Methods for predicting the lifespan of refractory materials

Abstract

The refractory life prediction method is a method for predicting the life of refractory in contact with molten metal while it is lined with a surface deposit in a molten steel container. The method includes: a step of acquiring measurement data representing the surface condition of the refractory measured by a sensor (S22); a step of calculating thickness data based on the difference between the measurement data and surface reference data representing the surface condition (S23); a step of comparing minimum thickness data representing the minimum thickness calculated based on past thickness data with the thickness data to calculate attachment thickness data representing the thickness of the deposit (S24); and a step of predicting the life of the refractory by inputting the attachment thickness data into a life prediction model (S26).

Description

Technical Field

[0001] This disclosure relates to a method for predicting the life of refractory materials. More particularly, this disclosure relates to a method for predicting the life of an unshaped refractory lining used in the ironmaking industry, taking into account the peeling off of molten steel containers used for conveying or holding molten steel and for various refining processes. Prior Technology

[0002] Refractory materials used in molten steel containers are divided into inner linings that come into contact with the high-temperature molten material and outer linings (permanent linings) such as sheet metal on the frame side. The inner lining is damaged by chemical melting due to contact with molten steel and slag, cracking caused by thermal load, and subsequent peeling after cracking. Therefore, the inner lining should be replaced with a new one before it completely disappears. Damage is assessed not only by visually monitoring the lining but also by monitoring the sheet metal temperature using thermocouples embedded in the refractory lining or infrared thermal imagers, and by measuring the remaining thickness using a 3D laser scanner.

[0003] For example, Patent Document 1 discloses a measurement system embedded in the inner surface of the outer wall of a metallurgical vessel and used to predict the future state of a refractory lining exposed to heat during exposure to molten metal. The system includes one or more laser scanners and a processor. The laser scanners perform multiple laser scans of the refractory lining while the metallurgical vessel is empty. At least one laser scanner performs a laser scan of the refractory lining before heating and collects data related to the structural state of the refractory lining before heating. At least one laser scanner collects data related to the structural state of the refractory lining after heating. The processor predicts the future state of the lining. [Existing technical documents] [Patent Literature]

[0004] Patent Document 1: Japanese Patent Publication No. 2022-550056 Summary of the Invention

[0005] [The problem that the invention aims to solve] Here, the technology described in Patent Document 1 does not envision a state where there are attachments on the surface of the amorphous refractory, making it difficult to estimate the peeling damage of such refractory.

[0006] The purpose of this disclosure, made in view of the aforementioned circumstances, is to provide a method for predicting the life of refractory linings on the inner surface of containers for transporting or holding molten metal, taking into account peeling damage. [Methods for solving problems]

[0007] (1) The life prediction method for a refractory in one embodiment disclosed herein Methods for predicting the lifespan of refractory materials in contact with molten metal while their surfaces are covered with deposits include: The step of acquiring measurement data representing the surface condition of the refractory, determined by a sensor; The step of calculating the thickness data based on the difference between the measured data and the surface reference data representing the surface condition of the reference; The step of comparing the minimum thickness data, calculated based on the previous thickness data, representing the minimum thickness, with the thickness data to calculate the attachment thickness data representing the thickness of the attachment; and The step of predicting the life of the refractory by inputting the adhesion thickness data into the life estimation model.

[0008] (2) As one embodiment of this disclosure, as described in (1), wherein, The life estimation model is generated by taking past adhesion thickness data as input, life data representing the past life of the refractory as output, and using machine learning to generate the model.

[0009] (3) As one embodiment of this disclosure, as described in (2), wherein, The life estimation model is generated by taking the surface temperature of the refractory and the operational data of manufacturing metal products from the molten metal as further inputs, and using machine learning.

[0010] (4) As one embodiment of this disclosure, as described in (2), wherein, The lifetime estimation model is generated by taking the image of the thickness data and the minimum thickness data as input and using machine learning.

[0011] (5) As one embodiment of this disclosure, as described in (4), wherein, The input image is a combination of multiple images formed by imagerizing the thickness data and the minimum thickness data from multiple charges.

