Quality factor estimation method, operation condition change method, model generation method, quality factor estimation device, and operation condition change device

Abstract

This quality factor estimation method comprises: a step for acquiring quality data indicating the quality of products for normal quality and m (m is equal to or greater than 2) types of quality defects; a step for acquiring operation data of a manufacturing process; a step for creating quality evaluation data for each of the m types of quality defects; a step for acquiring m types of models generated by using the quality evaluation data, and determining a variable importance indicating the magnitude of the contribution of the operation variable for each of the m types of models; and a step for identifying an operation variable estimated to be the main factor with respect to at least one of the m types of quality defects by comparing the variable importances of the m types of models.

Description

Quality factor estimation method, operation condition change method, model generation method, quality factor estimation device, and operation condition change device

[0001] The present disclosure relates to a quality factor estimation method, an operation condition changing method, a model generation method, a quality factor estimation device, and an operation condition changing device.

[0002] In manufacturing processes where operational conditions determine quality, there are known methods for predicting quality from operational conditions. For example, there are methods for predicting quality using a physical model constructed based on knowledge of the mechanisms by which quality defects occur, and methods for predicting quality by applying a multiple regression model or a decision tree model using operational data and quality data. In methods that use models, the operational data of the product is input into the model to calculate a predicted value for quality, and the model can be evaluated by evaluating the predicted value.

[0003] For example, Patent Document 1 discloses a method for modeling an operation variable space, which uses operation data as a basis vector, by dividing the operation variable space into several local regions based on operation data and quality data. An activity function is calculated from the operation data, which expresses the contribution rate, indicating the degree to which each local relational expression expressing the model affects the overall quality, as a function of coordinates in the operation variable space. A mathematical model expressing the relationship between the overall operation variables and quality is then derived. If the error of the mathematical model does not satisfy a set convergence test, the number of divisions of the operation variable space is increased, and operation variables to be used in the local relational expressions are selected using a stepwise method. The processes of increasing the number of divisions of the operation variable space, constructing the local relational expressions, selecting variables, and constructing the activity function are repeated until the convergence test is satisfied.

[0004] JP 2012-27683 A

[0005] The method of Patent Document 1 determines the number of divisions based on a convergence test calculated from the model error. The division points that determine the division pattern are determined, for example, by extracting data for one operational variable, dividing the data into multiple groups, determining the values ​​of the operational variables that form the boundaries between each group, and then calculating these values ​​for all operational variables. Specifically, a clustering method or values ​​manually set by an operator are used. However, the method of determining the number of divisions and division points disclosed in Patent Document 1 may generate local regions in the derived mathematical model that represents the relationship between the overall operational variables and quality, which may contradict the mechanism of occurrence of the quality defect being predicted. Because the operational variables selected within the local regions may also contradict the physical mechanism of occurrence, there is a problem that even if the mathematical model has high accuracy, the quality defect factors estimated from the constructed mathematical model do not match the mechanism of occurrence of the quality defect.

[0006] In view of the above circumstances, an object of the present disclosure is to provide a quality factor estimation method, an operating condition changing method, a model generation method, a quality factor estimation device, and an operating condition changing device that can estimate quality defect factors with high accuracy.

[0007] (1) A quality factor estimation method according to one embodiment of the present disclosure is a quality factor estimation method executed by a quality factor estimation device used in a manufacturing process for manufacturing a product, the quality factor estimation method including: a quality data acquisition step of acquiring quality data indicating quality of the product for normal quality and two or more m types of quality defects; an operation data acquisition step of acquiring operation data of the manufacturing process; a quality evaluation data creation step of creating, for each of the m types of quality defects, m types of quality evaluation data composed of the quality data and the operation data corresponding to the corresponding quality defect and the quality data and the operation data corresponding to the normal quality; a variable importance determination step of acquiring m types of models for predicting each of the m types of quality defects, the m types of models being generated using the quality evaluation data, with the product quality as a dependent variable and operation variables included in the operation data as explanatory variables, and determining variable importance indicating a magnitude of contribution of the operation variables for each of the m types of models; and a defect factor estimation step of identifying the operation variable estimated to be a major factor for at least one of the m types of quality defects by comparing the variable importance of the m types of models.

[0008] (2) As one embodiment of the present disclosure, in (1), the defect factor estimation step inputs the quality assessment data used in generating one model and the quality assessment data used in generating another model into one of the m types of models, and if the magnitude of the difference in accuracy rate of the prediction results is greater than a predetermined value, identifies the operational variable that is estimated to be the main cause of the quality defect corresponding to the one model.

