it can be seen that the variability of the estimated temperature is larger compared to Figure 16. In addition, in some plots, estimated values ​​(error of more than 2%) that differ significantly from the actual temperature and PGM concentration were output, indicating that misestimation occurred.

[0098] One possible cause of the above misestimation is that, within the range of the machine learning model's training, the jet shape may approximate even at different temperatures and PGM concentrations. For example, according to computational fluid dynamics simulation (see Figure 14), the jet diameter of molten glass with a PGM concentration of 0% is approximately 1.6 cm at 5 cm below the nozzle exit and approximately 1.2 cm at 10 cm below the nozzle exit when the temperature is 1325 K. Looking at molten glass with a PGM concentration of 3.5%, it can be seen that the jet diameter becomes similar to that of molten glass with a PGM concentration of 0% in the temperature range of 1475 to 1500 K. This is thought to be because the decrease in viscosity due to the increase in temperature of the molten glass and the increase in viscosity due to the increase in PGM concentration in the molten glass cancel each other out. In the comparative example, it is thought that the misestimation occurred because the comparative estimation model could not correctly determine the PGM concentration of the flowing molten glass.

[0099] As described above, the comparative estimation model created based on comparative training data including the three data sets (a) to (c) mentioned above, was found to be unable to estimate the temperature of the molten glass at the bottom of the melting furnace with an accuracy of within 2% error when using only jet shape data, and misestimations may occur.

[0100] <Example 1> In Example 1, in order to suppress the occurrence of the aforementioned misestimation, training data was used that included (d) a dataset containing temperature information of the molten glass (jet) flowing down from the nozzle, in addition to the data (a) to (c) described above. Figure 19 is a conceptual diagram of the training data for the example. This training data includes a dataset consisting of four data points: (a) training images of the molten glass flowing down from the nozzle, (b) the temperature of the molten glass near the nozzle inlet (in 25K intervals in the range of 1300 to 1500K), (c) the PGM concentration in the molten glass near the nozzle inlet (0% or 3.5%), and (d) the temperature distribution of the molten glass (jet) flowing down from the nozzle. The RGB values ​​of the training images were normalized according to the data preprocessing method described above to give the training images a color coding that shows the temperature distribution of the molten glass (jet). Then, an estimation model was created based on the training data for the example according to the model construction method described above.

[0101] The analytical images input to the estimation model are simulated images of molten glass flowing down from a nozzle. The simulated images of molten glass were also color-coded to show the temperature distribution of the molten glass (jet). Multiple analytical images with PGM concentrations of 0% and 3.5% were prepared, but when inputting them into the estimation model, the PGM concentration was treated as unknown. The above analytical images were then input into the estimation model, and the estimation model was made to output estimated values ​​for the temperature of the molten glass at the bottom of the melting furnace and the PGM concentration.

[0102] Figure 20 is a graph showing the estimation results using an estimation model created based on training data for the present invention example. (a) shows the estimation results for temperature, and (b) shows the estimation results for PGM concentration. Looking at Figure 20(a), it can be seen that the variability of the estimated temperature is smaller compared to Figure 18(a), with an error of 2% or less in the temperature range of 1300 to 1500 K, indicating that misestimation was suppressed. Also, looking at Figure 20(b), the error in the estimated value of the PGM concentration is smaller than in Figure 18(b) (generally 10% or less), indicating a good estimation result.

[0103] As described above, the estimation model created based on training data including the four data sets (a) to (d) mentioned above has been demonstrated to be able to estimate the temperature of the molten glass at the bottom of the furnace with an accuracy of less than 2% error and without the occurrence of misestimations. Furthermore, it has been demonstrated that, according to the present invention, it is possible to estimate the PGM concentration of the molten glass with good accuracy simultaneously with the estimation of the temperature of the molten glass at the bottom of the furnace.

[0104] <Example 2> In Example 1, under conditions where the specific metal elements were mainly platinum group metal elements, ruthenium (Ru, specific gravity approximately 12.4) and palladium (Pd, specific gravity approximately 12.0), the viscosity μ of the molten glass was measured. g The viscosity μ of molten glass was defined and a simulation image was created. In Example 2, the viscosity μ of molten glass was defined under conditions where the specific metal elements were mainly minor actinides such as neptunium (Np, specific gravity approximately 20.5), americium (Am, specific gravity approximately 13.7), and curium (Cm, specific gravity approximately 13.5). g We defined the parameters and created training images (simulation images) in the same manner as in Example 1.