[0012] (6) As one embodiment of this disclosure, as described in (2), wherein, The life estimation model is generated by taking the average and standard deviation calculated from the thickness data at each measurement point of the refractory as input and using machine learning. [The effects of the invention]

[0013] This invention provides a method for predicting the life of refractory linings on the inner surface of containers used for transporting or holding molten metal, taking into account peeling damage. Simple Explanation of the Diagram

[0014] Figure 1 is a schematic structural diagram of a refractory life prediction system that implements an embodiment of the refractory life prediction method disclosed herein. Figure 2 is a diagram showing an example of the structure of a molten steel container. Figure 3 is a diagram used to illustrate the temperature gradient of refractory materials. Figure 4 is a diagram used to illustrate peeling damage. Figure 5 shows the situation of measuring the molten steel container using a sensor. Figure 6 is a flowchart illustrating the process of generating an example lifetime estimation model. Figure 7 is a flowchart illustrating the processing of a refractory life prediction method according to an embodiment of the present invention. Implementation

[0015] The following describes a method for predicting the life of a refractory material according to one embodiment of the present invention, with reference to the accompanying drawings. These drawings are schematic and may sometimes differ from actual shapes, dimensions, etc.

[0016] Figure 1 is a schematic structural diagram of a refractory life prediction system implementing the refractory life prediction method of this embodiment. The refractory life prediction system includes a data processing unit 10 and a sensor 20. The refractory life prediction system may further include an operation database 21. The operation database 21 stores operation data when manufacturing metal products from molten metal. Additionally, the operation database 21 also stores various measurement data related to the manufacturing equipment. The data processing unit 10 includes a memory unit 11 and a calculation unit 12. The memory unit 11 stores various data used in the refractory life prediction method, which includes the life prediction model described later. The calculation unit 12 performs various calculations in the refractory life prediction method. The data processing unit 10 predicts the life of the refractory in contact with the molten metal when there are deposits on the surface of the molten steel container 30 (see Figure 2). The constituent components of the refractory life prediction system will be described below.

[0017] (Structure of molten steel container 30) Figure 2 is a diagram showing a structural example of a molten steel container 30. The molten steel container 30 is sometimes referred to as a ladle. The molten steel container 30 has a multi-layered structure with a steel shell and an inner lining and a permanent lining. As the inner lining, bricks (shaped refractories) are installed on the upper side, and refractories (unshaped refractories) are installed on the lower side. On the upper side, the refractories are damaged (melting loss) due to contact with molten slag. Therefore, dense bricks with excellent resistance to melting loss are used. On the lower side, since it is mainly in contact with molten steel, unshaped refractories with excellent workability are used. In this embodiment, the refractories for which life prediction is performed are the unshaped refractories installed on the lower side. Hereinafter, the unshaped refractories will be simply referred to as refractories.

[0018] The rotary kiln receives molten steel in a steel container 30. After secondary refining within the container, the molten steel is discharged from the bottom of the container for casting. Then, the slag remaining in the container is discharged by tilting it. The container remains empty until the rotary kiln receives the next batch of molten steel. One cycle (one heat) is completed from the time the rotary kiln receives molten steel until the next batch is received.

[0019] (Attached substances and wetting phase) The steel molten steel container 30 repeatedly receives molten steel from the rotary kiln until the refractory is disassembled and repaired. Due to use, a substance called buildup forms on the surface of the steel molten steel container 30. The buildup consists of slag and steel impurities mixed in with the alumina component of the steel precipitated from the refractory.

[0020] Furthermore, during the secondary refining process, slag penetrates into the interior of the refractory through the vent openings in the molten steel container 30. The refractory reacts chemically with the slag to form compounds. Here, Figure 3 illustrates the temperature gradient of the refractory. The refractory exhibits a temperature gradient where the working side in contact with the molten steel is hotter than the back side (the side furthest from the molten steel). The interior of the refractory is colder than the working side, thus reaching the solidification temperature at a predetermined location within the refractory. Inside the refractory, the slag reacts with the refractory, and the solidified region of the refractory is called the wetting phase. Additionally, the region of the refractory inside the refractory where slag has not penetrated and which retains the original structure from the initial construction is called the healthy phase. Here, since the thermal conductivity of the adhering material and the wetting phase differs from that of the healthy phase, the temperature gradient is not strictly linear; however, for convenience, a linear representation of the temperature gradient is used in Figure 3.