[0009] (3) As an embodiment of the present disclosure, in (1) or (2), the defect factor estimation step searches for an operation variable that commonly has a high variable importance in a corresponding model for two or more types of the m types of quality defects, and identifies the operation variable by regarding the searched operation variable as a common cause of the corresponding quality defects.

[0010] (4) As an embodiment of the present disclosure, in any one of (1) to (3), the defect factor estimation step selects, for at least one of the m types of quality defects, a candidate operational variable that is estimated to be a major factor based on the variable importance, and further uses performance data that indicates the relationship between the quality defects and the operational variables to probabilistically evaluate the relationship between the selected candidate variable and the quality defects, thereby identifying the operational variable that is estimated to be a major factor.

[0011] (5) An operating condition changing method according to an embodiment of the present disclosure includes an operating condition change instruction step of changing the operating conditions of the operating variables identified by the quality factor estimation method of (3) or (4) by a change instruction.

[0012] (6) A model generation method according to one embodiment of the present disclosure is a model generation method for generating a model used in any of the quality factor estimation methods of (1) to (4), comprising the steps of: using the quality evaluation data corresponding to the one quality defect as learning data in which the product quality is a dependent variable and operation variables included in the operation data are explanatory variables; and generating a model corresponding to the one quality defect by machine learning using the learning data; and generating the m types of models by executing the step of generating the model for each of the m types of quality defects.

[0013] (7) A quality factor estimation device according to an embodiment of the present disclosure is a quality factor estimation device used in a manufacturing process for manufacturing a product, comprising: a quality data acquisition unit that acquires quality data indicating quality of the product for normal quality and two or more m types of quality defects; an operation data acquisition unit that acquires operation data of the manufacturing process; a quality evaluation data creation unit that creates m types of quality evaluation data for each of the m types of quality defects, the quality evaluation data being composed of the quality data and the operation data corresponding to the corresponding quality defect and the quality data and the operation data corresponding to the normal quality; a variable importance determination unit that acquires m types of models that are models for predicting each of the m types of quality defects, the m types of models being generated using the quality evaluation data, with the product quality as a dependent variable and operation variables included in the operation data as explanatory variables, and determines variable importance indicating a magnitude of contribution of the operation variables for each of the m types of models; and a defect factor estimation unit that identifies the operation variable that is estimated to be a major factor for at least one of the m types of quality defects by comparing the variable importance of the m types of models.

[0014] (8) As an embodiment of the present disclosure, in (7), the defect factor estimation unit inputs the quality evaluation data used in generating one model and the quality evaluation data used in generating another model into one of the m types of models, and if the magnitude of the difference in accuracy rate of the prediction results is greater than a predetermined value, identifies the operational variable that is estimated to be the main cause of the quality defect corresponding to the one model.

[0015] (9) As an embodiment of the present disclosure, in (7) or (8), the defect factor estimation unit searches for an operation variable that has a common high variable importance in a corresponding model for two or more types of the m types of quality defects, and identifies the operation variable by regarding the searched operation variable as a common cause of the corresponding quality defects.

[0016] (10) As an embodiment of the present disclosure, in any one of (7) to (9), the defect factor estimation unit selects, for at least one of the m types of quality defects, a candidate operational variable that is estimated to be a major factor based on the variable importance, and further uses performance data that indicates a relationship between the quality defects and the operational variables to probabilistically evaluate the relationship between the selected candidate variable and the quality defects, thereby identifying the operational variable that is estimated to be a major factor.

[0017] (11) An operating condition change device according to an embodiment of the present disclosure includes an operating condition change instruction unit that changes the operating conditions of the operating variables identified by the quality factor estimation device of (9) or (10) by a change command.

[0018] According to the present disclosure, it is possible to provide a quality factor estimation method, an operating condition changing method, a model generation method, a quality factor estimation device, and an operating condition changing device that can estimate quality defect factors with high accuracy.

[0019] Fig. 1 is a block diagram showing an example configuration of a quality factor estimation device according to an embodiment of the present disclosure. Fig. 2 is a diagram illustrating changes in the accuracy rate of prediction. Fig. 3 is a diagram in which operational variables of a steelmaking model are arranged in order of variable importance. Fig. 4 is a diagram in which operational variables of a hot rolling model are arranged in order of variable importance. Fig. 5 is a diagram showing the relationship between the occurrence rate of steelmaking defects, the total length of the coil to be analyzed, and the factors of steelmaking defects and the estimated operational variables.

