Material property predictor, material property prediction method and program

Abstract

To provide a material property predictor or the like which can accurately predict a material property value over a substantially whole length of a steel material.SOLUTION: A material property predictor 100 which predicts a material property value over a substantially whole length of an object steel material as a prediction object, comprises a first material property value prediction model that inputs an operational result at each longitudinal position in an object steel material manufacturing process, and outputs probability distribution of the material property value at each longitudinal position of the object steel material. The first material property value prediction model is constructed by using known data that inputs, for example, an operational result at a prescribed longitudinal position in a manufacturing process for a reference steel material manufactured in the past, and outputs a material property value measured for a test piece acquired from the prescribed longitudinal position of the reference steel material. The material property predictor 100 outputs the probability distribution of the material property value at each longitudinal position of the object steel material over the substantially whole length of the object steel material from the first material property value prediction model, thereby predicting the material property value over the substantially whole length of the object steel material.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a material property prediction device, a material property prediction method, and a program for causing a computer to function as a material property prediction device, which are capable of accurately predicting material property values ​​such as tensile stress over substantially the entire length of a steel material such as a steel plate. [Background technology]

[0002] In order to guarantee the required quality, steel materials such as steel plates manufactured through known manufacturing processes such as heating and rolling processes are measured for material property values ​​such as tensile stress prior to delivery. The material property values ​​are measured by taking test pieces from the front or rear end of the steel material (including the front or rear end of the steel material after cutting to a length according to the product specifications) and using a specified testing machine. For this reason, it is not possible to measure the material property values ​​of the steel material over the entire length of the steel material.

[0003] The leading and trailing ends of steel products have a large deviation from the set values ​​for the steel product temperature and rolling load, and the material property values ​​often do not satisfy the requirements. For this reason, when the material property values ​​do not satisfy the requirements at the position where the test piece was taken, conventionally, based on the information on the steel product temperature and rolling load over the entire length of the steel product (temperature chart, load chart), the steel product is cut from the leading or trailing end to a position where the material property values ​​are thought to satisfy the requirements, and a test piece is taken again from the leading or trailing end of the cut steel product to check whether the measured material property values ​​satisfy the requirements.

[0004] However, since the material property values ​​of steel are determined by many operational records such as steel temperature, rolling load, and component values, it is difficult to appropriately determine the cutting position of the steel by only checking the temperature chart and load chart. Therefore, a technology that can accurately predict the material property values ​​over the entire length of the steel is desired. In addition, predicting the material property values ​​over the entire length of the steel is also useful from the viewpoint of quality assurance.

[0005] Conventionally, techniques for predicting material property values ​​over the entire length of a steel material have been proposed, for example, as described in Patent Documents 1 and 2. The technology described in Patent Document 1 is a technology in which a prediction model is created based on information correlating operational records, which are obtained by measuring operational conditions including temperature distribution over the entire area of ​​a steel material, with material property values ​​measured at a predetermined position on the steel material, and then the material property values ​​of the steel material are predicted from the material property prediction model based on the operational records of the steel material produced next. However, the technology described in Patent Document 1 does not take into account the uncertainty of the prediction model, and therefore it is not possible to predict the degree of variation in the material property values, making it difficult to apply the technology to optimizing the cutting position of the steel material or guaranteeing the material property values ​​over the entire length.

[0006] Moreover, the technology described in Patent Document 2 is a hybrid method of a metallurgical model and a machine learning model, and is a method of estimating manufacturing condition factors, which are input variables of the metallurgical model, from equipment output factors by machine learning. The technology described in Patent Document 2 has a problem that the estimation accuracy of the manufacturing condition factors deteriorates when the combination of equipment output factors does not exist in the past performance. In addition, since the metallurgical model does not take into account various disturbances in the steelmaking process, there is a problem that it is difficult to predict the material characteristic values ​​with high accuracy even if the manufacturing condition factors are calculated correctly.

[0007] Non-Patent Documents 1 and 2 describe heteroscedastic Gaussian processes (Gaussian process regression taking heteroscedasticity into account). [Prior art documents] [Patent documents]

[0008] [Patent Document 1] Patent No. 6086155 [Patent Document 2] Patent Publication No. 2022-48037 [Non-patent literature]

[0009] [Non-Patent Document 1] Alan D. Saul, James Hensman, Aki Vehtari, Neil D. Lawrence, "Chained Gaussian Processes", Proceedings of the 19th International Conference on Artificial Intelligence and Statistics, PMLR 51:1431-1440, 2016 [Non-Patent Document 2] "Heteroskedastic Likelihood and Multi-Latent GP", [online], [Retrieved September 19, 2022], Internet<URL:https: / / gpflow.github.io / GPflow / develop / notebooks / advanced / heteroskedastic.html> Summary of the Invention [Problem to be solved by the invention]

[0010] The present invention has been made to solve the problems of the conventional technology described above, and has an objective of providing a material property prediction device, a material property prediction method, and a program that are capable of accurately predicting material property values ​​(including not only expected values ​​but also variations) over almost the entire length of a steel material. [Means for solving the problem]

[0011] In order to solve the above-mentioned problems, the present invention provides, as a first means, a material property prediction device that predicts material property values ​​of a target steel material, which is the steel material to be predicted, over approximately the entire length of the target steel material, and is equipped with a first material property value prediction model that inputs operational results for each longitudinal position in the manufacturing process of the target steel material and outputs a probability distribution of material property values ​​for each longitudinal position of the target steel material, and outputs from the first material property value prediction model the probability distribution of material property values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material, thereby predicting the material property values ​​of the target steel material over approximately the entire length of the target steel material.

[0012] In the first aspect of the present invention, "operational results" refers to data that affect the quality of steel, and is a concept that includes measured values ​​and set values ​​obtained in the steel manufacturing process, as well as data obtained by processing these values ​​using physical formulas, etc. Examples of operational results include the temperature (surface temperature) of the steel in the manufacturing process, the load (rolling load) applied to the steel, the component values ​​of the steel, the temperature of the equipment (rolls, heating furnaces, etc.), and the size (thickness, width) of the steel. The same applies to the second aspect of the present invention described below. In the first aspect of the present invention, the "material characteristic value" refers to the mechanical properties of the steel material, and examples of such properties include tensile stress (TS), yield stress, surface hardness, bending strength, toughness, and fatigue life. The same applies to the second aspect of the present invention described below. In the first aspect of the present invention, the term "operational performance at each longitudinal position" refers to a concept including operation performance that is obtained sequentially at a predetermined pitch (for example, 1 m) in the longitudinal direction, operation performance that is obtained only at a plurality of pitches, or operation performance that is obtained only as a representative value for the entire steel material length, in which the same value as the obtained operation performance or an interpolated value is assigned to each pitch. The same applies to the second aspect of the present invention described later. Furthermore, in the first aspect of the present invention, "almost the entire length" is not limited to the exact entire length of the steel material, but includes the case where there is an unexpected part in the longitudinal direction of the steel material, as long as it is permissible from the viewpoint of quality assurance of the steel material. The same applies to the second aspect of the present invention described later. The first material characteristic value prediction model may be constructed by the material characteristic prediction device relating to the first means of the present invention, or the first material characteristic value prediction model constructed by another device may be stored in the material characteristic prediction device relating to the first means of the present invention.

