METHOD FOR IDENTIFYING FACTORS OF QUALITY CHANGE, APPARATUS ... SYSTEM FOR IDENTIFYING FACTORS OF QUALITY CHANGE, AND METAL MATERIAL PRODUCTION METHOD

Abstract

The method for identifying factors causing quality changes executed by a computer includes a quality prediction step of predicting quality and outputting it as a quality prediction value, a quality prediction error calculation step of calculating the difference between the actual quality value and the quality prediction value as a quality prediction error, a variable selection step of selecting, based on a first variable importance, manufacturing conditions that have a high contribution to the quality prediction error from among the manufacturing conditions for each group, and a factor identification step of identifying, based on a second variable importance, manufacturing conditions that have a high contribution to the quality prediction error from among the selected manufacturing conditions, as factors causing changes in product quality.

Description

[Technical Field]

[0001] The present disclosure relates to a quality change factor identifying method, a quality change factor identifying device, a quality change factor identifying device system, and a method for manufacturing a metal material. [Background technology]

[0002] In manufacturing processes such as steel production, it is generally necessary to go through many production lines to create a single product. Variations in product quality can occur due to daily improvement activities on each production line, operational changes, or aging of equipment. For example, changes in manufacturing conditions to improve yield and changes in equipment status (e.g., cleaning at another facility) can combine to result in changes in product quality (quality abnormalities).

[0003] For example, Patent Document 1 discloses a method that can present possible causes of quality abnormalities when they occur. The method in Patent Document 1 uses a quality prediction model with multiple manufacturing conditions as input variables and product quality as an output variable, calculates the difference between the quality prediction value and the quality evaluation value as a quality prediction error, calculates the quality contribution of each manufacturing condition, and presents the manufacturing condition that is the cause. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 7251687 Summary of the Invention [Problem to be solved by the invention]

[0005] For example, if a product is steel, it is manufactured through multiple manufacturing processes, such as steelmaking, hot rolling, pickling, cold rolling, annealing, and plating. Therefore, the number of manufacturing conditions that may affect product quality can reach several hundred. It is possible to use machine learning or statistical models (such as various regression models) to build a quality prediction model in which hundreds of manufacturing conditions are used as explanatory variables (input variables) and product quality is used as the target variable (output variable). However, the quality prediction model has excessive degrees of freedom, which can easily lead to the problem of overfitting (also known as excessive fitting or overlearning), in which the model is too faithfully adapted to the training data.

[0006] The method of Patent Document 1 can present possible causes of quality anomalies when they occur, but does not consider the problem of overfitting. Predictions using a quality prediction model that does not address the problem of overfitting make it difficult to correctly identify manufacturing conditions that are potential causes of quality anomalies. For example, noise or variability in the performance data used for learning can lead to overfitting for manufacturing conditions whose values ​​change within a specific period, resulting in the importance (quality contribution) of manufacturing conditions that are not actually correlated with quality being ranked higher. Furthermore, manufacturing conditions that are actually correlated with quality can be ranked lower in importance.

[0007] The purpose of the present disclosure, made in consideration of the above circumstances, is to provide a quality change factor identification method, a quality change factor identification device, a quality change factor identification device system, and a metal material manufacturing method that can accurately identify the factors behind quality changes in products manufactured through multiple manufacturing processes. [Means for solving the problem]

[0008] (1) A quality change factor identification method according to an embodiment of the present disclosure includes: A quality change factor identification method for causing a computer to identify a factor of quality change in a product manufactured through a plurality of manufacturing processes, comprising: a quality prediction step of inputting the manufacturing conditions into a quality prediction model having manufacturing conditions of the plurality of manufacturing processes as input variables and quality as an output variable, and causing the computer to execute a process of predicting the quality as a quality prediction value; a quality prediction error calculation step of causing the computer to execute a process of calculating a difference between a quality result value of an actual product manufactured through the plurality of manufacturing processes and the quality predicted value as a quality prediction error; a variable selection step of causing the computer to execute a process of dividing manufacturing conditions of the plurality of manufacturing processes into a plurality of groups, calculating a first variable importance of the manufacturing conditions for each group using a plurality of first prediction error analysis models generated using the manufacturing conditions of each group as input variables and the quality prediction error as an output variable, and selecting, from the manufacturing conditions for each group, a manufacturing condition that has a high contribution to the quality prediction error based on the first variable importance; and a factor identification step of calculating second variable importance of the selected manufacturing conditions using a second prediction error analysis model generated using only the selected manufacturing conditions as input variables and the quality prediction error as an output variable, and causing the computer to execute a process of identifying, from the selected manufacturing conditions, a manufacturing condition that has a high contribution to the quality prediction error as a factor of a change in product quality based on the second variable importance.

[0009] (2) As one embodiment of the present disclosure, in (1), In the variable selection step, the manufacturing conditions of the plurality of manufacturing processes are grouped for each of the same manufacturing steps.

[0010] (3) As one embodiment of the present disclosure, in (1), In the variable selection step, the manufacturing conditions of the plurality of manufacturing processes are classified into groups according to the physical phenomena that affect the quality, and the manufacturing conditions are grouped according to the manufacturing conditions that affect the physical phenomena.