[0020] A quality factor estimation method, an operational condition modification method, a model generation method, a quality factor estimation device 10 (see FIG. 1 ), and an operational condition modification device according to embodiments of the present disclosure will be described below with reference to the drawings. The quality factor estimation method and the quality factor estimation device 10 according to the present embodiment use a model (mathematical model) generated to predict product quality in a manufacturing process for producing a product. The model is generated (constructed) by the model generation method according to the present embodiment. The quality factor estimation method and the quality factor estimation device 10 according to the present embodiment can be used in all processes in which operations determine quality, and by showing the relationship between multiple operational conditions and product quality, they are useful in identifying the causes of product defects.

[0021] Fig. 1 is a block diagram showing an example of the configuration of a quality factor estimation device 10 according to this embodiment. The quality factor estimation device 10 is used in a manufacturing process for manufacturing a product. Furthermore, the processing flow indicated by arrows, with each block in Fig. 1 treated as a process, represents a method (quality factor estimation method) for estimating factors of product quality defects in the manufacturing process according to this embodiment. The quality factor estimation method is executed by the quality factor estimation device 10.

[0022] For example, if the manufacturing process is a steel manufacturing process, the operational data may include a thermocouple temperature in a continuous casting process and a furnace temperature in a hot rolling process. The operational data may be given as a continuous value, a discrete value, or a categorical value. For example, if the manufacturing process is a steel manufacturing process, the quality data may be the type of quality defect and may be given as a categorical value.

[0023] Here, the operation data and the quality data may be associated with one product, or may be associated with each of a plurality of parts of one product. In this embodiment, one product is divided into N regions. Each region is assigned a quality variable a as quality data and an operation variable b = [b １ , b ２ , ..., b ｐ ] data is given. The quality variable a is a label that indicates whether the product falls under m types of quality defects or does not fall under a quality defect (normal quality). m is 2 or more. That is, the m types are two or more types. Therefore, a data set corresponding to one product can be obtained by accumulating quality data and operation data in each of N regions, and can be represented by a matrix of N rows and (1 + p) columns. Here, it is not necessary to divide one product into multiple regions as described above; for example, N can be 1. In other words, N is an integer greater than or equal to 1. Furthermore, since there are multiple operation variables b, p is an integer greater than or equal to 2.

[0024] In each of the N regions, the quality variable a corresponds to the operation variable b, and can generally be expressed as a = f(b) using a mapping function f. In the present disclosure, poor quality or normal quality (good quality) is predicted by a model (F) that uses the operation variable b as an explanatory variable and the quality variable a as a target variable.

[0025] The model (F) used for prediction generally represents a complex function in the entire space consisting of operational data, and is not a relational expression that is easy for humans to understand. Therefore, in this disclosure, a model (F) that classifies normal quality and quality defects for each of m types of quality defects is proposed. １ , F ２ , ..., F ｍ ) is defined. One model (F ｉ ) is used to classify (predict) that one quality defect from normal quality. ｉ ) has a variable importance that indicates the magnitude of the contribution of each element of the operational variable b. Here, when indicating one of m types of quality defects, it may be expressed as quality defect (i). The model corresponding to quality defect (i) is model (F ｉ )

[0026] As shown in FIG. 1 , the quality factor estimation device 10 includes a quality data acquisition unit 100, an operation data acquisition unit 101, a quality evaluation data creation unit 102, a variable importance determination unit 103, and a defect factor estimation unit 104. As in the present embodiment, the quality factor estimation device 10 may include an operation condition change instruction unit 105. The quality factor estimation device 10 may further include a storage unit (storage device) that stores, for example, quality data, operation data, and a model. The storage unit may include any storage device such as a semiconductor storage device, an optical storage device, or a magnetic storage device. The quality factor estimation device 10 may further include a model generation unit that executes a model generation method when generating a model.

[0027] The quality factor estimation apparatus 10 acquires quality data and operation data from an operation data server 60. The operation data server 60 is capable of communicating with the quality factor estimation apparatus 10 via a network and may be realized, for example, by a computer that manages a manufacturing process. The network is, for example, the Internet. In this embodiment, the display unit 30 displays the model evaluation results and the contributions of explanatory variables output from the quality factor estimation apparatus 10. For example, images such as those shown in FIGS. 2 to 4 (described later) may be displayed on the display unit 30 as the model evaluation results and the contributions of explanatory variables. The quality factor estimation apparatus 10 may be configured as a computer separate from the operation data server 60. The display unit 30 may be a device through which an operator working in the manufacturing process or the like obtains information. The display unit 30 may be a display device such as a liquid crystal display (LCD) or an organic electroluminescence panel (OLED). The display unit 30 may also be, for example, a touch panel display, and may also be used as an input unit through which an operator or the like issues instructions to change operating conditions (described later) based on the displayed information. The display unit 30 may be realized by a display of a terminal device such as a smartphone or tablet. The terminal device may also be a small computer such as a PC (Personal Computer).

