Quality prediction device, quality prediction model generating device, quality prediction model generating method and program

Abstract

To generate a quality prediction model that can easily capture the relation between an operation variable and a quality variable even with multiple quality variables.SOLUTION: A quality prediction model generation device comprises: a quality prediction model generation unit that analyzes the relevance between operation data and quality data to generate a quality prediction model based on multiple data including operation data composed of multiple operation variables and quality data composed of multiple quality variables in a manufacturing process; and an output unit that outputs the quality prediction model. The operation variables include common operation variables that affect all the physical elements and individual operation variables that affect only some physical elements. The quality prediction model generation unit computes a common model relational expression using the common operation variables, also computes an individual model relational expression for each quality variable using the common operation variables and the individual operation variables, and further computes a relational expression, which expresses the relation between the operation data and the quality data in the whole area, as a quality prediction model based on the common model relational expression and the individual model relational expression.SELECTED DRAWING: Figure 13

Description

[Technical field]

[0001] The present invention relates to a quality prediction device, a quality prediction model generating device, a quality prediction model generating method, and a program for a manufacturing process, and in particular to a quality prediction device, a quality prediction model generating device, a quality prediction model generating method, and a program for a manufacturing process that are applicable even when the relationship between operation variables and quality has complex nonlinear characteristics and is multivariate in all processes in which quality is determined as an operation result. [Background technology]

[0002] Conventionally, in manufacturing processes where quality is determined based on operational conditions, well-known methods for predicting quality during the manufacturing process of a product include using a physical model created based on knowledge of the mechanism of quality defect occurrence, or a linear model obtained by applying multiple regression analysis to operational data and quality data (hereafter referred to as a multiple regression model). In methods that use such models, the operational data of the product is input into the model to calculate a predicted value of quality, and the predicted value is then evaluated to predict quality.

[0003] In addition, in the method disclosed in Patent Document 1, an operation variable space with operation data as a base vector is divided into several local regions based on operation data and quality data, and the relationship between the operation variables and quality in each local region is modeled using a linear formula that is easy for humans to intuitively understand. Then, an activity function that expresses the contribution rate, which is the weight when expressing the overall quality as a weighted linear sum of each local linear formula, as a function of the coordinates of the operation variable space is obtained from the operation data, and a mathematical model that expresses the relationship between the overall operation variables and quality is derived, thereby realizing an analysis method that presents the relationship between multivariate operation variables having complex nonlinear characteristics and quality in a format that is easy for humans to understand. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 5068637 [Non-patent literature]

[0005] [Non-Patent Document 1] Sakamoto, Yoshiyuki et al., Information Quantitative Statistics, Kyoritsu Shuppan (1983) [Non-Patent Document 2] Naoji Yoshitani, "Parameter Recursive Estimation Using Vector-Type Forgetting Factor and Its Application to Actual Plants," Transactions of the Society of Instrumentation, Control and Automation, vol. 25, No. 5, 579 / 585 (1989) [Non-Patent Document 3] Itoshi Kawaguchi, Introduction to Multivariate Analysis I, Morikita Publishing (1973), P.3-33 [Non-Patent Document 4] CM Bishop, Pattern Recognition and Machine Learning, Springer Japan Co., Ltd. (2007), pp. 142-144 [Non-Patent Document 5] Miyashita and Sasaki, Chemometrics: Chemical Pattern Recognition and Multivariate Analysis, Kyoritsu Shuppan, pp. 55-72 Summary of the Invention [Problem to be solved by the invention]

[0006] In the conventional method using multiple regression models, the analysis is performed by deriving correlation coefficients and regression models based on the premise that the operation and quality variables to be analyzed can be expressed by a single linear model, specifically a linear polynomial, in the entire operation range. Therefore, when analyzing operation variables and quality variables obtained from a nonlinear process in which multiple quality failure factors each have different characteristics, there was a problem that it was difficult to create a quality prediction model with the required accuracy and the relationship between the two could not be correctly grasped.

[0007] In the method disclosed in Patent Document 1, a plurality of division patterns are created to divide the entire region of the operation factor space or the local regions obtained by dividing the entire region, and for each division pattern, an activity function that represents a contribution rate that is a weight when expressing the relationship between the operation factors and quality in the entire operation factor space as a weighted linear sum of local relational expressions that express the relationship between the operation factors and quality in the local region, and an undetermined coefficient of the local relational expressions are calculated, and then a mathematical model of the entire operation factor space is constructed using the activity function and the local relational expressions. Then, for each division pattern, a model error that is the difference between the model prediction value of the mathematical model of the entire operation factor space and the quality data is calculated, and an optimal division pattern that minimizes the model error is selected.

[0008] However, in Patent Document 1, the quality variable to be predicted is limited to one numerical variable, i.e., a scalar quantity. When the quality variable to be predicted is composed of multiple numerical variables, i.e., a vector quantity, a quality prediction model is created for each numerical variable, i.e., each element of the vector. In this case, the division pattern of the operation factor space differs for each numerical variable of the quality variable, making it difficult to grasp the relationship between the operation variable and the quality variable.

[0009] Therefore, the present invention has been made in consideration of the above problems, and an object of the present invention is to provide a new and improved quality prediction device, quality prediction model generation device, quality prediction model generation method, and program that make it possible to easily grasp the relationship between operational variables and quality variables even when the quality variables to be predicted are multiple numerical variables, i.e., vector quantities. [Means for solving the problem]

[0010] In order to solve the above-mentioned problems, according to one aspect of the present invention, a quality prediction device for predicting product quality includes a quality prediction model creation unit that analyzes a relationship between operation data and quality data based on a plurality of data including operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product, and creates a quality prediction model that expresses the relationship, and a quality prediction unit that predicts product quality based on the quality prediction model, wherein the quality prediction model creation unit includes a division pattern candidate creation unit that creates a plurality of division pattern candidates that divide an entire region into a plurality of local regions, taking an area in which the operation variables of the operation data take values ​​as an entire region, and a local relational equation calculation unit that calculates, for each division pattern candidate, a local relational equation that expresses a relationship between the operation data and the quality data in each local region for each quality variable. and a quality prediction model determination unit that determines the relational equation of the division pattern candidate selected by the minimum error relational equation selection unit as a quality prediction model. The quality prediction model determination unit determines the relational equation of the division pattern candidate selected by the minimum error relational equation selection unit as a quality prediction model. The quality prediction model determination unit determines the relational equation of the division pattern candidate selected by the minimum error relational equation selection unit as a quality prediction model.

[0011] In order to solve the above-mentioned problems, according to another aspect of the present invention, a quality prediction model generation device for generating a quality prediction model includes a quality prediction model creation unit that analyzes a relationship between operation data and quality data based on a plurality of data including operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product, and creates a quality prediction model that represents the relationship, and an output unit that outputs the quality prediction model, in which the operation variables include common operation variables that affect all physical elements and individual operation variables that affect only some of the physical elements, and the quality prediction model creation unit creates a quality prediction model that represents the relationship based on a plurality of data including operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product. In one embodiment, a quality prediction model generation device is provided, the device having: a common model calculation unit that calculates, by using common operation variables, a common model relational equation common to quality variables that expresses the relationship between operation data and quality data common to physical elements; an individual model calculation unit that calculates, for each quality variable, by using the common operation variables and the individual operation variables, an individual model relational equation that expresses the relationship between operation data related to only some physical elements and the quality data; and a relational equation calculation unit that calculates, as a quality prediction model, a relational equation that expresses the relationship between operation data and quality data in a region where the common operation variables take values, based on the common model relational equation common to the quality variables and the individual model relational equation calculated for each quality variable.

[0012] For example, the manufacturing process may be a tandem mill, with the physical element being a single mill.

[0013] At least the common model equation may be a non-linear equation.

[0014] The non-linear equation may be, for example, an equation represented by a generalized linear model, a kernel method, or a neural network.

[0015] The quality prediction model generating device may further include an update processing unit that updates the individual model relational equations based on operation data and quality data of the product successively acquired from the manufacturing process.

[0016] The quality prediction model creation unit includes a division pattern candidate creation unit that creates a plurality of division pattern candidates by dividing an entire region into a plurality of local regions, the entire region being an area where the common operation variables take values; a minimum error relational equation selection unit that calculates a prediction error of the relational equation for each division pattern candidate based on quality prediction data calculated from the operation data using the relational equation calculated by the relational equation calculation unit and quality data corresponding to the operation data, and selects the relational equation of the division pattern candidate with the minimum prediction error; and a quality prediction model determination unit that determines the relational equation of the division pattern selected by the minimum error relational equation selection unit as a quality prediction model. The individual model calculation unit calculates, for each division pattern candidate created by the division pattern candidate creation unit, a local relational equation expressing a relationship between the operation data and the quality data in each local region, for each quality variable, using individual operation variables corresponding to physical elements to which the common operation variables in each local region and the quality index of the quality data correspond, and the common model calculation unit may calculate, for each division pattern candidate, an activity function common to the quality variables as a weight of each local relational equation in each local region of each division pattern candidate.

[0017] The quality prediction model determination unit determines the convergence of the prediction error of the selected division pattern, and if it determines that the prediction error has not converged, the division pattern candidate creation unit increases the number of divisions of the selected division pattern candidate and generates multiple new division pattern candidates, and if it determines that the prediction error has converged, it may determine the relational equation of the division pattern at the time of convergence as the quality prediction model.

[0018] The local relationship may be a linear polynomial with multiple operational variables as input variables.

[0019] The individual model calculation unit may calculate the coefficients of the local relational expressions by any one of multiple regression, stepwise method, sparse regression, and partial least squares method.

[0020] Here, when the quality prediction model generating device further includes an update processing unit that updates the individual model relational equations based on operation data and quality data of the product sequentially acquired from the manufacturing process, the update processing unit may include an operation data input unit that acquires operation data of the target product from the manufacturing process, a quality data input unit that acquires quality data of the target product from the manufacturing process, a common model calculation unit that calculates weights of each individual model relational equation based on the operation data input from the operation data input unit and the common model relational equation calculated in advance by the common model calculation unit, an individual model calculation unit that calculates individual model relational equations set in advance for each quality variable based on the operation data input from the operation data input unit, a quality prediction data calculation unit that calculates quality prediction data based on calculation results of each individual model relational equation by the individual model calculation unit and the weights of each individual model relational equation calculated by the common model calculation unit, and an individual model update unit that updates coefficients of each individual model relational equation based on the quality data input from the quality data input unit and the quality prediction data calculated by the quality prediction data calculation unit.

[0021] The individual model update unit also calculates a vector of data consisting of the product's operational variables used to update the linear polynomial in the local region k as v_ (k) (T), each element y of quality y_ in each local region k l The coefficients of the linear polynomial before updating corresponding to A l (k) As (T), the coefficient may be calculated using the following formula (101).

[0022]

number

[0023] Here, v_ is a subscript that has _ under v, T is a subscript that indicates that it is an updated coefficient or a product to be used for updating, T-1 is a subscript that indicates that it is a current coefficient, and A l (k) (T) is the coefficient of the updated linear polynomial, y l (T) is quality data, K l (k)(T) is the update rate matrix in the local region k and is expressed by the above formula (102). In formula (102), P l (k) (T-1) is the adaptive gain matrix for local region k.

[0024] Also, a quality prediction device may be provided that predicts the quality of a product based on a quality prediction model input from an output unit of the above-mentioned quality prediction model generation device.

[0025] Further, in order to solve the above-mentioned problems, according to another aspect of the present invention, there is provided a quality prediction model generating method for generating a quality prediction model, comprising: a quality prediction model generating step of analyzing a relationship between operation data and quality data based on a plurality of data each including operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product, and generating a quality prediction model representing the relationship; and an output step of outputting the quality prediction model, wherein the operation variables include common operation variables that affect all physical elements and individual operation variables that affect only some physical elements, and the quality prediction model generating step outputs a quality prediction model representing the relationship between the operation data and the quality data based on a plurality of data each including operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product. The quality prediction model generating method includes: a common model calculation unit step of calculating, using common operation variables, a common model relational equation common to quality variables expressing the relationship between operation data and quality data common to physical elements; an individual model calculation step of calculating, for each quality variable, an individual model relational equation expressing the relationship between operation data related to only some physical elements and the quality data, using the common operation variables and the individual operation variables; and a relational equation calculation step of calculating, as a quality prediction model, a relational equation expressing the relationship between operation data and quality data in a region where the common operation variables take values, based on the common model relational equation common to the quality variables and the individual model relational equation calculated for each quality variable.

[0026] In order to solve the above problem, according to another aspect of the present invention, there is provided a program for causing a computer to function as the above quality prediction device or quality prediction model generating device. Such a program is stored in a storage device of a computer, and is read and executed by a CPU of the computer to cause the computer to function as the above quality prediction device or quality prediction model generating device. There is also provided a computer-readable recording medium on which the program is recorded. The recording medium is, for example, a magnetic disk or an optical disk. Effect of the Invention

[0027] As described above, according to the present invention, even when the quality variables to be predicted are a plurality of numerical variables, i.e., vector quantities, it is possible to easily grasp the relationship between the operation variables and the quality variables. [Brief description of the drawings]

[0028] [Figure 1] 1 is a block diagram showing a configuration of a quality prediction device according to a first embodiment of the present invention. [Diagram 2] 11 is a flowchart showing a quality prediction process performed by the quality prediction device according to the embodiment. [Diagram 3] 11 is an explanatory diagram showing a process of creating a division pattern candidate by a division pattern candidate creating unit; FIG. [Figure 4] FIG. 11 is an explanatory diagram illustrating a procedure for creating division pattern candidates in a two-dimensional operation variable space. [Diagram 5] FIG. 13 is an explanatory diagram illustrating a procedure for creating division pattern candidates in a two-dimensional operation variable space divided into three parts. [Figure 6] 13 is a graph showing the distribution of local activity functions when a one-dimensional operational variable space is divided into four. [Figure 7] 13 is a graph showing the distribution of local activity functions when a two-dimensional operational variable space is divided into three. [Figure 8] FIG. 11 is a block diagram showing a configuration of a quality prediction device according to a second embodiment of the present invention. [Figure 9] 13 is a flowchart showing an update process of a local relational equation by the quality prediction device according to the embodiment. [Figure 10] FIG. 13 is a diagram showing a process of obtaining an activity function and a local relational equation using operation variables consisting of common operation variables and independent individual operation variables that are related to only one physical element among the individual operation variables and affect the quality variables, using a plurality of rolling mills as an example. [Figure 11] FIG. 13 is a diagram showing a process of obtaining an activity function and a local relational equation using operation variables consisting of common operation variables and common individual operation variables that are related to multiple physical elements among the individual operation variables and similarly affect the quality variables, using multiple rolling mills as an example. [Figure 12] 13 is a flowchart showing a quality prediction process performed by a quality prediction device according to a third embodiment of the present invention. [Figure 13] FIG. 13 is a block diagram showing a configuration of a quality prediction model generating device according to a fifth embodiment of the present invention. [Figure 14] 11 is a diagram for explaining the process of obtaining a common model relational expression and an individual model relational expression according to the embodiment in a simple case. FIG. [Figure 15] 11A and 11B are diagrams for explaining the process of obtaining a common model relational expression and an individual model relational expression according to the embodiment in a more complicated case. [Figure 16] 10 is a flowchart showing a quality prediction model generation process performed by the quality prediction model generation device according to the embodiment; [Figure 17] 1 is a block diagram showing an example of a hardware configuration of a quality prediction device and a quality prediction model generation device according to an embodiment of the present invention. [Figure 18] FIG. 1 is an explanatory diagram showing an outline of a hot rolling process for a steel sheet. [Figure 19] 1 is a graph showing the results of evaluating the RMSE of the model by cross-validation in Example A. [Figure 20] FIG. 11 is a diagram showing regression coefficients of a local relational expression of a quality prediction model constructed by the quality prediction device according to the first embodiment in Example B. [Figure 21]FIG. 13 is a diagram showing regression coefficients of a local relational expression of a quality prediction model constructed by a quality prediction device according to the third embodiment (or the fifth embodiment) in Example B. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0029] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In this specification and the drawings, components having substantially the same functional configurations are designated by the same reference numerals, and duplicated explanations will be omitted.