[0011] (4) A quality change factor identification device according to an embodiment of the present disclosure includes: A quality change factor identifying device that identifies factors that cause quality changes in a product manufactured through a plurality of manufacturing processes, a quality prediction unit that inputs the manufacturing conditions of the plurality of manufacturing processes into a quality prediction model having the manufacturing conditions as input variables and quality as an output variable, predicts the quality as a quality prediction value, and calculates a difference between the quality performance value of an actual product manufactured through the plurality of manufacturing processes and the quality prediction value as a quality prediction error; a variable selection unit that divides manufacturing conditions of the plurality of manufacturing processes into a plurality of groups, calculates a first variable importance of the manufacturing conditions for each group using a plurality of first prediction error analysis models generated using the manufacturing conditions of each group as input variables and the quality prediction error as an output variable, and selects, from the manufacturing conditions for each group, a manufacturing condition that has a high degree of contribution to the quality prediction error based on the first variable importance; and a factor identification unit that calculates second variable importance of the selected manufacturing conditions using a second prediction error analysis model generated using the selected manufacturing conditions as input variables and the quality prediction error as an output variable, and identifies, from the selected manufacturing conditions, a manufacturing condition that has a high contribution to the quality prediction error as a factor of a change in product quality based on the second variable importance.

[0012] (5) As an embodiment of the present disclosure, in (4), The apparatus further includes a first output unit that outputs, as an output signal, information including the specific manufacturing condition that is the cause of the quality change identified by the cause identification unit.

[0013] (6) As an embodiment of the present disclosure, in (4) or (5), a quality prediction model regeneration unit that regenerates and updates a quality prediction model in which the specific manufacturing conditions that are factors of the quality change identified by the factor identification unit are added to input variables; The system further includes a specific manufacturing condition control range determination unit that determines the control range of the specific manufacturing condition using the updated quality prediction model.

[0014] (7) As an embodiment of the present disclosure, in (6), The manufacturing method further includes a second output section that outputs the control range of the specific manufacturing condition determined by the specific manufacturing condition control range determination section as an output signal.

[0015] (8) A quality change factor identification device system according to an embodiment of the present disclosure includes: A quality change factor identification device system comprising a quality change factor identification device according to (4) or (6), a performance data acquisition device, and an information output device, for identifying factors of quality change in a product manufactured through a plurality of manufacturing processes, the performance data acquisition device acquires data including manufacturing conditions and quality performance values ​​of the plurality of manufacturing processes; The quality change factor identification device uses the data to identify information on specific manufacturing conditions that are manufacturing conditions that cause quality changes in the product, The information output device outputs information relating to the specific manufacturing conditions as an output signal.

[0016] (9) A method for producing a metal material according to an embodiment of the present disclosure includes: Predicting the quality of a product using a quality prediction model that includes, as input variables, specific manufacturing conditions identified by any one of the quality change factor identification methods (1) to (3) and that has the quality as an output variable; determining manufacturing conditions identified as factors affecting product quality so that the predicted product quality falls within a predetermined range; Manufacture steel products in accordance with specified manufacturing conditions. [Effects of the Invention]

[0017] According to the present disclosure, it is possible to provide a quality change factor identification method, a quality change factor identification device, a quality change factor identification device system, and a metal material manufacturing method that can accurately identify the factors that cause quality changes in products manufactured through multiple manufacturing processes. [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 is a block diagram showing the configuration of a quality change factor identification device according to an embodiment of the present disclosure. [Figure 2] FIG. 2 is a flowchart showing the flow of processing in a quality change factor identification method according to an embodiment of the present disclosure. [Figure 3] FIG. 3 is a diagram showing the procedure of the embodiment. [Figure 4] FIG. 4 is a diagram showing the relationship between the quality prediction value and the quality actual value in the quality prediction model of the embodiment. [Figure 5] FIG. 5 is a diagram showing the relationship between the quality prediction value and the quality actual value in the quality prediction model of the embodiment. [Figure 6] FIG. 6 is a graph showing the variable importance of the manufacturing conditions used as input variables calculated for each group and arranged in descending order. [Figure 7A] FIG. 7A shows a prediction error analysis of an example. [Figure 7B] FIG. 7B is a diagram showing a prediction error analysis of a comparative example. [Figure 8A] FIG. 8A is a diagram for explaining an outline of a conventional method for identifying a cause of a quality change. [Figure 8B] FIG. 8B is a diagram for explaining an outline of a conventional method for identifying a cause of a quality change. [Figure 8C] FIG. 8C is a diagram for explaining an outline of a conventional method for identifying a cause of a quality change. [Figure 8D] FIG. 8D is a diagram for explaining an outline of a conventional method for identifying a cause of a quality change. [Figure 9A] FIG. 9A is a diagram for explaining an overview of the quality change factor identification method according to this embodiment. [Figure 9B] FIG. 9B is a diagram for explaining an overview of the quality change factor identification method according to this embodiment. [Figure 10] FIG. 10 is a diagram showing a schematic configuration of a quality change factor identifying device system according to this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0019] Hereinafter, a quality change factor identification method, a quality change factor identification device, a quality change factor identification device system, and a metal material manufacturing method according to an embodiment of the present disclosure will be described with reference to the drawings. First, an overview of a conventional quality change factor identification method and an overview of a quality change factor identification method according to the present embodiment will be described.

[0020] 8A to 8D are diagrams for explaining an outline of a conventional method for identifying quality change factors. As shown in FIG. 8A, for example, when a product is steel, the product is manufactured through many manufacturing steps in an actual manufacturing process (actual plant). Therefore, the number of all manufacturing conditions (a1, a2, ... ai, b1, b2, ... bi, c1, c2 ... ci, d1, d2 ... di) that may affect a certain quality (actual measured value y) of the product may reach several hundred.