[0013] The material property prediction device according to the first means of the present invention includes a first material property value prediction model that receives as input the operation results for each longitudinal position in the manufacturing process of a target steel material, which is a steel material to be predicted, and outputs a probability distribution of material property values ​​for each longitudinal position of the target steel material. This first material property value prediction model is constructed by using known data that receives as input the operation results in the manufacturing process at the longitudinal positions (e.g., the front end, rear end, and center of the steel material (corresponding to the front end or rear end of the steel material after cutting to a length according to the product specifications)) where a test piece is obtained from a reference steel material, which is a steel material manufactured in the past, and outputs the material property values ​​measured for the test piece obtained from the reference steel material. By using such known data as teacher data and constructing the first material property value prediction model by machine learning, for example, the first material property value prediction model can accurately calculate and output the probability distribution of material property values ​​for each longitudinal position (e.g., every 1 m) of the target steel material, in other words, the probability distribution of material property values ​​including the longitudinal positions other than the longitudinal positions where the test piece is obtained. Therefore, by outputting the probability distribution of material characteristic values ​​for each longitudinal position of the target steel material from the first material characteristic value prediction model over approximately the entire length of the target steel material, it is possible to accurately predict the material characteristic values ​​of the target steel material over approximately the entire length of the target steel material. The material property prediction device according to the first aspect of the present invention outputs the probability distribution of the material property values ​​of the target steel material, in other words, the expected value and variation of the material property values, as prediction results, and therefore it is expected that the prediction accuracy of the material property values ​​(especially the variation) will be improved even for operational records of the target steel material that have no past operational records. It is also expected that the prediction accuracy of the material property values ​​will be improved even for operational records of the target steel material in which it is difficult to control the material property values ​​and variation in the material property values ​​is likely to occur. This makes it possible to guarantee the material property values ​​over almost the entire length of the steel material.

[0014] In the first method of the present invention, preferably, the first material characteristic value prediction model is reconstructed by adding and using new known data, in which the operational results at the longitudinal position where the test piece was obtained in the manufacturing process of the target steel are used as input, and the material characteristic values ​​measured on the test piece obtained from the target steel are used as output.

[0015] According to the above preferred configuration, not only known data about the reference steel material but also new known data about the target steel material are used as known data for constructing the first material characteristic value prediction model, so that it is expected that the prediction accuracy of the material characteristic value will be improved.

[0016] In addition, in order to solve the above-mentioned problem, the present invention provides, as a second means, a material property prediction device that predicts material property values ​​of a target steel material, which is the steel material to be predicted, over approximately the entire length of the target steel material, and is equipped with a second material property value prediction model that receives as input operational performance for each longitudinal position in the manufacturing process of the target steel material, material property values ​​measured on a test piece obtained from the target steel material, and the longitudinal position of the target steel material from which the test piece was obtained, and outputs a probability distribution of material property values ​​for each longitudinal position of the target steel material, and predicts the material property values ​​of the target steel material over approximately the entire length of the target steel material by outputting from the second material property value prediction model the probability distribution of material property values ​​for each longitudinal position of the target steel material. The second material property value prediction model is constructed by using known data, for example, which takes as input the operational performance of the manufacturing process at the longitudinal position where a test piece for model output was obtained from a reference steel material which is a steel material manufactured in the past, the longitudinal position where the test piece for model output was obtained, the material property values ​​measured for a standard test piece obtained from the reference steel material, the operational performance of the manufacturing process at the longitudinal position of the reference steel material where the standard test piece was obtained, and the longitudinal position where the standard test piece was obtained, and outputs the material property values ​​measured for the test piece for model output.

[0017] When the second material characteristic value prediction model in the second means of the present invention is constructed using known data as described above, the "test piece for model output" means a test piece obtained from a reference steel material, the measured material characteristic values ​​of which are used as output from the second material characteristic value prediction model in the known data when constructing the second material characteristic value prediction model. Furthermore, when the second material characteristic value prediction model is constructed using known data as described above, the "reference test piece" refers to a test piece obtained from a predetermined reference position (a reference longitudinal position that includes a longitudinal position different from the longitudinal position at which the model output test piece was obtained) of the reference steel material. The reference test piece may include a test piece obtained from a longitudinal position different from the model output test piece, as well as a test piece obtained from the same longitudinal position as the model output test piece. Furthermore, when the second material characteristic value prediction model is constructed using known data as described above, when constructing the second material characteristic value prediction model, the "longitudinal position at which the model output test piece was acquired" and the "longitudinal position at which the reference test piece was acquired" used as inputs to the second material characteristic value prediction model in the known data may be used individually as inputs, or the relative position of the longitudinal position at which the model output test piece was acquired, based on the longitudinal position (reference position) at which the reference test piece was acquired (the difference or the absolute value of the difference between the longitudinal position at which the model output test piece was acquired and the reference position) may be used as an input. Furthermore, when the second material property value prediction model is constructed using known data as described above, the known data used to construct the second material property value prediction model is data that takes as input the operational performance of the manufacturing process at the longitudinal position where the test piece for model output was obtained from the reference steel, the longitudinal position where the test piece for model output was obtained from the reference steel, the material property values ​​measured for the standard test piece obtained from the reference steel, the operational performance of the manufacturing process at the longitudinal position of the reference steel where the standard test piece was obtained from the reference steel, and the longitudinal position where the standard test piece was obtained from the reference steel, and outputs the material property values ​​measured for the test piece for model output obtained from the reference steel, but is not limited to this. In addition to the known data on the reference steel, it is also possible to use data as known data, which is input with the operation results in the manufacturing process at the longitudinal position where the test piece for model output was obtained from the target steel, the longitudinal position where the test piece for model output was obtained from the target steel, the material property value measured for the reference test piece obtained from the target steel, the operation results in the manufacturing process at the longitudinal position of the target steel where the reference test piece was obtained from the target steel, and the longitudinal position where the reference test piece was obtained from the target steel, and output with the material property value measured for the test piece for model output obtained from the target steel. In this way, when data on both the reference steel and the target steel are used as known data used to construct the second material property value prediction model, the longitudinal position where the test piece for model output was obtained from the reference steel and the longitudinal position where the test piece for model output was obtained from the target steel may be the same position or different positions. Similarly, the longitudinal position at which the reference test piece is obtained from the reference steel material and the longitudinal position at which the reference test piece is obtained from the target steel material may be the same or different. The second material characteristic value prediction model may be constructed by the material characteristic prediction device relating to the second means of the present invention, or the second material characteristic value prediction model constructed by another device may be stored in the material characteristic prediction device relating to the second means of the present invention.