[0028] As described above, the quality factor estimation device 10 can be realized by, for example, a computer. The computer includes, for example, a memory, a hard disk drive (storage device), and a CPU (processing device). The program can be stored in the hard disk drive and, when executed by the CPU, is read from the hard disk drive to the memory. The quality data acquisition unit 100, the operation data acquisition unit 101, the quality evaluation data creation unit 102, the variable importance determination unit 103, and the defect factor estimation unit 104 can be realized by, for example, a CPU that reads and executes a program. Furthermore, when the quality factor estimation device 10 includes a model generation unit, the model generation unit can be realized by, for example, a CPU that reads and executes a program.

[0029] The quality data acquiring unit 100 executes a process of acquiring quality data indicating the quality of the product for m types of quality defects (quality data acquiring step). In this embodiment, the quality data acquiring unit 100 receives information from a database (operation data server 60) in which quality data associated with the above-mentioned arbitrary areas is accumulated, and collects the data as one database or tabular data.

[0030] The operation data acquisition unit 101 executes a process of acquiring operation data of the manufacturing process (operation data acquisition step). In this embodiment, the operation data acquisition unit 101 receives information from a database (operation data server 60) in which operation data associated with the above-mentioned arbitrary area is accumulated, and collects the data as one database or tabular data.

[0031] The quality evaluation data creation unit 102 executes a process of creating quality evaluation data for each of the m types of quality defects (quality evaluation data creation step). The quality evaluation data for one quality defect is configured as a group of operation data corresponding to that one quality defect. In this embodiment, the quality evaluation data includes operation data corresponding to quality data indicating that the quality corresponds to that one quality defect and operation data corresponding to quality data indicating that the quality does not correspond to that one quality defect (normal quality). For example, if the quality defect is a defect, the quality evaluation data is configured to include operation data when an area (an area into which the product is divided) that has a defect and is determined to be of poor quality is manufactured, and operation data when an area without a defect and is determined to be of normal quality is manufactured.

[0032] The variable importance determination unit 103 acquires m types of models and executes a process (variable importance determination step) of determining the variable importance indicating the magnitude of the contribution of the operation variable b for each of the m types of models. In other words, the variable importance determination unit 103 determines the variable importance for each quality defect. The m types of models are models for predicting each of the m types of quality defects, and are generated using quality evaluation data with the product quality as the objective variable and the operation variable b included in the operation data as the explanatory variable. The variable importance determination unit 103 may acquire the m types of models by, for example, reading out the m types of models stored in the storage unit. The m types of models are generated by using m models (F １ , F ２ , ..., F ｍ ) and the quality assessment data (D １ , D ２ , ..., D ｍ ) and variable importance (FI １ , F.I. ２ , ..., FI ｍ ) model (F ｉ ) is the quality evaluation data corresponding to the quality evaluation data (D ｉ ) is sometimes expressed as quality evaluation data (D ｉ ) is a model (F １ , F ２ , ..., F ｍ ) and is used in the evaluation of the model (F ｉ ) is also used as model generation data (learning data) to generate the model (F ｉ ) is calculated as the variable importance (FI ｉ ) is sometimes expressed as follows.

[0033] Here, the generation of the model may be performed by a model generation device different from the quality factor estimation device 10, or may be performed by the quality factor estimation device 10. As described above, when the quality factor estimation device 10 generates a model, the quality factor estimation device 10 may further include a model generation unit. For example, the model generation device or the model generation unit generates quality evaluation data (D ｉ) is used as learning data with product quality as the objective variable and operation variable b included in the operation data as the explanatory variable. The model generation device or model generation unit uses the learning data to generate a model (F ｉ The model generation device or the model generation unit generates a model (F ｉ ) for each of the m types of quality defects, thereby generating m types of models (F １ , F ２ , ..., F ｍ ), where the model may be a machine learning model including a deep learning model, a decision tree, etc.

[0034] In addition, variable importance (FI ｉ ) is the p number of operational conditions, that is, the operational variable b = [b １ , b ２ , ..., b ｐ ] indicates the degree of influence on the prediction of quality defects (i) (quality judgment). ｉ ) may be calculated using a general-purpose prediction importance factor calculation method that does not depend on the algorithm for creating a numerical model, such as Permutation Importance. ｉ ) may be calculated in other ways. ｉ ) may be a trained model by machine learning as in this embodiment, but is not limited to this. ｉ ) is, for example, a linear regression model, ｉ ) may be calculated from the partial regression coefficients and the values ​​of the explanatory variables. The variable importance determination unit 103 calculates the variable importance for the p operating conditions by any method, and determines the order of magnitude of the variable importance (the magnitude of the influence on the quality judgment) (see FIGS. 3 and 4). Here, the variable importance may be any numerical value that allows a relative comparison of magnitude, and may be a value in arbitrary units.