[0021] As shown in FIG. 8B, the quality prediction model for predicting quality uses N manufacturing conditions (x1, x2, ... xN) that have a high influence on the quality to be predicted from among all manufacturing conditions (a1, a2, ... ai, b1, b2, ... bi, c1, c2 ... ci, d1, d2 ... di). The quality prediction model selects, for example, N manufacturing conditions that have a high metallurgical influence, and uses them as explanatory variables (input variables), and outputs a predicted value y^ of the target quality. Hereinafter, the predicted value of quality may be referred to as a "predicted quality value." Furthermore, the actual measured value of quality may be referred to as a "measured quality value."

[0022] For example, the method of Patent Document 1 calculates the quality contribution of N manufacturing conditions (x1, x2, ... xN) input to a quality prediction model when the quality prediction error (y^-y), which is the difference between the actual quality value and the predicted quality value, becomes large. Then, the cause of the product quality abnormality is identified from among the N manufacturing conditions.

[0023] When attempting to identify the cause of a product quality abnormality from among all manufacturing conditions, including N manufacturing conditions, using the method of Patent Document 1, a prediction error analysis model is generated as shown in FIG. 8C. That is, a prediction error analysis model is generated in which all manufacturing conditions (a1, a2, ... ai, b1, b2, ... bi, c1, c2 ... ci, d1, d2 ... di) are used as input variables and the quality prediction error (y^-y) is used as the output variable. When there are hundreds of manufacturing conditions, even if variable importance (feature importance) is calculated based on the generated prediction error analysis model as shown in FIG. 8D, the large number of explanatory variables increases the degree of freedom of the model, which can lead to overfitting (overlearning). Therefore, it is difficult to identify the manufacturing condition that is the true cause of the quality change. Hereinafter, a large quality prediction error is simply referred to as a large quality prediction error. That is, unless otherwise specified, "a large quality prediction error" means that the absolute value of the quality prediction error is large. The quality change includes, but is not limited to, a quality abnormality, and includes any abnormal fluctuation in the quality of the target. Specific examples of variable importance will be described later.

[0024] The inventors came up with the idea of ​​analyzing quality prediction errors in two steps to resolve the problem of overfitting in prediction error analysis models and identify the manufacturing conditions that are the true cause of quality changes. In the first step, all manufacturing conditions are narrowed down to those that have a high contribution to the quality prediction error. In the second step, the quality prediction error for the narrowed-down manufacturing conditions is analyzed. These two steps resolve the problem of overfitting and make it possible to identify the manufacturing conditions that are the true cause of quality changes.

[0025] 9A and 9B are diagrams for explaining an overview of a quality change factor identification method according to this embodiment. As shown in FIG. 9A, in a first step, manufacturing conditions (a1, a2, ... ai, b1, b2, ... bi, c1, c2 ... ci, d1, d2 ... di) of multiple manufacturing processes are divided into multiple groups. In the example of FIG. 9A, the number of groups is four, but is not limited to a specific number. For each group, a prediction error analysis model is generated using the manufacturing conditions as input variables and the quality prediction error (ŷ-y) as an output variable. In other words, multiple prediction error analysis models are generated. In the example of FIG. 9A, a first prediction error analysis model A, a first prediction error analysis model B, a first prediction error analysis model C, and a first prediction error analysis model D (hereinafter referred to as "first prediction error analysis models A, B, C, and D") are generated. Then, using the first prediction error analysis models A, B, C, and D, the variable importance (first variable importance) of the manufacturing conditions is calculated for each group. In addition, based on the first variable importance, the manufacturing conditions for each group that have a high contribution to the quality prediction error (y^-y) (the manufacturing condition group x in the example of FIG. 9A) are selected. A , x B , x C , x D ) are selected.

[0026] As shown in FIG. 9B, the selected manufacturing conditions (x A , x B , x C , x D ) as an input variable and a prediction error analysis model (second prediction error analysis model) with the quality prediction error (y^-y) as an output variable is generated. Then, using the second prediction error analysis model, the selected manufacturing conditions (x A , x B , x C , x D ) is calculated. Based on the second variable importance, the manufacturing conditions (x in the example of FIG. 9B) that have a high contribution to the quality prediction error (y^-y) are calculated. res ) are identified as manufacturing conditions that cause quality changes in the product. Details of the quality change factor identifying device, quality change factor identifying method, quality change factor identifying device system, and metal material manufacturing method according to this embodiment will be described below.

[0027] <Quality change factor identification device> FIG. 1 is a block diagram illustrating the configuration of a quality change factor identification device according to an embodiment of the present disclosure. The quality change factor identification device identifies factors that cause quality changes in a product manufactured through multiple manufacturing processes. The quality change factor identification device diagnoses processes that cause quality changes in a manufacturing process, such as steel manufacturing. In this embodiment, the product is steel. As shown in FIG. 1 , the quality change factor identification device includes a data input unit 11, a performance database 12, a quality prediction unit 13, a variable selection unit 14, a factor identification unit 15, and a first output unit 16 as its main components. As in this embodiment, the quality change factor identification device may further include a quality prediction model regeneration unit 17, a specific manufacturing condition control range determination unit 18, and a second output unit 19. In this embodiment, the quality change factor identification device executes a quality change factor identification method and is used in a manufacturing method for metal materials. Here, the arrows in FIG. 1 simply indicate the flow of data or processing executed by the components of the quality change factor identification device.

[0028] (Data entry section) The data input unit 11 acquires, as actual operation data, the manufacturing conditions of multiple processes (multiple manufacturing steps) for manufacturing a product for which quality change factors are to be identified. The actual operation data also includes actual quality values. In this embodiment, the data input unit 11 receives the data via a network. The data input unit 11 stores (accumulates) the received actual operation data in a performance database 12. It is desirable that the performance database 12 be configured to be able to store large amounts of data for a long period of time. The performance database 12 may be stored in a storage unit included in (or accessible from) the quality change factor identification device. The storage unit is, for example, one or more memories. The memory is, for example, a semiconductor memory, a magnetic memory, an optical memory, or the like, but is not limited to these and can be any memory.