[0030] <1. First embodiment> [1-1. Overview] First, an overview of a quality prediction device according to a first embodiment of the present invention will be described. The quality prediction device according to this embodiment is a device that performs quality prediction based on a nonlinear model that represents the relationship between an input variable u_ (u_ is assumed to have _ under u) such as an operating condition and an output variable y such as a quality index. The features of this embodiment in constructing the nonlinear model will be described below in comparison with the conventional method according to Patent Document 1.

[0031] In the method disclosed in Patent Document 1, the input variable u is u1 to u p A p-dimensional vector (u1u 2 ··· u p ) T (The superscript T denotes the transpose of a matrix.) The input variable space is divided into M local regions. In each local region, the relationship between the input u_ and the output y is modeled linearly as shown in the following equation (2). This is called the local model, and y^ (k) (y^ is expressed as y with a ^ placed above it.)

[0032] The global model y^ is calculated by dividing each local model y^ by the following equation (1): (k) Then, a nonlinear activation function Φ (k) The activity function Φ (k) is a function of the points u_ in the input variable space, and the kth local model y^(k) In the area where it is dominant, it takes a value of 1, and in the area where it has little effect, it takes a value close to 0. This allows us to determine the local model y^ at point u_. (k) It is possible to express how much each contributes to the overall model y^.

[0033] Activity function Φ (k) Any function system can be used as long as it satisfies the normalization condition of the following formula (3) (i.e., the sum of all activity functions is 1 for any u). For example, the activity function can be a function obtained by substituting the normal distribution function expressed by the following formula (4) into formula (3).

[0034]

number

[0035] Here, in the method of determining model space division in Patent Document 1, processing starts from an initial state where no region division is performed, and the following processing [1] to [4] are repeated to search for a model with good accuracy. Then, when the modeling error becomes equal to or smaller than a predetermined error, or when the number of divisions reaches a predetermined number of divisions, the processing ends. Process [1]: Create multiple models by dividing the area into two for each operating factor. Process [2]: Select the most accurate two-part model. Step [3]: Select the area with the poorest accuracy as the subdivision area and create a three-division model. Process [4]: ​​The process of selecting from a plurality of division candidates and selecting a re-division region is repeated in sequence.

[0036] In process [1], the local model y^ (k) Coefficient a_ (k) =(a0 (k) a1 (k) a p (k) ) T is the actual value y of the quality index shown in the following formula (5). i and the local model prediction y^ (k) (u_ i) and the squared error J (k) is determined so that y^ is minimized. (k) (u_ i ) is expressed as the following formula (6): Here, N is the number of data items of the input variables u and output variables y used to construct the model, and i is a subscript representing this.

[0037]

number

[0038] On the other hand, in process [2], the actual value y of the quality index shown in the following formula (7) i and the local model prediction y (k) (u_ i The model that minimizes the square error J with y^ (k) (u_ i ) is expressed as the following equation (8).

[0039]

number

[0040] Here, in Patent Document 1, the output variable y is a scalar variable, but when applied to a manufacturing process having multiple quality indicators that are closely related to each other, it is necessary to create a quality prediction model for each quality indicator separately. For example, a correction coefficient for correcting the predicted value of the rolling load of each rolling stand of a rolling mill consisting of multiple rolling stands is set taking into consideration the operating conditions of the preceding and following rolling stands. In addition, for example, a correction coefficient for the predicted temperature of a steel plate in multiple cooling zones in a rolling process is also set taking into consideration the cooling conditions in other cooling zones. In this way, when creating a quality prediction model for a manufacturing process having multiple quality indicators that are closely related to each other, there is a problem that the division pattern of the operation factor space differs for each numerical variable of the quality variable, making it difficult to grasp the relationship between the operation variable and the quality variable.

[0041] Therefore, the inventors of the present application have considered treating a plurality of quality indicators as an output variable y_ of a vector. That is, the output variable y_ is expressed as a set of numerical variables y1 to y q A q-dimensional vector (y1y 2 ··· y q ) T Note that the output variable y_ does not necessarily correspond to the numerical variables y1 to y q It is not necessary to express it as a vector, and the quality data of the present invention may be composed of a plurality of quality variables.

[0042] In each local region, the input variable u_ and the output variable y l The relationship between (l=1,2,...,q) is expressed by the local linear model y l ^ (k) In addition, the overall model y^_=(y1^ y2^ ··· y q ^) T Each component y l ^ is the local model y l ^ (k) Then, a nonlinear activation function Φ (k) The activity function Φ (k) is y1^~y q ^ is common. In addition, Φ (k) (u_), μ (k) (u_) is expressed by the following equations (11) and (12), respectively.

[0043]

number

[0044] By combining the above equations (9) and (10), the overall model y^_ can be expressed as the following equation (13).

[0045]

number

[0046] In the method of determining the model space division according to this embodiment, similar to Patent Document 1, the process starts from an initial state where no region division is performed, and the following processes [1] to [4] are repeated to search for a model with good accuracy. The process ends when the modeling error falls below a predetermined error or when the number of divisions reaches a predetermined number of divisions. Process [1]: Create multiple models by dividing the area into two for each operating factor. Process [2]: Select the most accurate two-part model. Step [3]: Select the area with the poorest accuracy as the subdivision area and create a three-division model. Process [4]: ​​The process of selecting from a plurality of division candidates and selecting a re-division region is repeated in sequence.

[0047] In the above process [1], the local model y^ (k) =(y l ^ (k) y2^ (k) y q ^ (k) ) T Coefficient a of jl (k) The determination of (j=1, , p, l=1, , q) involves the actual value of the quality index y i _ and the local model predicted value y l ^ (k) (u_ i ) the following equation (14) can be used as the evaluation function.

[0048]

number

[0049] In addition, in process [2], the actual value of the quality index y i _ and the overall model predicted value y^_(u_ i ) to minimize the error with the model obtained by minimizing equation (16).

[0050]

number

[0051] Here, the L2 norm (|| || 2) was used to evaluate the model error. However, if the importance differs depending on the elements of the objective variable y_, the weighting between the elements w_ = (w1w2... w q ) T Introducing the norm (|| || w ) can also be used.

[0052]

number

[0053] In this manner, in the present embodiment, when the quality variables to be predicted are multiple numerical variables, i.e., vector quantities, a quality prediction model is created with a vector as the objective variable. In this case, in the present embodiment, a local relational expression expressing the relationship between the operation variables and the quality variables in the local regions obtained by dividing the operation factor space is created for each quality variable as shown in the above formula (10), thereby expressing the difference in the relationship between the operation variables that differ for each quality variable. On the other hand, the activity function expressing the weight of each local relational expression when calculating the relationship between the operation data and the quality data in the entire operation factor space is made common to each variable of the quality variables as shown in the above formula (11) based on the division pattern of the operation factor space, making it easier to grasp the relationship between the operation variables and the quality variables that change depending on the operation variables.

[0054] The configuration and processing of the quality prediction device according to this embodiment will be specifically described below.

[0055] [1-2. Configuration of quality prediction device] First, the configuration of a quality prediction device in a manufacturing process according to a first embodiment of the present invention will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of a quality prediction device 100 according to this embodiment.

[0056] The quality prediction device 100 in the manufacturing process according to the present embodiment is a device that predicts the quality of an object to be subjected to quality prediction. The quality prediction device 100 according to the present embodiment is suitable for predicting multiple quality indexes that are closely related to each other. For example, in the case of a steel process, the quality predicted by the quality prediction device 100 includes a correction coefficient for predicted rolling loads at multiple rolling mills and a correction coefficient for predicted temperatures at multiple cooling zones in hot rolling and cold rolling of thin plates.

[0057] 1, the quality prediction device 100 includes a data extraction unit 110, a division pattern candidate creation unit 120, an activity function calculation unit 130, a local relational expression calculation unit 140, a relational expression calculation unit 150, a minimum error relational expression selection unit 160, a quality prediction model determination unit 170, a quality prediction unit 180, and a database 190. The division pattern candidate creation unit 120, the activity function calculation unit 130, the local relational expression calculation unit 140, the relational expression calculation unit 150, the minimum error relational expression selection unit 160, and the quality prediction model determination unit 170 constitute a quality prediction model creation unit 102. The quality prediction model creation unit 102 analyzes the association between the operation data and the quality data based on a plurality of data including operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product, and creates a quality prediction model that represents the association. Based on the quality prediction model created by the quality prediction model creation unit 102, the quality prediction unit 180 predicts the quality of the product.

[0058] The data extraction unit 110 extracts multiple pieces of operation data in a manufacturing process and quality data corresponding to the operations from the database 190. The operation data includes multiple operation variables. For example, in a steelmaking process, the operation variables include the amounts of various elements in molten steel measured in a refining process, the rolling load and rolling speed in hot rolling or cold rolling, and the line speed and steel plate temperature in a cooling process. The quality data also includes multiple quality variables. For example, the quality variables include correction coefficients for predicting rolling loads in multiple rolling mills and correction coefficients for predicting temperatures in multiple cooling zones in hot rolling or cold rolling of thin plates. The data extraction unit 110 outputs the data extracted from the database 190 to the division pattern candidate creation unit 120.

[0059] The division pattern candidate creation unit 120 divides the entire region, which is the region in which the operation variables of the operation data take values, into multiple local regions, and performs processing to create multiple division pattern candidates. That is, the division pattern candidate creation unit 120 divides the entire region of the operation data input from the data extraction unit 110, and creates multiple division pattern candidates.

[0060] The division pattern candidate creating unit 120 according to the present embodiment sets a plurality of division pattern candidates generated by dividing the operation variables constituting the entire region as division pattern candidates for the entire region. Then, the division pattern candidate creating unit 120 outputs the generated plurality of division pattern candidates to the activity function calculating unit 130.

[0061] The activity function calculation unit 130 calculates an activity function for each of all the division pattern candidates input from the division pattern candidate creation unit 120, based on the division coordinate information of the operation variables representing the boundaries of the division regions. A local relational expression calculation unit 140, which will be described later, calculates each local relational expression in each local region of each division pattern candidate, and the activity function calculation unit 130 calculates an activity function common to the quality variables of the quality data as a weight of the local relational expression. The activity function calculation unit 130 outputs the calculated activity function to the local relational expression calculation unit 140 and the relational expression calculation unit 150.

[0062] The local relational equation calculation unit 140 creates a relational equation (local relational equation) that expresses the relationship between the operation data and the quality data for each local region in a mathematical expression. In this embodiment, a local relational equation based on a linear polynomial is used. When a linear polynomial is used, the undetermined coefficients of the linear polynomial are found by a method described below, and the local relational equation can be determined relatively easily compared to using a general mathematical expression. The local relational equation calculation unit 140 forms a data matrix for each local region, finds the undetermined coefficients of the local relational equation for each local region, and determines the local relational equation for each local region. The local relational equation calculation unit 140 outputs the determined local relational equation to the relational equation calculation unit 150.

[0063] The relational equation calculation unit 150 creates a relational equation between the operation data and the quality data in the entire region, using the activity function calculated by the activity function calculation unit 130 and the local relational equation obtained by the local relational equation calculation unit 140. The relational equation calculation unit 150 outputs the created relational equation between the operation data and the quality data to the minimum error relational equation selection unit 160.

[0064] The minimum error relational equation selection unit 160 calculates the prediction error of the relational equation for each division pattern candidate based on the quality prediction data calculated using the relational equation calculated by the relational equation calculation unit 150 and the quality data corresponding to the operation data, and selects the relational equation of the division pattern candidate that minimizes the prediction error. The prediction error is calculated as the difference between the error calculated from the relational equation between the operation data and the quality data created for each of all the division pattern candidates, and the error calculated from the relational equation calculated by the relational equation calculation unit 150. The minimum error relational equation selection unit 160 outputs the selected relational equation to the quality prediction model determination unit 170.

[0065] The quality prediction model determination unit 170 performs a convergence judgment of the prediction error of the relational expression selected by the minimum error relational expression selection unit 160 (hereinafter referred to as the "relational expression with the minimum error"), and determines the relational expression of the division pattern when the prediction error has converged as the quality prediction model. The quality prediction model determination unit 170 compares the prediction error of the relational expression with the minimum error with a preset evaluation reference value to judge whether or not a relational expression with sufficient accuracy has been constructed. When the quality prediction model determination unit 170 judges that the prediction error has converged, it outputs the relational expression with the minimum error as the quality prediction model to the quality prediction unit 180. At this time, the quality prediction model determination unit 170 may output the activity function and the coefficient of the local relational expression, which are information for expressing the relational expression representing the determined quality prediction model, to the quality prediction unit 180. On the other hand, when it is judged that the convergence of the prediction error is insufficient, the quality prediction model determination unit 170 instructs the division pattern candidate creation unit 120 to further divide the division pattern selected by the minimum error relational expression selection unit 160 to create a new division pattern candidate.

[0066] The quality prediction unit 180 predicts the quality of the product based on the quality prediction model input from the quality prediction model determination unit 170. The quality prediction unit 180 calculates quality prediction data using, for example, operation data separately input to the quality prediction model, and outputs the calculated quality prediction data to the outside. The quality prediction data output from the quality prediction unit 180 can be used, for example, as guidance to an operator or an input signal to a process control system. The quality prediction unit 180 may also represent the quality prediction model itself as visual information and display it on a display or the like. This makes it possible to visually present the association between operation data consisting of a plurality of operation variables and operation data consisting of a plurality of quality variables.

[0067] The database 190 is a storage unit that stores past operation data and quality data in the manufacturing process, manufacturing time in each manufacturing process, order information such as destination, product number for identifying the product, and the like, in association with each other. This information is input from outside at a predetermined timing to the database 190 and recorded. The information stored in the database 190 is extracted by the data extraction unit 110 and used to construct a quality prediction model.

[0068] [1-3. Quality prediction processing using quality prediction device] Next, the quality prediction process by the quality prediction device 100 according to this embodiment will be described in detail with reference to Figs. 2 to 7. Fig. 2 is a flowchart showing the quality prediction process by the quality prediction device according to this embodiment. Fig. 3 is an explanatory diagram showing the creation process of a split pattern candidate by the split pattern candidate creation unit 120. Fig. 4 is an explanatory diagram showing a procedure for creating a split pattern candidate in a two-dimensional operation variable space. Fig. 5 is an explanatory diagram showing a procedure for creating a split pattern candidate in a two-dimensional operation variable space divided into three. Fig. 6 is a diagram showing a local activity function distribution when a one-dimensional operation variable space is divided into four. Fig. 7 is a diagram showing a local activity function distribution when a two-dimensional operation variable space is divided into three.

[0069] p operational variables u1, u2, . . . , u p Given N pieces of operational data with p columns and n rows, the operational data is an N-by-p matrix. Corresponding to the operational data, q quality variables y1, y2, . . . , y q If N quality data are given, the quality data will be an N-row, q-column matrix. If operational data, which is an N-row, p-column matrix, and quality data, which is an N-row, q-column matrix, are given, the operational variables are expressed by linear algebra theory as u1~u p The quality variable y_=(y1y 2 ··· y q ) T is y1~y qIt can be considered as N points distributed in a q-dimensional quality variable space based on . Therefore, if quality is represented by the symbol y, the operational variables and quality variables can generally be expressed by the following equation (19) via a vector mapping function f(·).

[0070]

number

[0071] In the case of a normal manufacturing process, the mapping function f(·) is a nonlinear and multivariate complex function, and it is difficult to find an appropriate function formula that spans the entire operation variable space. Therefore, in this embodiment, multiple division pattern candidates are created to divide the entire domain consisting of the operation variables into multiple local domains, and a local relational formula y1^ that represents the relationship between the operation variables and the quality variables in each local domain is created. (k) (u_), and the activity function Φ, which represents the weight of the local relation equation when calculating the predicted value at each point in the entire region as a weighted linear sum of the local relation equations. (k) The overall relationship y^_=(y1^ y2^ ··· y q ^) T In addition, in the formula (20), Σ is the sum of the terms, and M is the number of local regions (number of divisions).