[0035] The defect factor estimation unit 104 compares the variable importance of the m types of models to identify an operation variable b that is estimated to be a major factor for at least one of the m types of quality defects. ｉ ) a specific operating condition among p operating conditions is ranked high in the variable importance, and the model (F ｉ When a specific operational condition is not ranked highly in the variable importance in any model other than model (F), the specific operational condition can be estimated as a major factor in the quality defect (i). Furthermore, the defect factor estimation unit 104 may search for an operational variable that is commonly ranked highly in the variable importance of the corresponding model for two or more types of quality defects among the m types of quality defects, and may identify the operational variable b by regarding the searched operational variable as a common factor in the corresponding quality defects. ｉ ) a specific operating condition among p operating conditions is ranked high in the variable importance, and the model (F ｉ When a specific operating condition ranks high in the variable importance even in models other than the model (model 101), the specific operating condition can be estimated as a common factor. The defect factor estimation unit 104 can estimate the quality defect factors with high accuracy by comparing the variable importance of m types of models. Here, the process in which the defect factor estimation unit 104 estimates the factors of quality defects by comparing the variable importance of m types of models is sometimes referred to as a defect factor estimation step.

[0036] Here, the defect factor estimation unit 104 generates a model (F ｉ ) may be evaluated. The evaluation index may be, for example, the accuracy rate (Accuracy) used to evaluate a machine learning model. The accuracy rate (Accuracy) is calculated by defining a confusion matrix based on the pass / fail classification indicating actual pass / fail products and the truth or falsity of predictions by the model, and using each classification of true positive (TP), false positive (FP), false negative (FN), and true negative (TN). The calculation formula is Accuracy = ((TP + TN) / (TP + TN + FP + FN)).

[0037] The defect factor estimation unit 104 uses the model (F ｉ ) for the model (F ｊ The quality assessment data (D ｊ) to obtain the difference in the evaluation index (ΔR). In this embodiment, the evaluation index is the accuracy rate. The difference in the evaluation index (ΔR) is calculated by inputting the model (F ｉ ) to model (F ｉ The quality assessment data (D ｉ ) from the accuracy rate when the model (F ｉ ) to another model (F ｊ The quality assessment data (D ｊ ) is input. In other words, the difference in the evaluation index (ΔR) is the accuracy rate when the model (F ｉ ) is the difference in the accuracy rate of the prediction result for the quality defect (i) by the model (F). Here, j is an integer between 1 and m, different from i. When the magnitude of the difference in the evaluation index (ΔR) is greater than a predetermined value, the defect factor estimation unit 104 ｉ ) may be specified. The predetermined value is not limited to a specific value, but when the evaluation index is the accuracy rate as in this embodiment, 20% may be used as an example. The larger the absolute value of the difference (ΔR) in the evaluation index, the more likely it is that a factor specific to the quality defect (i) is included in the model (F ｉ ) variable importance (FI １ , F.I. ２ , ..., FI ｍ In other words, the smaller the absolute value of the difference in the evaluation index (ΔR), the more likely it is that the factor common to quality defect (i) and quality defect (j) is in the model (F ｉ ) variable importance (FI １ , F.I. ２ , ..., FI ｍ ), that is, it is considered difficult to identify a factor specific to the quality defect (i). The defect factor estimation unit 104 performs such processing using m models (F １ , F ２ , ..., F ｍ ) may be performed for each of the

[0038] Furthermore, the defect factor estimation unit 104 may further use performance data indicating the relationship between the m types of quality defects included in the operation data and the operation variable b to identify the operation variable b that is estimated to be the main factor for at least one of the m types of quality defects. The performance data indicating the relationship between the quality defects and the operation variable b may be accumulated, for example, in a database of the operation data server 60 and acquired by the quality factor estimation device 10 as part of the operation data. By using the performance data, the defect factor estimation unit 104 can estimate the quality defect factors with even greater accuracy. Furthermore, the defect factor estimation unit 104 may output the model evaluation results and the contribution of the explanatory variables to the display unit 30.