[0018] The second material property value prediction model provided in the material property prediction device according to the second means of the present invention outputs a probability distribution of material property values ​​for each longitudinal position of the target steel material, similar to the first material property value prediction model provided in the material property prediction device according to the first means. However, unlike the first material property value prediction model, the second material property value prediction model receives as input not only the operational results for each longitudinal position in the manufacturing process of the target steel material, but also material property values ​​measured on a test piece obtained from the target steel material and the longitudinal position of the target steel material from which the test piece was obtained. Furthermore, as described above, since the input of the second material property value prediction model differs from the input of the first material property value prediction model, the second material property value prediction model is constructed using known data that is different from the known data used to construct the first material property value prediction model, and that takes as inputs, for example, the operational results of the manufacturing process at the longitudinal position where the test piece for model output was obtained from the reference steel, the longitudinal position where the test piece for model output was obtained from the reference steel, the material property values ​​measured for the standard test piece obtained from the reference steel, the operational results of the manufacturing process at the longitudinal position of the reference steel where the standard test piece was obtained from the reference steel, and the longitudinal position where the standard test piece was obtained from the reference steel, and outputs the material property values ​​measured for the test piece for model output obtained from the reference steel. As described above, the second material property value prediction model differs from the first material property value prediction model in that it receives as input not only the operational performance of the target steel material, but also the material property values ​​measured on the test piece obtained from the target steel material, and the longitudinal position of the target steel material from which the test piece was obtained. Therefore, the material property prediction device relating to the second means having the second material property value prediction model produces prediction results that are more strongly influenced by the material property values ​​measured on the target steel material than the material property prediction device relating to the second means having the first material property value prediction model.

[0019] The present invention also provides, as a third means, a material property prediction device that includes both a first material property value prediction model and a second material property value prediction model and outputs a combination of the outputs of the prediction models. Specifically, the present invention provides a material property prediction device as a third means, which comprises the first material property value prediction model and the second material property value prediction model, and predicts the material property values ​​of the target steel material over approximately the entire length of the target steel material by outputting a probability distribution that is a combination of the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the first material property value prediction model and the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the second material property value prediction model as the final probability distribution of material property values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material.

[0020] According to the material property prediction device relating to the third means of the present invention, the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the first material property value prediction model and the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the second material property value prediction model are combined to obtain the final probability distribution of material property values ​​for each longitudinal position of the target steel material, thereby improving the prediction accuracy of the material property values.

[0021] In the first to third means of the present invention, preferably, a longitudinal cutting position of the target steel material is calculated so as to minimize the expected manufacturing cost based on the predicted material property values ​​of the target steel material over approximately the entire length of the target steel material.

[0022] According to the above-mentioned preferred configuration, the cutting position in the longitudinal direction of the target steel material is calculated so as to minimize the expected value of the manufacturing cost based on the accurately predicted material property values ​​of the target steel material over substantially the entire length of the target steel material, so that it is expected that the cutting position of the target steel material will be appropriately determined, thereby making it possible to improve the yield of the target steel material.

[0023] In addition, in order to solve the above-mentioned problems, the present invention is also provided as a material property prediction method for predicting material property values ​​of a target steel material, which is the steel material to be predicted, over approximately the entire length of the target steel material, using a first material property value prediction model that inputs operational results for each longitudinal position in the manufacturing process of the target steel material and outputs a probability distribution of material property values ​​for each longitudinal position of the target steel material, and outputs from the first material property value prediction model the probability distribution of material property values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material, thereby predicting the material property values ​​of the target steel material over approximately the entire length of the target steel material. The present invention is also provided as a material property prediction method for predicting material property values ​​of a target steel material, which is the steel material to be predicted, over approximately the entire length of the target steel material, using a second material property value prediction model that takes as input the operational performance for each longitudinal position in the manufacturing process of the target steel material, the material property values ​​measured on a test piece obtained from the target steel material, and the longitudinal position of the target steel material from which the test piece was obtained, and outputs a probability distribution of material property values ​​for each longitudinal position of the target steel material, and outputs from the second material property value prediction model the probability distribution of material property values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material, thereby predicting the material property values ​​of the target steel material over approximately the entire length of the target steel material. The present invention also provides a material property prediction method, which uses the first material property value prediction model and the second material property value prediction model, and predicts the material property values ​​of the target steel material over approximately the entire length of the target steel material by outputting a probability distribution obtained by combining the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the first material property value prediction model and the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the second material property value prediction model as the final probability distribution of material property values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material.

[0024] Furthermore, in order to achieve the above object, the present invention also provides a program for causing a computer to function as the material property prediction device according to any one of the first to third aspects of the present invention. Effect of the Invention

[0025] According to the present invention, it is possible to accurately predict material characteristic values ​​over substantially the entire length of a steel material. [Brief description of the drawings]