[0029] (Quality Prediction Department) The quality prediction unit 13 generates a quality prediction model that predicts quality using past actual operation data (manufacturing conditions and quality results) acquired from the performance database 12. The quality prediction model uses manufacturing conditions of multiple manufacturing processes as input variables and quality as an output variable. The quality prediction model may be generated using machine learning including statistical methods such as PLS regression or random forest.

[0030] The quality prediction unit 13 uses the generated quality prediction model to predict the quality of a target product for which the factors of quality change have been identified. The target product may be, for example, the latest product (a product currently being manufactured). In other words, the quality prediction unit 13 inputs the manufacturing conditions into a quality prediction model that uses manufacturing conditions of multiple manufacturing processes as input variables and quality as an output variable, and predicts the quality as a predicted value.

[0031] In this embodiment, the quality prediction unit 13 calculates the difference between the actual quality value of an actual product manufactured through multiple manufacturing processes and the predicted quality value as a quality prediction error. The calculated quality prediction error may be output to the variable selection unit 14, the factor identification unit 15, and the first output unit 16. The calculated quality prediction error may be output only when the quality prediction error exceeds a predetermined range.

[0032] (Variable selection section) The variable selection unit 14 divides manufacturing conditions of a plurality of manufacturing processes into a plurality of groups, and generates a plurality of first prediction error analysis models using the manufacturing conditions of each group as input variables and the quality prediction error as an output variable. The variable selection unit 14 also calculates the first variable importance of the manufacturing conditions for each group using the first prediction error analysis model, and selects manufacturing conditions that have a high degree of contribution to the quality prediction error from the manufacturing conditions for each group based on the first variable importance.

[0033] In this embodiment, the variable selection unit 14 generates a plurality (N) of models (first prediction error analysis models) for analyzing the quality prediction error calculated by the quality prediction unit 13.

[0034] As described above, each first prediction error analysis model has manufacturing conditions as input variables and quality prediction errors as output variables. The manufacturing conditions used as input variables for one first prediction error analysis model are not all of the manufacturing processes, but are selected so that the number of manufacturing conditions is not more than the number necessary to avoid overfitting. Here, there may be overlap in the manufacturing conditions used as input variables for multiple models, as long as the number of overlaps is not more than the number necessary to avoid overfitting for each model. However, in this embodiment, the N first prediction error analysis models will be described as having different input variables.

[0035] Here, the number of input variables of the first prediction error analysis model differs depending on, for example, the product, but is preferably selected to be, for example, 50 or less (50 cases) in order to avoid overfitting.

[0036] The first prediction error analysis model has many input variables. Therefore, in the case of a linear model, the first prediction error analysis model is preferably generated by principal component regression (PCR) or partial least squares regression (PLS) in consideration of multicollinearity. In addition, in the case of a nonlinear model, the first prediction error analysis model is preferably generated by machine learning of a tree model (decision tree) such as random forest or gradient boosting (LightGBM). The number of data (manufacturing conditions and quality prediction errors) used to generate the first prediction error analysis model is preferably at least 80 times the number of input variables.

[0037] The manufacturing conditions may be grouped based on the product, quality details, characteristics of the manufacturing line, etc. Here, it is preferable to group the manufacturing conditions by the same manufacturing process. That is, it is preferable that the variable selection unit 14 group the manufacturing conditions of multiple manufacturing processes by the same manufacturing process. By classifying the models by manufacturing process, it is possible to extract the influence of each process and clarify the manufacturing conditions that have a large influence (contribution) on the quality of the product in each manufacturing process. Since it is possible to clarify the equipment and its manufacturing conditions that are important for quality control for each manufacturing process, an improvement in the quality control level can be expected. In particular, grouping by manufacturing process is preferable when the equipment itself of the manufacturing process has changed due to repair work or other factors.

[0038] As another example, the variable selection unit 14 may classify the manufacturing conditions of a plurality of manufacturing processes into factors (physical phenomena) that affect quality, and group the manufacturing conditions according to the factors (physical phenomena). This type of grouping can also be expected to improve the level of quality control.

[0039] Taking the example of tensile strength as the quality of a metallic material, the strengthening mechanisms (one example of a physical phenomenon) of metallic materials are classified into solid solution strengthening, precipitation strengthening, microstructure strengthening (dislocation strengthening), grain refinement, etc. Then, metallic materials are grouped according to the manufacturing conditions that affect the strengthening mechanisms (physical phenomena). For example, manufacturing conditions can be grouped into a group of variables (manufacturing conditions) related to solid solution strengthening, a group of variables (manufacturing conditions) related to precipitation strengthening, a group of variables (manufacturing conditions) related to microstructure strengthening, etc.

[0040] Furthermore, if the product is steel, the elemental composition values ​​have a significant effect on quality. The elemental composition values ​​can be grouped separately from other manufacturing conditions.

[0041] Furthermore, if it is clear that the manufacturing conditions have been changed within the period in which the data (manufacturing conditions and quality prediction errors) for generating the first prediction error analysis model were measured, it is preferable that the changed manufacturing conditions be collectively classified into one group.

[0042] The variable selection unit 14 calculates the variable importance (first variable importance) of the manufacturing conditions for each group with respect to the quality prediction error using a plurality (N) of first prediction error analysis models. Based on the first variable importance, the manufacturing conditions for each group that have a high contribution to the quality prediction error are selected.