[0105] Using the training images created in Example 2, an estimation model was created in the same manner as in Example 1, and the temperature of the molten glass and the concentration of specific metal elements in the analysis images (simulation images) were estimated. Similar to Example 1, the temperature of the molten glass at the bottom of the furnace could be estimated with an accuracy of less than 2% error and without any misestimations. Also, similar to Example 1, the concentration of specific metal elements in the molten glass could be estimated with good accuracy.

[0106] <Example 3> In Example 3, the viscosity μ of the molten glass was measured under conditions where the specific metal elements were mainly lanthanides such as lanthanum (La, specific gravity approximately 6.2), cerium (Ce, specific gravity approximately 6.8), and neodymium (Nd, specific gravity approximately 7.0). g We defined the parameters and created training images (simulation images) in the same manner as in Example 1.

[0107] Using the training images created in Example 3, an estimation model was created in the same manner as in Example 1, and the temperature of the molten glass and the concentration of specific metal elements in the analysis images (simulation images) were estimated. Similar to Example 1, the temperature of the molten glass at the bottom of the furnace could be estimated with an accuracy of less than 2% error and without any misestimations. Also, similar to Example 1, the concentration of specific metal elements in the molten glass could be estimated with good accuracy. [Industrial applicability]

[0108] The method for manufacturing glass in a container, the manufacturing apparatus, and the estimation method according to the present invention are useful for producing glass solidified bodies for geological treatment containing high-level radioactive waste liquid.

[0109] Furthermore, the method for manufacturing containerized glass, the manufacturing apparatus, and the estimation method according to the present invention are not limited to the treatment of radioactive waste liquid. The present invention may also be applicable in general when the metal element-containing raw material contains a metal element that has a specific gravity greater than that of the glass raw material in its solid state. In such cases, a colored glass product can be considered as a vitrified body, in which a metal material is added to the glass raw material as a coloring raw material. [Explanation of Symbols]

[0110] 1. Manufacturing apparatus for glass containers 2 Raw materials containing metal elements 3. Glass raw materials 5. Molten glass 6. Flowing molten glass 10. Melting furnace 11a, 11b Main electrode 11c bottom electrode 12 Flow nozzle 14 Firebricks 15 Raw material charging port 16 Exhaust vents 20 Control Unit 21 Imaging Department 30 Storage Units 31 Container (canister) 32 bogies

Claims

1. The process involves heating metal element-containing raw materials and glass raw materials in a melting furnace to form molten glass. The process involves flowing the formed molten glass down from the flow nozzle of the melting furnace, and A method for manufacturing containerized glass, which includes placing the flowing molten glass into a container, Metal element-containing raw materials include metal elements (hereinafter also referred to as "specific metal elements") that have a specific gravity greater than that of the glass raw material in a solid state. The manufacturing method further includes obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle by the estimation method described below, and controlling the heating based on the obtained estimated value so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature: ・Estimation method Prepare an analytical image obtained by imaging the molten glass flowing down from the flow nozzle of the melting furnace, the analytical image including temperature information of the flowing molten glass. Inputting the aforementioned analysis images into a machine learning-based estimation model, and An estimation method comprising obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle from the estimation model into which the aforementioned analytical image has been input: However, the estimation model described above is a machine learning model created by machine learning the correlation between data based on training data including the following data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the inlet of the nozzle in (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a).

2. The temperature information included in the analytical image is at least one of the temperature at a predetermined point in the flowing molten glass in the analytical image and the temperature distribution in the flowing molten glass in the analytical image. The manufacturing method according to claim 1, wherein the temperature information of the data (d) is at least one of the temperature of a predetermined point in the flowing molten glass in the training image and the temperature distribution in the flowing molten glass in the training image.

3. The manufacturing method according to claim 1 or 2, wherein the metal element-containing raw material contains a specific metal element having a specific gravity of 7 or more in a solid state.