[0026] [Figure 1] 1 is a block diagram showing a schematic configuration of a material property prediction device according to an embodiment of the present invention. [Diagram 2] 2 is a flowchart showing an outline of the procedure of a material property prediction method executed by using the material property prediction apparatus 100 shown in FIG. [Diagram 3] 2 is a diagram showing an example of a manufacturing record acquired by a manufacturing record acquiring unit 10 shown in FIG. 1; [Figure 4] FIG. 11 is a diagram showing an example of known data (teacher data) for constructing a second material characteristic value prediction model. [Diagram 5] 1 is a diagram for explaining an example of constructing a first material characteristic value prediction model in comparison with an example of constructing a conventional machine learning model. FIG. [Figure 6] 2 is a diagram showing an example of operation results acquired by an operation results acquisition unit 40 shown in FIG. 1. FIG. [Figure 7] 2 is a diagram showing an example of mechanical test values ​​and mechanical test positions acquired by the material characteristic value acquisition unit 50 shown in FIG. 1. FIG. [Figure 8] FIG. 11 is a diagram showing an example of input data to a second material characteristic value prediction model. [Figure 9] 4 is a schematic diagram illustrating a combination of a probability distribution output from a first material characteristic value prediction model and a probability distribution output from a second material characteristic value prediction model. FIG. [Figure 10] 2 is a diagram showing an example of mechanical test values ​​of a target steel material over substantially the entire length of the target steel material, predicted by the material characteristic value prediction unit 70 shown in FIG. 1. FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] Hereinafter, with reference to the accompanying drawings as appropriate, a material property prediction device according to one embodiment of the present invention will be described using an example in which the steel material is a steel plate. Fig. 1 is a block diagram showing a schematic configuration of a material property prediction device according to the present embodiment. Fig. 2 is a flow chart showing a schematic procedure of a material property prediction method executed by the material property prediction device shown in Fig. 1. The material property prediction device 100 according to this embodiment is a device for predicting material property values ​​of a target steel material over substantially the entire length of the target steel material, which is the steel material to be predicted. As shown in Fig. 1, the material property prediction device 100 includes a manufacturing record acquisition unit 10, a material property value prediction model construction unit 20, a material property value prediction model storage unit 30, an operation record acquisition unit 40, a material property value acquisition unit 50, a material property value prediction model acquisition unit 60, a material property value prediction unit 70, and a steel material cutting position calculation unit 80.

[0028] The material property prediction device 100 includes one or more hardware processors such as a CPU (Central Processing Unit), one or more memories such as a RAM (Random Access Memory) and a ROM (Read Only Memory), and executes various calculations by executing one or more programs stored in the memory by the one or more hardware processors. As a result, the material property prediction device 100 functions as a manufacturing record acquisition unit 10, a material property value prediction model construction unit 20, a material property value prediction model storage unit 30, an operation record acquisition unit 40, a material property value acquisition unit 50, a material property value prediction model acquisition unit 60, a material property value prediction unit 70, and a steel cutting position calculation unit 80. The material property prediction device 100 may be a PLC (Programmable Logic Controller), or may be realized by dedicated hardware such as an ASIC (Application Specific Integrated Circuit). Each unit 10 to 80 included in the material property prediction device 100 will be described below in order.

[0029] <Manufacturing Record Acquisition Department 10> The manufacturing history acquisition unit 10 executes step ST1 shown in Fig. 2. Specifically, the manufacturing history acquisition unit 10 acquires, from a predetermined database (not shown), manufacturing history in the manufacturing process at a longitudinal position (also referred to as "mechanical test position" in this embodiment) where a test piece was acquired from a reference steel material, which is a steel material manufactured in the past. The manufacturing history at the mechanical test position is a combination of data of the operation history at the mechanical test position, material characteristic values ​​(also referred to as "mechanical test values" in this embodiment) measured for the test piece acquired from the mechanical test position, and the mechanical test position. Fig. 3 is a diagram showing an example of manufacturing results acquired by the manufacturing results acquisition unit 10. In the example shown in Fig. 3, for reference steel No. 001, mechanical test values ​​are measured at mechanical test positions 3 m, 751 m, and 1235 m from the tip in the longitudinal direction, and for reference steel No. 002, mechanical test values ​​are measured at mechanical test positions 6 m, 482 m, and 1036 m from the tip in the longitudinal direction. In addition to the mechanical test values ​​and mechanical test positions, the size (plate thickness, plate width, etc.) of the reference steel, component values, and rolling results (heating temperature, rolling load, etc.) are acquired as operational results at each mechanical test position.

[0030] <Material characteristic value prediction model construction unit 20> The material characteristic value prediction model construction unit 20 executes step ST2 shown in FIG. 2. Specifically, the material characteristic value prediction model construction unit 20 constructs a material characteristic value prediction model using known data. For example, the material characteristic value prediction model construction unit 20 may construct a material characteristic value prediction model by machine learning using known data as teacher data. In this embodiment, a material characteristic value prediction model using a Gaussian regression process is constructed in order to quantify the uncertainty of prediction due to the lack of past operating results, but the present invention is not limited to this. For example, the material characteristic value prediction model construction unit 20 may construct a material characteristic value prediction model by statistical processing such as quantile regression using known data, or may construct a material characteristic value prediction model by other machine learning such as a Bayesian neural network. The Gaussian regression process used in this embodiment will be outlined below. In the following equations, variables written in bold represent vectors or matrices. When the output vector f, expressed by the following equation (2), corresponding to the input x, expressed by the following equation (1), follows a Gaussian distribution N(0,K) with mean 0 and covariance matrix K, expressed by the following equation (3), the output f is said to follow a Gaussian process, and can be written as shown in the following equation (4).

number

number

[0031] Although it is possible to use a normal Gaussian regression process as a material characteristic value prediction model, the normal Gaussian regression process cannot take into account the change in the variation due to the difference in the operation record. Therefore, in this embodiment, Gaussian processes gP(0,K1) and gP(0,K2) that individually estimate the expected value (average value) and the variance (standard deviation) expressed by the following formulas (6) to (10) are defined, and a non-uniform variance Gaussian process (Gaussian process considering non-uniform variance) is used in which the final prediction result is a Gaussian distribution with the output of each Gaussian process as a parameter.

number

[0032] The material characteristic value prediction model constructing unit 20 of this embodiment constructs a material characteristic value prediction model using the non-uniform variance Gaussian process described above. Specifically, the material characteristic value prediction model constructing unit 20 constructs a first material characteristic value prediction model and a second material characteristic value prediction model as the material characteristic value prediction models. Each material characteristic value prediction model will be described in order below.

[0033] [First material property value prediction model] The first material characteristic value prediction model is a model that takes operational results for each longitudinal position in the manufacturing process of the target steel as input, and outputs the probability distribution of mechanical test values ​​for each longitudinal position of the target steel. The first material property value prediction model is constructed by using known data, as shown in Figure 3, in which the operational results of the manufacturing process at each mechanical test position of the reference steel are used as input, and the mechanical test values ​​measured on test pieces obtained from each mechanical test position of the reference steel are used as output. The first material characteristic value prediction model can be reconstructed (re-learned when constructed by machine learning) by adding and using new known data about the target steel material, in addition to the known data about the reference steel material as shown in FIG. 3. Specifically, the model can be reconstructed by adding and using known data in which the operation records in the manufacturing process at each mechanical test position of the target steel material acquired by the operation record acquisition unit 40 described later are used as input, and the mechanical test values ​​measured on test pieces acquired from each mechanical test position of the target steel material are used as output. In this way, by using not only the known data about the reference steel material, but also the known data about the target steel material, it is expected that the prediction accuracy of the mechanical test values ​​will be improved.

