Quality prediction model generation method, metal material quality prediction method, metal material manufacturing method, metal material manufacturing condition presentation method, quality prediction model generation device, metal material quality prediction device, metal material manufacturing condition presentation device, and metal material manufacturing system

Abstract

This quality prediction model generation method includes: an acquisition step (S1) for acquiring an explanatory variable selected from a manufacturing condition of each process, and an objective variable which is a state of a quality defect of a manufactured metal material; a storage step (S3) for storing the explanatory variable and the objective variable in association with each other as training data; calculation steps (S4 to S6) for dividing the training data into groups to inspect whether there is a significant difference in the state of the quality defect; a retrieval step (S7) for retrieving the grouping having the highest significance; and a generation step (S8) for generating a quality prediction model by machine learning with a group conforming to the retrieved grouping having the highest significance.

Description

Quality prediction model generation method, quality prediction method for metallic materials, manufacturing method for metallic materials, manufacturing condition presentation method for metallic materials, quality prediction model generation device, quality prediction device for metallic materials, manufacturing condition presentation device for metallic materials, and manufacturing system for metallic materials

[0001] The present disclosure relates to a quality prediction model generation method, a quality prediction method for metallic materials, a manufacturing method for metallic materials, a method for presenting manufacturing conditions for metallic materials, a quality prediction model generation device, a quality prediction device for metallic materials, a manufacturing condition presentation device for metallic materials, and a manufacturing system for metallic materials.

[0002] A known method for predicting quality for any required condition is to calculate the distance between a plurality of past observation conditions stored in a performance database and the desired required condition, and predict quality using the calculated distance. For example, Patent Document 1 discloses a method for calculating weights of observation data (performance data) from the calculated distance, creating a function that fits the neighborhood of the required condition from the calculated weight, and predicting quality for the required condition using the created function.

[0003] Patent No. 7207547

[0004] The quality prediction method of Patent Document 1 associates the manufacturing conditions of each process with the quality of the metal material manufactured under those manufacturing conditions for each predetermined range, accumulates actual data, and generates a quality prediction model that predicts the quality of the metal material for each predetermined range. The predetermined range is identified taking into account, for example, the cutting position of the metal material. Here, while the quality prediction method of Patent Document 1 can predict the quality for any manufacturing condition with high accuracy, further improvement in prediction accuracy is desired. For example, further improvement in prediction accuracy is expected by objectively classifying the data to correspond to changes in the quality of the metal material and generating a quality prediction model for each classification.

[0005] In view of the above circumstances, the purpose of the present disclosure is to provide a quality prediction model generation method, a quality prediction method for metal materials, a manufacturing method for metal materials, a quality prediction model generation device, and a quality prediction device for metal materials that enable highly accurate prediction of the quality of metal materials.

[0006] (1) A quality prediction model generation method according to one embodiment of the present disclosure is a quality prediction model generation method for generating a quality prediction model for a metal material manufactured through one or more processes, and includes: an acquisition step for acquiring explanatory variables selected from manufacturing conditions for each process and a dependent variable that is a state of quality defects in the manufactured metal material; a storage step for associating the explanatory variables with the dependent variable and storing them as learning data; a calculation step for dividing the learning data into groups and testing whether there is a significant difference in the state of the quality defects; a search step for searching for the most significant grouping; and a generation step for generating the quality prediction model by machine learning using a group according to the most significant grouping found.

[0007] (2) As one embodiment of the present disclosure, in (1), the explanatory variables are acquired as numerical variables or categorical variables.

[0008] (3) As one embodiment of the present disclosure, in (2), when the explanatory variables are acquired as categorical variables, the calculation step performs a numerical transformation to classify the categorical variables by assigning different binary numerical values, and calculates an index to evaluate the significant difference based on the assigned numerical values ​​and the state of the quality defect.

[0009] (4) As an embodiment of the present disclosure, in (3), the calculation step calculates the index for each of a plurality of groupings.

[0010] (5) As one embodiment of the present disclosure, in any one of (1) to (4), the calculation step, after the search step is performed, further grouping and the test are performed on one of the groups according to the most significant grouping found.

[0011] (6) As an embodiment of the present disclosure, in any one of (1) to (5), the machine learning techniques include at least linear regression, local regression, principal component regression, PLS regression, neural network, regression tree, random forest, LightGBM, and XGBoost.

[0012] (7) As an embodiment of the present disclosure, in any one of (1) to (6), the quality prediction model is regenerated when the prediction accuracy of the quality prediction model falls outside a predetermined range or at predetermined intervals.

[0013] (8) A quality prediction method for a metal material according to one embodiment of the present disclosure predicts the state of quality defects in the metal material using a quality prediction model generated by any of the quality prediction model generation methods (1) to (7).

[0014] (9) A method for manufacturing a metal material according to one embodiment of the present disclosure predicts the state of quality defects in the metal material using the metal material quality prediction method of (8), and if it is predicted that the metal material has a quality defect, changes at least a portion of the manufacturing conditions.