[0039] The method for identifying an operation variable b as a major factor for at least one quality defect (i) among m types of quality defects is performed as follows. First, multiple operation variables are selected as candidate factors based on the variable importance obtained by the defect factor estimation unit 104, which has a factor specific to the quality defect (i) ranked high. The relationship between these operation variables and the quality defect (i) is investigated, and one operation variable that can most likely explain the mechanism by which the quality defect (i) is generated is selected. The criteria for selecting the candidate factors are not limited. For example, a predetermined number of operation variables may be selected from the top of the importance rankings as candidate factors (candidate variables). The number of selected operation variables may be determined as a ratio, for example, 0.05 * p, where p is the number of operation variables. However, one or more variables must be selected. The method for investigating the relationship between the operation variables as candidate factors and the quality defect (i) is not limited. For example, correlation analysis between the probability of occurrence of the quality defect and the operation variables may be used. In other words, a probabilistic evaluation may be performed.

[0040] The quality factor estimation device 10 may further include an operational condition change instruction unit 105. The operational condition change instruction unit 105 outputs an instruction to change the operational variables identified by the defect factor estimation unit 104 as defect factor candidates for the quality defect (i). When the quality factor estimation device 10 includes the operational condition change instruction unit 105, the quality factor estimation device 10 may be referred to as an operational condition change device, focusing on the function of changing the operational conditions by issuing a change instruction to a control device 70 such as a process computer. When the quality factor estimation device 10 is referred to as an operational condition change device, a quality factor estimation method further including an operational condition change instruction step executed by the operational condition change device and changing the operational conditions by issuing a change instruction may be referred to as an operational condition change method.

[0041] The operational condition change instruction unit 105 outputs a change command when it receives an operational condition change instruction from an operator (e.g., an operator). The operational condition change instruction can be input using an input device (e.g., a keyboard, a pointer such as a mouse, or a touch panel) connected to a terminal device functioning as the display unit 30. In addition to a display for identifying operational variables that are candidate defect factors, the display unit 30 may display an input interface for specifying the operational variables to be changed and the method of change. The operator can issue a change command to the control device 70, such as a process computer, by inputting the operational variables to be changed and the method of change into the terminal device while referring to the displayed content. When changing the operational variables, the change method may be specified in advance for each operational variable. For example, a table may be created that defines the amount of change per operation for each target quality defect and each operational variable. When an operational variable is specified, the operational condition change instruction unit 105 may reference the table and output a change command. Furthermore, even for a single operational variable, the settings may be stratified according to manufacturing conditions such as the type of manufactured product or the size of the product. In this case, the input interface may be configured so that specific manufacturing conditions can be specified.

[0042] Here, the change of the operational variables by the operational condition change instruction unit 105 may be used only for temporary changes of the manufacturing conditions (during manufacturing tests). After the manufacturing conditions are determined, the program of the control device 70 such as a process computer may be modified and used in actual operation as a permanent determination of the operational variables.

[0043] EXAMPLES The effects of the present disclosure will be specifically described below based on examples, but the present disclosure is not limited to the contents of the examples.

[0044] An example will be described in which 33 operating conditions related to steelmaking in a steelmaking process, 15 operating conditions related to hot rolling, 5 operating conditions related to cold rolling, and 5 operating conditions related to surface treatment are defined as the operating variables b, and surface defects of automotive outer panels are used as the quality data to be predicted. The quality data is a categorical variable (categorical value) consisting of four types: "steelmaking defects," "hot rolling defects," "other defects" representing defects other than steelmaking or hot rolling, and "normal" representing no defects. Here, the occurrence mechanisms of steelmaking defects and hot rolling defects have been estimated based on past knowledge. It is believed that the occurrence mechanisms of steelmaking defects and hot rolling defects are different. In this example, two defects are individually identified, which corresponds to the case where the above m types are two.

[0045] The analysis targets a total of 477,236 records of operation data recorded for each meter of 164 coils collected via the process computer (operation data server 60). In the analysis, the collected operation data was standardized so that the mean and variance across the entire range of the data were 0 and 1, respectively.

[0046] 485 records in which steelmaking defects were recorded were extracted, and 485 records were randomly sampled from 473,178 records in which normal data were recorded, and the two sets of data were combined to create steelmaking defect prediction data (one set of quality evaluation data). Similarly, 732 records in which hot ductility defects were recorded were extracted, and 732 records were randomly sampled from 473,178 records in which normal data was recorded, and the two sets of data were combined to create hot ductility defect prediction data (another set of quality evaluation data). The operating conditions (operation variable b) included in the steelmaking defect prediction data and the hot ductility defect prediction data are common.

[0047] 2 is a diagram showing changes in the accuracy rate of prediction. The accuracy rate of a steelmaking model (one model) generated using steelmaking defect prediction data and the accuracy rate when hot ductility defect prediction data is input to the steelmaking model are shown. The accuracy rate of a hot rolling model (another model) generated using the hot ductility defect prediction data and the accuracy rate when steelmaking defect prediction data is input to the hot rolling model are also shown.