[0021] (Thickness shift) The thickness of the refractory material in the molten steel container 30 decreases with use. The factors contributing to this thickness reduction are spalling damage and melting loss. Spalling damage, in particular, can occur suddenly. In the prior art, spalling damage is not presumed. If spalling damage occurs, there is a risk of steel leakage due to the rapid thinning of the refractory material; therefore, presuming spalling damage is important for life management aspects such as refractory repair planning. Furthermore, as mentioned above, deposits form on the surface of the refractory. The thickness of these deposits increases with use.

[0022] (Peeling damage) Figure 4 is a diagram illustrating peeling damage and shows a partial cross-section of the molten steel container 30. As mentioned above, the refractory includes an adhering phase, a wetting phase, and a healthy phase. Due to the effects of heating and cooling and differences in physical properties, as shown in Figure 4, a portion of the working surface (the surface in contact with the molten steel) of the refractory may sometimes experience peeling damage from the interface between the wetting phase and the healthy phase.

[0023] (Presumed peel damage) In this embodiment, the refractory life prediction system uses sensor 20 to measure the surface of the molten steel container 30 before the rotary kiln receives the next batch of molten steel. Here, the surface of the molten steel container 30 is the working surface of the refractory in the molten steel container 30 (see Figure 4). Furthermore, sensor 20 is not limited to any specific device capable of measuring surface shape; for example, it could be a three-dimensional laser scanner or an infrared thermal imager. In this embodiment, we will assume that sensor 20 is a three-dimensional laser scanner for explanation.

[0024] Figure 5 illustrates the measurement of the molten steel container 30 using a three-dimensional laser scanner. The three-dimensional laser scanner measures the surface condition of the refractory in each heat before the rotary kiln receives the next batch of molten steel. Furthermore, as shown in Figure 5, the surface condition is measured at multiple measurement points on the refractory. In this embodiment, the three-dimensional laser scanner also measures the molten steel container 30 before the refractory is installed (permanently lined only). The measurement data of the molten steel container 30 before the refractory is installed is used as surface reference data to represent the surface condition. Additionally, in this embodiment, all data measured by the three-dimensional laser scanner is stored in the operation database 21.

[0025] The calculation unit 12 of the data processing device 10 acquires measurement data representing the surface condition of the refractory being measured from the operating database 21 or a three-dimensional laser scanner. Additionally, the calculation unit 12 acquires surface reference data representing the surface condition of the molten steel container 30. The calculation unit 12 can calculate thickness data based on the difference between the measurement data and the surface reference data. The thickness data represents the thickness of each measurement point of the refractory. Here, as shown in FIG. 4, the calculated thickness is the total thickness of the refractory and the attached material. Hereinafter, the thickness of the attached material alone will be referred to as the attachment thickness.

[0026] As described above, due to sudden peeling damage, the thickness of the refractory material is sometimes thinner than the initial thickness. Therefore, even if the thickness of the refractory material changes after peeling damage occurs, the adhesion thickness cannot be accurately calculated using the initial thickness of the refractory material. Therefore, in this embodiment, minimum thickness data representing the minimum thickness is used, calculated based on past thickness data. In detail, the thickness data after refractory material construction is stored in the memory unit 11 as a measurement history in chronological order, and the calculation unit 12 determines the minimum thickness of the refractory material (minimum thickness) by comparing it with past thickness data. For example, in the example of FIG4, the initial thickness of the refractory material is the minimum thickness (left figure), and peeling damage is assumed to have occurred. In this case, the minimum thickness is updated by the calculation unit 12 (right figure). By updating the minimum thickness in this way, even if peeling damage occurs, the calculation unit 12 can compare the minimum thickness data with the thickness data, thereby accurately calculating the adhesion thickness data.

[0027] The calculation unit 12 predicts the lifespan of the refractory by inputting the adhesion thickness data into the lifespan estimation model. Before performing the process of predicting the lifespan of the refractory, the calculation unit 12 generates the lifespan estimation model and stores it in the memory unit 11. Here, in this embodiment, the lifespan estimation model is a model generated by machine learning, taking past adhesion thickness data as input and lifespan data representing the past lifespan of the refractory as output. In other words, the lifespan estimation model of this embodiment is a learning model generated by machine learning, using stored past adhesion thickness data as explanatory variables and lifespan data representing the lifespan of the refractory corresponding to the adhesion thickness data as target variables. The lifespan data can be classified (labeled) for example by whether or not it has been peeled off after a furnace. Here, the method for generating the lifespan estimation model is not limited to a specific method, for example, any of the following first to fourth methods can be performed.