[0015] (10) A method for presenting manufacturing conditions for a metal material according to one embodiment of the present disclosure predicts the state of a quality defect in the metal material using the quality prediction method for a metal material of (8), and if a quality defect in the metal material is predicted, searches for and presents at least a partial change in the manufacturing conditions that will eliminate the quality defect.

[0016] (11) A method for manufacturing a metallic material according to one embodiment of the present disclosure manufactures a metallic material based on the changes in manufacturing conditions presented by the method for presenting manufacturing conditions for a metallic material according to (10).

[0017] (12) A quality prediction model generation device according to one embodiment of the present disclosure is a quality prediction model generation device that generates a quality prediction model for a metal material manufactured through one or more processes, and includes: an acquisition unit that acquires explanatory variables selected from manufacturing conditions for each process and a dependent variable that is a state of quality defects in the manufactured metal material; a storage unit that associates the explanatory variables with the dependent variable and stores them as learning data; a calculation unit that divides the learning data into groups and tests whether there is a significant difference in the state of the quality defects; a search unit that searches for the most significant grouping; and a generation unit that generates the quality prediction model by machine learning using a group according to the most significant grouping found.

[0018] (13) A quality prediction device for a metallic material according to one embodiment of the present disclosure predicts the state of quality defects in the metallic material using a quality prediction model generated by the quality prediction model generation device of (12).

[0019] (14) A manufacturing condition presentation device for metal materials according to one embodiment of the present disclosure presents changes to at least some of the manufacturing conditions for metal materials based on the presence or absence of quality defects in the metal material predicted by the quality prediction device for metal materials of (13).

[0020] (15) A metal material manufacturing system according to one embodiment of the present disclosure controls manufacturing equipment for each process of manufacturing metal material in accordance with the changes in manufacturing conditions output from the output unit of the metal material manufacturing condition presentation device of (14).

[0021] According to the present disclosure, it is possible to provide a quality prediction model generation method, a quality prediction method for metal materials, a manufacturing method for metal materials, a quality prediction model generation device, and a quality prediction device for metal materials that enable highly accurate prediction of the quality of metal materials.

[0022] Fig. 1 is a block diagram showing an example configuration of a quality prediction model generation device according to an embodiment of the present disclosure. Fig. 2 is a flowchart showing processing of a quality prediction model generation method according to an embodiment of the present disclosure. Fig. 3 is a diagram for explaining grouping. Fig. 4 is a diagram showing combinations of two groups and the presence or absence of quality defects. Fig. 5 is a diagram illustrating a slab casting process.

[0023] Hereinafter, a quality prediction model generation method, a quality prediction method for a metallic material, a manufacturing method for a metallic material, a method for presenting manufacturing conditions for a metallic material, a quality prediction model generation device 10 (see FIG. 1 ), a quality prediction device for a metallic material, a manufacturing condition presentation device for a metallic material, and a manufacturing system for a metallic material according to an embodiment of the present disclosure will be described with reference to the drawings. The quality prediction model generation method and the quality prediction model generation device 10 according to the present embodiment generally acquire manufacturing conditions and quality data for each process of manufacturing a metallic material, test the acquired data, and divide the data into groups with significant differences in quality. The quality prediction model generation method and the quality prediction model generation device 10 then generate a quality prediction model that predicts the quality for each group. The quality prediction method for a metallic material according to the present embodiment predicts the state of quality defects in the metallic material using the quality prediction model generated by the quality prediction model generation method or the quality prediction model generation device 10 according to the present embodiment. The state of quality defects is classified into a first state and a second state. For example, the first state may correspond to "present," "a predetermined number or more," or "distributed over a predetermined range," while the second state may correspond to "absent," "less than a predetermined number," or "distributed below a predetermined range," etc. In the following description, the first state will be described as "present" and the second state as "absent," but the "presence or absence" of a quality defect is an example of the "state" of the quality defect, and "presence or absence" can be replaced with "state." The manufacturing method for a metallic material according to this embodiment is a method for manufacturing a metallic material, in which the presence or absence of a quality defect in the metallic material is predicted by the metallic material quality prediction method according to this embodiment, and at least some of the manufacturing conditions are changed if a quality defect in the metallic material is predicted. The metallic material quality prediction method according to this embodiment predicts the presence or absence of a quality defect in the metallic material using a quality prediction model generated by the quality prediction model generation method or quality prediction model generation device 10 according to this embodiment. Furthermore, the metallic material manufacturing condition suggestion method according to this embodiment suggests changes to at least some of the manufacturing conditions for the metallic material based on the result of the presence or absence of a quality defect in the metallic material predicted by the metallic material quality prediction method or quality prediction device according to this embodiment.