[0034] [Second material property value prediction model] The second material property value prediction model is a model that takes as input the operating results for each longitudinal position in the manufacturing process of the target steel, the mechanical test values ​​measured on test pieces obtained from each mechanical test position of the target steel, and each mechanical test position, and outputs the probability distribution of the mechanical test values ​​for each longitudinal position of the target steel. FIG. 4 is a diagram showing an example of known data for constructing the second material characteristic value prediction model. The second material characteristic value prediction model is, for example, as shown in Figure 4, based on the operational results in the manufacturing process at the mechanical test position where the test pieces for model output were obtained from the reference steel and the target steel (corresponding to "size", "components" and "rolling results" among the "operational results" shown in Figure 4), the mechanical test position where the test pieces for model output were obtained from the reference steel and the target steel (corresponding to the "mechanical test position" shown in Figure 4, which is input to the second material characteristic value prediction model as a "relative position" as described below), and the mechanical test values ​​measured for the standard test pieces obtained from the reference steel and the target steel (hereinafter, the mechanical test values ​​measured for the standard test pieces will be referred to as "standard mechanical test values" as appropriate, and the "standard mechanical test values" shown in Figure 4 correspond to this). The model is constructed by using known data that takes as inputs the operational results in the manufacturing process at the mechanical test position where standard test pieces were obtained from the reference steel and the target steel (hereinafter, the mechanical test position where the standard test pieces were obtained will be referred to as the "standard mechanical test position" as appropriate) (corresponding to "Rolling results (standard)" among the "operational results" shown in Figure 4), and the mechanical test position where standard test pieces were obtained from the reference steel and the target steel (this is the standard mechanical test position, and in the example shown in Figure 4, this is input into the second material characteristic value prediction model as a "relative position" as described below), and outputs the mechanical test values ​​measured on the model output test pieces obtained from the reference steel and the target steel (corresponding to the "mechanical test values" shown in Figure 4). In the example shown in FIG. 4, the known data includes data on both the reference steel material and the target steel material, but the present invention is not limited thereto, and the known data may include at least data on the reference steel material. In the example shown in FIG. 4, the reference test piece includes only test pieces obtained from a longitudinal position different from that of the model output test piece (in other words, the known data is input and used when the "relative position" described below is not 0), but the present invention is not limited thereto, and the reference test piece may include test pieces obtained from the same longitudinal position as the model output test piece (in other words, the known data may be input when the "relative position" described below is 0). In addition, in the above-mentioned known data, the operation results (corresponding to the "operation results" shown in FIG. 4) in the manufacturing process at the mechanical test position where the model output test piece and the reference test piece are obtained from the target steel material are obtained by the operation results acquisition unit 40 described below. In addition, the mechanical test values ​​measured for the model output test pieces and reference test pieces obtained from the target steel material, and the mechanical test positions at which the model output test pieces and reference test pieces were obtained from the target steel material, are obtained by the material characteristic value acquisition unit 50 described below.

[0035] In the example shown in Fig. 4, as in the case described with reference to Fig. 3, for example, for the reference steel No. 001, mechanical test values ​​are measured at each mechanical test position of 751 m, 1235 m, and 3 m in the longitudinal direction from the tip. In the example shown in Fig. 4, each mechanical test position is not input as an individual mechanical test position at which the model output test piece and the reference test piece are obtained (instead of inputting the absolute position, which is the distance from the tip or rear end, individually), but as a preferred embodiment, a relative position is input. Specifically, for example, as surrounded by a thick frame in Fig. 4, when the mechanical test position of the model output test piece is 751 m, a relative position of 748 m when the position of 3 m is set as the reference mechanical test position, and a relative position of 484 m when the position of 1235 m is set as the reference mechanical test position are input as known data. Since the length of the steel and the mechanical test position (especially the mechanical test position located at the center of the steel) generally differ for each steel, in order to reduce the influence of the difference on the prediction accuracy, it is preferable to input the relative position as each mechanical test position as described above. However, if the length and mechanical test position of the reference steel and the target steel are equivalent, there is no problem even if the absolute position is input individually. In addition, when the mechanical test position of the model output test piece is 751 m, the mechanical test value as the input of the known data is not the mechanical test value of 1015 MPa at the mechanical test position 751 m (i.e., not the mechanical test value of the model output test piece), but the mechanical test value at the reference mechanical test position 3 m (i.e., the reference mechanical test value which is the mechanical test value of the reference test piece) of 1020 MPa and the mechanical test value at the reference mechanical test position 1235 m (reference mechanical test value) of 990 MPa are used, as shown in the thick frame in FIG. 4. However, the present invention is not limited to this, and it is also possible to add the mechanical test value of 1015 MPa of the model output test piece at the mechanical test position 751 m as an input of known data (in other words, the model output test piece may be included in the reference test piece). Furthermore, when the mechanical test position of the model output test piece is 751 m, as shown in a thick frame in FIG. 4, the rolling results (operational results) as an input of known data are the rolling results (heating temperature T2 [°C], rolling load (here, rolling load per unit width) at the mechanical test position 751 m where the model output test piece was obtained.In addition to the rolling results (heating temperature T1 [℃], rolling load P1 [MPa / mm]) at the reference mechanical test position 3m where the reference test piece was obtained, and the rolling results (heating temperature T3 [℃], rolling load P3 [0.25 MPa / mm]) at the reference mechanical test position 1235m where the reference test piece was obtained are also used. When the mechanical test position of the model output test piece is 751m, as shown in the bold frame in Figure 4, the mechanical test value of 1015MPa of the model output test piece at the mechanical test position 751m is used as the output of known data. The same applies to other mechanical test positions of steel No. 001 and other steels. The known data is not limited to the example shown in Fig. 4, and various modifications are possible. For example, it is also possible to use relative values ​​as the heating temperature and rolling load in the operational results, similar to the mechanical test positions.

[0036] The first material characteristic value prediction model described above does not use mechanical test values ​​of the reference steel and the target steel as input, but models the causal relationship between operational performance and the resulting mechanical test values. On the other hand, the second material property value prediction model also uses the mechanical test values ​​of at least the reference steel (in the example shown in Figure 4, both the reference steel and the target steel) as input, and is therefore a model that is more strongly influenced by the mechanical test values ​​measured at least on the reference steel (in the example shown in Figure 4, both the reference steel and the target steel) than by the relationship between operational performance and mechanical test values.

