JP6881137B2 - Product status predictors and methods, manufacturing process control systems, and programs - Google Patents

Description

The present invention relates to a product state predictor and method for predicting the state of a product manufactured by a manufacturing process, a manufacturing process control system, and a program.

In the manufacturing process of manufacturing a product, it is required to control the state of the product so as to be in a desired state.
In order to achieve this, a machine learning method is widely used in which the past achievements of product states and manufacturing conditions are learned and the learning results are added to the control of the manufacturing process.
For example, in the cooling of hot rolling of a steel sheet, the temperature of the steel sheet is calculated using a model based on the heat transfer theory, and the amount of cooling water and the amount of cooling water transferred so that the temperature on the outlet side of the cooling device (cooling stop temperature) becomes a desired value. Set the speed. However, there are some parts that cannot be expressed by the model, such as differences in heat transfer coefficient due to the surface properties of each steel type and changes over time in equipment, which may cause errors. Machine learning is used to empirically predict and compensate for these errors from actual results.
Further, in the width control of hot rolling of a steel sheet, the target value of the width on the rough rolling machine output side (rough output side width), which is the operation amount, is set so that the product width becomes a desired value. However, in the subsequent finish rolling process, "width shrinkage" occurs due to tension and heat shrinkage between the rolling mills. It is necessary to predict this amount of width shrinkage in order to obtain the desired product width on the rough side, but a general model showing the relationship between the steel type and tension and the amount of width shrinkage has not yet been established. .. Therefore, machine learning is used to empirically predict the amount of width shrinkage from actual results and reflect it in the product width.

Stratified learning is known as a method of machine learning. In stratified learning, values are stored in a stratified table classified according to manufacturing conditions (for example, steel plate size, composition, temperature, etc.). Then, the value of the category corresponding to the material to be predicted is output as the predicted value of the objective variable, and the value of the category corresponding to the product subjected to actual rolling is updated with the actual value of the objective variable.
However, stratified learning has the following problems, and the prediction accuracy is limited.
-If the stratification is too coarse, products with different tendencies of objective variables will be learned in the same category, and the prediction accuracy will not improve.
・ If the stratification is too fine, the number of achievements that fall into one category will decrease, learning will not proceed, and prediction accuracy will not improve.
-If the number of manufacturing conditions to be considered is increased, the stratification becomes too detailed and learning does not proceed, so the number of manufacturing conditions that can be considered is at most about 4 conditions. Therefore, even if there are other manufacturing conditions that are likely to cause fluctuations in the objective variable, they cannot be taken into consideration, and the prediction accuracy does not improve.
-Since the objective variable is treated as a single value within the same category regardless of the manufacturing conditions, the relationship between the manufacturing conditions and the objective variable cannot be expressed, and the prediction accuracy does not improve.

Therefore, there is known a method of improving the prediction accuracy of the objective variable by generating a regression model based on the actual data.
In Patent Document 1, in order to improve the accuracy of temperature control when cooling of thick steel sheets is stopped, past operation results are stored in a database, and before the next and subsequent cooling controls are performed, rolling results, cooling conditions, etc. The actual data whose manufacturing conditions are similar to those of the thick steel sheet are extracted, and the error and heat transfer coefficient correction amount related to the cooling stop temperature prediction model of the thick steel sheet are estimated from the extracted past data, and this error and heat transfer coefficient are estimated. A configuration is disclosed in which the correction amount and the temperature prediction model are combined to determine the water amount and the steel plate transport speed of the water cooling device in the cooling zone.

In Patent Document 1, when the manufacturing conditions to extract the actual data that is similar (referred to as the neighborhood teacher data), defining the similarity weighted Euclidean distance d _i, as represented by the formula (1). Here, x _i = (x _{i, 1} , ..., x _{i, j} , ..., x _{i, m} ) ^T is an explanatory variable (manufacturing condition) of the past data, and i is a subscript indicating the data No. i = 0: Prediction target material), j is a subscript indicating the explanatory variable No, m is the number of types of manufacturing conditions, and q _j is the weight.

Then, based on the neighborhood teacher data, a local regression model for obtaining the predicted value of the objective variable is generated. In Patent Document 1, as a local regression method, the predicted value y ^ of the objective variable (^ is assumed to be attached above y) is represented by a linear combination of the regression coefficients β and x, and the error from the actual value y. Linear multiple regression (Equations (2) and (3)) that minimizes the sum of squares and PLS (Partial Least Squares) that generate new variables that eliminate the correlation between manufacturing conditions are shown. y _i is the objective variable of the data No. _{i, β reg} is the regression coefficient that minimizes the sum of squared errors, and n is the number of neighboring teacher data.

In this way, the method of generating a regression model that expresses the relationship between the objective variable and the manufacturing conditions based on the neighborhood teacher data does not have a fixed division as in stratified learning, but exists in the manufacturing condition space. It can be said that it is synonymous with the fact that the size of the division changes according to the density of the operation results. Therefore, it is unlikely that products with different tendencies of objective variables will be learned in the same category because the stratification is too coarse, or that learning will not proceed because the stratification is too fine. Moreover, the manufacturing conditions that can be considered are theoretically infinite. For the above reasons, better prediction accuracy of the objective variable than stratified learning can be obtained.