[0024] FIG. 1 is a block diagram showing an example configuration of a quality prediction model generation device 10 according to this embodiment. The quality prediction model generation device 10 is used in a manufacturing process for producing metallic materials. The metallic materials may be, for example, steel products, including semi-finished products such as slabs and finished products such as steel plates produced by rolling slabs. The quality prediction model generation device 10 generates a quality prediction model for metallic materials manufactured through one or more processes. The quality prediction model generation device 10 executes a quality prediction model generation method (see FIG. 2). In the example of FIG. 1, the quality prediction model generation device 10 also functions as a metallic material quality prediction device that uses the generated quality prediction model to predict the presence or absence of quality defects in the metallic material. The metallic material quality prediction device executes the metallic material quality prediction method (see FIG. 2). Here, the quality prediction model generation device 10 may only have the function of generating a quality prediction model. In this case, the metallic material quality prediction device may be configured as a device separate from the quality prediction model generation device 10 that can acquire the quality prediction model from the quality prediction model generation device 10. 1 , the quality prediction model generation device 10 also functions as a manufacturing condition presentation device for metallic materials that presents changes to at least some of the manufacturing conditions for metallic materials based on the presence or absence of quality defects in the metallic material predicted by the prediction method for metallic materials. The manufacturing condition presentation device for metallic materials executes the manufacturing condition presentation method for metallic materials. The manufacturing condition presentation device may be configured as a device separate from the quality prediction model generation device 10.

[0025] As shown in FIG. 1 , the quality prediction model generation device 10 includes a communication unit 11, a storage unit 12, and a control unit 13. The control unit 13 includes an acquisition unit 131, a storage unit 132, a calculation unit 133, a search unit 134, a generation unit 135, a prediction unit 136, a manufacturing condition presentation unit 137, and an output unit 138. The quality prediction model generation device 10 may be realized as a hardware configuration using a general-purpose computer (information processing device) such as a personal computer or a workstation. The computer includes a processor such as a CPU (Central Processing Unit). The computer also includes, as main components, a memory such as a RAM (Random Access Memory) and a ROM (Read Only Memory). Details of the components of the quality prediction model generation device 10 will be described later.

[0026] The quality prediction model generation device 10 acquires quality data and operation data from the operation data server 20. The operation data includes manufacturing conditions for each process of manufacturing a metal material. In the example of FIG. 1, the quality prediction model generation device 10 acquires quality data and operation data stored in a performance database provided in the operation data server 20. The operation data server 20 is capable of communicating with the quality prediction model generation device 10 via a network and may be realized, for example, by a computer that manages the manufacturing process. The network is, for example, the Internet. In this embodiment, the display unit 30 displays the prediction results output from the quality prediction device for metal materials (integrated with the quality prediction model generation device 10 in the example of FIG. 1).

[0027] Here, the manufacturing conditions stored in the performance database may include at least one of the components, temperature, pressure, plate thickness, and plate threading speed of the metal material in each process. Furthermore, the quality data stored in the performance database may include at least one of material properties such as tensile strength, the presence or absence of surface defects in the material, the presence or absence of internal defects in the material, the number of defects, and the location of the defects. The manufacturing conditions and quality data are obtained as operational performance data, operational target values, analytical values, or quality judgment results from, for example, sensors, manufacturing equipment, control devices, or quality control devices installed in a factory. For example, the presence or absence of internal defects in a steel plate (an example of a metal material) may be obtained as quality data using an ultrasonic flaw detector (an example of a sensor and a quality control device). The performance database is a database in which data on the quality of the manufactured metal material is associated with the manufacturing conditions, operational performance data, various analytical values, and the like for each corresponding process.

[0028] Here, although the quality prediction model generation device 10 is a separate computer from the operation data server 20 in the example of FIG. 1 , it may be configured as the same computer as the operation data server 20. The display unit 30 may be a device through which an operator working in the manufacturing process can obtain information. The display unit 30 may also be realized by the display of a terminal device such as a smartphone or tablet. When a quality defect in the metal material is predicted, a manufacturing method for the metal material may be executed in which at least some of the manufacturing conditions are changed based on instructions from the operator. The display unit 30 may be a display device such as a liquid crystal display (LCD) or an organic electroluminescence panel (OLED).

[0029] The components of the quality prediction model generation device 10 will be described in detail below. The communication unit 11 includes one or more communication modules connected to a network. The communication unit 11 may include a communication module compatible with mobile communication standards such as 4G (4th Generation) and 5G (5th Generation). The communication unit 11 may include a communication module compatible with a wired or wireless LAN standard.

[0030] The storage unit 12 is one or more memories. The memory may be, for example, a semiconductor memory, a magnetic memory, an optical memory, or the like, but is not limited to these, and may be any memory. The storage unit 12 is, for example, built into the quality prediction model generation device 10, but may also be configured to be externally accessed by the quality prediction model generation device 10 via any interface.

[0031] The storage unit 12 stores various data used in various calculations performed by the control unit 13. The storage unit 12 may also store results and intermediate data of various calculations performed by the control unit 13. In this embodiment, the storage unit 12 stores learning data and generated models.

[0032] The control unit 13 is one or more processors. The processor may be, for example, a general-purpose processor or a dedicated processor specialized for a specific process, but is not limited to these and may be any processor. The control unit 13 controls the overall operation of the quality prediction model generation device 10.

[0033] Here, the quality prediction model generation device 10 may have the following software configuration: One or more programs used to control the operation of the quality prediction model generation device 10 are stored in the storage unit 12. When the programs stored in the storage unit 12 are read by the processor of the control unit 13, they cause the control unit 13 to function as an acquisition unit 131, a storage unit 132, a calculation unit 133, a search unit 134, a generation unit 135, and a prediction unit 136.