[0037] FIG. 5 is a diagram for explaining an example of constructing a first material characteristic value prediction model in comparison with an example of constructing a conventional machine learning model. Specifically, for convenience of illustration, FIG. 5 shows an example of constructing a model that predicts tensile stress (TS) as a mechanical test value by plotting only the annealing temperature of the steel material among multiple parameters constituting the operating conditions. FIG. 5(a) shows an example of constructing a conventional machine learning model, and FIG. 5(b) shows an example of constructing a first material characteristic value prediction model. Data plotted with "x" in FIG. 5 is known data. In the conventional machine learning model shown in Fig. 5(a), the average value of the mechanical test value (TS) is output as the prediction result for the input operating conditions (annealing temperature). Therefore, in the area surrounded by the dashed line in Fig. 5(a) where the mechanical test values ​​vary widely even under the same operating conditions, there is a problem that the prediction accuracy of the mechanical test value is poor. In contrast, the first material characteristic value prediction model shown in FIG. 5(b) outputs a probability distribution of mechanical test values ​​as a prediction result. Specifically, the first material characteristic value prediction model of this embodiment is a model using a non-uniform variance Gaussian process, and outputs a mean value and a standard deviation as a prediction result. The hatched area shown in FIG. 5(b) is a 95% prediction range (95% confidence interval) obtained from the mean value and the standard deviation. In the first material characteristic value prediction model, as shown by the arrow A1, an area with a large variation in mechanical test values ​​(an area corresponding to the area surrounded by a dashed line in FIG. 5(a)) is included in the probability distribution (95% prediction range), and is reflected as an output. In addition, in the first material characteristic value prediction model, as shown by the arrows A2 and A3, a probability distribution (95% prediction range) is output even in an area with no operating history (an area corresponding to the hatched area in FIG. 5(a)), so that the uncertainty of the prediction is taken into consideration.

[0038] <Material characteristic value prediction model storage unit 30> The material characteristic value prediction model storage unit 30 executes step ST3 shown in Fig. 2. Specifically, the material characteristic value prediction model storage unit 30 stores (memorizes) the first material characteristic value prediction model and the second material characteristic value prediction model constructed by the material characteristic value prediction model construction unit 20.

[0039] <Operational Performance Acquisition Department 40> The operation record acquiring unit 40 executes step ST4 shown in Fig. 2. Specifically, the operation record acquiring unit 40 acquires operation records for each longitudinal position in the manufacturing process of the target steel material, which are obtained after the manufacturing process of the target steel material, from a predetermined database (not shown). Note that in this specification, the manufacturing process of steel material means a process up to a process that adds value to the steel material product (continuous hot-dip galvanizing process, continuous annealing process, etc.), and does not include a finishing process in which the steel material is cut and inspected. Fig. 6 is a diagram showing an example of operation results acquired by the operation results acquisition unit 40. In the example shown in Fig. 6, operation results are acquired every 1 m in the longitudinal direction from the tip of steel material No. 101, which is the target steel material. Note that in the example shown in Fig. 6, similar to the example shown in Fig. 4 described above, mechanical test values ​​are measured at positions 3 m, 654 m, and 1301 m in the longitudinal direction from the tip of the hatched target steel material (steel material No. 101).

[0040] <Material characteristic value acquisition unit 50> The material characteristic value acquisition unit 50 executes step ST5 shown in Fig. 2. Specifically, the material characteristic value acquisition unit 50 acquires the mechanical test values ​​measured for the test pieces acquired from each mechanical test position of the target steel material after the manufacturing process and each mechanical test position from a predetermined database (not shown). Note that the material characteristic value acquisition unit 50 and step ST5 are necessary when the material characteristic value prediction unit 70 described later predicts the mechanical test value of the target steel material using only the second material characteristic value prediction model and when the material characteristic value prediction unit 70 predicts the mechanical test value of the target steel material using both the first material characteristic value prediction model and the second material characteristic value prediction model, but are not necessary when the material characteristic value prediction unit 70 predicts the mechanical test value of the target steel material using only the first material characteristic value prediction model. Fig. 7 is a diagram showing an example of mechanical test values ​​and mechanical test positions acquired by the material characteristic value acquisition unit 50. As described above with reference to Fig. 6, in the example shown in Fig. 7, mechanical test values ​​are measured at each of the mechanical test positions of 3m located at the top end (TOP) of the target steel material, 654m located at the middle (MID), and 1301m corresponding to the bottom end (BOT). For example, if the required lower limit value (lower limit value of required specifications) of the tensile stress (TS) of the target steel material is 980 MPa, in the example shown in FIG. 7, the mechanical test value of 975 MPa at the mechanical test position 3 m is below the required lower limit value. In this way, the steel material having a portion (failed portion) that does not satisfy the required specifications is cut at the failed portion of the mechanical test value and becomes a waste product (scrap). Conventionally, as described above, the cutting position was determined while checking a temperature chart, etc., and there were problems with its accuracy and personal nature. In contrast, in the material property prediction device 100 according to this embodiment, as described later, the material property value prediction unit 70 can accurately predict the mechanical test value of the target steel material over approximately the entire length of the target steel material, so that the steel cutting position calculation unit 80 appropriately determines the cutting position of the target steel material.

[0041] <Material characteristic value prediction model acquisition unit 60> The material characteristic value prediction model acquisition unit 60 executes step ST6 shown in FIG. 2. Specifically, the material characteristic value prediction model acquisition unit 60 acquires either the first material characteristic value prediction model or the second material characteristic value prediction model stored in the material characteristic value prediction model storage unit 30, or both the first material characteristic value prediction model and the second material characteristic value prediction model. More specifically, when the material characteristic value prediction unit 70 described later predicts the mechanical test value of the target steel material using only the first material characteristic value prediction model, the material characteristic value prediction model acquisition unit 60 acquires the first material characteristic value prediction model. When the material characteristic value prediction unit 70 predicts the mechanical test value of the target steel material using only the second material characteristic value prediction model, the material characteristic value prediction model acquisition unit 60 acquires the second material characteristic value prediction model. When the material characteristic value prediction unit 70 predicts the mechanical test values ​​of the target steel material using both the first material characteristic value prediction model and the second material characteristic value prediction model, the material characteristic value prediction model acquisition unit 60 acquires both the first material characteristic value prediction model and the second material characteristic value prediction model.