[0048] The accuracy rate of the steelmaking model was 90% when the prediction results were evaluated using data for predicting steelmaking defects, but was 61% when the prediction results were evaluated using data for predicting hot ductility defects. Because the accuracy rate dropped significantly to 29%, it is believed that the parameters in the steelmaking model are optimized to increase the contribution of explanatory variables that are significantly related only to steelmaking defects to the prediction. Furthermore, the accuracy rate of the hot rolling model was 87% when the prediction results were evaluated using data for predicting hot ductility defects, but was 73% when the prediction results were evaluated using data for predicting steelmaking defects. Although the accuracy rate dropped by 14%, the drop was smaller than that of the steelmaking model.

[0049] A process for estimating the causes of steelmaking defects was performed. Figure 3 shows the variable importance of the steelmaking model. As described above, in the variable importance of the steelmaking model, operational conditions related only to steelmaking defects are ranked highly in importance. Figure 4 shows the variable importance of the hot rolling model. Operational condition 1, which has the highest variable importance in the steelmaking model, is ranked ninth in the hot rolling model. Operational condition 2, which has the second highest variable importance in the steelmaking model, is ranked second in the hot rolling model. Furthermore, operational condition 3, which has the third highest variable importance in the steelmaking model, is ranked third in the hot rolling model. Because operational conditions 1 to 3 have high variable importance in both the steelmaking model and the hot rolling model, they were estimated to be common factors for steelmaking defects and hot rolling defects, rather than unique factors for steelmaking defects. Furthermore, operational conditions 4 to 6 were ranked 4th to 6th in variable importance in the steelmaking model, but were highly likely to be factors causing hot ductility defects based on past knowledge (performance data showing the relationship between quality defects and operational variable b), and were therefore excluded from the list of candidates for steelmaking defect factors. Operational conditions 7 and 8, which were ranked 7th and 8th in variable importance in the steelmaking model, were not ranked highly in variable importance in the hot rolling model and were not excluded based on past knowledge, so they were determined to be factors particularly related to steelmaking defects. Although it is possible to identify factors related to hot ductility defects using a similar method, in this example, only steelmaking defects were evaluated. Here, operational condition 8 was a specific water volume condition. Because specific water volume is an item that can be experimented with, additional verification was performed.

[0050] Figure 5 shows the frequency distribution according to the specific water content in a bar graph, and the relationship with the occurrence rate of steelmaking defects in a line graph. Here, since the operation data is recorded for each meter of the coil, the total length of the coil (right vertical axis) is used as the frequency. The occurrence rate of steelmaking defects corresponds to the left vertical axis. It was shown that when the specific water content was 148 [L / t], the occurrence rate of steelmaking defects increased sharply to 4%.

[0051] According to past findings, excessive cooling water is sprayed on the edge portions in the width direction of the slab compared to the center portion, which can lead to overcooling. Overcooling can cause slab buckling cracks, which are steelmaking defects. The estimation that the specific water content obtained in this example is particularly related to steelmaking defects was consistent with past findings.

[0052] Based on the analysis results, an experiment was conducted to optimize the cooling pattern in the manufacturing process of an automotive outer plate, in which the amount of cooling water sprayed on the edge portions in the width direction of the slab was reduced and the amount of cooling water sprayed on the center portion in the width direction of the slab was increased. That is, as a manufacturing test, the operational variables were changed by the operational condition change instruction unit 105, and the manufacturing conditions were finally determined. As a result, it was confirmed that overcooling on the edge portions was suppressed and the incidence of steelmaking defects was reduced.

[0053] As described above, the quality factor estimation method, operational condition modification method, model generation method, quality factor estimation device 10, and operational condition modification device according to the present embodiment can estimate quality defect factors with high accuracy by comparing the variable importance of m types of models. Furthermore, the quality factor estimation method, operational condition modification method, model generation method, quality factor estimation device 10, and operational condition modification device according to the present embodiment can present the model evaluation results and the contribution of the explanatory variables to the operator.

[0054] Although the embodiments of the present disclosure have been described based on the drawings and examples, it should be noted that the present disclosure is not limited to the above embodiments, and 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 be logically inconsistent, and multiple components or steps can be combined or divided into one. The embodiments of the present disclosure can also be realized as a storage medium on which a program executed by a processor included in an apparatus is recorded. It should be understood that these are also included within the scope of the present disclosure.