[0034] The acquisition unit 131 acquires manufacturing conditions and quality data from a performance database. The performance database stores a large amount of quality data, such as time-series data, including manufacturing conditions for each process and measurements of the presence or absence of quality defects in the metal material manufactured under those manufacturing conditions. The acquisition unit 131 may selectively acquire explanatory variables and objective variables that can be used as learning data for machine learning, which will be described later, from the performance database. In this embodiment, the acquisition unit 131 acquires explanatory variables selected from the manufacturing conditions for each process and an objective variable that indicates the presence or absence of quality defects in the manufactured metal material.

[0035] The storage unit 132 associates the explanatory variables, which are data acquired by the acquisition unit 131, with the objective variables and stores them as training data. That is, the storage unit 132 associates (links) the manufacturing conditions for each process acquired as the explanatory variables with quality data, including measurements of the presence or absence of quality defects in the metal material manufactured under those manufacturing conditions, to create the training data. The training data may be stored in, for example, the memory unit 12. If the manufacturing conditions for each process are acquired over a long period of time or at high frequency and represent a large amount of data, the storage unit 132 may thin out or convert the data into a representative value, such as the average or median, to reduce the data volume to an appropriate amount before storing the training data. Furthermore, as shown in FIG. 3 , the storage unit 132 may associate the "manufacturing conditions for each process and the presence or absence of quality defects (data group)" for each predetermined range of the metal material. In the example of FIG. 3 , the training data is associated for each predetermined range from the beginning of the metal material transport direction and divided into data group 1, data group 2, data group 3, data group 4, ..., and data group N. Here, N is an integer equal to or greater than 2. Furthermore, the predetermined range may be determined based on the cutting position when, for example, a metal material is cut during manufacturing, but is not limited to this determination method.

[0036] The calculation unit 133 divides the learning data into groups and performs a test (significant difference test) to determine whether there is a significant difference in the presence or absence of quality defects. For example, the calculation unit 133 divides the data groups 1 to N into group X and group Y, as shown in patterns 1 to M in FIG. 3 . Here, M is an integer equal to or greater than 2. The calculation unit 133 performs multiple groupings and performs a significant difference test for each of the multiple groupings. The calculation unit 133 may perform comprehensive grouping for all combinations, or may perform grouping for only some combinations. For example, a constraint such as an upper limit on calculation time may be imposed, and the calculation unit 133 may perform grouping for combinations that can be calculated within the constraint. Furthermore, a priority order for grouping may be determined in advance based on manufacturing knowledge, etc.

[0037] In this embodiment, the calculation unit 133 uses Fisher's exact test. However, the method of the significant difference test is not limited. In this embodiment, the calculation unit 133 uses a P-value as an index for evaluating a significant difference. The P-value is a value indicating the probability of a combination resulting in a difference greater than or equal to the observed difference among all possible combinations. Therefore, the smaller the P-value, the higher the significance (the higher the probability of a significant difference). As shown in Figure 4, suppose the numbers corresponding to the combinations of two groups and the presence or absence of quality defects are a, b, c, and d. In this case, the P-value is calculated using the following formula: n is the total number of learning data.

[0038]

[0039] Here, the grouping is performed based on explanatory variables. The explanatory variables are acquired as numerical variables or categorical variables. For example, when the explanatory variables used for grouping are acquired as numerical variables, the calculation unit 133 may classify the group as group X if the explanatory variables are equal to or greater than a threshold, and as group Y if the explanatory variables are less than the threshold.

[0040] For example, when an explanatory variable used for grouping is acquired as a categorical variable, the calculation unit 133 may perform a numerical conversion to classify the categorical variable by assigning different binary numerical values. The different binary numerical values ​​may be, for example, 0 or 1, and in the present embodiment, a description will be given assuming that 0 or 1 is used. Furthermore, one-hot encoding may be used, for example, as the numerical conversion, but is not limited to this. Then, the calculation unit 133 may calculate an index (P value) for evaluating a significant difference based on the assigned 0 or 1 and the presence or absence of a quality defect. That is, the calculation unit 133 may classify a group as group X if the explanatory variable after the numerical conversion is 1 (i.e., if it belongs to a certain category), and may classify a group as group Y if it is 0 (i.e., if it does not belong to a certain category).

[0041] Here, the explanatory variables are selected from the manufacturing conditions of each process. The manufacturing conditions include many variables such as the composition of the metal material and the temperature in various processes. Therefore, there are various ways to select explanatory variables to be used for grouping. For example, if n explanatory variables are used for grouping, then 2 ｎThe P-values ​​may be calculated for all of the patterns, or for some of the patterns. In addition, candidates for explanatory variables to be used for grouping may be determined in advance based on manufacturing knowledge.

[0042] The search unit 134 searches for the most significant grouping. In this embodiment, the search unit 134 identifies the grouping with the lowest P value. In the example of FIG. 3 , the calculation unit 133 calculates a P value for each of patterns 1 to M, and the search unit 134 identifies the pattern with the lowest P value (e.g., pattern 2) as the most significant grouping. In this example, the search unit 134 determines that group X, which includes data groups 1 and 2, and group Y, which includes data groups 3 to N, have the most significant difference.