[0072]

number

[0073] In the quality prediction process by the quality prediction device 100 according to this embodiment, first, the data extraction unit 110 extracts operation data and quality data related to a target for quality prediction from the database 190 (step S100), as shown in Fig. 2. The data extraction unit 110 outputs the extracted operation data and quality data to the division pattern candidate creation unit 120.

[0074] Next, the division pattern candidate creation unit 120 creates a plurality of division pattern candidates from the entire region consisting of the operation variables in the operation data (step S102). In this embodiment, the division pattern candidates are created from the region of the operation variables. As shown in FIG. 3, the division pattern candidate creation unit 120 performs a process of dividing the undivided two-dimensional operation variable space into two local regions (M=2). Here, the division of the operation variable space is set by dividing the axis parallel to each operation variable. That is, as shown in FIG. 4, the two-dimensional operation variable space is made up of two operation variables u1 and u2. Therefore, the region 1-1 is divided parallel to the operation variable u1 axis and parallel to the operation variable u2 axis. In this way, the region 1-1 is divided into two local regions 2-1 and 2-2. By changing the division point setting, a plurality of division pattern candidates are created as shown in FIG. 3.

[0075] Also, if the operation variable space has already been divided into several local regions, the error of each local relational equation is calculated using equation (21), which is an error evaluation function, and the local region with the largest error is divided into two. As an example, FIG. 5 shows a procedure for dividing an operation variable space that has already been divided into three into four (M=4). As shown in FIG. 5, if the region with the largest error is region 3-2, a division point is set so as to divide region 3-2 into two on an axis parallel to the operation variable u1 or u2 axis. At this time, the remaining regions 3-1 and 3-3 are not divided.

[0076]

number

[0077] In Fig. 3, when the operation variable space is divided into two, if the area with the largest error is area 2-1, the division point is set so as to divide area 2-1 into two along an axis parallel to the operation variable u1 or u2 axis. In this way, for the operation variables, a plurality of new division pattern candidates are generated from the current division pattern.

[0078] The value of the division candidate point required for creating the division pattern candidate of the operation variable can be, for example, extracted from data of one operation variable, divided into a plurality of groups, and the value of the operation variable that is the boundary between each group is calculated for all operation variables, and used as the division candidate point. Specifically, for example, the data of the operation variable is divided into a plurality of groups using a clustering method, and the minimum and maximum values ​​of the data values ​​included in each group are calculated. Then, the average value of the maximum value of the group with the smaller data value and the minimum value of the group with the larger operation variable value among the adjacent groups is set as the division point value. Alternatively, if an operation operator or person in charge can set the range of values ​​of the data of the operation variables of the groups that can be considered to have the same operation level for the operation variable data, a manually set division candidate point may be used.

[0079] When the division pattern candidates are created, the activity functions, local relational expressions, and global relational expressions are calculated for each of the division pattern candidates (steps S104 to S108).

[0080] First, the activity function calculation unit 130 calculates the activity function for each of all the division pattern candidates obtained by the division pattern candidate creation unit 120 (step S104). As the activity function, any function that satisfies the normal condition expressed by the following formula (22) can be used.

[0081]

number

[0082] Specifically, for example, a normal distribution function μ (k) Based on this, the normal membership function defined by equation (24) can be used as the activity function.

[0083]

number

[0084] where c (k) j is the center point of the local region, σ (k) j represents the standard deviation of the normal distribution function. Figure 6 shows an example of the normal distribution function and activity function when a one-dimensional operation variable space is divided into four local regions. Figure 7 shows an example when a two-dimensional operation variable space is divided into three local regions. If we focus on the boundary region between the local regions, the activity functions in the regions on both sides of the boundary line smoothly overlap, so it is possible to express a situation in which multiple local relational expressions overlap to determine quality. Note that the activity function is common to each quality variable of the quality data.

[0085] Next, the local relational expression calculation unit 140 creates a relational expression that expresses the relationship between the operation data and the quality data for each local region in a mathematical formula (step S106). In this embodiment, the local relational expression calculation unit 140 calculates a data matrix X (k) The data matrix X is created. (k) is expressed by the following equation (25).

[0086]

number

[0087] Then, the local relational equation calculation unit 140 calculates the regression coefficients A of the regression equation for each local region k represented by the following equation (26) (i.e., the undetermined coefficients of the local relational equation) (k) In addition, the undetermined coefficient A of the local relational equation of the local region k is calculated. (k) is expressed by the following equation (27).

[0088]

number

[0089] Undetermined coefficient A (k) is calculated, for example, by multiple regression expressed by the following formula (28): The matrix Y is an N-row, q-column matrix representing the quality data as shown in formula (29).

[0090]

number

[0091] In addition, the undetermined coefficient A (k) In addition to the multiple regression represented by the above formula (28), for example, the stepwise method, sparse regression, partial least squares method, etc. may be used to calculate. The stepwise method is a method of creating a model by repeatedly adding and removing input variables of a linear model, i.e., input variables of a linear regression equation, by setting a certain test standard in order to reduce the number of input variables as much as possible and to make the prediction error of the linear regression equation small enough to be practically acceptable (see Non-Patent Document 3 above). In addition, sparse regression is a method of determining a regression coefficient by minimizing the weighted sum of a first term representing the magnitude of the prediction error of the linear regression equation and a second term representing the sum of the absolute values ​​of the regression coefficients in a linear model, and is a method of preventing the regression coefficient from becoming an unnecessarily large value and setting the regression coefficient of an input variable with a small influence to 0, thereby substantially excluding it from the linear model (see Non-Patent Document 4 above). Partial Least Squares (PLS) is a method for creating a model using a linear regression equation in which multiple latent variables are linearly combined with input variables in a linear model so that the latent variables are independent of each other, and the latent variables are used as input variables (see Non-Patent Document 5 above).

[0092] Returning to the explanation of FIG. 2, the relational expression calculation unit 150 uses the activity function and the local relational expression obtained by the activity function calculation unit 130 and the local relational expression calculation unit 140 to construct a relational expression between the operation data and the quality data shown in formula (20) for all the division patterns (S108). When the relational expression between the operation data and the quality data of formula (20) is constructed in the relational expression calculation unit 150, the minimum error relational expression selection unit 160 selects a relational expression with the smallest error from among the multiple division pattern candidates created by the division pattern candidate creation unit 120 (S110). A relational expression between the operation data and the quality data shown in formula (20) is created for all the multiple division pattern candidates created. The minimum error relational expression selection unit 160 uses this to evaluate the error with the error evaluation formula defined by the above formula (7). Then, the minimum error relational expression selection unit 160 selects the relational expression of the division pattern with the smallest error calculated by formula (7).

[0093] When one relational equation is selected by the minimum error relational equation selection unit 160, the quality prediction model determination unit 170 compares the error of the relational equation with the smallest error selected by the minimum error relational equation selection unit 160 with a preset evaluation reference value to determine whether learning is sufficient, i.e., whether the error has converged. Methods of convergence determination include, for example, a method of comparing the error of the relational equation with a convergence determination variable (evaluation reference value), a method of comparing the amount of change in the relational equation error with respect to an increment in local region division with a convergence determination variable (evaluation reference value), and a method of using an index that takes into account the number of divisions and the error, for example, an index that evaluates not only the learning error but also the number of local regions, such as Akaike's information index described in Non-Patent Document 1, and discontinuing division at the point when the index increases with respect to an increase in division.

[0094] The quality prediction model determination unit 170 evaluates the error using such a method of convergence judgment, and if the error is greater than the evaluation reference value, it judges that a relational equation that can explain the data with sufficient accuracy has not yet been constructed, and the convergence of the error is insufficient. In this case, the quality prediction model determination unit 170 instructs the division pattern candidate creation unit 120 to increase the division number M of the division pattern by one and perform a process of creating a division pattern candidate. Then, the process from the division pattern process creation to the calculation of the relational equation between operation and quality is repeated until it is judged by the learning error evaluation that a relational equation that can explain the data with sufficient accuracy has been constructed.

[0095] On the other hand, if the error is equal to or smaller than the evaluation reference value, the quality prediction model determination unit 170 determines that a relational expression capable of explaining the data with sufficient accuracy has been constructed, and determines the relational expression as a quality prediction model to be used in the quality prediction unit 180. Then, the quality prediction model determination unit 170 extracts the obtained quality prediction model, or an activity function, which is information for expressing the quality prediction model, and a coefficient of the local relational expression, and outputs them to the quality prediction unit 180. The quality prediction unit 180 uses the quality prediction model and separately input operation data to calculate quality prediction data from equation (20), and outputs the data to the outside. The quality prediction data calculated by the quality prediction unit 180 can be used, for example, as guidance to a quality prediction operator or an input signal to a process control system.

[0096] The quality prediction device 100 according to the first embodiment of the present invention and the quality prediction method using the same have been described above. According to the quality prediction device 100 according to this embodiment, when the quality variables to be predicted are multiple numerical variables, i.e., vector quantities, a quality prediction model is created in which the objective variable is a vector. In this case, the activity function representing the weight of each local relational expression when calculating the relationship between the operation data and the quality data is made common to each of the quality variables, making it easier to grasp the relationship between the operation variables and the quality variables.

[0097] <2. Second embodiment> Next, a quality prediction device 200 according to a second embodiment of the present invention will be described. The quality prediction device 200 according to this embodiment is connected to a manufacturing process 300, and can update the local relational equation according to the state of the manufacturing process 300. This makes it possible to quickly respond to any change in the relationship between the operation data and the quality data, and significantly shorten the time required for the prediction accuracy to be improved.

[0098] Hereinafter, the configuration of the quality prediction device 200 according to this embodiment and the update process of the local relational equation performed by this will be described in detail with reference to Figs. 8 and 9. Fig. 8 is a block diagram showing the configuration of the quality prediction device 200 according to this embodiment. Fig. 9 is a flowchart showing the update process of the local relational equation performed by the quality prediction device according to this embodiment. Fig. 8 shows only the processing units necessary for the update process of the local relational equation (i.e., the processing units constituting the update processing unit). Therefore, the quality prediction device 200 according to this embodiment can also include other processing units as described in the first embodiment.

[0099] [2-1. Configuration of quality prediction device] The quality prediction device 200 of this embodiment includes an operation data input unit 210, an activity function memory unit 220, an activity function calculation unit 230, a local relational equation calculation unit 240, a quality prediction data calculation unit 250, a quality prediction data output unit 260, a quality data input unit 270, and a local relational equation update unit 280, as processing units that constitute an update processing unit that updates the local relational equation based on product operation data and quality data successively acquired from the manufacturing process, as shown in FIG. 8.

[0100] The operation data input unit 210 receives operation data of the product collected from the manufacturing process 300 connected to the quality prediction device 200. Specifically, the operation data input unit 210 may be, for example, a keyboard, an OCR for reading a data sheet, or a computer for sequentially collecting and storing measurement signals from sensors installed in a factory and importing data via a LAN or the like at a preset timing. The operation data input by the operation data input unit 210 includes a plurality of operation variables. The operation data is subjected to calculation processing by each processing unit described below, whereby a calculation equivalent to Equation (20) is performed, and quality prediction data is calculated. The operation data input from the operation data input unit 210 is output to the activity function calculation unit 230 and the local relational expression calculation unit 240.

[0101] The activity function storage unit 220 stores the activity function Φ i That is, the activity function storage unit 220 stores information about the division pattern selected in the first embodiment.

[0102] The activity function calculation unit 230 calculates the weights of the local relational expressions when calculating quality prediction data, which is a weighted linear sum of the local relational expressions. The activity function calculation unit 230 calculates the weights of the local relational expressions using the operation data input from the operation data input unit 210 and the activity function represented by equation (24) stored in the activity function storage unit 220. Then, the activity function calculation unit 230 outputs the calculated weights of the local relational expressions to the quality prediction data calculation unit 250.

[0103] The local relational expression calculation unit 240 calculates the local relational expression based on the operation data input from the operation data input unit 210. The local relational expression calculation unit 240 stores the coefficients of the local relational expression in the local region k represented by the formula (2), and calculates the local relational expression based on the operation data as the operation variables u1 to u2 of the local relational expression in the formula (2) in the local region k. pto perform a calculation process of the local relational equation. The coefficient values ​​of this local relational equation are updated by a local relational equation update unit 280, which will be described later. When an equation other than a linear polynomial is used for the local relational equation, the local relational equation can be calculated by setting the numerical values ​​of the operation data to the variables corresponding to the operation variables of the equation. The local relational equation calculation unit 240 outputs the calculation results of the local relational equation calculated for each local region k to the quality prediction data calculation unit 250. The local relational equation is created for each quality variable of the quality data.

[0104] The quality prediction data calculation unit 250 calculates the relational equation between the operation data and the quality data, i.e., quality prediction data based on the quality prediction model, based on the weight of the local relational equation input from the activity function calculation unit 230 and the calculation result of the local relational equation for each local region k input from the local relational equation calculation unit 240. The quality prediction data calculation unit 250 calculates the product of the weight of the local relational equation and the calculation result of the local relational equation, and further calculates the sum of these products calculated for all local regions, thereby calculating quality prediction data based on the relational equation between the operation data and the quality data of equation (20). Then, the quality prediction data calculation unit 250 outputs the calculated quality prediction data to the quality prediction data output unit 260 and the local relational equation update unit 280.

[0105] The quality prediction data output unit 260 functions as an interface unit that outputs the quality prediction data calculated by the quality prediction data calculation unit 250 to the outside. The quality prediction data output unit 260 displays the quality prediction data, for example, as guidance information, or outputs it to a control system of the manufacturing process to be used for quality control. By displaying the quality prediction data, an operation operator can determine operation conditions to obtain a desired quality, and an inspector can prevent the shipment of defective products to customers by performing focused inspection of products that may be of poor quality. In addition, when using it for quality control, the control system calculates target values ​​of operation variables that are the operation terminals so that the quality is within a desired range based on the quality prediction data, and reflects the results in the operation of the manufacturing process, thereby suppressing the occurrence of quality defects.

[0106] The quality data input unit 270 inputs quality data extracted from the manufacturing process at the time when the quality of the product was measured. The quality data input unit 270 can use, for example, a keyboard, an OCR that reads a data sheet, or a computer that sequentially collects measurement signals from a sensor and transmits them via a LAN or the like to input the quality data. The quality data input unit 270 outputs the input quality data to the local relational equation update unit 280.

[0107] The local relational equation updating unit 280 updates the coefficients of the local relational equations used in the local relational equation calculation unit 240. The local relational equation updating unit 280 calculates a prediction error, which is the difference between the quality data and the quality prediction data, and updates the values ​​of the coefficients of the local relational equations based on this prediction error and the weights of the local relational equations for the operation data. Details of the update process of the local relational equations will be described later. After updating the coefficients of the local relational equations, the local relational equation updating unit 280 outputs the updated local relational equations to the local relational equation calculation unit 240.

[0108] [2-2. Local relational equation update process by quality prediction device] Next, the update process of the coefficient for the operation variables of the local relational equation according to this embodiment will be described with reference to FIG. 9. Here, it is assumed that the local relational equation uses the equation (2) composed of the operation variables. In the first process of the linear polynomial, the processes of steps S200 to S208 in FIG. 9 are performed. At this time, the update process of the local relational equation in step S212 is not performed, so the calculation of the local relational equation is performed based on the initial setting. At this time, it is assumed that the local relational equation is a relational equation with a predetermined accuracy. First, the operation data of the product collected from the manufacturing process 300 is input to the operation data input unit 210 of the quality prediction device 200 (step S200). For example, it is assumed that the first operation data related to the first product is input to the operation data input unit 210. The operation data input unit 210 outputs the input first operation data to the activity function calculation unit 230 and the local relational equation calculation unit 240.

[0109] The local relational expression calculation unit 240 performs calculation processing of the local relational expression based on the first operation data input from the operation data input unit 210 and the coefficients of the local relational expressions in each local region k stored in advance (step S202). The coefficients of the local relational expressions used in the local relational expression calculation unit 240 are updated by the local relational expression update unit 280, but when the quality prediction data of the manufacturing process 300 is calculated for the first time by the quality prediction device 200, the coefficients after update are not calculated by the local relational expression update unit 280, so that the coefficients of the local relational expressions are not updated. In this case, the initial values ​​of the coefficients of the local relational expressions in the local region k may be calculated, for example, by the method of the first embodiment so that the local relational expressions have a predetermined accuracy. The calculation results of the local relational expressions calculated by the local relational expression calculation unit 240 are output to the quality prediction data calculation unit 250.