[0028] (First Method) As a descriptive variable, an unfolded image of the thickness data and minimum thickness data, visualized using grayscale, can be used. That is, the image of the thickness data and minimum thickness data can be further used as input, and a lifetime estimation model can be generated using machine learning. In this case, it can be assumed that modeling by distinguishing the thickness of the infiltrated phase from the healthy phase will improve the accuracy of lifetime prediction. Here, convolutional neural networks can be used as a machine learning method.

[0029] (Second Method) As a descriptive variable, multiple unfolded images (e.g., three) that are imaged using grayscale to represent thickness and minimum thickness data can be combined. That is, the input image can be a combination of multiple images representing thickness and minimum thickness data from multiple passes. Here, as a machine learning method, a convolutional neural network can be used.

[0030] (Third method) As explanatory variables, surface temperature data of the refractory and operational data during steelmaking can be used. That is, the surface temperature of the refractory and operational data during the manufacture of metal products from molten metal can be further used as inputs, and a life prediction model can be generated using machine learning. It can be argued that by using, for example, the highest temperature as input for the refractory's surface temperature, the prediction accuracy of the life will be improved. Furthermore, it can be argued that by using, for example, the composition of the molten steel, which is highly correlated with the chemical reaction of the refractory, as operational data, the prediction accuracy of the life will be improved. Here, as a machine learning method, convolutional neural networks can be used.

[0031] (Fourth Method) As explanatory variables, the mean and standard deviation calculated based on the thickness of the refractory at each measurement point can be used. That is, the mean and standard deviation calculated based on the thickness data of the refractory at each measurement point can be further used as input, and a life estimation model can be generated through machine learning. Here, gradient boosting trees can be used as the machine learning method.

[0032] Here, the data processing device 10 can acquire a lifetime estimation model generated by another device and store it in the memory unit 11. However, it is preferable, as in this embodiment, that the arithmetic unit 12 has the function of generating a lifetime estimation model (learning function). By having the learning function, the lifetime estimation model stored in the memory unit 11 can be updated, thereby improving performance.

[0033] The performance of a lifetime estimation model can be evaluated, for example, using test data and the error in the predicted values. The error in the predicted values ​​used in the evaluation can be, for example, the mean absolute error (MAE). If the error in the predicted values ​​exceeds a predetermined threshold, the performance is deemed insufficient, and the computation unit 12 can perform relearning of the lifetime estimation model. The threshold can be determined based on the calculated prediction performance. Furthermore, the lifetime estimation model can be evaluated using reproducibility and fitness rate. Here, reproducibility is an indicator that the model successfully identifies a peeling event as a peeling event that actually occurs. Fitness rate is an indicator that represents the percentage of peeling events predicted by the model that actually occur.

[0034] Machine learning methods are not limited to convolutional neural networks and gradient boosting trees. Furthermore, lifetime estimation models can be structures that combine multiple models. Additionally, data from various sensors can be added to the explanatory variables.

[0035] The data processing device 10 can be implemented, for example, by a computer. A computer may include a memory device such as memory, a central processing unit (CPU), and other control devices such as processors. In this case, the memory unit 11 can be implemented by the computer's memory device. The arithmetic unit 12 can be implemented in the computer's control device by executing a program. The program can be stored in the memory device and read from the control device when it needs to be executed.

[0036] (Methods for generating lifetime estimation models) Figure 6 is a flowchart illustrating the process of generating the life estimation model. The generation of the life estimation model is performed before the life prediction method for refractory is executed. The calculation unit 12 acquires surface reference data (step S11). Additionally, the calculation unit 12 acquires past measurement data (step S12). The calculation unit 12 calculates thickness data based on the difference between the measurement data and the surface reference data (step S13). The calculation unit 12 compares the minimum thickness data with the thickness data to calculate the adhesion thickness data (step S14). As described above, by comparing with past thickness data, the minimum thickness data is updated in the event of peeling damage. Furthermore, the calculation unit 12 acquires life data corresponding to the past measurement data, representing the past life of the refractory (step S15). The calculation unit 12 generates a life estimation model using machine learning (step S16). The calculation unit 12 evaluates the generated life estimation model using test data and calculates the error of the predicted value (step S17). If the error is greater than a predetermined threshold (Yes in step S18), the performance is deemed insufficient, and the computation unit 12 returns to the processing in step S16 to perform relearning of the lifetime estimation model. If the error is less than or equal to a predetermined threshold (No in step S18), the performance is deemed sufficient, and the computation unit 12 stores the generated lifetime estimation model in the memory unit 11, ending the series of processes.