[0043] Here, the calculation unit 133 and the search unit 134 may perform multi-stage grouping starting with the group with the lowest P value. That is, after the search unit 134 searches for the most significant grouping, the calculation unit 133 may further group one of the groups based on that grouping and perform a significance test (calculate a P value). In the example of FIG. 3 , after the search unit 134 identifies pattern 2 as the most significant grouping, the calculation unit 133 may further group grouping and calculate a P value for group Y including data groups 3 to N. Then, the search unit 134 may search for the most significant grouping for group Y and determine, for example, that group Y1 including data groups 3 to i and group Y2 including data groups i to N have the most significant difference. Here, i is an integer greater than or equal to 3. The number of times (repetitions) the multi-stage grouping is performed is not limited, but it is necessary that the number of training data included in each group not be too small so as not to reduce the accuracy of the generated quality prediction model.

[0044] Here, for multiple explanatory variables, those with a P-value equal to or less than a predetermined value (e.g., 1 / n where n is the total number of learning data) may be determined to be important variables, and grouping may be performed using the important variables. Furthermore, of the important variables, only one that is easy to explain the relationship with product quality may be selected, or two or more important variables may be selected. Furthermore, when the accuracy of predictions using a quality prediction model generated in the past is obtained, the important variables may be selected taking that accuracy into consideration.

[0045] The generation unit 135 generates a quality prediction model by machine learning using a group according to the most significant grouping found by the search unit 134. The machine learning method is not particularly limited, and may be, for example, linear regression, local regression, principal component regression, PLS regression, neural network, regression tree, random forest, LightGBM, or XGBoost. In other words, the machine learning methods that the generation unit 135 can use include at least linear regression, local regression, principal component regression, PLS regression, neural network, regression tree, random forest, LightGBM, and XGBoost. For example, in the above example in which the calculation unit 133 and the search unit 134 perform multi-stage grouping, the generation unit 135 generates three quality prediction models using the learning data for group X, group Y1, and group Y2. The generation unit 135 stores the generated quality prediction models in, for example, the storage unit 12.

[0046] The prediction unit 136 uses the generated quality prediction model to predict whether or not there will be a quality defect in a metal material manufactured under any manufacturing conditions. The prediction result is displayed, for example, on the display unit 30. If a quality defect is predicted in the metal material, at least a portion of the manufacturing conditions is changed based on instructions from the operator, thereby improving the quality of the manufactured metal material.

[0047] In order to maintain the prediction accuracy of the presence or absence of quality defects in the metallic material, the prediction unit 136 may output a command to the calculation unit 133, the search unit 134, and the generation unit 135 to regenerate the quality prediction model for the metallic material. That is, the prediction unit 136 may cause the calculation unit 133 to regroup the learning data and retest whether there is a significant difference in the state of quality defects. The prediction unit 136 may output a command to the search unit 134 to re-search for the most advantageous grouping and to the generation unit 135 to re-generate the quality prediction model by machine learning using the group according to the most advantageous grouping found. The prediction unit 136 may output a command to regenerate the quality prediction model for the metallic material, for example, when the prediction accuracy of the presence or absence of quality defects in the metallic material falls outside a predetermined range (is not within the predetermined range) or at predetermined intervals (e.g., every three months). The quality prediction model may then be regenerated based on the command.

[0048] Based on the presence or absence of quality defects in the metal material predicted by the prediction unit 136, the manufacturing condition presentation unit 137 searches for at least some of the manufacturing conditions for each process that will eliminate the quality defects in the prediction result, and presents the changes to the found manufacturing conditions via the display unit 30. The manufacturing condition presentation unit 137 changes at least some of the operable manufacturing conditions for each process from their current values ​​to different values, while leaving the current values ​​for the other manufacturing conditions unchanged, and inputs these as explanatory variables of the quality prediction model generated by the generation unit 135. The manufacturing condition presentation unit 137 then predicts the presence or absence of quality defects in the metal material. If the prediction result is "absence of quality defects," the changed manufacturing conditions are presented via the display unit 30. If the prediction result is "presence of quality defects," the values ​​of the operable manufacturing conditions for each process are further changed to different values ​​until the prediction result of the quality prediction model becomes "absence of quality defects." For example, an optimization method such as an iterative method may be used. The manufacturing condition presenting unit 137 may use the quality prediction model generated by the generation unit 135 through machine learning to calculate variable importance (feature importance) for a target variable, which is the presence or absence of quality defects in the manufactured metal material. Then, based on the variable importance, the manufacturing condition presenting unit 137 may select, as a manufacturing condition to be changed, a manufacturing condition that has a high contribution to the target variable from among the manufacturing conditions of each process, which are explanatory variables. The manufacturing condition presenting unit 137 may calculate the variable importance (feature importance) by applying a mathematical method such as SHAP (Shapley Additive Explanations), Gini impurity, frequency, or gain (the amount of reduction in the objective function due to the use or non-use of the feature).