Japanese Patent No. 5682484

However, when a regression model is generated and the predicted value of the objective variable is obtained as described above, a value deviating from the actual value of the objective variable may be output as a predicted value (hereinafter, referred to as a deviated predicted value). ).
As shown in FIG. 6, the present inventor outputs the deviation prediction value when the value of the prediction target material (☆ in the figure) is outside the range R taken by the neighborhood teacher data group under certain manufacturing conditions. I found that it was a case of a certain extrapolation. FIG. 6 shows the relationship between the explanatory variables (manufacturing conditions) and the objective variables. As shown in FIG. 6, the local regression model (solid line in the figure) may deviate from the true model (dotted line in the figure) due to the influence of noise such as measurement error and operation disturbance. In this case, when the manufacturing condition of the predicted material is extrapolated to the neighborhood teacher data (the manufacturing condition of the predicted material is larger than the maximum value or smaller than the minimum value of the manufacturing condition of the neighborhood teacher data). If there is, the influence of noise may become large, and the predicted value of the objective variable may be far from the true model.

The present invention has been made in view of the above points, and an object of the present invention is to suppress the occurrence of an outlier predicted value when predicting the state of a product as an objective variable using manufacturing conditions in a manufacturing process.

The gist of the present invention for solving the above problems is as follows.
[1] A product state prediction device that predicts the state of a product as an objective variable using manufacturing conditions in a manufacturing process.
In response to the input of multiple types of manufacturing conditions for the material to be predicted, the database that accumulates the actual data including the actual manufacturing conditions in actual operation and the actual values of the objective variable is converted into a distance function that expresses the similarity of the manufacturing conditions. Based on this, an extraction means for extracting actual data similar to the production conditions of a plurality of types of the predicted target material as neighborhood teacher data, and
When there is a manufacturing condition outside the range of the manufacturing condition of the neighborhood teacher data extracted by the extraction means for at least one of the manufacturing conditions of the plurality of types of the predicted target material, the neighborhood teacher data is obtained from the database. A correction means for modifying the neighborhood teacher data so that it is included in the above range by additionally extracting the data.
When it is extracted by the extraction means and the neighborhood teacher data is modified by the modification means, a regression model for obtaining a predicted value of the objective variable is generated based on the modified neighborhood teacher data, and the generated regression model and the prediction A product state prediction device including a calculation means for calculating a predicted value of the objective variable using the manufacturing conditions of the target material.
[2] When the neighborhood teacher data is additionally extracted, the correction means is based on at least one of a distance function representing the similarity of manufacturing conditions, the novelty of actual data, and matching of predetermined conditions. The product state prediction device according to [1], wherein the neighborhood teacher data is extracted.
[3] When additionally extracting the neighborhood teacher data by the correction means, at least a part of the neighborhood teacher data extracted by the extraction means is canceled in response to the input of the production conditions of a plurality of types of the predicted target material. The state prediction device for a product according to [1] or [2], which comprises a canceling means.
[4] The canceling means is the oldness of the neighborhood teacher data extracted by the extraction means in response to the input of the distance function representing the similarity of the manufacturing conditions and the manufacturing conditions of a plurality of types of the predicted target material. It is characterized in that at least a part of the neighborhood teacher data extracted by the extraction means is canceled in response to the input of the production conditions of a plurality of types of the predicted target material based on at least one of them [3]. Product status predictor described in.
[5] The product state prediction device according to any one of [1] to [4], wherein the product is a steel plate.
[6] The product state prediction device according to any one of [1] to [5] and the product state prediction device.
A manufacturing process control system including a control means for controlling equipment used in the manufacturing process based on a predicted value of an objective variable calculated by the calculation means.
[7] A method for predicting the state of a product by using the manufacturing conditions in the manufacturing process to predict the state of the product as an objective variable.
In response to the input of multiple types of manufacturing conditions for the material to be predicted, the database that accumulates the actual data including the actual manufacturing conditions in actual operation and the actual values of the objective variable is converted into a distance function that expresses the similarity of the manufacturing conditions. Based on this, an extraction step of extracting performance data similar to the production conditions of a plurality of types of the predicted target material as neighborhood teacher data, and
When there is a manufacturing condition outside the range of the manufacturing condition of the neighborhood teacher data extracted in the extraction step for at least one of the manufacturing conditions of the plurality of types of the predicted target material, the neighborhood teacher data is obtained from the database. A modification step that modifies the neighborhood teacher data so that it falls within the above range by additionally extracting,
When the data is extracted in the extraction step and the neighborhood teacher data is modified in the modification step, a regression model for obtaining the predicted value of the objective variable is generated based on the modified neighborhood teacher data, and the generated regression model and the prediction are generated. A method for predicting the state of a product, which comprises a calculation step of calculating a predicted value of the objective variable using the manufacturing conditions of the target material.
[8] A program for predicting the state of a product as an objective variable using manufacturing conditions in a manufacturing process.
In response to the input of multiple types of manufacturing conditions for the material to be predicted, the database that accumulates the actual data including the actual manufacturing conditions in actual operation and the actual values of the objective variable is converted into a distance function that expresses the similarity of the manufacturing conditions. Based on this, an extraction means for extracting actual data similar to the production conditions of a plurality of types of the predicted target material as neighborhood teacher data, and
When there is a manufacturing condition outside the range of the manufacturing condition of the neighborhood teacher data extracted by the extraction means for at least one of the manufacturing conditions of the plurality of types of the predicted target material, the neighborhood teacher data is obtained from the database. A correction means for modifying the neighborhood teacher data so that it is included in the above range by additionally extracting the data.
When it is extracted by the extraction means and the neighborhood teacher data is modified by the modification means, a regression model for obtaining a predicted value of the objective variable is generated based on the modified neighborhood teacher data, and the generated regression model and the prediction are generated. A program for operating a computer as a calculation means for calculating a predicted value of the objective variable using the manufacturing conditions of the target material.