[0043] When the first prediction error analysis model is generated by principal component regression or partial least quadratic regression, the variable selection unit 14 may calculate the first variable importance using the principal component score of the first principal component.

[0044] Furthermore, when the first prediction error analysis model is generated by machine learning, the variable selection unit 14 may calculate the first variable importance using a SHAP (Shapley Additive exPlanations) value or permutation importance. Permutation importance is a method used to quantify the importance of features in a machine learning model.

[0045] The variable selection unit 14 sorts the manufacturing conditions in descending order of first variable importance, and selects a predetermined number of variables for each group in descending order (top-rank) as variables having a high degree of contribution to the quality prediction error. In this embodiment, the variables are manufacturing conditions. Also, in this embodiment, it is assumed that the predetermined number is M. Therefore, the variable selection unit 14 selects M manufacturing conditions having a high degree of contribution to the quality prediction error for each of the multiple (N) first prediction error analysis models, and therefore selects a total of N×M variables (manufacturing conditions).

[0046] The total of N×M manufacturing conditions selected by the variable selection unit 14 are used as input variables for a model (second prediction error analysis model) in the factor identification unit 15. Therefore, in order to avoid overfitting of the second prediction error analysis model, it is desirable that the number of N×M manufacturing conditions be 50 (50 items) or less.

[0047] Furthermore, the variable selection unit 14 may set a threshold for the variable importance and select variables (manufacturing conditions) corresponding to variable importance equal to or greater than the set threshold. The threshold for variable importance may be set in advance depending on the manufacturing process and the quality of the prediction target.

[0048] (Cause identification part) The factor identification unit 15 generates a second prediction error analysis model using the N×M manufacturing conditions selected by the variable selection unit 14 as input variables and the quality prediction error calculated by the quality prediction unit 13 as an output variable. The factor identification unit 15 calculates second variable importance of the selected N×M manufacturing conditions using the generated second prediction error analysis model. Then, based on the second variable importance, the factor identification unit 15 identifies, from the selected manufacturing conditions, a manufacturing condition that has a high contribution to the quality prediction error as a factor of a change in product quality.

[0049] As described above, the number of input variables (N × M) of the second prediction error analysis model is set to 50 or less. Therefore, it is possible to avoid overfitting and identify important factors that are quality change factors from the manufacturing conditions selected by the variable selection unit 14. In addition to the important factors, it is also possible to identify indirect factors that change the quality correlation.

[0050] (First output section) The first output unit 16 outputs, as an output signal, information including the manufacturing conditions identified by the factor identification unit 15 as factors of a change in product quality. The output signal may include a quality prediction error calculated by the quality prediction unit 13. Furthermore, the first output unit 16 may output an output signal when the quality prediction error calculated by the quality prediction unit 13 exceeds the range between the upper and lower control limits and it is determined that a quality abnormality has occurred. If the output signal is output to an equipment control device, the output signal may be a control signal that automatically stops the target equipment that sets the manufacturing conditions that are factors of the quality change identified by the factor identification unit 15. Furthermore, if the output signal is output to an operator's operation terminal, the output signal may be operator guidance information to be displayed as guidance on the screen.

[0051] (Quality prediction model regeneration part) The quality prediction model regeneration unit 17 regenerates and updates the quality prediction model by adding the manufacturing conditions (hereinafter referred to as "specific manufacturing conditions") that are factors of quality changes identified by the factor identification unit 15 to the input variables of the quality prediction model generated by the quality prediction unit 13. The method for regenerating the quality prediction model is the same as the generation method by the quality prediction unit 13.

[0052] (Specific Manufacturing Conditions Management Range Determination Department) The specific manufacturing condition control range determination unit 18 uses the updated quality prediction model to determine the control range of the specific manufacturing condition (in other words, a new target range for the specific manufacturing condition). The specific manufacturing condition control range determination unit 18 inputs the reference value of the specific manufacturing condition and other manufacturing conditions into the updated quality prediction model based on past actual operation data (manufacturing conditions and quality result values) acquired from the performance database 12, and predicts a quality prediction value. The specific manufacturing condition control range determination unit 18 then calculates the difference between the quality result value and the quality prediction value as the quality prediction error. The range of change from the reference value of the specific manufacturing condition, within which the calculated quality prediction error falls within the upper and lower control limits of the quality prediction error, can be determined as the control range using mathematical programming such as the branch and bound method.

[0053] (Second output section) The second output unit 19 outputs the control range of the specific manufacturing condition determined by the specific manufacturing condition control range determination unit 18 as an output signal.

[0054] When the output signal is output to a control device of equipment, the output signal may be a control signal for the equipment that sets specific manufacturing conditions.

[0055] Furthermore, when the output signal is output to an operation terminal of an operator, the output signal may be operator guidance information to be displayed on the screen as guidance.

[0056] <How to identify quality change factors> 2 is a flowchart showing the flow of processing of a quality change factor identification method executed by a quality change factor identification device according to this embodiment. The quality change factor identification method includes at least a quality prediction step, a quality prediction error calculation step, a variable selection step, and a factor identification step.

[0057] In the quality prediction step, the quality prediction unit 13 predicts the quality of the target product for which the factors of quality change are identified, using the generated quality prediction model. Next, in the quality prediction error calculation step, the quality prediction unit 13 calculates the difference between the actual quality value of the actual product manufactured through multiple manufacturing processes and the quality predicted value as the quality prediction error.

[0058] Next, in the variable selection step, the variable selection unit 14 divides the manufacturing conditions of the multiple manufacturing processes into multiple groups (N), selects M manufacturing conditions that have a high contribution to the quality prediction error for each group, and selects M × N manufacturing conditions overall.