[0049] The quality prediction model generating device 10 may also include an output unit 138 that outputs (transmits) via a network changes to the manufacturing conditions presented by the manufacturing condition presenting unit 137 to a control computer that controls manufacturing equipment for each process of manufacturing the metal material. The control computer controls the manufacturing equipment for each process of manufacturing the metal material in accordance with the transmitted changes to the manufacturing conditions, and the quality prediction model generating device 10 and the control computer may constitute a manufacturing system for the metal material. The control computer may also accept modifications to the manufacturing conditions by an operator and control the manufacturing equipment for each process.

[0050] The output unit 138 may display the changes in the manufacturing conditions presented by the manufacturing condition presenting unit 137 on a display device (for example, a liquid crystal display) connected via a network.

[0051] FIG. 2 is a flowchart showing the processing of the quality prediction model generation method executed by the quality prediction model generation device 10 according to this embodiment. In this embodiment, the quality prediction model generation device 10 also functions as a quality prediction device for metallic materials, predicting the presence or absence of quality defects in the metallic material. FIG. 2 is also a flowchart showing the processing of the quality prediction method for metallic materials. In this embodiment, the quality prediction model generation device 10 also functions as a manufacturing condition presentation device for metallic materials, presenting at least some changes to the manufacturing conditions for the metallic material. FIG. 2 is also a flowchart showing the processing of the manufacturing condition presentation method for metallic materials. In this embodiment, the quality prediction model generation device 10, together with a control computer that controls the manufacturing equipment that produces the metallic material, constitutes a metallic material manufacturing system. FIG. 2 is also a flowchart showing the processing of the manufacturing method for metallic materials. Here, the prediction of the presence or absence of quality defects in the metallic material (step S9), regeneration of the quality prediction model (step S10), presentation of manufacturing conditions for the metallic material (step S11), and manufacturing of the metallic material (step S12) may be performed consecutively with the generation of the quality prediction model (step S8). Furthermore, steps S9 to S12 may be performed a certain time after the generation of the quality prediction model.

[0052] The acquiring unit 131 acquires manufacturing conditions and quality data from the performance database (step S1). Step S1 corresponds to an acquiring step.

[0053] The storage unit 132 eliminates multicollinearity from the data acquired by the acquisition unit 131 (step S2), associates the explanatory variables with the target variables, and stores the data as training data (step S3). A known method may be used to eliminate multicollinearity. Step S2 may be omitted. Step S3 corresponds to a storage step.

[0054] If the explanatory variables used for grouping are acquired as categorical variables (Yes in step S4), the calculation unit 133 performs a numerical conversion such as one-hot encoding (step S5). If the explanatory variables used for grouping are acquired as numerical variables (No in step S4), the calculation unit 133 proceeds to the processing of step S6. The calculation unit 133 divides the learning data into groups and performs a significance test (step S6). Steps S4 to S6 correspond to calculation steps.

[0055] The search unit 134 searches for the most significant grouping (step S7). Step S7 corresponds to a search step.

[0056] The generating unit 135 generates a quality prediction model using groups according to the grouping found (step S8). Step S8 corresponds to the generating step.

[0057] The prediction unit 136 uses the generated quality prediction model to predict the presence or absence of quality defects in the metal material (step S9).

[0058] The prediction unit 136 may output a command to regenerate the quality prediction model for the metallic material in order to maintain the prediction accuracy of the presence or absence of quality defects in the metallic material (step S10). The prediction unit 136 may output a command to regenerate the quality prediction model for the metallic material, for example, when the prediction accuracy of the presence or absence of quality defects in the metallic material falls outside a predetermined range (is not within the predetermined range), or at predetermined intervals. If the quality prediction model for the metallic material is to be regenerated (Yes in step S10), the process returns to step S6. If the quality prediction model for the metallic material is not to be regenerated (No in step S10), the process proceeds to step S11.

[0059] Based on the predicted presence or absence of quality defects in the metallic material, the manufacturing condition presentation unit 137 searches for at least some of the manufacturing conditions for each process that will result in a prediction result of "no quality defects" and presents the found changes to the manufacturing conditions (step 11). The found changes to the manufacturing conditions are transmitted as a manufacturing condition change command via the output unit 138 to the control computer that controls the manufacturing equipment in each process of manufacturing the metallic material, and are displayed on the display unit 30.

[0060] Based on the searched changes in the manufacturing conditions, the metal material is manufactured (step S12). The changes in the manufacturing conditions are transmitted to the control computer, and the metal material is manufactured in accordance with the transmitted changes in the manufacturing conditions. Based on the display on the display unit 30, the operator may set the changes in the manufacturing conditions in the control computer.

[0061] The effects of the present disclosure will be specifically described below based on examples, but the present disclosure is not limited to these examples.

[0062] In this example, the quality prediction method for metallic materials according to the above embodiment was applied to predicting the presence or absence of defects in a structural steel plate. The structural steel plate in this example was manufactured from a slab, which is an intermediate product manufactured in a casting process (continuous casting process) shown in Fig. 5.