[0110] On the other hand, the activity function calculation unit 230, to which the first operation data is input from the operation data input unit 210, calculates the weight of the local relational equation using the first operation data and the activity function (step S204). When the function formula defined by formula (24) is used for the activity function, the activity function calculation unit 230 sets numerical values ​​corresponding to the operation variables in the first operation data input from the operation data input unit 210, and calculates the weight of the local relational equation by performing the calculation of formula (23). The weight of the local relational equation calculated by the activity function calculation unit 230 is output to the quality prediction data calculation unit 250.

[0111] Thereafter, the quality prediction data calculation unit 250 calculates quality prediction data based on the relational equation between the first operation data and the first quality data for the first product based on the calculation result of the local relational equation for each local region k calculated in step S202 and the weight of the local relational equation calculated in step S204 (step S206). The quality prediction data calculation unit 250 calculates the product of the weight of the local relational equation and the calculation result of the local relational equation, and further calculates the quality prediction data based on the relational equation between the operation data and the quality data of equation (20) by taking the sum of these products calculated for all local regions. Then, the quality prediction data calculation unit 250 outputs the calculated quality prediction data to the quality prediction data output unit 260 and the local relational equation update unit 280.

[0112] The quality prediction data output unit 260 outputs the quality prediction data calculated by the quality prediction data calculation unit 250 to an output device or the like that notifies the external manufacturing process 300 or an operator (step S208). On the other hand, when the first quality data on the first product is input to the quality data input unit 270 (step S210), the local relational equation update unit 280 calculates the coefficients of a new local relational equation using the first quality data and the quality prediction data of the first product calculated in step S206, and updates the local relational equation (step S212). The local relational equation update unit 280 calculates a prediction error, which is the difference between the first quality data and the quality prediction data, and calculates the coefficient value of the new local relational equation based on the prediction error and the weight of each local relational equation for the first operational data.

[0113] Now, the product operation variable u_(T)=(u1(T) u2(T) … u p (T) T The vector constructed by the following equation (30) from the value of the activation function v_ (k) (T), each element y of quality y_ in local region k l Let A be the vector of coefficients of the linear polynomial before updating. l (k)The quality data of the product used for updating input from the quality data input unit 270 is represented as y_(T), and each element of the quality data is represented as y l (T), where Φ (k) (u_(T)) is the weight of the local relation equation in the local region k of the operational data u_(T) used for updating. l (k) (T) is expressed as in equation (31).

[0114]

number

[0115] In this case, each element y of quality y_ in local region k l The coefficient A of the local relation corresponding to l (k) (T) is updated by the following equation (32), where K l (k) (T) is the update rate matrix, calculated by equation (33).

[0116]

number

[0117] Here, P l (k) (T-1) is an adaptive gain matrix. The above formulas (32) and (33) are update formulas for the recursive least squares method described in, for example, Non-Patent Document 2. Note that the adaptive gain matrix P l (k) (T-1) is the next A l (k) In preparation for updating (T), it is updated using the following equation (34).

[0118]

number

[0119] where I is the identity matrix and Λ l (k)(T) is the coefficient vector A l (k) Forgetting factor λ corresponding to each element of (T) jl (k) It is a matrix expressed by equation (35) whose diagonal elements are the reciprocals of (j=1, ..., p, k=1, ..., M, l=1, ..., q). The forgetting factor is a coefficient for setting the extent to which the correction based on the latest operation and quality data is reflected in the correction amount when calculating the correction amount for updating the coefficient, and is calculated by equation (36) in this embodiment.

[0120]

number

[0121] Forgetting factor λ jl (k) (T) takes values ​​in the range of 0 to 1, and λ jl (k) The smaller the value of (T), the greater the forgetting effect, and the update matrix K l (k) (T) is highly evaluated. On the other hand, λ jl (k) When (T)=1, the forgetting effect is zero, and the coefficients are not updated regardless of the magnitude of the prediction error. jl is an adjustment coefficient for individually setting the forgetting effect for each operation variable and each quality variable, and may be empirically set based on physical knowledge for each operation variable and each quality variable, or may be adjusted while performing actual operation so as to improve the quality prediction accuracy.

[0122] When a function other than the linear polynomial of formula (2) is used for the local relational equation, the coefficients can be updated by performing calculations in the same manner as above using an update equation corresponding to the function of the local relational equation employed. When the relationship between the operation data and the quality data changes, the coefficients of the local relational equation are quickly updated by such an update equation for the local relational equation. Therefore, it is possible to significantly shorten the time required for the prediction accuracy to be improved. In this manner, the coefficients of the local relational equation can be updated. The output of the quality prediction data of the first product and the update process of the coefficients of the local relational equation by the processes of steps S200 to S212 have been described above.

[0123] Subsequently, when the second operation data and the second quality data regarding the second product are input, the quality prediction device 200 performs the same processing. When the second operation data is input from the manufacturing process 300 to the operation data input unit 210 (step S200), the operation data input unit 210 outputs the second operation data to the activity function calculation unit 230 and the local relational expression calculation unit 240. The local relational expression calculation unit 240 performs calculation processing of the local relational expression using the input second operation data and the local relational expression whose coefficient is updated by the first operation data and the first quality data of the first product (step S202). The calculation result of the local relational expression calculated by the local relational expression calculation unit 240 is output to the quality prediction data calculation unit 250. On the other hand, the activity function calculation unit 230 calculates the weight of the local relational expression using the second operation data and the activity function, and outputs it to the quality prediction data calculation unit 250 (step S204).

[0124] The quality prediction data calculation unit 250 calculates quality prediction data for the second product based on the calculation result of the local relational equation for each local region k calculated in step S202 and the weight of the local relational equation calculated in step S204 (step S206). The quality prediction data calculation unit 250 outputs the calculated quality prediction data to the quality prediction data output unit 260 and the local relational equation update unit 280.

[0125] The quality prediction data output unit 260 outputs the quality prediction data calculated by the quality prediction data calculation unit 250 to an external manufacturing process 300 or an output device that notifies an operator (step S208). On the other hand, when the second quality data on the second product is input to the quality data input unit 270 (step S210), the local relational equation update unit 280 calculates the coefficients of a new local relational equation using the second quality data and the second quality prediction data calculated in step S206, and updates the local relational equation (step S212). The local relational equation update unit 280 calculates a prediction error, which is the difference between the second quality data and the quality prediction data of the second product, and calculates the coefficient value of the new local relational equation based on the prediction error and the weight of each local relational equation for the second operation data. In this way, the coefficient of the local relational equation is updated using the quality prediction data of the second product calculated from the local relational equation updated based on the second operation data, the second quality data, and information on the first product.

[0126] Thereafter, when the third quality data and the quality data of the third product are input to the quality prediction device 200, the quality prediction data is calculated in the same manner, and the coefficients of the local relational expressions are updated. In this manner, by repeating the process of FIG. 9, the quality prediction data in the manufacturing process 300 can be accurately evaluated. Here, the first quality data of the first product is not necessarily input before the quality data of the second product is predicted. For example, it may be input at a time when the quality data of the hundredsth product is predicted, which is much later, so it is preferable to update the local relational expressions based on the latest quality data, quality prediction data, and operation data obtained at the time when the quality prediction data is calculated.

[0127] <3. Third embodiment> Next, an overview of a quality prediction device according to a third embodiment of the present invention will be described. The quality prediction device according to this embodiment has a similar configuration to the quality prediction device 100 according to the first embodiment shown in Fig. 1, but is different from the first embodiment in the process of creating a division pattern candidate by the division pattern candidate creating unit 120 and the process of creating a local relational expression by the local relational expression calculating unit 140. The quality prediction device according to this embodiment and the quality prediction method using the same will be described below, focusing on the differences from the first embodiment.

[0128] [3-1. Configuration of quality prediction device] First, the configuration of the quality prediction device in the manufacturing process according to this embodiment will be described. The quality prediction device according to this embodiment has a similar configuration to the quality prediction device 100 shown in Fig. 1 in the first embodiment, so the configuration of the quality prediction device 100 in this embodiment will also be described based on Fig. 1.

[0129] The quality prediction device 100 in the manufacturing process according to this embodiment is a device that predicts the quality of an object to be subjected to quality prediction. As in the first embodiment, the quality prediction device 100 according to this embodiment is suitable for predicting multiple quality indexes that are closely related to each other. In particular, the quality prediction device 100 according to this embodiment is suitable for a case in which multiple quality indexes each correspond to a physical element, and the operation variables used for prediction include a common operation variable that affects all physical elements, and an individual operation variable that affects only a part (one or more) of the physical elements.

[0130] For example, in the case of a steel process, the quality prediction device according to the present embodiment can be applied to quality prediction in hot rolling, cold rolling, etc. of thin plates. For example, when predicting correction coefficients for predicted rolling loads in multiple rolling mills as the quality predicted by the quality prediction device 100, the physical element is the rolling mill, the common operation variables are product dimensions and various element components, etc., and the individual operation variables are reduction, rolling speed, rolling load, etc. In addition, when predicting correction coefficients for predicted temperatures in multiple cooling zones as the quality predicted by the quality prediction device 100, the physical element is the cooling zone, the common operation variables are product dimensions and various element components, etc., and the individual operation variables of each cooling zone are the amount of cooling water, cooling speed, etc.

[0131] As shown in FIG. 1, such a quality prediction device 100 includes a data extraction unit 110, a division pattern candidate creation unit 120, an activity function calculation unit 130, a local relational equation calculation unit 140, a relational equation calculation unit 150, a minimum error relational equation selection unit 160, a quality prediction model determination unit 170, a quality prediction unit 180, and a database 190.

[0132] The data extraction unit 110 extracts a plurality of pieces of operation data in a manufacturing process and quality data corresponding to the operations from the database 190. The data extraction unit 110 functions in the same manner as the data extraction unit 110 in the first embodiment. The data extraction unit 110 outputs the data extracted from the database 190 to the split pattern candidate creation unit 120.

[0133] The division pattern candidate creation unit 120 divides the entire region, which is a region in which common operation variables that affect all physical elements among the operation variables of the operation data take values, into multiple local regions, and performs a process of creating multiple division pattern candidates. That is, the division pattern candidate creation unit 120 divides the entire region consisting of the common operation variables of the operation data input from the data extraction unit 110, and creates multiple division pattern candidates. The division pattern candidate creation unit 120 regards the multiple division pattern candidates generated by dividing the common operation variables that constitute the entire region as division pattern candidates for the entire region. Then, the division pattern candidate creation unit 120 outputs the created multiple division pattern candidates to the activity function calculation unit 130.

[0134] The activity function calculation unit 130 calculates an activity function for each of all the division pattern candidates input from the division pattern candidate creation unit 120 based on the division coordinate information of the common operation variables. A local relational expression calculation unit 140 (described later) calculates each local relational expression in each local region of each division pattern candidate, and the activity function calculation unit 130 calculates an activity function common to the quality variables of the quality data as a weight of the local relational expression. The activity function calculation unit 130 outputs the calculated activity function to the local relational expression calculation unit 140 and the relational expression calculation unit 150.

[0135] The local relational expression calculation unit 140 creates a relational expression (local relational expression) that expresses the relationship between the operation data and the quality data for each local region in a mathematical expression. The local relational expression calculation unit 14 according to this embodiment creates a local relational expression using only common operation variables that affect all physical elements among the operation variables of the operation data, and individual operation variables that affect the physical elements to which the quality index of the quality data corresponds. In this embodiment, the local relational expression can be determined relatively easily compared to using a general mathematical expression, so the local relational expression may be a linear polynomial. The local relational expression calculation unit 140 forms a data matrix for each local region, obtains undetermined coefficients of the local relational expression for each local region, and determines the local relational expression for each local region. The local relational expression calculation unit 140 outputs the determined local relational expression to the relational expression calculation unit 150.

[0136] The relational equation calculation unit 150 creates a relational equation between the operation data and the quality data in the entire region by using the activity function calculated by the activity function calculation unit 130 and the local relational equation obtained by the local relational equation calculation unit 140. The relational equation calculation unit 150 functions in the same manner as the relational equation calculation unit 150 in the first embodiment. The relational equation calculation unit 150 outputs the created relational equation between the operation data and the quality data to the minimum error relational equation selection unit 160.

[0137] The minimum error relational equation selection unit 160 calculates the prediction error of the relational equation for each division pattern candidate based on the quality prediction data calculated using the relational equation calculated by the relational equation calculation unit 150 and the quality data corresponding to the operation data, and selects the relational equation of the division pattern candidate that minimizes the prediction error. The minimum error relational equation selection unit 160 functions in the same way as the minimum error relational equation selection unit 160 in the first embodiment. The minimum error relational equation selection unit 160 outputs the selected relational equation to the quality prediction model determination unit 170.

[0138] The quality prediction model determination unit 170 performs a convergence judgment of the prediction error of the relational equation with the minimum error selected by the minimum error relational equation selection unit 160, and determines the relational equation of the division pattern when the prediction error has converged as a quality prediction model. The quality prediction model determination unit 170 functions in the same manner as the quality prediction model determination unit 170 of the first embodiment. When the quality prediction model determination unit 170 judges that the prediction error has converged, it outputs the relational equation with the minimum error as a quality prediction model to the quality prediction unit 180. On the other hand, when it judges that the convergence of the prediction error is insufficient, the quality prediction model determination unit 170 instructs the division pattern candidate creation unit 120 to further divide the division pattern selected by the minimum error relational equation selection unit 160 to create a new division pattern candidate.

[0139] The quality prediction unit 180 predicts the quality of the product based on the quality prediction model input from the quality prediction model determination unit 170. The quality prediction unit 180 functions similarly to the quality prediction unit 180 of the first embodiment. The quality prediction unit 180 calculates quality prediction data using, for example, operation data separately input to the quality prediction model, and outputs the calculated quality prediction data to the outside.

[0140] As in the first embodiment, the database 190 is a storage unit that stores past operation data and quality data in the manufacturing process, manufacturing time in each manufacturing process, order information such as destination, product number for identifying the product, and the like, in association with each other. These pieces of information are input from outside to the database 190 at a predetermined timing and recorded. The information stored in the database 190 is extracted by the data extraction unit 110 and used to construct a quality prediction model.

[0141] [3-2. Quality prediction processing using quality prediction device] Next, the quality prediction process by the quality prediction device 100 according to this embodiment will be described in detail with reference to Figs. 10 to 12. Fig. 10 is a diagram for explaining the process of obtaining an activity function and a local relational expression according to this embodiment in a simple case. Fig. 11 is a diagram for explaining the process of obtaining an activity function and a local relational expression according to this embodiment in a more complicated case. Fig. 12 is a flowchart showing the quality prediction process by the quality prediction device according to this embodiment.

[0142] In this embodiment, several individual operational variables affect the multiple quality indexes via the corresponding physical elements. In other words, the numerical variables y1 to y, which are the elements of the output y_ of the vector representing the multiple quality indexes, q For each of the individual operational variables, u (1) ~u_ (q) is defined. The common operation variables are vectors u_ (0) It is expressed as:

[0143] In a simple case, as shown in FIG. 10, the operational variables of the present embodiment are a common operational variable vector u_ (0) and a vector of individual operational variables u_ that affect each physical element individually (1) ~u_ (q) In a more complicated case, as shown in FIG. 11, the operational variables of the present embodiment include a common operational variable vector u_ (0) and a vector of individual operational variables u_ that affect each physical element individually. (1) ~u_ (q) Here, "individually affecting some physical elements" means that some physical elements are commonly affected, but other physical elements are not affected. Also, as shown in the figure, individual operational variables may have overlapping effects on the same physical element.