[0037] (Methods for predicting the lifespan of refractory materials) Figure 7 is a flowchart illustrating the processing of the refractory life prediction method of this embodiment. The calculation unit 12 acquires surface reference data (step S21). Additionally, the calculation unit 12 acquires measurement data (step S22). The calculation unit 12 calculates thickness data based on the difference between the measurement data and the surface reference data (step S23). The calculation unit 12 compares the minimum thickness data with the thickness data to calculate the adhesion thickness data (step S24). As described above, by comparing with past thickness data, the minimum thickness data is updated in the event of peeling damage. Furthermore, the calculation unit 12 acquires a life estimation model from the memory unit 11 (step S25). The calculation unit 12 predicts the life of the refractory by inputting the adhesion thickness data into the life estimation model (step S26).

[0038] Here, the processing of the lifetime estimation model generation method in Figure 6 and the lifetime prediction method in Figure 7 are examples. For example, when the first method is used, in the lifetime estimation model generation and lifetime prediction method processing, the calculation unit 12 further uses the thickness data and minimum thickness data as inputs (explanation variables) of the lifetime estimation model.

[0039] (Example) The inner surface of the molten steel container 30 is cylindrical, with an opening of 4.5 m in diameter and a height of 4 m. A three-dimensional laser scanner is installed at a height of 2.8 m above the opening plane of the molten steel container 30 to measure the surface shape of the refractory. The thickness of the refractory is calculated by the data processing device 10 as described above to predict the refractory life. Here, in the generation of the life prediction model, 20,000 images of the refractory thickness data are prepared and classified into 14,000 training data and 6,000 test data. The images are labeled based on whether there is peeling after one heat. When the life prediction model is evaluated using the test data, cross-validation is used to confirm that there is no overlearning and that its usefulness is confirmed.

[0040] As described above, the refractory life prediction method of this embodiment can consider (estimate) the spalling damage of the refractory used as the lining of the molten steel container 30. By estimating the spalling damage of the refractory, the risk of steel leakage can be reduced, and the repair timing of the refractory can be accurately set.

[0041] Although embodiments of this disclosure have been described based on various drawings and examples, these embodiments illustrate methods for concretizing the technical ideas of this disclosure and are not definitive structures. Those skilled in the art should note that various modifications or alterations can be easily made based on this disclosure. Therefore, it should be understood that such modifications or alterations are included within the scope of this disclosure.

[0042] 10: Data processing device 11: Memory Department 12: Arithmetic Department 20: Sensor 21: Operation Database 30: Steel molten container S11, S12, S13, S14, S15, S16, S17, S18, S21, S22, S23, S24, S25, S26: Steps

Claims

1. A method for predicting the life of a refractory, specifically a method for predicting the life of a refractory in contact with molten metal while its surface is covered with deposits, lining a molten steel container, comprising: The step of acquiring measurement data representing the surface condition of the refractory, determined by a sensor; The steps include: calculating thickness data based on the difference between the measured data and surface reference data representing the surface condition of the reference; calculating attachment thickness data representing the thickness of the attachment by comparing the minimum thickness data representing the minimum thickness calculated based on the past thickness data with the thickness data; and predicting the life of the refractory by inputting the attachment thickness data into a life estimation model.

2. The method for predicting the life of refractory materials as described in claim 1, wherein, The life estimation model is a model generated by taking past adhesion thickness data as input, life data representing the past life of the refractory as output, and using machine learning.

3. The method for predicting the life of refractory materials as described in claim 2, wherein, The life estimation model is generated by taking the surface temperature of the refractory and the operational data of manufacturing metal products from the molten metal as further inputs, and using machine learning.

4. The method for predicting the life of refractory materials as described in claim 2, wherein, The lifetime estimation model is generated by taking the image of the thickness data and the minimum thickness data as input and using machine learning.

5. The method for predicting the life of refractory materials as described in claim 4, wherein, The input image is a combination of multiple images formed by imagerizing the thickness data and the minimum thickness data from multiple furnaces.

6. The method for predicting the life of refractory materials as described in claim 2, wherein, The life estimation model is generated by taking the average and standard deviation calculated from the thickness data at each measurement point of the refractory as input and using machine learning.

Images