[0063] In the casting process shown in Figure 5, molten steel in a ladle is poured into a mold via a tundish. The molten steel in the mold undergoes primary cooling in the mold, solidifying from the surface to form a slab. The slab is then pulled out of the mold and undergoes secondary cooling with cooling water below the mold, finally solidifying completely. The slab is then cut into slabs at predetermined lengths by a cutter. The slabs are rolled in the next process, and are then processed into products. Here, the slabs produced are numbered from 1 to N in the cutting order from top to bottom for each slab in the slab-cutting order.

[0064] The objective variable (quality) for quality prediction in this example is the presence or absence of defects determined by ultrasonic testing of products (structural steel plates, steel for pressure vessels, steel for marine structures). The explanatory variables (manufacturing conditions) include the chemical composition of the metallic material in the smelting process. The explanatory variables also include the temperature of the metallic material in the casting process, the casting speed, the cooling conditions, and the slab cutting order. The explanatory variables also include the temperature of the metallic material in the heating process, the temperature of the metallic material in the hot rolling process, and the temperature of the metallic material in the cooling process. The explanatory variables also include the temperature of the metallic material in the cold rolling process and the cutting dimensions.

[0065] As in the quality prediction model generation method according to the above embodiment, data was acquired from a performance database and a quality prediction model, which is a machine learning model, was generated. LightGBM (Light Gradient Boosting Machine) was used as the machine learning method. Fisher's exact test was used as the significant difference test method. An arbitrary division point for a continuous variable was determined by dividing the difference between the maximum and minimum values ​​of the variable ("maximum minus minimum") by 25 to find the smallest P-value, and the division point corresponding to the smallest P-value was adopted as the division point for that variable.

[0066] Furthermore, the slab cutting order is a type of categorical variable. Data corresponding to cutting order 1 was categorized as data group 1, data corresponding to cutting orders 2 to N-1 were categorized as data groups 2 to N-1, and data corresponding to the last cutting order (N) was categorized as data group N. Each data group was classified by binarization using one-hot encoding. In this example, it was determined that the most significant difference was found between the first group including data groups 1 and 2, the second group including data groups N and N-1, and the third group including the other data groups.

[0067] A quality prediction model was generated for each of these groups using LightGBM. Furthermore, the presence or absence of quality defects was predicted using these quality prediction models. Here, the number of samples in the performance database was 307, and the number of explanatory variables was 260. As a result of the prediction, the accuracy rate of "quality defects present" using this method was 89%. The accuracy rate of "quality defects present" using the conventional method was 68%, confirming a significant increase in the accuracy rate. Here, the conventional method generates one quality prediction model using all of the training data without performing grouping using a significance test.

[0068] As described above, the quality prediction model generation method, metallic material quality prediction method, metallic material manufacturing method, quality prediction model generation device 10, metallic material quality prediction device, and metallic material manufacturing system according to the present embodiment perform appropriate grouping based on significant differences when generating a quality prediction model. Therefore, by using the generated quality prediction model, it is possible to make highly accurate predictions about the quality of metallic materials.

[0069] In the model generation of this embodiment, the operating conditions during each manufacturing process of a product manufactured in the past and the quality of the product manufactured under those operating conditions are acquired as performance data, and the corresponding operating condition data and quality data are linked and stored. Then, multiple explanatory variables are grouped, and the grouping with the highest significance for the quality data is identified. This type of grouping has not been performed in the past. The method disclosed herein enables objectively meaningful grouping even for workers who do not have deep knowledge or sufficient experience of the manufacturing process. The grouping according to the method disclosed herein uses significance testing to classify the manufacturing conditions and quality data to correspond to changes in the quality of the metal material (to reflect factors contributing to quality). Therefore, a quality prediction model generated according to such grouping enables highly accurate predictions.

[0070] Furthermore, the method and device for proposing manufacturing conditions for metallic materials according to this embodiment search for and present manufacturing conditions that improve internal defects (defects detected by ultrasonic flaw detection) in structural steel plates. In this example, a quality prediction model generated using the LightGBM was used. Among the manipulable manufacturing conditions in the refining, casting, heating, hot rolling, cooling, and cold rolling processes, the heating temperature of the heating process, which has a large gain in LightGBM and contributes greatly to the presence or absence of internal defects (the target variable), was selected as the condition to be changed. Using the quality prediction model, the presence or absence of quality defects in the metallic material was predicted, and a heating temperature that resulted in a prediction of "no quality defects" was searched for using an iterative method (golden section method) within the range from the maximum to the minimum heating temperature values ​​obtained from the operation data server 20. The searched heating temperature change for the heating process was then input into the control computer, and metallic materials were produced. As a result, the heating temperature increased by an average of +30°C, and the occurrence rate of quality defects was 9%. Before applying the method for presenting manufacturing conditions for metal materials according to this embodiment (conventional), the rate of occurrence of quality defects was 10%, and it was confirmed that the rate of occurrence of quality defects was improved.