4. The manufacturing method according to claim 3, wherein the metal element-containing raw material contains a platinum group metal element.

5. The manufacturing method according to claim 1 or 2, further comprising obtaining, in addition to the estimated temperature, an estimated value of the concentration of the specific metal element in the molten glass near the inlet of the flow nozzle.

6. The manufacturing method according to claim 1 or 2, wherein the metal element-containing raw material is radioactive waste liquid.

7. The manufacturing method according to claim 1 or 2, wherein the analytical image is a video or a still image.

8. A melting furnace equipped with a flow nozzle, which heats a metal element-containing raw material and a glass raw material to form molten glass, A control unit that controls heating by the melting furnace, An imaging unit for imaging the molten glass flowing down from the aforementioned flow nozzle, A manufacturing apparatus for containerized glass, comprising a receiving section for containing the flowing molten glass in a container, Metal element-containing raw materials include metal elements (hereinafter also referred to as "specific metal elements") that have a specific gravity greater than that of the glass raw material in a solid state. The control unit obtains an estimated value of the temperature of the molten glass near the inlet of the flow nozzle using the following estimation method, and controls the heating based on the obtained estimated value so that the temperature of the molten glass near the inlet of the flow nozzle is maintained at a predetermined set temperature, in a manufacturing apparatus: ・Estimation method Prepare an analytical image obtained by imaging the molten glass flowing down from the flow nozzle of the melting furnace with the imaging unit, the analytical image including temperature information of the flowing molten glass. Inputting the aforementioned analysis images into a machine learning-based estimation model, and An estimation method comprising obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle from the estimation model into which the aforementioned analytical image has been input: However, the estimation model described above is a machine learning model created by machine learning the correlation between data based on training data including the following data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the inlet of the nozzle in (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a).

9. The temperature information included in the analytical image is at least one of the temperature at a predetermined point in the flowing molten glass in the analytical image and the temperature distribution in the flowing molten glass in the analytical image. The manufacturing apparatus according to claim 8, wherein the temperature information of the data (d) is at least one of the temperature of a predetermined point in the flowing molten glass in the training image and the temperature distribution in the flowing molten glass in the training image.

10. The manufacturing apparatus according to claim 8 or 9, wherein the metal element-containing raw material includes a specific metal element having a specific gravity of 7 or more in a solid state.

11. The manufacturing apparatus according to claim 10, wherein the metal element-containing raw material includes a platinum group metal element.

12. The manufacturing apparatus according to claim 8 or 9, wherein the estimation method further includes obtaining an estimated value of the concentration of the specific metal element in the molten glass near the inlet of the flow nozzle, in addition to the estimated value of the temperature.

13. The manufacturing apparatus according to claim 8 or 9, wherein the metal element-containing raw material is radioactive waste liquid.

14. The manufacturing apparatus according to claim 8 or 9, wherein the analysis image is a video or a still image.

15. A method for estimating the temperature of molten glass near the inlet of the flow nozzle, used in the method for manufacturing a containerized glass according to claim 1, Prepare an analytical image obtained by imaging the molten glass flowing down from the flow nozzle of the melting furnace, the analytical image including temperature information of the flowing molten glass. Inputting the aforementioned analysis images into a machine learning-based estimation model, and An estimation method comprising obtaining an estimated value of the temperature of the molten glass near the inlet of the flow nozzle from the estimation model into which the aforementioned analytical image has been input: However, the estimation model described above is a machine learning model created by machine learning the correlation between data based on training data including the following data (a) to (d). (a) Learning images of molten glass flowing down from the nozzle, (b) Temperature of molten glass near the inlet of the nozzle in (a), (c) The concentration of a specific metal element in the molten glass near the inlet of the nozzle in (a), (d) Temperature information of the molten glass flowing down from the nozzle in (a).

16. The temperature information included in the analytical image is at least one of the temperature at a predetermined point in the flowing molten glass in the analytical image and the temperature distribution in the flowing molten glass in the analytical image. The estimation method according to claim 15, wherein the temperature information of the data (d) is at least one of the temperature of a predetermined point in the flowing molten glass in the training image and the temperature distribution in the flowing molten glass in the training image.

Images