[0059] Next, in the factor identification step, the factor identification unit 15 identifies a manufacturing condition that is a factor of the quality change from the M×N manufacturing conditions (variables) selected in the variable selection step.

[0060] <Metal material manufacturing method> Using the above-described quality change factor identification method, a metal material manufacturing method is carried out in which a metal material including steel is manufactured as a product. That is, the quality of the product is predicted using a quality prediction model that includes the specific manufacturing conditions identified by the quality change factor identification method as input variables and has quality as an output variable. Furthermore, the manufacturing conditions identified as factors of product quality are determined so that the predicted product quality falls within a predetermined range. Then, steel products are manufactured in accordance with the identified manufacturing conditions.

[0061] In a manufacturing method for metallic materials, if a specific manufacturing condition is not included in the manufacturing conditions that are explanatory variables of the quality prediction model, the quality prediction model is regenerated and updated. The quality prediction model is regenerated and updated by adding the specific manufacturing condition to the input variables (explanatory variables). Here, if the specific manufacturing condition is included as an explanatory variable of the quality prediction model, the quality prediction model may be regenerated so as to minimize the quality prediction error without making any changes to the explanatory variables.

[0062] Next, a control range for the specific manufacturing conditions is determined so that the product quality predicted using the regenerated quality prediction model falls within a predetermined range. The control range for the specific manufacturing conditions can be determined by mathematical programming such as the branch and bound method.

[0063] Next, the specific manufacturing conditions are adjusted so that they fall within the determined control range, and steel products are manufactured.

[0064] The effects of the present disclosure will be specifically described below based on examples, but the present disclosure is not limited to these examples.

[0065] An example of the method for identifying quality change factors according to this embodiment will be described below. In this example, quality change factors were identified for the tensile strength of a highly formable, high-strength cold-rolled steel sheet, which is a type of cold-rolled steel sheet. Figure 3 shows the procedure of this example.

[0066] Cold-rolled steel sheets are manufactured through a steelmaking process, a hot rolling process, a cold rolling process, and an annealing process. In the steelmaking process, the elemental component values ​​are the representative operating conditions. In the hot rolling process, the material temperature is the representative operating condition. In the cold rolling process, the reduction amount is the representative operating condition. In the annealing process, the material temperature is the representative operating condition. In this example, manufacturing conditions (variables) that are candidates for quality change factors were obtained from these manufacturing processes (STEP 1). There were a total of approximately 200 types of manufacturing conditions.

[0067] In this example, a quality prediction model was generated for predicting tensile strength for the same product (products with the same manufacturing specifications or dimensional size) based on data on manufacturing conditions and actual tensile strength measurements (48,000 samples) from the past three years. In this example, the quality prediction model was generated using partial least squares regression (PLS). From approximately 200 types of operating conditions, 20 manufacturing conditions that are considered to have a particularly high correlation with tensile strength from a metallurgical perspective were used as input variables (explanatory variables). Figure 4 shows the relationship between the quality prediction value in the generated quality prediction model and the actual tensile strength value (actual quality value).

[0068] Then, data (4000 samples) on manufacturing conditions and actual measured values ​​(actual quality values) of tensile strength for the most recent six months, which differ from the data used to generate the quality prediction model, were input into the generated quality prediction model to calculate a predicted value of tensile strength (predicted quality value). In addition, the difference between the calculated predicted quality value and the actual quality value was calculated as a quality prediction error (STEP 2). STEP 2 corresponds to the quality prediction step and the quality prediction error calculation step.

[0069] Figure 5 shows the relationship between the quality prediction value in the quality prediction model and the actual measured value of tensile strength (actual quality value) in STEP 2. In Figure 5, the quality prediction error is larger than in Figure 4 (the predicted quality value is off from the actual quality value), indicating that the quality characteristics have changed in the last six months.

[0070] Next, four prediction error analysis models (first prediction error analysis models) were generated to predict the quality prediction error of tensile strength in STEP 2 as an output variable. In this example, the first prediction error analysis models were generated using the gradient boosting (LightGBM) method.

[0071] In this example, the approximately 200 types of manufacturing conditions (variables) selected in STEP 1 were grouped into four manufacturing process groups. That is, the variables of the steelmaking process (Group A), the variables of the hot rolling process (Group B), the variables of the cold rolling process (Group C), and the variables of the annealing process (Group D) were each set as input variables (explanatory variables) of the first prediction error analysis model (STEP 3). STEP 3 corresponds to a variable selection step.

[0072] FIG. 6 shows a graph in which the variable importance of the manufacturing conditions used as input variables was calculated for each group using the first prediction error analysis model and arranged in descending order. In this example, the top five manufacturing conditions (variables) were selected for each of the four groups. That is, from the approximately 200 manufacturing conditions selected in STEP 1, a total of 20 manufacturing conditions were selected for the four groups for prediction error analysis. The horizontal axis of each graph in FIG. 6 represents variable importance, and in this example, the SHAP value was used as the variable importance. That is, each bar graph indicates the magnitude of the SHAP value for each manufacturing condition.

[0073] A prediction error analysis model (second prediction error analysis model) with the quality prediction error from STEP 2 as the output variable was generated using the gradient boosting method. At this time, the input variables (explanatory variables) were the 20 manufacturing conditions selected in STEP 3 (STEP 4). STEP 4 corresponds to the factor identification step.

[0074] Figure 7A shows the prediction error analysis of this example. Using the second prediction error analysis model, the variable importance (SHAP value) of 20 types of manufacturing conditions used as input variables was calculated, and the manufacturing conditions were sorted in descending order of variable importance. The manufacturing condition that caused the quality change was identified as "a1" of the steelmaking process (Group A), which has the highest variable importance in Figure 7A.