[0144] For example, the operational variable representing the tension between adjacent rolling mills in the rolling direction (e.g., F1 stand) and i+1 (e.g., F2 stand) is the operational variable u_ (i) is an individual operational variable vector, and the operational variable representing the tension between adjacent rolling mills in the rolling direction (e.g., F2 stand) and i+2 (e.g., F3 stand) is the operational variable u_ (i+1) Let be the individual operation variable vector. Then, u_ (i) andu_ (i+1) are individual operation variables that overlappingly affect the same physical element, such as rolling mill i+1. In the following, an individual operation variable that affects only one physical element shown in FIG. 10 is referred to as an "independent individual operation variable." In addition, an individual operation variable that affects multiple physical elements shown in FIG. 11 is referred to as a "common individual operation variable." When simply referring to an "individual operation variable," it is intended to include at least one of an "independent individual operation variable" and a "common individual operation variable."

[0145] In this embodiment, the operation variable and the quality variable yl The local relation y l ^ (k) The common operational variable vector u_ (0) and,y l The individual operational variable vector u_ that affects (l) In addition, the activity function Φ represents the weight of the local relational expressions when calculating the predicted value at each point in the entire region as a weighted linear sum of the local relational expressions. (k) is the common operational variable vector u_ (0) Therefore, the quality variable y l The overall relationship y^ that predicts l is expressed by the following formula (20').

[0146]

number

[0147] In addition, when modeling the local relational equation in each local region with a linear formula, the input u_ (0) =(u1 (0) u2 (0) ··· u p0 (0) ) T and u_ (l) =(u1 (l) u2 (l) ··· u pl (l) ) T and output y l ^ (k) The relationship between p0 and p0 is expressed by the following formula (2'). l are the common operation variables u_ (0) The number of individual operation variables u_ (l) The coefficients of the linear equation are u_ (0) The corresponding one is a jl (k) and u_ (l) The corresponding one is a~ jl (k) It is distinguished by expressing it as:

[0148]

number

[0149] In the quality prediction process by the quality prediction device 100 according to this embodiment, first, the data extraction unit 110 extracts operation data and quality data related to a target for quality prediction from the database 190 (step S300), as shown in Fig. 12. The data extraction unit 110 outputs the extracted operation data and quality data to the division pattern candidate creation unit 120.

[0150] Next, the division pattern candidate creation unit 120 selects common operation variables u_ among the operation variables u_ that affect all physical elements in common. (0) (Step S302), and the common operation variable u_ (0) (Step S304) The process of creating the split pattern candidates may be performed in the same manner as in step S102 in the first embodiment, and therefore a detailed description thereof will be omitted here.

[0151] When the division pattern candidates are created, the activity functions, local relational expressions, and global relational expressions are calculated for each of the division pattern candidates (steps S306 to S312).

[0152] First, the activity function calculation unit 130 calculates an activity function for each of all the split pattern candidates obtained by the split pattern candidate creation unit 120 (step S306). As in the first embodiment, any function that satisfies the normal condition expressed by the above formula (22) can be used as the activity function. Specifically, for example, a normal distribution function μ (k) Based on this, the normal membership function defined by the above formula (24) can be used as the activity function.

[0153] Next, the local relational expression calculation unit 140 creates a relational expression that mathematically expresses the relationship between the operation data and the quality data for each local region. In this embodiment, the quality variable y lThe operation variables used in the local relational equation differ depending on the quality variable y l For each, the common operation variable u_ (0) and the quality variable y l Individual operational variables u_ that affect (l) From these operational data and the activity function, the quality variable y l The data matrix X for each local region corresponding to l (k) is created (step S308). l (k) is expressed by the following equation (25').

[0154]

number

[0155] Then, the local relational expression calculation unit 140 calculates the quality variable y l For each local region k, the regression coefficients of the regression equation (i.e., the undetermined coefficients of the local relation equation) A l (k) (Step S310). Note that the undetermined coefficient A l (k) is expressed by the following equation (27').

[0156]

number

[0157] Undetermined coefficient A l (k) is calculated by multiple regression expressed by the following formula (28'). l is a column vector representing the quality data as shown in equation (29').

[0158]

number

[0159] In addition, the undetermined coefficient Al (k) may be calculated using a method such as a stepwise method, a sparse regression method, or a partial least squares method in addition to the multiple regression represented by the above formula (28').

[0160] Returning to the description of FIG. 12, the relational expression calculation unit 150 uses the activity function and the local relational expression calculated by the activity function calculation unit 130 and the local relational expression calculation unit 140 to construct a relational expression between the operation data and the quality data shown in the formula (20') for all the division patterns (S312). When the relational expression between the operation data and the quality data of the formula (20') is constructed in the relational expression calculation unit 150, the minimum error relational expression selection unit 160 selects a relational expression with the smallest error from among the multiple division pattern candidates created by the division pattern candidate creation unit 120 (S314). The relational expression between the operation data and the quality data shown in the formula (20') is created for all the multiple division pattern candidates. The minimum error relational expression selection unit 160 uses this to evaluate the error with the error evaluation formula defined by the above formula (7). Then, the minimum error relational expression selection unit 160 selects the relational expression of the division pattern with the smallest error calculated by the formula (7).

[0161] When one relational equation is selected by the minimum error relational equation selection unit 160, the quality prediction model determination unit 170 compares the error of the relational equation with the minimum error selected by the minimum error relational equation selection unit 160 with a preset evaluation reference value to determine whether learning is sufficient, that is, whether the error has converged (step S316). The convergence determination may be performed in the same manner as in the first embodiment.

[0162] If the error is greater than the evaluation reference value, the quality prediction model determination unit 170 determines that a relational equation that can explain the data with sufficient accuracy has not yet been constructed and that the convergence of the error is insufficient. In this case, the quality prediction model determination unit 170 instructs the division pattern candidate creation unit 120 to increase the division number M of the division pattern by one and perform a process of creating a division pattern candidate (step S318). Then, the process from the division pattern process creation to the calculation of the relational equation between operation and quality is repeated until it is determined by the learning error evaluation that a relational equation that can explain the data with sufficient accuracy has been constructed.

[0163] On the other hand, if the error is equal to or smaller than the evaluation reference value, the quality prediction model determination unit 170 judges that a relational expression capable of explaining the data with sufficient accuracy has been constructed, and determines the relational expression as a quality prediction model to be used in the quality prediction unit 180 (step S320). Then, the quality prediction model determination unit 170 extracts the obtained quality prediction model, or an activity function, which is information for expressing the quality prediction model, and a coefficient of the local relational expression, and outputs them to the quality prediction unit 180. The quality prediction unit 180 uses the quality prediction model and separately input operation data to calculate quality prediction data from equation (20'), and outputs it to the outside. The quality prediction data calculated by the quality prediction unit 180 can be used, for example, as guidance to a quality prediction operator or an input signal to a process control system.

[0164] The quality prediction device 100 according to the third embodiment of the present invention and the quality prediction method using the device have been described above. According to the quality prediction device 100 according to this embodiment, when the quality variables to be predicted are multiple numerical variables (i.e., vector quantities), each of which corresponds to multiple physical elements, and when the operation variables used for prediction also correspond to the physical elements, a quality prediction model is created with the objective variable being a vector. In this case, in this embodiment, the operation variables are divided into common operation variables that affect all the quality variables and individual operation variables that affect only some of the quality variables, and the domain division of the operation factor space is performed only using the common operation variables, and the local relational expressions are created only using the common operation variables and the individual operation variables that affect the quality variables. This makes it possible to prevent the creation of a physically unreasonable prediction model by using operation variables corresponding to other physical elements (e.g., rolling mills that are not adjacent) in the prediction of a quality variable corresponding to a physical element such as a rolling mill. By preventing the prediction model from becoming physically unreasonable, it becomes possible to present information that can appropriately grasp the relationship between the operation variables and the quality variables.

[0165] <4. Fourth embodiment> Next, a quality prediction device according to a fourth embodiment of the present invention will be described. The quality prediction device according to this embodiment is connected to the manufacturing process 300, similar to the quality prediction device 200 according to the second embodiment, and can update the local relational equation according to the state of the manufacturing process 300. This makes it possible to quickly respond to any changes in the relationship between the operation data and the quality data, and significantly shorten the time required for the prediction accuracy to be improved.

[0166] The quality prediction device according to this embodiment has a similar configuration to the quality prediction device 200 according to the second embodiment shown in Fig. 8, but differs from the second embodiment in the weight calculation process of the local relational equation by the activity function calculation unit 230, the calculation process of the local relational equation by the local relational equation calculation unit 240, and the update process of the local relational equation by the local relational equation update unit 280. The configuration of the quality prediction device according to this embodiment and the update process of the local relational equation performed thereby will be described in detail below, focusing on the differences from the second embodiment.

[0167] [4-1. Configuration of quality prediction device] First, the configuration of the quality prediction device in the manufacturing process according to this embodiment will be described. Note that the quality prediction device according to this embodiment has a similar configuration to the quality prediction device 200 shown in Fig. 8 in the second embodiment, so the configuration of the quality prediction device 200 in this embodiment will also be described based on Fig. 8.

[0168] The quality prediction device 200 of this embodiment includes an operation data input unit 210, an activity function memory unit 220, an activity function calculation unit 230, a local relational equation calculation unit 240, a quality prediction data calculation unit 250, a quality prediction data output unit 260, a quality data input unit 270, and a local relational equation update unit 280, as processing units that constitute an update processing unit that updates the local relational equation based on product operation data and quality data successively acquired from the manufacturing process, as shown in FIG. 8.

[0169] The operation data input unit 210 receives operation data of the product collected from the manufacturing process 300 connected to the quality prediction device 200. The operation data input unit 210 functions in the same manner as the operation data input unit 210 of the second embodiment. The operation data input by the operation data input unit 210 consists of a plurality of operation variables. The operation data is subjected to calculation processing by each processing unit described below, whereby a calculation equivalent to equation (20') is performed and quality prediction data is calculated. The operation data input from the operation data input unit 210 is output to the activity function calculation unit 230 and the local relational equation calculation unit 240.

[0170] The activity function storage unit 220 stores the activity function Φ i That is, the activity function storage unit 220 stores information about the division pattern selected in the third embodiment.

[0171] The activity function calculation unit 230 calculates the weights of the local relational expressions when calculating quality prediction data, which is a weighted linear sum of the local relational expressions. The activity function calculation unit 230 calculates the weights of the local relational expressions using the operation data of the common operation variables among the operation data input from the operation data input unit 210 and the activity function represented by equation (24) stored in the activity function storage unit 220. Then, the activity function calculation unit 230 outputs the calculated weights of the local relational expressions to the quality prediction data calculation unit 250.

[0172] The local relational expression calculation unit 240 calculates the local relational expression based on the operation data input from the operation data input unit 210. The local relational expression calculation unit 240 stores the coefficients of the local relational expression in the local region k represented by the formula (2'), and calculates the local relational expression based on the operation data input from the operation data input unit 210. (0) ~u p0 (0) and each element y of quality y_ l Individual operation variable u1 corresponding to (l) ~u pl (l) to perform a calculation process of the local relational equation. The coefficient values ​​of this local relational equation are updated by a local relational equation update unit 280, which will be described later. When an equation other than a linear polynomial is used for the local relational equation, the local relational equation can be calculated by setting numerical values ​​of operation data to variables corresponding to the common operation variables and individual operation variables of the equation. The local relational equation calculation unit 240 outputs the calculation results of the local relational equation calculated for each local region k to the quality prediction data calculation unit 250. The local relational equation is created for each quality variable of the quality data.

[0173] The quality prediction data calculation unit 250 calculates the relational expression between the operation data and the quality data, that is, the quality prediction data based on the quality prediction model, based on the weight of the local relational expression input from the activity function calculation unit 230 and the calculation result of the local relational expression for each local region k input from the local relational expression calculation unit 240. The quality prediction data calculation unit 250 functions in the same manner as the quality prediction data calculation unit 250 similar to the second embodiment. The quality prediction data calculation unit 250 calculates the product of the weight of the local relational expression and the calculation result of the local relational expression, and further calculates the sum of these products calculated for all the local regions, thereby calculating the quality prediction data based on the relational expression between the operation data and the quality data of the formula (20'). Then, the quality prediction data calculation unit 250 outputs the calculated quality prediction data to the quality prediction data output unit 260 and the local relational expression update unit 280.

[0174] The quality prediction data output unit 260 functions as an interface unit that outputs the quality prediction data calculated by the quality prediction data calculation unit 250 to the outside. The quality prediction data output unit 260 functions in the same manner as the quality prediction data output unit 260 of the second embodiment. The quality prediction data output unit 260 displays the quality prediction data as, for example, guidance information, or outputs the quality prediction data to a control system of the manufacturing process to be used for quality control.

[0175] The quality data input unit 270 inputs quality data extracted from the manufacturing process at the time when the quality of the product is measured. The quality data input unit 270 functions in the same manner as the quality data input unit 270 in the second embodiment. The quality data input unit 270 outputs the input quality data to the local relational equation update unit 280.

[0176] The local relation equation updating unit 280 updates the coefficients of the local relation equations used in the local relation equation calculation unit 240. The local relation equation updating unit 280 functions in the same manner as the local relation equation updating unit 280 of the second embodiment. After updating the coefficients of the local relation equations, the local relation equation updating unit 280 outputs the updated local relation equations to the local relation equation calculation unit 240.

[0177] [4-2. Local relational equation update process by quality prediction device] Next, the update process of the coefficient for the operation variables of the local relational equation according to this embodiment will be described. In this embodiment, the process is performed in the same manner as the update process of the coefficient for the operation variables of the local relational equation according to the second embodiment shown in Fig. 9. Therefore, hereinafter, the update process of the coefficient for the operation variables of the local relational equation according to this embodiment will be described based on Fig. 9.

[0178] Here, it is assumed that the local relational expression uses the equation (2') composed of operational variables. In the first processing of the linear polynomial, the processing of steps S200 to S208 in FIG. 9 is performed. The processing of steps S200 to S208 is performed in the same manner as in the second embodiment. Note that in the first processing of the linear polynomial, the update processing of the local relational expression in step S212 is not performed, and therefore the calculation of the local relational expression is performed based on the initial setting. At this time, it is assumed that the local relational expression is a relational expression with a predetermined accuracy.

[0179] First, product operation data collected from the manufacturing process 300 is input to the operation data input unit 210 of the quality prediction device 200 (step S200). For example, it is assumed that first operation data related to a first product is input to the operation data input unit 210. The operation data input unit 210 outputs the input first operation data to the activity function calculation unit 230 and the local relational equation calculation unit 240.

[0180] The local relational expression calculation unit 240 performs calculation processing of the local relational expression based on the first operation data input from the operation data input unit 210 and the coefficients of the local relational expressions in each local region k stored in advance (step S202). The coefficients of the local relational expressions used in the local relational expression calculation unit 240 are updated by the local relational expression update unit 280, but when the quality prediction data of the manufacturing process 300 is calculated for the first time by the quality prediction device 200, the coefficients after update are not calculated by the local relational expression update unit 280, so that the coefficients of the local relational expressions are not updated. In this case, the initial values ​​of the coefficients of the local relational expressions in the local region k may be calculated, for example, by the method of the third embodiment so that the local relational expressions have a predetermined accuracy. The calculation results of the local relational expressions calculated by the local relational expression calculation unit 240 are output to the quality prediction data calculation unit 250.

[0181] On the other hand, the activity function calculation unit 230, to which the first operation data is input from the operation data input unit 210, calculates the weight of the local relational equation using the first operation data and the activity function (step S204). When the function formula defined by formula (24) is used for the activity function, the activity function calculation unit 230 sets numerical values ​​corresponding to the operation variables in the first operation data input from the operation data input unit 210, and calculates the weight of the local relational equation by performing the calculation of formula (23). The weight of the local relational equation calculated by the activity function calculation unit 230 is output to the quality prediction data calculation unit 250.

[0182] Thereafter, the quality prediction data calculation unit 250 calculates quality prediction data based on the relational equation between the first operation data and the first quality data for the first product based on the calculation result of the local relational equation for each local region k calculated in step S202 and the weight of the local relational equation calculated in step S204 (step S206). The quality prediction data calculation unit 250 calculates the product of the weight of the local relational equation and the calculation result of the local relational equation, and further calculates the quality prediction data based on the relational equation between the operation data and the quality data of equation (20) by taking the sum of these products calculated for all local regions. Then, the quality prediction data calculation unit 250 outputs the calculated quality prediction data to the quality prediction data output unit 260 and the local relational equation update unit 280.