[0055] 10 Quality factor estimation device (operation condition change device) 30 Display unit 60 Operation data server 70 Control device 100 Quality data acquisition unit 101 Operation data acquisition unit 102 Quality evaluation data creation unit 103 Variable importance determination unit 104 Defect factor estimation unit 105 Operation condition change instruction unit

Claims

1. A quality factor estimation method executed by a quality factor estimation device used in a manufacturing process for manufacturing a product, comprising: a quality data acquisition step of acquiring quality data indicating quality of the product for normal quality and two or more m types of quality defects; an operation data acquisition step of acquiring operation data of the manufacturing process; a quality evaluation data creation step of creating, for each of the m types of quality defects, m types of quality evaluation data consisting of the quality data and the operation data corresponding to the relevant quality defect and the quality data and the operation data corresponding to the normal quality; a variable importance determination step of acquiring m types of models for predicting each of the m types of quality defects, the m types of models being generated using the quality evaluation data with the quality of the product as a response variable and operation variables included in the operation data as explanatory variables, and determining variable importance indicating the magnitude of contribution of the operation variables for each of the m types of models; and a defect factor estimation step of identifying the operation variable estimated to be a main factor for at least one of the m types of quality defects by comparing the variable importance of the m types of models.

2. The quality factor estimation method according to claim 1, wherein the defect factor estimation step inputs the quality evaluation data used in generating one model and the quality evaluation data used in generating another model into one model out of the m types of models, and identifies the operational variable that is estimated to be a major factor of the quality defect corresponding to the one model when the magnitude of the difference in accuracy of the prediction results is greater than a predetermined value.

3. The quality factor estimating method according to claim 1 or 2, wherein the defect factor estimating step searches for an operation variable that has a common high variable importance in a corresponding model for two or more of the m types of quality defects, and identifies the operation variable by regarding the searched-for operation variable as a common factor of the corresponding quality defects.

4. The quality factor estimating method according to any one of claims 1 to 3, wherein the defect factor estimating step selects, for at least one type of quality defect out of the m types of quality defects, a candidate operation variable estimated to be a major factor based on the variable importance, and further uses performance data showing a relationship between the quality defect and the operation variable to probabilistically evaluate a relationship between the selected candidate variable and the quality defect, thereby identifying the operation variable estimated to be a major factor.

5. An operational condition changing method comprising an operational condition change instruction step of changing the operational conditions by a change command for the operational variables identified by the quality factor estimation method according to claim 3 or 4.

6. A model generation method for generating a model used in the quality factor estimation method according to any one of claims 1 to 4, comprising the steps of: treating the quality evaluation data corresponding to the one quality defect as learning data having the product quality as a target variable and operation variables included in the operation data as explanatory variables; and generating a model corresponding to the one quality defect by machine learning using the learning data; and generating the m types of models by executing the step of generating the model for each of the m types of quality defects.

7. A quality factor estimation device used in a manufacturing process for manufacturing a product, comprising: a quality data acquisition unit that acquires quality data indicating quality of the product for normal quality and two or more m types of quality defects; an operation data acquisition unit that acquires operation data for the manufacturing process; a quality evaluation data creation unit that creates m types of quality evaluation data for each of the m types of quality defects, the quality evaluation data and the operation data corresponding to the relevant quality defect and the quality data and the operation data corresponding to the normal quality; a variable importance determination unit that is a model for predicting each of the m types of quality defects, the variable importance determination unit acquiring m types of models generated using the quality evaluation data with the quality of the product as a response variable and operation variables included in the operation data as explanatory variables, and determining variable importance indicating a magnitude of contribution of the operation variables for each of the m types of models; and a defect factor estimation unit that identifies the operation variable estimated to be a main factor for at least one of the m types of quality defects by comparing the variable importance of the m types of models.

8. The quality factor estimation device described in claim 7, wherein the defect factor estimation unit inputs the quality evaluation data used in generating one model and the quality evaluation data used in generating another model into one of the m types of models, and identifies the operational variable that is estimated to be a major factor of quality defects corresponding to the one model when the magnitude of the difference in accuracy of the prediction results is greater than a predetermined value.

9. A quality factor estimation device according to claim 7 or 8, wherein the defect factor estimation unit searches for an operation variable that commonly has a high variable importance in a corresponding model for two or more of the m types of quality defects, and identifies the operation variable by regarding the searched-for operation variable as a common factor of the corresponding quality defects.

10. A quality factor estimation device according to any one of claims 7 to 9, wherein the defect factor estimation unit selects candidates for operation variables estimated to be major factors for at least one of the m types of quality defects based on the variable importance, and further uses performance data indicating the relationship between the quality defects and the operation variables to probabilistically evaluate the relationship between the selected candidate variables and the quality defects, thereby identifying the operation variables estimated to be major factors.

11. An operational condition change device comprising an operational condition change instruction unit for changing operational conditions by a change command for the operational variables identified by the quality factor estimation device according to claim 9 or 10.

Images