[0042] <Material characteristic value prediction unit 70> The material characteristic value prediction unit 70 executes step ST7 shown in Fig. 2. Specifically, the material characteristic value prediction unit 70 predicts mechanical test values ​​of the target steel material over substantially the entire length of the target steel material using either one of the first material characteristic value prediction model and the second material characteristic value prediction model, or both the first material characteristic value prediction model and the second material characteristic value prediction model. Below, the cases where each prediction model is used will be described in order.

[0043] [When using the first material property value prediction model] When the first material characteristic value prediction model is used, the material characteristic value prediction unit 70 inputs the operation results for each longitudinal position in the manufacturing process of the target steel material as shown in Fig. 6, acquired by the operation result acquisition unit 40, to the first material characteristic value prediction model. The pitch (e.g., 1 m) at which the operation results for each longitudinal position are obtained is much smaller than the pitch of the mechanical test positions. Then, the material property value prediction unit 70 predicts the mechanical test values ​​of the target steel material over approximately the entire length of the target steel material by outputting the probability distribution of the mechanical test values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material from the first material property value prediction model.

[0044] [When using the second material property value prediction model] When the second material characteristic value prediction model is used, the material characteristic value prediction unit 70 creates input data for the second material characteristic value prediction model based on the operational results for each longitudinal position in the manufacturing process of the target steel material acquired by the operational result acquisition unit 40, and the mechanical test values ​​measured on test pieces acquired from each mechanical test position of the target steel material acquired by the material characteristic value acquisition unit 50 and each mechanical test position. FIG. 8 is a diagram showing an example of input data to the second material characteristic value prediction model. The second material characteristic value prediction model inputs the operation results for each longitudinal position in the manufacturing process of the target steel (in the example shown in FIG. 8, every 1 m in the longitudinal direction from the tip of the target steel), the mechanical test values ​​measured on the test pieces obtained from each mechanical test position of the target steel (in the example shown in FIG. 8, positions 3 m, 654 m, and 1301 m in the longitudinal direction from the tip of the target steel), and each mechanical test position, as shown in FIG. 8. Since the mechanical test values ​​cannot be obtained at longitudinal positions over almost the entire length of the target steel, the mechanical test values ​​measured at positions close to each longitudinal position of the target steel (for example, mechanical test positions located at the same tip, center, or rear end as each longitudinal position) are assigned to each longitudinal position. Also, in the example shown in FIG. 8, instead of inputting each mechanical test position as it is, as a preferred embodiment, the relative positions are input. Specifically, for example, when the longitudinal position of the target steel is 1 m at the top (TOP), the relative position of 2 m is input when the mechanical test position of 3 m at the top (TOP) is set as the reference mechanical test position, and when the longitudinal position of the target steel is 653 m at the center (MID), the relative position of 1 m is input when the mechanical test position of 654 m at the center (MID) is set as the reference mechanical test position. Also, in the example shown in Figure 8, not only the rolling results at each longitudinal position of the target steel, but also the rolling results at the mechanical test positions corresponding to each longitudinal position are input as the rolling results (operational results). Specifically, for example, when the longitudinal position of the target steel is 4m located at the top (TOP), not only the rolling results at the 4m position (heating temperature T7 [℃], rolling load P3 [MPa / mm]) but also the rolling results at the mechanical test position 3m corresponding to the 4m position (heating temperature T4 [℃], rolling load P1 [MPa / mm]) are input as the rolling results. The same applies to other positions of the target steel. The input data to the second material characteristic value prediction model is not limited to the example shown in FIG. 8, and various modifications are possible. For example, the reference mechanical test position can be set arbitrarily, and it is also possible to use relative values ​​as the heating temperature and rolling load in the operational results, similar to the mechanical test position. In addition, it is also possible to prepare a plurality of different input data to the second material characteristic value prediction model, and to use a probability distribution obtained by synthesizing the probability distributions output when each input data is input to the second material characteristic value prediction model as the probability distribution of the mechanical test value for each longitudinal position of the final target steel material. The material property value prediction unit 70 predicts the mechanical test values ​​of the target steel material over approximately the entire length of the target steel material by outputting the probability distribution of the mechanical test values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material from the second material property value prediction model.

[0045] [When using both the first material property value prediction model and the second material property value prediction model] When the first material property value prediction model and the second material property value prediction model are used, the material property value prediction unit 70 combines the probability distribution of the mechanical test values ​​for each longitudinal position of the target steel material output from the first material property value prediction model and the probability distribution of the mechanical test values ​​for each longitudinal position of the target steel material output from the second material property value prediction model to obtain the final probability distribution of the mechanical test values ​​for each longitudinal position of the target steel material.Then, the material property value prediction unit 70 predicts the mechanical test values ​​of the target steel material over approximately the entire length of the target steel material by outputting this final probability distribution of the mechanical test values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material. 9 is a schematic diagram for explaining the synthesis of the probability distribution output from the first material characteristic value prediction model and the probability distribution output from the second material characteristic value prediction model. As shown in FIG. 9, the probability distribution output from the first material characteristic value prediction model is expressed as N(μ1, σ1 2 ), and the probability distribution output from the second material characteristic value prediction model is N(μ2, σ2 2 ), the probability distribution N(μ 1+2 ,σ 1+2 2) can be calculated by the formula (11) and the formula (12). i (i=1,2,1+2) is a parameter corresponding to loc(x) shown in the above equation (8), and σ i (i=1, 2, 1+2) is a parameter corresponding to the above-mentioned scale(x).

[0046] FIG. 10 is a diagram showing an example of the mechanical test value (tensile stress TS) of the target steel material (steel material No. 101 shown in FIG. 8) over substantially the entire length of the target steel material predicted by the material characteristic value prediction unit 70. FIG. 10(a) shows the result of prediction using the first material characteristic value prediction model. FIG. 10(b) shows the result of prediction using the second material characteristic value prediction model. FIG. 10(c) shows the result of prediction using both the first material characteristic value prediction model and the second material characteristic value prediction model. FIG. 10(d) shows the result of prediction using both the first material characteristic value prediction model and the second material characteristic value prediction model reconstructed (re-learned) by adding and using the mechanical test value measured at the cutting position (136 m) described later as known data. In FIG. 10, the data plotted with "●" are the mechanical test values ​​measured for the test piece obtained from the target steel material. As shown in FIG. 10, the prediction results are strongly influenced by the mechanical test values ​​measured on the target steel and the 95% prediction range narrows in the following order: when the first material characteristic value prediction model is used (FIG. 10(a)), when the second material characteristic value prediction model is used (FIG. 10(b)), and when both the first material characteristic value prediction model and the second material characteristic value prediction model are used (FIGS. 10(c) and 10(d)). As shown in FIG. 10(c), the teacher data for the first material characteristic value prediction model and the second material characteristic value prediction model, and the mechanical test values ​​measured at the cutting position (136 m) described below (data plotted with "x" in FIG. 10(c)), which were not used as input for the second material characteristic value prediction model when predicting the mechanical test values ​​of the target steel, are within the 95% prediction range, indicating that predictions are made with high accuracy.