[0183] The quality prediction data output unit 260 outputs the quality prediction data calculated by the quality prediction data calculation unit 250 to an output device or the like that notifies the external manufacturing process 300 or an operator (step S208). On the other hand, when the first quality data on the first product is input to the quality data input unit 270 (step S210), the local relational equation update unit 280 calculates the coefficients of a new local relational equation using the first quality data and the quality prediction data of the first product calculated in step S206, and updates the local relational equation (step S212). The local relational equation update unit 280 calculates a prediction error, which is the difference between the first quality data and the quality prediction data, and calculates the coefficient value of the new local relational equation based on the prediction error and the weight of each local relational equation for the first operational data.

[0184] Now, the product operation variable u_(T)=(u1(T) u2(T) … u p (T) T From each element of quality y_, l The common operational variables u_ (0) (T) and the quality variable y l Individual operation variables u_ (l) (T) and the value of the activity function are used to construct a vector v_ l (k) (T), each element y of quality y_ in local region kl Let A be the vector of coefficients of the linear polynomial before updating. l (k) The quality data of the product used for updating input from the quality data input unit 270 is represented as y_(T), and each element of the quality data is represented as y l (T), where Φ (k) (u_(T)) is the weight of the local relation equation in the local region k of the operational data u_(T) used for updating. l (k) (T) is expressed as equation (31').

[0185]

number

[0186] In this case, each element y of quality y_ in local region k l The coefficient A of the local relation corresponding to l (k) (T) is updated by the following equation (32'), where K l (k) (T) is the update rate matrix, calculated by equation (33').

[0187]

number

[0188] Here, P l (k) (T-1) is an adaptive gain matrix. The above formulas (32') and (33') are update formulas for the recursive least squares method described in, for example, Non-Patent Document 2. Note that the adaptive gain matrix P l (k) (T-1) is the next A l (k) In preparation for updating (T), it is updated using the following equation (34').

[0189]

number

[0190] where I is the identity matrix and Λ l (k) (T) is the coefficient vector A l (k) Forgetting factor λ corresponding to each element of (T) jl (k) It is a matrix expressed by equation (35') whose diagonal elements are the reciprocals of (j=1, ..., p, k=1, ..., M, l=1, ..., q). The forgetting factor is a coefficient for setting the extent to which the correction based on the latest operation and quality data is reflected in the correction amount when calculating the correction amount for updating the coefficients, and is calculated by equation (36') in this embodiment.

[0191]

number

[0192] Forgetting factor λ jl (k) (T) takes values ​​in the range of 0 to 1, and λ jl (k) The smaller the value of (T), the greater the forgetting effect, and the update matrix K l (k) (T) is highly evaluated. On the other hand, λ jl (k) When (T)=1, the forgetting effect is zero, and the coefficients are not updated regardless of the magnitude of the prediction error. jl is an adjustment coefficient for individually setting the forgetting effect for each operation variable and each quality variable, and may be empirically set based on physical knowledge for each operation variable and each quality variable, or may be adjusted while performing actual operation so as to improve the quality prediction accuracy.

[0193] When a function other than the linear polynomial of formula (2') is used for the local relational equation, the coefficients can be updated by performing calculations in the same manner as above using an update equation corresponding to the function of the local relational equation employed. When the relationship between the operation data and the quality data changes, the coefficients of the local relational equation are quickly updated by such an update equation for the local relational equation. Therefore, it is possible to significantly shorten the time required for the prediction accuracy to be improved. In this manner, the coefficients of the local relational equation can be updated. The output of the quality prediction data of the first product and the update process of the coefficients of the local relational equation by the processes of steps S200 to S212 have been described above.

[0194] Subsequently, when the second operation data and the second quality data regarding the second product are input, the quality prediction device 200 performs the same processing. When the second operation data is input from the manufacturing process 300 to the operation data input unit 210 (step S200), the operation data input unit 210 outputs the second operation data to the activity function calculation unit 230 and the local relational expression calculation unit 240. The local relational expression calculation unit 240 performs calculation processing of the local relational expression using the input second operation data and the local relational expression whose coefficient is updated by the first operation data and the first quality data of the first product (step S202). The calculation result of the local relational expression calculated by the local relational expression calculation unit 240 is output to the quality prediction data calculation unit 250. On the other hand, the activity function calculation unit 230 calculates the weight of the local relational expression using the second operation data and the activity function, and outputs it to the quality prediction data calculation unit 250 (step S204).

[0195] The quality prediction data calculation unit 250 calculates quality prediction data for the second product based on the calculation result of the local relational equation for each local region k calculated in step S202 and the weight of the local relational equation calculated in step S204 (step S206). The quality prediction data calculation unit 250 outputs the calculated quality prediction data to the quality prediction data output unit 260 and the local relational equation update unit 280.

[0196] The quality prediction data output unit 260 outputs the quality prediction data calculated by the quality prediction data calculation unit 250 to an external manufacturing process 300 or an output device that notifies an operator (step S208). On the other hand, when the second quality data on the second product is input to the quality data input unit 270 (step S210), the local relational equation update unit 280 calculates the coefficients of a new local relational equation using the second quality data and the second quality prediction data calculated in step S206, and updates the local relational equation (step S212). The local relational equation update unit 280 calculates a prediction error, which is the difference between the second quality data and the quality prediction data of the second product, and calculates the coefficient value of the new local relational equation based on the prediction error and the weight of each local relational equation for the second operation data. In this way, the coefficient of the local relational equation is updated using the quality prediction data of the second product calculated from the local relational equation updated based on the second operation data, the second quality data, and information on the first product.

[0197] Thereafter, when the third quality data and the quality data of the third product are input to the quality prediction device 200, the quality prediction data is calculated in the same manner, and the coefficients of the local relational expressions are updated. In this manner, by repeating the process of FIG. 9, the quality prediction data in the manufacturing process 300 can be accurately evaluated. Here, the first quality data of the first product is not necessarily input before the quality data of the second product is predicted. For example, it may be input at a time when the quality data of the hundredsth product is predicted, which is much later, so it is preferable to update the local relational expressions based on the latest quality data, quality prediction data, and operation data obtained at the time when the quality prediction data is calculated.

[0198] <5. Fifth embodiment> Next, a quality prediction model generating device and a quality prediction model generating method according to a fifth embodiment of the present invention will be described. The quality prediction model generating device according to this embodiment is a generalized device of the quality prediction device 100 according to the third embodiment and the quality prediction device 200 according to the fourth embodiment.

[0199] That is, the quality prediction model generating device in the manufacturing process according to this embodiment is a device that generates a quality prediction model that predicts the quality of an object to be subjected to quality prediction, and generates a quality prediction model that is suitable for predicting multiple quality indexes that are closely related to each other. In particular, the quality prediction model generated by the quality prediction model generating device according to this embodiment is suitable for the case where multiple quality indexes correspond to physical elements, as in the above third and fourth embodiments, and the operation variables used for prediction include common operation variables that affect all physical elements and individual operation variables that affect only some (one or more) physical elements.

[0200] For example, in the case of a steel process, the quality prediction model generating device according to the present embodiment can be applied to quality prediction in hot rolling, cold rolling, etc. of thin plates. For example, when predicting a correction coefficient for a rolling load prediction in a plurality of rolling mills as the quality predicted by the quality prediction model, the physical element is the rolling mill, the common operation variables are the product dimensions and various element components, etc., and the individual operation variables are the reduction ratio, rolling speed, rolling load, etc. Also, for example, when predicting a correction coefficient for a predicted temperature in a plurality of cooling zones as the quality predicted by the quality prediction model, the physical element is the cooling zone, the common operation variables are the product dimensions and various element components, etc., and the individual operation variables of each cooling zone are the cooling water volume, cooling speed, etc.

[0201] In the third and fourth embodiments, the operation variable and the quality variable y l The local relation y l ^ (k) is expressed as a linear equation as shown in the above equation (2'), but in this embodiment, it may be either a linear equation or a nonlinear equation. l The overall relationship for predicting the activity function Φ (k) However, in this embodiment, it may be expressed in either a linear or non-linear manner. The quality prediction model generating device and the quality prediction model generating method using the same according to this embodiment will be described below.

[0202] [5-1. Configuration of the quality prediction model generation device] First, the configuration of a quality prediction model generating device according to this embodiment will be described with reference to Fig. 13. Fig. 13 is a block diagram showing the configuration of a quality prediction model generating device according to a fifth embodiment of the present invention. As shown in Fig. 13, the quality prediction model generating device 500 according to this embodiment includes a quality prediction model creating unit 510 and an output unit 520. The quality prediction model generating device 500 may also include a quality prediction unit 530 and an update processing unit 540.

[0203] The quality prediction model creation unit 510 analyzes the association between the operation data and the quality data based on a plurality of data including operation data consisting of a plurality of operation variables in the manufacturing process and quality data consisting of a plurality of quality variables related to the manufactured product, and creates a quality prediction model that represents the association. The quality prediction model creation unit 510 according to this embodiment corresponds to the quality prediction model creation unit 102 in FIG. 1.

[0204] As shown in FIG. 13, the quality prediction model creation unit 510 has a common model calculation unit 511, an individual model calculation unit 513, and a relational expression calculation unit 515.

[0205] The common model calculation unit 511 calculates a common model relational expression common to quality variables that express the relationship between operation data and quality data common to all physical elements, using the common operation variables. The common model calculation unit 511 corresponds to the activity function calculation unit 130 in FIG. 1. The common model relational expression is expressed as a linear or nonlinear expression. For example, a weighted average using the activity function shown in the above formula (20') is also an example of the common model relational expression. In addition, for example, a nonlinear expression expressed using a generalized linear model, a kernel method, a neural network including deep learning, or the like may be used as the common model relational expression. The common model calculation unit 511 outputs the calculated common model relational expression to the individual model calculation unit 513 and the relational expression calculation unit 515.

[0206] The individual model calculation unit 513 calculates an individual model relational expression that expresses the relationship between the operation data and the quality data related to only some physical elements for each quality variable using the common operation variables and the individual operation variables. In addition, only the individual operation variables that affect the target quality variable are used to calculate the individual model relational expression. The individual model calculation unit 513 corresponds to the local relational expression calculation unit 140 in FIG. 1. The individual model relational expression is expressed by a linear expression or a nonlinear expression. For example, the linear expression shown in the above formula (2') is also an example of the individual model relational expression. In addition, for example, a nonlinear expression expressed by using a generalized linear model, a kernel method, a neural network including deep learning, or the like may be used as the individual model relational expression. The individual model calculation unit 513 outputs the calculated individual model relational expression to the relational expression calculation unit 515.

[0207] The relational expression calculation unit 515 calculates a relational expression representing the relationship between the operation data and the quality data in a region where the common operation variables take values, based on the common model relational expression common to the quality variables and the individual model relational expressions calculated for each quality variable. The relational expression represents a quality prediction model. The relational expression calculation unit 515 corresponds to the relational expression calculation unit 150 in FIG. 1. The relational expression calculation unit 515 outputs the created relational expression between the operation data and the quality data to the output unit 520 as a quality prediction model.

[0208] The output unit 520 outputs the quality prediction model created by the quality prediction model creation unit 510 to various functional units or an external device. For example, the output unit 520 outputs the quality prediction model to the quality prediction unit 530 or an external quality prediction device 600.

[0209] The quality prediction unit 530 predicts the quality of the product based on the quality prediction model input from the output unit 520. The quality prediction unit 530 according to this embodiment corresponds to the quality prediction unit 180 in Fig. 1. The quality prediction unit 530 functions similarly to the quality prediction unit 180 in Fig. 1. The quality prediction unit 530 calculates quality prediction data using, for example, operation data separately input to the quality prediction model, and outputs the calculated quality prediction data to the outside.

[0210] 13, the quality prediction unit 530 is illustrated as one functional unit constituting the quality prediction model generating device 500, but the present invention is not limited to this example. For example, instead of the quality prediction unit 530, a quality prediction device 600 separate from the quality prediction model generating device 500 may be provided, and the quality prediction device 600 may have the function of the quality prediction unit 530. In this case, the quality prediction device 600 acquires the quality prediction model created by the quality prediction model generating device 500, and executes a process of predicting the quality of the product using the quality prediction model.

[0211] The quality prediction model generating device 500 according to this embodiment is connected to a database 190 that stores past operation data and quality data in a manufacturing process, order information such as manufacturing time in each manufacturing process and destination, product number for identifying a product, and the like, in association with each other. As shown in Fig. 1, the database 190 may be configured as a storage unit of the quality prediction model generating device 500. The operation data in the manufacturing process and the quality data corresponding to the operation recorded in the database 190 are extracted by a quality prediction model creating unit 510 and used to construct a quality prediction model.

[0212] Furthermore, when the quality prediction model generating device 500 according to this embodiment uses the above formula (20') as the common model relational equation and the above formula (2') as the individual model function, it functions similarly to the quality prediction device 100 according to the third embodiment. In this case, the quality prediction model creating unit 510 may include a division pattern candidate creating unit 120, a minimum error relational equation selecting unit 160, and a quality prediction model determining unit 170, similar to the quality prediction device 100 shown in Fig. 1. Note that these functional units function similarly to the third embodiment, and therefore detailed explanations are omitted here.

[0213] Returning to the description of FIG. 13, the update processing unit 540 updates the individual model relational equation based on the operation data and quality data of the product sequentially acquired from the manufacturing process 300. The update processing unit 540 is connected to the manufacturing process 300, and updates the individual model relational equation according to the state of the manufacturing process 300. This allows a quick response even when the relationship between the operation data and the quality data changes, and makes it possible to significantly shorten the time required to improve the prediction accuracy. Specifically, the update processing unit 540 functions in the same manner as the quality prediction device 200 shown in FIG. 8. Therefore, the update processing unit 540 may be a device separate from the quality prediction model generation device 500 shown in FIG. 13. In this case, the quality prediction model generation device 500 shown in FIG. 13 does not necessarily need to be connected to the manufacturing process 300.

[0214] [5-2. Quality prediction model generation process by the quality prediction model generation device] Next, the quality prediction model generation process by the quality prediction model generation device 500 according to this embodiment will be described in detail with reference to Figs. 14 to 16. Fig. 14 is a diagram for explaining the process of obtaining the common model relational equation and the individual model relational equation according to this embodiment in a simple case. Fig. 15 is a diagram for explaining the process of obtaining the common model relational equation and the individual model relational equation according to this embodiment in a more complicated case. Fig. 16 is a flowchart showing the quality prediction model generation process by the quality prediction model generation device 500 according to this embodiment.

[0215] In this embodiment, similarly to the third embodiment, several individual operational variables affect the multiple quality indexes via the corresponding physical elements. In other words, the numerical variables y1 to y, which are the elements of the output y_ of the vector representing the multiple quality indexes, q For each of the individual operational variables, u (1) ~u_ (q) is defined. The common operation variables are vectors u_ (0) It is expressed as:

[0216] In a simple case, as shown in FIG. 14, the operational variables of this embodiment are a common operational variable vector u_ (0) and a vector of individual operational variables u_ that affect each physical element individually (1) ~u_ (q) In a more complicated case, as shown in FIG. 15, the operational variables of this embodiment include a common operational variable vector u_ (0) and a vector of individual operational variables u_ that affect each physical element individually. (1) ~u_ (q) Here, "individually affecting some physical elements" means that some physical elements are commonly affected, but other physical elements are not affected. Also, as shown in the figure, individual operational variables may have overlapping effects on the same physical element.

[0217] 14 and 15, like FIG. 10 and FIG. 11, show a tandem rolling mill having multiple rolling mills as an example. That is, one rolling mill is a physical element, and the tandem rolling mills are a manufacturing process. In FIG. 14 and FIG. 15, the undetermined coefficient A of the local relational expression of the local region k shown in FIG. 10 and FIG. l (k) are shown as individual model elements, and the activation function Φ (k) are shown as common model elements.