[0075] FIG. 7B shows the prediction error analysis of the comparative example. In the comparative example, STEP 3 (variable selection) was not performed, and a prediction error analysis model (see FIG. 8C) was generated using approximately 200 types of manufacturing conditions selected in STEP 1 as input variables and the quality prediction error in STEP 2 as output variables. Then, in the comparative example, this prediction error analysis model was used to calculate the variable importance (SHAP value) of the approximately 200 types of manufacturing conditions used as input variables, and the variables were sorted in descending order as shown in FIG. 7B. In the comparative example, the manufacturing condition with the highest variable importance was indicated as "dN" for the annealing process (Group D), resulting in an incorrect result. Furthermore, in the comparative example, "a1" for the steelmaking process (Group A), which was the true cause, was indicated second.

[0076] As described above, the quality change factor identification method, quality change factor identification device, and metal material manufacturing method according to this embodiment can accurately identify the factors behind quality changes in products manufactured through multiple manufacturing processes using the above-described configurations and processes.

[0077] That is, according to the present disclosure, even when there are hundreds of variables (manufacturing conditions) that can affect quality, it is possible to group the manufacturing conditions of multiple manufacturing processes. Using multiple models (first prediction error analysis models) that represent the relationship between the manufacturing conditions of each group and the quality prediction error, it is possible to select the manufacturing conditions (variables) that contribute most to the quality prediction error for each group. Furthermore, it is possible to identify quality change factors using a model (second prediction error analysis model) that represents the relationship between all the manufacturing conditions (variables) selected for each group and the quality prediction error. According to the present disclosure, as described above, it is possible to avoid overfitting, eliminate the influence of manufacturing conditions (variables) with low correlation, and easily identify important factors that cause quality change.

[0078] Furthermore, according to the present disclosure, it is possible to group products by the same manufacturing process. By grouping products in this way, it is possible to clarify the manufacturing conditions that have a large impact (contribution) on the quality of the product in each manufacturing process, which is expected to improve the level of quality control.

[0079] Furthermore, according to the present disclosure, manufacturing conditions can be grouped by physical phenomena (factors) that affect quality. By grouping in this way, important manufacturing processes, equipment, and their manufacturing conditions can be clarified for each physical phenomenon (factor), which is expected to improve the level of quality control.

[0080] Furthermore, according to the present disclosure, in a method for manufacturing a metal material that is manufactured through multiple manufacturing processes, it is possible to determine manufacturing conditions that contribute to quality so that they fall within a control range, thereby enabling the manufacturing of a high-quality metal material.

[0081] Although the 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, the 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.

[0082] Here, the quality change factor identification device (e.g., a computer) may not be a single device, but may be composed of multiple devices located in multiple locations and capable of sending and receiving data to and from each other via a network. In other words, multiple devices connected via a network may function as a quality change factor identification device as a whole. Therefore, for example, the quality change factor identification device may be composed of a single computer as a hardware configuration, or may be composed of multiple computers connected via a network. When composed of multiple computers, the performance database 12 may be stored in a shared memory accessible by each computer.

[0083] 1 is merely an example. For example, the data input unit 11 and performance database 12 may be provided as a performance data acquisition device outside the quality change factor identification device, and the performance database 12 may be accessible by the quality change factor identification device. The performance data acquisition device acquires data including manufacturing conditions and quality performance values ​​for multiple manufacturing processes. In this configuration, the performance database 12 may accumulate actual operation data from the data input unit 11 provided in, for example, a host computer. The host computer may be, for example, a process computer that manages the manufacturing process of metal materials.

[0084] The first output unit 16 and the second output unit 19 may be provided as information output devices external to the quality change factor identification device, and may output, as an output signal, information regarding specific manufacturing conditions identified by the quality change factor identification device as a cause of a change in product quality. The first output unit 16 has been described as outputting information including the specific manufacturing conditions as an output signal, which, when output to an equipment control device, may be a control signal for automatically stopping the equipment for which the specific manufacturing conditions are set. The second output unit 19 has been described as outputting the management range of the specific manufacturing conditions as an output signal, which, when output to an equipment control device, may be a control signal for the equipment. Here, the output signal regarding the specific manufacturing conditions refers to the output signal of at least one of the first output unit 16 and the second output unit 19. The output signal may also include various information obtained by processing the quality change factor identification method. For example, the output signal may include manufacturing conditions that contribute highly to quality prediction errors, or may include information sorted using a graph or the like, showing the variable importance of each manufacturing condition, as shown in FIG. 9B. At this time, the information contained in the output signal may be displayed on a display device such as a liquid crystal display (LCD) or an organic electroluminescence panel (OLED). The display device may be a device that can be viewed by an operator of the metal material manufacturing process. The display device may also be a display on an operation terminal such as a smartphone or tablet used by the operator. The operator can confirm or make a final decision on changes to the manufacturing conditions based on the operator guidance information (such as manufacturing conditions with high contribution rates and sorted manufacturing conditions) displayed on the display device.

[0085] Furthermore, for example, a quality change factor identification device, a performance data acquisition device, and an information output device including an operation terminal may constitute a quality change factor identification device system. As shown in FIG. 10 , the quality change factor identification device system 20 is configured such that a performance data acquisition device 21, an information output device 22, and a quality change factor identification device 23 can communicate with each other via a network. Of the quality change factor identification device system 20, the data input unit 11 and the information output device 22 included in the performance data acquisition device 21 are located within the steelworks. The performance database 12 and the quality change factor identification device 23 included in the performance data acquisition device 21 are realized, for example, by a server on the cloud. The quality change factor identification device system 20 may further be configured with a host computer such as a process computer. The devices constituting the quality change factor identification device system can communicate with each other via a network. The network is, for example, the Internet. For example, the network may be configured to include a LAN (Local Area Network) in part.