[0047] <Steel material cutting position calculation unit 80> The steel material cutting position calculation unit 80 executes step ST8 shown in Fig. 2. Specifically, the steel material cutting position calculation unit 80 calculates the cutting position in the longitudinal direction of the target steel material so as to minimize the expected value of the manufacturing cost, based on the mechanical test values ​​of the target steel material over substantially the entire length of the target steel material predicted by the material characteristic value prediction unit 70. More specifically, this is as follows. If the pass rate of the mechanical test value at the longitudinal position l [m] of the target steel product is p(l), p(l) can be calculated using the following equation (13).

number

[0048] Then, if the additional process cost after cutting the target steel material (cost incurred in the event of rejection) is W1 and the price of the steel material per meter is W2, the expected manufacturing cost J(l) when cutting the steel material at the longitudinal position l [m] can be calculated using the following equation (14).

number

[0049] In this embodiment, a prediction model using a non-uniform variance Gaussian process is used as the material property value prediction model (first material property value prediction model and second material property value prediction model), but the present invention is not limited to this. As the material property value prediction model, a prediction model using a Bayesian neural network, quantile regression, or bootstrap method can also be used.

[0050] In addition, in this embodiment, no mention is made of the width direction of the steel material (the direction perpendicular to the longitudinal direction), but if the width direction of the steel material can be divided into specified sections and operational performance and mechanical test values ​​can be obtained for each section, it is also possible to predict mechanical test values ​​over almost the entire length and width of the steel material by constructing a material property value prediction model for each section. [Explanation of symbols]

[0051] 10. Manufacturing Performance Acquisition Department 20. Material characteristic value prediction model construction section 30 Material property value prediction model storage section 40. Operational Performance Acquisition Department 50 Material characteristic value acquisition unit 60 Material characteristic value prediction model acquisition unit 70...Material characteristic value prediction section 80... Steel material cutting position calculation section 100 Material property prediction device

Claims

1. A material property prediction device for predicting material property values ​​of a target steel material over substantially the entire length of the target steel material, the material property prediction device comprising: A first material characteristic value prediction model is provided, which receives an operation result for each longitudinal position in a manufacturing process of the target steel material as an input and outputs a probability distribution of a material characteristic value for each longitudinal position of the target steel material, A probability distribution of material property values ​​for each longitudinal position of the target steel material is output from the first material property value prediction model over approximately the entire length of the target steel material, thereby predicting the material property values ​​of the target steel material over approximately the entire length of the target steel material. Material property prediction device.

2. The first material characteristic value prediction model is constructed by using known data in which an operation record in a manufacturing process at a longitudinal position where a test piece was obtained from a reference steel material, which is a steel material manufactured in the past, is input, and a material characteristic value measured on the test piece obtained from the reference steel material is output. The material property prediction device according to claim 1 .

3. The first material characteristic value prediction model is reconstructed by adding and using new known data in which the operational performance at the longitudinal position where the test piece was obtained in the manufacturing process of the target steel is input, and the material characteristic value measured for the test piece obtained from the target steel is output. The material property prediction device according to claim 2 .

4. A material property prediction device for predicting material property values ​​of a target steel material over substantially the entire length of the target steel material, the material property prediction device comprising: a second material characteristic value prediction model that receives as input operational results for each longitudinal position in a manufacturing process of the target steel material, material characteristic values ​​measured on a test piece obtained from the target steel material, and the longitudinal position of the target steel material from which the test piece was obtained, and outputs a probability distribution of material characteristic values ​​for each longitudinal position of the target steel material; A probability distribution of material property values ​​for each longitudinal position of the target steel material is output from the second material property value prediction model over approximately the entire length of the target steel material, thereby predicting the material property values ​​of the target steel material over approximately the entire length of the target steel material. Material property prediction device.

5. The second material property value prediction model is constructed by using known data, which takes as inputs the operational performance in the manufacturing process at the longitudinal position where a test piece for model output was obtained from a reference steel material which is a steel material manufactured in the past, the longitudinal position where the test piece for model output was obtained, the material property values ​​measured for a standard test piece obtained from the reference steel material, the operational performance in the manufacturing process at the longitudinal position of the reference steel material where the standard test piece was obtained, and the longitudinal position where the standard test piece was obtained, and outputs the material property values ​​measured for the test piece for model output. The material property prediction device according to claim 4 .

6. The first material characteristic value prediction model according to claim 1 ; The second material characteristic value prediction model according to claim 4 or 5, a probability distribution obtained by combining the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the first material property value prediction model and the probability distribution of material property values ​​for each longitudinal position of the target steel material output from the second material property value prediction model is output as a final probability distribution of material property values ​​for each longitudinal position of the target steel material over approximately the entire length of the target steel material, thereby predicting material property values ​​of the target steel material over approximately the entire length of the target steel material; Material property prediction device.

7. Calculating a cutting position in the longitudinal direction of the target steel material so as to minimize an expected value of manufacturing cost based on the predicted material property values ​​of the target steel material over substantially the entire length of the target steel material. The material property prediction device according to any one of claims 1 to 5.

8. Calculating a cutting position in the longitudinal direction of the target steel material so as to minimize an expected value of manufacturing cost based on the predicted material property values ​​of the target steel material over substantially the entire length of the target steel material. The material property prediction device according to claim 7 .

9. A material property prediction method for predicting material property values ​​of a target steel material over substantially the entire length of the target steel material, the method comprising: A first material characteristic value prediction model is used, which inputs operational results for each longitudinal position in the manufacturing process of the target steel material and outputs a probability distribution of material characteristic values ​​for each longitudinal position of the target steel material, A probability distribution of material property values ​​for each longitudinal position of the target steel material is output from the first material property value prediction model over approximately the entire length of the target steel material, thereby predicting the material property values ​​of the target steel material over approximately the entire length of the target steel material. Materials property prediction methods.

10. A program for causing a computer to function as the material property prediction device according to any one of claims 1 to 5.

Images