[0218] The individual model elements are expressed as individual model relational expressions, which are linear or nonlinear. For example, the linear expression shown in the above formula (2') is an example of an individual model relational expression. Also, the nonlinear expression g shown in the following formula (2'') l (k) is also an example of an individual model relational equation. l (k) is expressed using, for example, a generalized linear model, a kernel method, a neural network including deep learning, or the like.

[0219]

number

[0220] The common model elements are also expressed as a common model relational expression of a linear or nonlinear expression. For example, the linear expression shown in the above formula (20') is an example of a common model relational expression. In addition, the nonlinear expression f shown in the following formula (20'') is also an example of a common model relational expression. The nonlinear expression f is expressed using, for example, a generalized linear model, a kernel method, a neural network including deep learning, or the like.

[0221]

number

[0222] In the quality prediction model generation process by the quality prediction model generation device 500 according to this embodiment, first, as shown in Fig. 16, the common model calculation unit 511 calculates a common model relational equation common to quality variables expressing the relationship between operation data and quality data common to all physical elements, using common operation variables (step S31). Operation data and quality data related to the target for which quality prediction is performed are extracted from the database 190. The common model calculation unit 511 outputs the common model relational equation calculated based on the extracted operation data and quality data to the individual model calculation unit 513 and the relational equation calculation unit 515.

[0223] Next, the individual model calculation unit 513 calculates an individual model relational equation that expresses the relationship between the operation data related to only some physical elements and the quality data for each quality variable, using the common operation variables and the individual operation variables (step S33). Operational data and quality data related to the target for which quality prediction is performed are extracted from the database 190. The operation data extracted here is only the operation data including the individual operation variables that affect the quality variable of the target. The individual model calculation unit 513 outputs the individual model relational equation calculated based on the extracted operation data and quality data to the relational equation calculation unit 515.

[0224] Then, the relational expression calculation unit 515 calculates a relational expression representing the relationship between the operation data and the quality data in the region where the common operation variables take values, based on the common model relational expression common to the quality variables and the individual model relational expressions calculated for each quality variable (step S35). The relational expression represents a quality prediction model. The relational expression calculation unit 515 outputs the created relational expression between the operation data and the quality data to the output unit 520 as a quality prediction model. The output unit 520 outputs the quality prediction model input from the relational expression calculation unit 515 to the quality prediction unit 530 or the quality prediction device 600 (step S37).

[0225] Thereafter, the quality prediction unit 530 or the quality prediction device 600 predicts the quality of the product based on the quality prediction model input from the output unit 520. The quality prediction unit 530 or the quality prediction device 600 calculates quality prediction data using, for example, operation data separately input to the quality prediction model, and outputs the calculated quality prediction data to the outside.

[0226] The quality prediction model generation process shown in Fig. 16 is a higher-level conceptualization of the quality prediction model generation process in the quality prediction process shown in Fig. 12. For example, when the above formula (20') is used as the common model relational expression and the above formula (2') is used as the individual model function, the quality prediction model generation process shown in Fig. 16 is substantially the same as the quality prediction model generation process in the quality prediction process shown in Fig. 12. At this time, the processes of steps S31 and S33 are represented by the processes of steps S300 to S318 in Fig. 12, and the process of step S35 is represented by the process of step S320 in Fig. 12.

[0227] In addition, when the quality prediction model generating device 500 according to this embodiment includes an update processing unit 540, the update processing of the individual model relational equation can be performed similarly to the fourth embodiment. For example, when the above formula (20') is used as the common model relational equation and the above formula (2') is used as the individual model function, the local relational equation update processing shown in FIG. 9 is substantially performed similarly to the fourth embodiment.

[0228] The quality prediction model generating device 500 according to the fifth embodiment of the present invention and the quality prediction model generating method using the same have been described above. According to the quality prediction model generating device 500 according to this embodiment, when the quality variables to be predicted are multiple numerical variables (i.e., vector quantities), each of which corresponds to multiple physical elements, and when the operation variables used for prediction also include those corresponding to the physical elements, a quality prediction model is generated in which the objective variable is a vector. In this case, in this embodiment, the operation variables are divided into common operation variables that affect all the quality variables and individual operation variables that affect only some of the quality variables, and a common model relational expression is created using the common operation variables, and an individual model relational expression is created using the common operation variables and the individual operation variables, to generate one quality prediction model.

[0229] By generating a quality prediction model by such multitask learning, the effects of operation variables of multiple related physical elements on quality are simultaneously learned, and the quality prediction accuracy can be improved. As a result, as in the third and fourth embodiments, it is possible to prevent the creation of a physically unreasonable prediction model, in which an operation variable corresponding to another physical element (e.g., a rolling mill that is not adjacent) is used to predict a quality variable corresponding to a physical element such as a rolling mill. By preventing the prediction model from becoming physically unreasonable, it is possible to present information that can appropriately capture the relationship between the operation variables and the quality variables.

[0230] <6. Hardware configuration> The quality prediction apparatuses 100, 200, 600 and the quality prediction model generating apparatus 500 in the manufacturing process according to the embodiments of the present invention can be realized by a computer. Fig. 17 shows an example of the configuration of a computer system 400 that can function as the quality prediction apparatus and the quality prediction model generating apparatus according to each embodiment of the present invention. The computer system 400 includes a CPU 401, a ROM 402, a RAM 403, a keyboard controller (KBC) 405, a CRT controller (CRTC) 406, a disk controller (DKC) 407, and a network interface controller (NIC) 408, which are connected to each other via a system bus 404 so as to be able to communicate with each other.

[0231] The CPU 401 executes software stored in the ROM 402 or HD 411, or software supplied from the FD 412, and generally controls each component connected to the system bus 404. That is, the CPU 401 reads out and executes a processing program according to a predetermined processing sequence from the ROM 402, HD 411, or FD 412, and performs control for realizing the functions of the quality prediction devices 100, 200, and 600 and the quality prediction model generation device 500 of this embodiment.

[0232] The RAM 403 functions as a main memory or a work area for the CPU 401. The KBC 405 controls instruction input from the KB 409 or a pointing device (not shown). The CRTC 406 controls the display of the CRT 410 which is a display unit. The DKC 407 controls access to the HD 411 and FD 412 which store the boot program, various applications, edit files, user files, network management programs, and predetermined processing programs in this embodiment. The NIC 408 exchanges data bidirectionally with devices or systems on the LAN 420.

[0233] In addition, the scope of the present invention also includes implementation by supplying a software program to a computer that describes the processes for realizing the functions of the means of the quality prediction device and quality prediction model generation device, which are embodiments of the present invention, and each process of the quality prediction device and quality prediction model generation device, and operating various devices in accordance with the program stored in the computer.

[0234] In this case, the software program itself realizes the processing functions of the quality prediction devices 100, 200, 600 and the quality prediction model generation device 500 of the present embodiment, and the program itself is included in the scope of the present invention. Note that a communication medium such as a computer network system that propagates and supplies a program as an electrical signal can also be used as a transmission medium for the program code.

[0235] Furthermore, means for supplying a program to a computer, such as a storage medium storing such a program, is also included in the scope of the present invention. Examples of storage media for storing such program code include flexible disks, hard disks, optical disks, magneto-optical disks, CD-ROMs, CD-Rs, magnetic tapes, non-volatile memory cards, ROMs, etc. EXAMPLES

[0236] Example A An example will be described below in which quality prediction is performed for a hot rolling process of thin steel plates, which are steel products, using the correction amount of the rolling load of each rolling mill in finish rolling as the quality, and the target plate width, target plate thickness, rolling conditions, and other operational conditions of the rolled material as operational variables.

[0237] Fig. 18 is a diagram showing an outline of the hot rolling process for thin steel sheets. After being extracted from a heating furnace, the base material to be rolled, called a slab, is rolled into a bar shape having a thickness of 40 to 50 mm in a roughing mill 10. The rolled material is then continuously rolled in a finishing mill 20 consisting of a plurality of rolling stands arranged in the sheet passing direction, each of which has a pair of work rolls 22 and a pair of backup rolls 24 supporting the work rolls 22, and then wound into a coil by a winder 30.

[0238] The finishing mill 20 is usually composed of about seven rolling stands, and the rolling load of each rolling stand is predicted using a rolling model so that the rolling load of each rolling stand is within an appropriate range for operation. However, since the accuracy of the rolling model is insufficient, a correction coefficient to be multiplied by the predicted load of the rolling model based on the rolling load in the past rolling results has been calculated in the past to ensure the accuracy of the predicted rolling load.

[0239] In this example, the prediction accuracy of the rolling load was compared between a quality prediction model for determining the correction coefficient constructed by the method according to the first embodiment (Example) and a quality prediction model constructed by the method of Patent Document 1 (Comparative Example).

[0240] Since the finishing mill 20 of the target hot rolling process is composed of seven rolling stands, there are seven correction coefficient values ​​for one rolled material. In the comparative example (the method of Patent Document 1), one quality prediction model was created for one correction coefficient, as shown in the following formula (37), and a total of seven quality prediction models were created.

[0241]

number

[0242] On the other hand, in the embodiment, as shown in the following formula (38), one quality prediction model can be created by treating the seven correction coefficients as a vector consisting of seven elements.

[0243]

number

[0244] The results of comparing the prediction accuracy of the rolling load between the working example and the comparative example are shown in Fig. 19. Fig. 19 shows the RMSE (root mean square error) of the model evaluated by cross-validation. The RMSE is calculated by the following formula (39).

[0245]

number

[0246] As shown in Fig. 19, the optimal number of divisions that minimizes the RMSE is 8 divisions in the Example, whereas it is 10 divisions in the Comparative Example. In addition, the RMSE value at the optimal number of divisions is also smaller in the Example than in the Comparative Example, and it was confirmed that a highly accurate quality prediction model can be obtained. This shows that a highly accurate quality prediction model can be obtained with a smaller number of divisions in the Example than in the Comparative Example.

[0247] Example B Next, coefficients of local relational expressions were examined for the quality prediction model constructed by the quality prediction device according to the first embodiment of the present invention and the quality prediction model constructed by the quality prediction device according to the third embodiment of the present invention (or the fifth embodiment of the present invention). As with Example A above, these quality prediction models are models that target the hot rolling process of thin steel plates, with the correction amount of the rolling load of each rolling mill in finish rolling being the quality, and with operational variables being the target plate width, target plate thickness, rolling conditions, and other operational conditions of the rolled material.

[0248] FIG. 20 shows the regression coefficients of the local relational equation (equation (2)) of the quality prediction model constructed by the quality prediction device according to the first embodiment. FIG. 21 shows the regression coefficients of the local relational equation (equation (2')) of the quality prediction model constructed by the quality prediction device according to the third embodiment (or the fifth embodiment of the present invention). In FIG. 20 and FIG. 21, the coefficients a_ (k) A portion of the above is shown.

[0249] 20 and 21, the number of regression coefficients is the same for the quality prediction model of the first embodiment and the quality prediction model of the third embodiment (or the fifth embodiment of the present invention). However, in the quality model of the first embodiment, sparse learning was performed, and as shown in FIG. 20, when looking at the regression coefficients corresponding to the quality variables of the F1 stand, some of the coefficients of the individual operation variables of the F2 stand that do not affect the F1 stand were non-zero. Also, when looking at the regression coefficients corresponding to the quality variables of the F2 stand, some of the coefficients of the individual operation variables of the F1 stand that do not affect the F2 stand were non-zero. In this way, the quality prediction model of the first embodiment may not be a physically rational model.

[0250] On the other hand, in the quality model of the third embodiment (or the fifth embodiment of the present invention), as shown in FIG. 21, when looking at the regression coefficients corresponding to the quality variables of the F1 stand, the coefficients of the individual operation variables of the F2 stand that do not affect the F1 stand are all 0. Also, when looking at the regression coefficients corresponding to the quality variables of the F2 stand, the coefficients of the individual operation variables of the F1 stand that do not affect the F2 stand are also all 0. In this way, the quality prediction model of the third embodiment (or the fifth embodiment of the present invention) is a physically rational model.

[0251] The quality prediction model according to the third embodiment (or the fifth embodiment of the present invention) is a physically rational model, and therefore, it is considered that, for example, by applying it to a process in which physical models are important, it is easy to analyze deviations from normal operation.

[0252] Although the preferred embodiment of the present invention has been described in detail above with reference to the accompanying drawings, the present invention is not limited to such an example. It is clear that a person having ordinary knowledge in the technical field to which the present invention pertains can conceive of various modified or altered examples within the scope of the technical ideas described in the claims, and it is understood that these also naturally belong to the technical scope of the present invention.

[0253] For example, in the above-mentioned embodiments, the quality prediction device is realized as a program on a computer, but the present invention is not limited to such an example. For example, it may be configured by hardware that combines a calculation device, a memory, etc. Furthermore, the operation-quality relationship analysis device of the present invention may be configured by a plurality of devices or may be configured by a single device.

[0254] Furthermore, in the above embodiment, the division pattern candidate generating unit divides the local region into two regions, but the present invention is not limited to this example, and the local region may be divided into three regions, for example.

[0255] The following configurations also fall within the technical scope of the present invention: The following configurations (1), (4)-(6), (10)-(12) correspond to the first embodiment, the following configurations (2) and (3) correspond to the third embodiment, and the following configurations (7)-(9) correspond to the second or fourth embodiment.

[0256] (1) A quality prediction device for predicting product quality, a quality prediction model creation unit that analyzes a relationship between operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to the manufactured product, based on the plurality of data, and creates a quality prediction model that represents the relationship; a quality prediction unit that predicts the quality of the product based on the quality prediction model; Equipped with The quality prediction model creation unit includes: a division pattern candidate creation unit that creates a plurality of division pattern candidates for dividing an area in which the operation variables of the operation data take values ​​into a total area into a plurality of local areas; a local relational equation calculation unit that calculates, for each of the division pattern candidates, a local relational equation that represents a relationship between the operation data and the quality data in each of the local regions, for each of the quality variables; an activity function calculation unit that calculates, for each of the division pattern candidates, an activity function common to the quality variables as a weight of each of the local relational expressions in each local region of the division pattern candidate; a relational expression calculation unit that calculates a relational expression that represents a relationship between operation data and quality data in the entire region based on the activity function common to the quality variables and a local relational expression calculated for each of the quality variables; a minimum error relational equation selection unit that calculates a prediction error of the relational equation for each of the division pattern candidates based on quality prediction data calculated using the relational equation calculated by the relational equation calculation unit from the operation data and the quality data corresponding to the operation data, and selects a relational equation for the division pattern candidate that minimizes the prediction error; a quality prediction model determination unit that determines the relational equation of the division pattern selected by the minimum error relational equation selection unit as the quality prediction model; The quality prediction device has

[0257] (2) The operational variables include common operational variables that affect all physical elements and individual operational variables that affect only some of the physical elements; the division pattern candidate creation unit creates a plurality of division pattern candidates for dividing an area in which the common operation variable takes a value into a total area into a plurality of local areas; The quality prediction device according to (1), wherein the local relational equation calculation unit calculates, for each of the division pattern candidates, the local relational equation for each of the quality variables by using, as the operation data, the common operation variables in each of the local regions and the individual operation variables corresponding to physical elements to which quality indexes of quality data correspond.

[0258] (3) The quality prediction device described in (2) above, wherein the individual operation variables include at least one of an independent individual operation variable that affects only one physical element and a common individual operation variable that affects multiple physical elements.

[0259] (4) The quality prediction model determination unit determining whether or not the prediction error of the division pattern selected by the minimum error relational expression selection unit has converged; When it is determined that the prediction error has not converged, the division pattern candidate creation unit increases the number of divisions of the division pattern candidate selected by the minimum error relational equation selection unit to generate a plurality of new division pattern candidates; The quality prediction device according to any one of (1) to (3) above, wherein, when it is determined that the prediction error has converged, a relational expression of the division pattern at the time of convergence is determined as the quality prediction model.

[0260] (5) The quality prediction device according to any one of (1) to (4) above, wherein the local relational expression is a linear polynomial expression having a plurality of the operational variables as input variables.