[0086] In the above embodiment, the quality prediction unit 13 executes the processes of the quality prediction step and the quality prediction error calculation step. Here, the quality change factor identification device may be configured to include a quality prediction error calculation unit that executes the process of the quality prediction error calculation step, separate from the quality prediction unit 13. In this case, the quality prediction unit 13 may execute only the quality prediction step. [Explanation of symbols]

[0087] 11 Data entry section 12 Performance database 13 Quality Prediction Department 14 Variable Selection Section 15. Factor Identification Section 16 First output section 17 Quality prediction model regeneration unit 18. Specific Manufacturing Conditions Control Range Determination Department 19 Second output section 20 Quality change factor identification device system 21 Performance data acquisition device 22 Information output device 23 Quality change factor identification device

Claims

1. A quality change factor identification method for causing a computer to identify a factor of quality change in a product manufactured through a plurality of manufacturing processes, comprising: a quality prediction step of inputting the manufacturing conditions into a quality prediction model having manufacturing conditions of the plurality of manufacturing processes as input variables and quality as an output variable, and causing the computer to execute a process of predicting the quality as a quality prediction value; a quality prediction error calculation step of causing the computer to execute a process of calculating a difference between a quality result value of an actual product manufactured through the plurality of manufacturing processes and the quality predicted value as a quality prediction error; a variable selection step of causing the computer to execute a process of dividing manufacturing conditions of the plurality of manufacturing processes into a plurality of groups, calculating a first variable importance of the manufacturing conditions for each group using a plurality of first prediction error analysis models generated using the manufacturing conditions of each group as input variables and the quality prediction error as an output variable, and selecting, from the manufacturing conditions for each group, a manufacturing condition that has a high degree of contribution to the quality prediction error based on the first variable importance; a factor identification step of calculating second variable importance of the selected manufacturing conditions using a second prediction error analysis model generated using only the selected manufacturing conditions as input variables and the quality prediction error as an output variable, and causing the computer to execute a process of identifying, from the selected manufacturing conditions, a manufacturing condition that has a high contribution to the quality prediction error as a factor of the change in quality of the product based on the second variable importance.

2. 2. The quality change factor identification method according to claim 1, wherein the manufacturing conditions of the plurality of manufacturing processes are grouped for each of the same manufacturing processes in the variable selection step.

3. 2. The quality change factor identification method according to claim 1, wherein in the variable selection step, the manufacturing conditions of the plurality of manufacturing processes are classified into groups according to physical phenomena that affect the quality.

4. A quality change factor identifying device that identifies factors that cause quality changes in a product manufactured through a plurality of manufacturing processes, a quality prediction unit that inputs the manufacturing conditions of the plurality of manufacturing processes into a quality prediction model having the manufacturing conditions as input variables and quality as an output variable, predicts the quality as a quality prediction value, and calculates a difference between the quality performance value of an actual product manufactured through the plurality of manufacturing processes and the quality prediction value as a quality prediction error; a variable selection unit that divides manufacturing conditions of the plurality of manufacturing processes into a plurality of groups, calculates a first variable importance of the manufacturing conditions for each group using a plurality of first prediction error analysis models generated using the manufacturing conditions of each group as input variables and the quality prediction error as an output variable, and selects, from the manufacturing conditions for each group, a manufacturing condition that has a high degree of contribution to the quality prediction error based on the first variable importance; a factor identifying unit that calculates second variable importance of the selected manufacturing conditions using a second prediction error analysis model generated with the selected manufacturing conditions as input variables and the quality prediction error as an output variable, and identifies, from the selected manufacturing conditions, a manufacturing condition that has a high contribution to the quality prediction error as a factor of the change in product quality based on the second variable importance.

5. 5. The quality change factor identifying device according to claim 4, further comprising: a first output unit that outputs, as an output signal, information including the specific manufacturing condition that is the factor of the quality change identified by the factor identifying unit.

6. a quality prediction model regeneration unit that regenerates and updates a quality prediction model in which the specific manufacturing conditions that are factors of the quality change identified by the factor identification unit are added to input variables; 6. The quality change factor identifying device according to claim 4, further comprising: a specific manufacturing condition control range determining unit that determines a control range of the specific manufacturing condition using the updated quality prediction model.

7. 7. The quality change factor identifying device according to claim 6, further comprising a second output unit that outputs the control range of the specific manufacturing condition determined by the specific manufacturing condition control range determining unit as an output signal.

8. A quality change factor identifying system comprising the quality change factor identifying device according to claim 4, a performance data acquiring device, and an information output device, for identifying factors of quality change in a product manufactured through a plurality of manufacturing processes, the performance data acquisition device acquires data including manufacturing conditions and quality performance values ​​of the plurality of manufacturing processes; The quality change factor identification device uses the data to identify information on specific manufacturing conditions that are manufacturing conditions that cause quality changes in the product, The information output device outputs information relating to the specific manufacturing conditions as an output signal.

9. predicting product quality using a quality prediction model that includes, as input variables, specific manufacturing conditions identified by the method for identifying quality change factors according to any one of claims 1 to 3 and that uses the quality as an output variable; determining manufacturing conditions identified as factors affecting product quality so that the predicted product quality falls within a predetermined range; A method for manufacturing metal materials that produces steel products according to specified manufacturing conditions.

Images