[0071] Although embodiments of the present disclosure have been described based on the drawings and examples, it should be noted that those skilled in the art would easily be able to make various modifications or alterations based on the present disclosure. Therefore, it should be noted that these modifications and alterations are included within the scope of the present disclosure. For example, functions included in each component or step can be rearranged so as not to cause logical inconsistencies, and multiple components or steps can be combined or divided into one. The embodiments of the present disclosure can also be realized as a program executed by a processor included in an apparatus or a storage medium on which a program is recorded. It should be understood that these are also included within the scope of the present disclosure.

[0072] Here, as described above, the quality prediction model generating device 10 does not need to have the function of a quality prediction device for metallic materials. Specifically, the quality prediction model generating device 10 may have only the function of generating a quality prediction model, and the prediction unit 136 may be included in a quality prediction device for metallic materials that is a computer (information processing device) different from the quality prediction model generating device 10. In this case, the quality prediction device for metallic materials obtains, via a network, the model generated by the quality prediction model generating device 10 and stored in the memory unit 12, and predicts the presence or absence of quality defects in the metallic material.

[0073] As another example, the display unit 30 may be included in the quality prediction model generation device 10 .

[0074] REFERENCE SIGNS LIST 10 Quality prediction model generating device (quality prediction device for metal materials) 11 Communication unit 12 Storage unit 13 Control unit 20 Operation data server 30 Display unit 131 Acquisition unit 132 Storage unit 133 Calculation unit 134 Search unit 135 Generation unit 136 Prediction unit 137 Manufacturing condition presentation unit 138 Output unit

Claims

1. A quality prediction model generation method for generating a quality prediction model of a metal material manufactured through one or more processes, comprising: an acquisition step for acquiring explanatory variables selected from manufacturing conditions for each process and a dependent variable which is a state of quality defects in the manufactured metal material; a storage step for associating the explanatory variables and the dependent variable and storing them as learning data; a calculation step for dividing the learning data into groups and testing whether there is a significant difference in the state of the quality defects; a search step for searching for the most significant grouping; and a generation step for generating the quality prediction model by machine learning using a group according to the most significant grouping found.

2. The quality prediction model generating method according to claim 1, wherein the explanatory variables are obtained as numerical variables or categorical variables.

3. The quality prediction model generating method according to claim 2, wherein the calculation step, when the explanatory variables are acquired as categorical variables, performs a numerical conversion to classify the categorical variables by assigning different binary numerical values, and calculates an index to evaluate the significant difference based on the assigned numerical values ​​and the state of the quality defect.

4. The quality prediction model generating method according to claim 3, wherein said calculation step calculates said index for each of a plurality of groupings.

5. A quality prediction model generating method as described in any one of claims 1 to 4, wherein the calculation step, after the search step is performed, further grouping and the test are performed on one of the groups according to the most significant grouping found.

6. A quality prediction model generation method according to any one of claims 1 to 5, wherein the machine learning techniques include at least linear regression, local regression, principal component regression, PLS regression, neural network, regression tree, random forest, LightGBM and XGBoost.

7. A quality prediction model generating method according to any one of claims 1 to 6, further comprising the step of regenerating the quality prediction model when the prediction accuracy of the quality prediction model falls outside a predetermined range or at predetermined intervals.

8. A quality prediction method for a metallic material, which predicts the state of quality defects of the metallic material using a quality prediction model generated by the quality prediction model generation method described in any one of claims 1 to 7.

9. A method for manufacturing a metallic material, which predicts the state of a quality defect of the metallic material using the metallic material quality prediction method described in claim 8, and changes at least a part of the manufacturing conditions when it is predicted that the metallic material has a quality defect.

10. A method for presenting manufacturing conditions for metallic materials, which predicts the state of a quality defect in the metallic material using the quality prediction method for metallic materials described in claim 8, and, if a quality defect in the metallic material is predicted, searches for and presents at least a partial change to the manufacturing conditions that will eliminate the quality defect.

11. A method for producing a metallic material, which produces a metallic material based on changes in the manufacturing conditions presented by the method for presenting manufacturing conditions for a metallic material as described in claim 10.

12. A quality prediction model generating device that generates a quality prediction model of a metal material manufactured through one or more processes, comprising: an acquisition unit that acquires explanatory variables selected from manufacturing conditions of each process and a dependent variable that is a state of quality defects in the manufactured metal material; a storage unit that associates the explanatory variables and the dependent variable and stores them as learning data; a calculation unit that divides the learning data into groups and tests whether there is a significant difference in the state of the quality defects; a search unit that searches for the most significant grouping; and a generation unit that generates the quality prediction model by machine learning using a group according to the most significant grouping found.

13. A quality prediction device for metallic materials, which predicts the state of quality defects in the metallic material using a quality prediction model generated by the quality prediction model generating device according to claim 12.

14. A manufacturing condition suggestion device for metallic materials, which suggests at least some changes to the manufacturing conditions for metallic materials based on the presence or absence of quality defects in metallic materials predicted by the quality prediction device for metallic materials described in claim 13.

15. A metallic material manufacturing system which controls manufacturing equipment for each process of manufacturing metallic materials in accordance with the changes in manufacturing conditions output from the output section of the metallic material manufacturing condition presentation device described in claim 14.

Images