[0261] (6) The quality prediction device according to any one of (1) to (5) above, wherein the local relational equation calculation unit calculates coefficients of the local relational equation for each of the local regions by any one of multiple regression, stepwise regression, sparse regression, and partial least squares.

[0262] (7) The quality prediction device according to any one of (1) to (6) above, further comprising an update processing unit that updates the local relational expression based on product operation data and quality data successively acquired from the manufacturing process.

[0263] (8) The update processing unit: an operation data input unit for acquiring operation data of a target product from the manufacturing process; a quality data input unit for acquiring quality data of a target product from the manufacturing process; an activity function calculation unit that calculates a weight of each of the local relational expressions based on the operation data input from the operation data input unit and the activity function calculated in advance by the activity function calculation unit; a local relational equation calculation unit that calculates the local relational equation that is preset for each of the local regions based on the operation data input from the operation data input unit; a quality prediction data calculation unit that calculates quality prediction data based on a calculation result of each of the local relational expressions by the local relational expression calculation unit and a weight of each of the local relational expressions calculated by the activity function calculation unit; a local relational equation updating unit that updates coefficients of the local relational equations based on the quality data input from the quality data input unit and the quality prediction data calculated by the quality prediction data calculation unit; The quality prediction device according to (7) above,

[0264] (9) The local relation update unit calculates a vector of data consisting of operational variables of a product used to update the linear polynomial in the local region k as v_ (k) (T), each element y of quality y_ in each local region k l The coefficients of the linear polynomial before updating corresponding to A l (k) The quality prediction device according to (8) above, wherein the coefficient (T) is calculated using the following formula (101).

[0265]

number

[0266] (10) A quality prediction method for predicting product quality, comprising: a quality prediction model creation step of analyzing a relationship between the operation data and the quality data based on a plurality of data including the operation data consisting of a plurality of operation variables in a manufacturing process and the quality data consisting of a plurality of quality variables related to the manufactured product, and creating a quality prediction model that represents the relationship; a quality prediction step of predicting a quality of the product based on the quality prediction model; Including, The quality prediction model creating step includes: a division pattern candidate creating step of creating a plurality of division pattern candidates for dividing an area in which the operation variables of the operation data take values ​​into a total area into a plurality of local areas; a local relational equation calculation step of calculating, for each of the division pattern candidates, a local relational equation that represents a relationship between the operation data and the quality data in each of the local regions, for each of the quality variables; an activity function calculation step of calculating, for each of the division pattern candidates, an activity function common to the quality variables as a weight of each of the local relational expressions in each local region of the division pattern candidate; a relational equation calculation step of calculating a relational equation representing a relationship between operation data and quality data in the entire region based on the activity function common to the quality variables and a local relational equation calculated for each of the quality variables; a minimum error relational equation selection step of calculating a prediction error of the relational equation for each of the division pattern candidates based on quality prediction data calculated from the operation data using the relational equation calculated in the relational equation calculation step and the quality data corresponding to the operation data, and selecting a relational equation of the division pattern candidate with the minimum prediction error; a quality prediction model determination step of determining the relational equation of the division pattern selected by the minimum error relational equation selection step as the quality prediction model; A quality prediction method comprising:

[0267] (11) A program for causing a computer to function as a quality prediction device for predicting product quality, comprising: a quality prediction model creation unit that analyzes a relationship between operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to the manufactured product, based on the plurality of data, and creates a quality prediction model that represents the relationship; a quality prediction unit that predicts the quality of the product based on the quality prediction model; Equipped with The quality prediction model creation unit includes: a division pattern candidate creation unit that creates a plurality of division pattern candidates for dividing an area in which the operation variables of the operation data take values ​​into a total area into a plurality of local areas; a local relational equation calculation unit that calculates, for each of the division pattern candidates, a local relational equation that represents a relationship between the operation data and the quality data in each of the local regions, for each of the quality variables; an activity function calculation unit that calculates, for each of the division pattern candidates, an activity function common to the quality variables as a weight of each of the local relational expressions in each local region of the division pattern candidate; a relational expression calculation unit that calculates a relational expression that represents a relationship between operation data and quality data in the entire region based on the activity function common to the quality variables and a local relational expression calculated for each of the quality variables; a minimum error relational equation selection unit that calculates a prediction error of the relational equation for each of the division pattern candidates based on quality prediction data calculated using the relational equation calculated by the relational equation calculation unit from the operation data and the quality data corresponding to the operation data, and selects a relational equation for the division pattern candidate that minimizes the prediction error; a quality prediction model determination unit that determines the relational equation of the division pattern selected by the minimum error relational equation selection unit as the quality prediction model; A program having

[0268] (12) A computer-readable storage medium having a program recorded thereon for causing a computer to function as a quality prediction device for predicting product quality, a quality prediction model creation unit that analyzes a relationship between operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to the manufactured product, based on the plurality of data, and creates a quality prediction model that represents the relationship; a quality prediction unit that predicts the quality of the product based on the quality prediction model; Equipped with The quality prediction model creation unit includes: a division pattern candidate creation unit that creates a plurality of division pattern candidates for dividing an area in which the operation variables of the operation data take values ​​into a total area into a plurality of local areas; a local relational equation calculation unit that calculates, for each of the division pattern candidates, a local relational equation that represents a relationship between the operation data and the quality data in each of the local regions, for each of the quality variables; an activity function calculation unit that calculates, for each of the division pattern candidates, an activity function common to the quality variables as a weight of each of the local relational expressions in each local region of the division pattern candidate; a relational expression calculation unit that calculates a relational expression that represents a relationship between operation data and quality data in the entire region based on the activity function common to the quality variables and a local relational expression calculated for each of the quality variables; a minimum error relational equation selection unit that calculates a prediction error of the relational equation for each of the division pattern candidates based on quality prediction data calculated using the relational equation calculated by the relational equation calculation unit from the operation data and the quality data corresponding to the operation data, and selects a relational equation for the division pattern candidate that minimizes the prediction error; a quality prediction model determination unit that determines the relational equation of the division pattern selected by the minimum error relational equation selection unit as the quality prediction model; A computer-readable storage medium having a program recorded thereon. [Explanation of symbols]

[0269] 100 Quality Prediction Device 102 Quality Prediction Model Creation Department 110 Data Extraction Unit 120 Division pattern candidate creation unit 122 Numerical division creation part 124 Code division creation section 130 Activity function calculation unit 140 Local relational expression calculation unit 150 Relational equation calculation section 160 Minimum error relational expression selection section 170 Quality Prediction Model Decision Unit 180 Quality Prediction Department 190 Databases 200 Quality Prediction Device 210 Operation Data Input Section 220 Activity function memory unit 230 Activity function calculation unit 240 Local Relational Expression Calculation Unit 250 Quality Prediction Data Calculation Department 260 Quality Prediction Data Output Unit 270 Quality Data Entry Section 280 Local Relation Update Section 300 Manufacturing Process 500 Quality prediction model generator 510 Quality Prediction Model Creation Department 511 Common Model Calculation Unit 513 Individual Model Calculation Unit 515 Relational equation calculation section 520 Output section 530 Quality Prediction Department 540 Update processing section 600 Quality Prediction Device

Claims

1. A quality prediction device for predicting product quality, a quality prediction model creation unit that analyzes a relationship between operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to the manufactured product, based on the plurality of data, and creates a quality prediction model that represents the relationship; a quality prediction unit that predicts the quality of the product based on the quality prediction model; Equipped with The quality prediction model creation unit includes: a division pattern candidate creation unit that creates a plurality of division pattern candidates for dividing an area in which the operation variables of the operation data take values ​​into a total area into a plurality of local areas; a local relational equation calculation unit that calculates, for each of the division pattern candidates, a local relational equation that represents a relationship between the operation data and the quality data in each of the local regions, for each of the quality variables; an activity function calculation unit that calculates, for each of the division pattern candidates, an activity function common to the quality variables as a weight of each of the local relational expressions in each local region of the division pattern candidate; a relational expression calculation unit that calculates a relational expression that represents a relationship between operation data and quality data in the entire region based on the activity function common to the quality variables and a local relational expression calculated for each of the quality variables; a minimum error relational equation selection unit that calculates a prediction error of the relational equation for each of the division pattern candidates based on quality prediction data calculated using the relational equation calculated by the relational equation calculation unit from the operation data and the quality data corresponding to the operation data, and selects a relational equation for the division pattern candidate that minimizes the prediction error; a quality prediction model determination unit that determines the relational equation of the division pattern selected by the minimum error relational equation selection unit as the quality prediction model; The quality prediction device has

2. A quality prediction model generation device for generating a quality prediction model, comprising: a quality prediction model creation unit that analyzes a relationship between operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product, based on the plurality of data, and creates a quality prediction model that represents the relationship; an output unit that outputs the quality prediction model; Equipped with The operational variables include common operational variables that affect all physical elements and individual operational variables that affect only some of the physical elements; The quality prediction model creation unit includes: a common model calculation unit that calculates a common model relational expression common to the quality variables that represents a relationship between the operation data and the quality data that is common to all physical elements, using the common operation variables; an individual model calculation unit that calculates an individual model relational equation that represents a relationship between the operation data and the quality data related to only some physical elements, for each of the quality variables, by using the common operation variables and the individual operation variables; a relational expression calculation unit that calculates, as the quality prediction model, a relational expression that represents a relationship between operation data and quality data in a range in which the common operation variable takes a value, based on the common model relational expression common to the quality variables and the individual model relational expressions calculated for each of the quality variables; A quality prediction model generating device comprising:

3. 3. The quality prediction model generating device according to claim 2, wherein the manufacturing process is a tandem rolling mill consisting of a plurality of rolling mills, and the physical element is one rolling mill.

4. The quality prediction model generating device according to claim 2 or 3, wherein at least the common model relational expression is a nonlinear expression.

5. The quality prediction model generating device according to claim 4 , wherein the nonlinear expression is represented by a generalized linear model, a kernel method, or a neural network.

6. 6. The quality prediction model generating device according to claim 2, further comprising an update processing unit that updates the individual model relational equations based on operation data and quality data of products sequentially acquired from the manufacturing process.

7. The quality prediction model creation unit includes: a division pattern candidate creation unit that creates a plurality of division pattern candidates for dividing an area in which the common operation variable takes a value into a total area into a plurality of local areas; a minimum error relational equation selection unit that calculates a prediction error of the relational equation for each of the division pattern candidates based on quality prediction data calculated using the relational equation calculated by the relational equation calculation unit from the operation data and the quality data corresponding to the operation data, and selects a relational equation for the division pattern candidate that minimizes the prediction error; a quality prediction model determination unit that determines the relational equation of the division pattern selected by the minimum error relational equation selection unit as the quality prediction model; having the individual model calculation unit calculates, for each of the division pattern candidates created by the division pattern candidate creation unit, a local relational expression expressing a relationship between the operation data and the quality data in each of the local regions, for each of the quality variables, by using the individual operation variables corresponding to physical elements to which the common operation variables and quality indexes of quality data in each of the local regions correspond; 4. The quality prediction model generating device according to claim 2, wherein the common model calculation unit calculates, for each of the division pattern candidates, an activity function common to the quality variables as a weight of each of the local relational expressions in each local region of each of the division pattern candidates.

8. The quality prediction model determination unit A convergence judgment is performed for the prediction error of the selected division pattern. When it is determined that the prediction error has not converged, the division pattern candidate creation unit increases the number of divisions of the selected division pattern candidate to generate a plurality of new division pattern candidates; 8. The quality prediction model generating device according to claim 7, wherein, when it is determined that the prediction error has converged, a relational expression of the division pattern at the time of convergence is determined as the quality prediction model.

9. The quality prediction model generating device according to claim 7 or 8, wherein the local relational expression is a linear polynomial expression having a plurality of the operational variables as input variables.

10. The quality prediction model generating device according to claim 9 , wherein the individual model calculation unit calculates coefficients of the local relational expressions by any one of multiple regression, a stepwise method, a sparse regression, and a partial least squares method.

11. an update processing unit that updates the individual model relational equation based on operation data and quality data of a product sequentially acquired from the manufacturing process; The update processing unit: an operation data input unit for acquiring operation data of a target product from the manufacturing process; a quality data input unit for acquiring quality data of a target product from the manufacturing process; a common model calculation unit that calculates a weight of each of the individual model relational expressions based on the operation data input from the operation data input unit and the common model relational expressions calculated in advance by the common model calculation unit; an individual model calculation unit that calculates the individual model relational equation that is preset for each of the quality variables based on the operation data input from the operation data input unit; a quality prediction data calculation unit that calculates quality prediction data based on a calculation result of each of the individual model relational expressions by the individual model calculation unit and a weight of each of the individual model relational expressions calculated by the common model calculation unit; an individual model updating unit that updates coefficients of the individual model relational expressions based on the quality data input from the quality data input unit and the quality prediction data calculated by the quality prediction data calculation unit; The quality prediction model generating device according to any one of claims 7 to 10, comprising:

12. The individual model update unit calculates a vector of data consisting of operational variables of a product used to update the linear polynomial in the local region k as v_ （ｋ） (T), each element y of quality y in each local region k ｌ The coefficients of the linear polynomial before updating corresponding to A ｌ （ｋ） The quality prediction model generating device according to claim 11, wherein the coefficient (T) is calculated using the following formula (101): [0010] Here, v_ is a subscript that indicates that the coefficient is updated or the product to be used for updating, T-1 is a subscript that indicates that the coefficient is the current coefficient, and A ｌ （ｋ） (T) is the coefficient of the updated linear polynomial, y ｌ (T) is the quality data, K ｌ （ｋ） (T) is the update rate matrix in the local region k, and is expressed by the above formula (102). In formula (102), P ｌ （ｋ） (T-1) is the adaptive gain matrix for local region k.

13. A quality prediction device that predicts quality of the product based on the quality prediction model input from the output unit of the quality prediction model generating device according to any one of claims 2 to 12.

14. A quality prediction model generation method for predicting product quality, comprising: a quality prediction model creation step of analyzing a relationship between the operation data and the quality data based on a plurality of data including the operation data consisting of a plurality of operation variables in a manufacturing process and the quality data consisting of a plurality of quality variables related to the manufactured product, and creating a quality prediction model that represents the relationship; an output step of outputting the quality prediction model; Including, The operational variables include common operational variables that affect all physical elements and individual operational variables that affect only some of the physical elements; The quality prediction model creating step includes: a common model calculation unit step of calculating a common model relational expression common to the quality variables, which expresses a relationship between the operation data and the quality data common to all physical elements, using the common operation variables; an individual model calculation step of calculating, for each of the quality variables, an individual model relational equation that represents a relationship between the operation data and the quality data related to only some physical elements, using the common operation variables and the individual operation variables; a relational equation calculation step of calculating, as the quality prediction model, a relational equation representing a relationship between operation data and quality data in a range in which the common operation variable takes a value, based on the common model relational equation common to the quality variables and the individual model relational equations calculated for each of the quality variables; A quality prediction model generation method comprising:

15. A program for causing a computer to function as a quality prediction model generating device for generating a quality prediction model, a quality prediction model creation unit that analyzes a relationship between operation data consisting of a plurality of operation variables in a manufacturing process and quality data consisting of a plurality of quality variables related to a manufactured product, based on the plurality of data, and creates a quality prediction model that represents the relationship; an output unit that outputs the quality prediction model; Equipped with The operational variables include common operational variables that affect all physical elements and individual operational variables that affect only some of the physical elements; The quality prediction model creation unit includes: a common model calculation unit that calculates a common model relational expression common to the quality variables that represents a relationship between the operation data and the quality data that is common to all physical elements, using the common operation variables; an individual model calculation unit that calculates an individual model relational equation that represents a relationship between the operation data and the quality data related to only some physical elements, for each of the quality variables, by using the common operation variables and the individual operation variables; a relational expression calculation unit that calculates, as the quality prediction model, a relational expression that represents a relationship between operation data and quality data in a range in which the common operation variable takes a value, based on the common model relational expression common to the quality variables and the individual model relational expressions calculated for each of the quality variables; A program for causing a computer to function as a quality prediction model generating device, comprising:

Images