According to the present invention, when the state of a product is predicted as an objective variable by using the manufacturing conditions in the manufacturing process, it is possible to suppress the occurrence of an outlier predicted value.

It is a figure which shows the structural example of the control system of the manufacturing process which concerns on this embodiment. It is a figure which shows the example of the relationship between the explanatory variable and the objective variable when the neighborhood teacher data is corrected. It is a flowchart which shows the processing example by the neighborhood teacher data correction part. It is a figure which shows the functional structure example for a numerical experiment. It is a figure which shows the example of the relationship between the actual value and the predicted value in an Example and a comparative example. It is a figure which shows the example of the relationship between the explanatory variable and the objective variable when the value of the predicted material is extrapolated.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.
FIG. 1 shows a configuration example of a control system for the manufacturing process according to the present embodiment. In this embodiment, an example of predicting the state of a product as an objective variable using the manufacturing conditions in the manufacturing process will be described.
As shown in FIG. 1, the manufacturing process control system according to the present embodiment includes equipment 1, a higher-level computer 2, a computer 3, and a process computer 4.
Equipment 1 is used in a manufacturing process, for example, a steel sheet cooling device or a rough rolling mill. The equipment 1 operates based on the amount of operation input from the process computer 4.

The host computer 2 is a business computer above the process computer 4, and outputs the manufacturing conditions of the material to be predicted. In the case of cooling and width control of steel sheets, the manufacturing conditions of the material to be predicted are factors that affect cooling and width control, such as the size, composition, and temperature of the steel sheet to be predicted. The production conditions of the material to be predicted may be output from the computer for product production planning.

The computer 3 functions as a product state predictor, generates a local regression model using the state of the steel plate manufactured by the manufacturing process as an objective variable and manufacturing conditions as explanatory variables, and uses the local regression model to set the objective variable. Predict. The objective variable is, for example, a model error of the cooling shutdown temperature in the case of cooling the steel sheet, and, for example, the amount of width shrinkage in the case of width control. The computer 3 includes a database 31, a data extraction unit 32, a neighborhood teacher data correction unit 33, and a local regression model generation unit 34, which will be described in detail later. In the present embodiment, the database 31, the data extraction unit 32, the neighborhood teacher data correction unit 33, and the local regression model generation unit 34 function as the database, extraction means, correction means, and calculation means, respectively, as referred to in the present invention. ..

The process computer 4 functions as a control means for controlling the equipment 1. The process computer 4 functions as a setting calculation unit that sets the operation amount of the equipment 1 based on the manufacturing conditions of the material to be predicted and the predicted value of the objective variable input from the local regression model generation unit 34 of the computer 3. .. The amount of operation of the equipment 1 is the amount of cooling water, the transport speed, etc. for cooling the steel sheet, and the target value for the width on the coarse side in the case of controlling the width of the steel sheet.

Next, the details of the computer 3 will be described.
The database 31 accumulates actual data including actual values of manufacturing conditions and objective variables in actual operation. Information on the operation date and time is associated with the actual data.

The data extraction unit 32 extracts neighborhood teacher data from the database 31.
The data extraction unit 32 receives the input of the production conditions of a plurality of types of the prediction target material from the host computer 2, and receives the production conditions of the plurality of types of the prediction target material based on the distance function representing the similarity of the production conditions. The actual data similar to the above is extracted as the neighborhood teacher data and output to the neighborhood teacher data correction unit 33. Here, as the distance function, a function indicating the distance between two points in the variable space can be widely used. Generally, the distance function is a function that shows a smaller value as the distance between two points becomes smaller, and the data extraction unit 32 extracts a predetermined number of actual data from the database 31 as neighborhood teacher data in ascending order of the value of the distance function. .. It is also possible to make the distance function a function that shows a larger value as the distance between two points becomes smaller. In that case, a predetermined number of actual data should be extracted from the database 31 in descending order of the value of the distance function. Just do it. The distance function is not limited to the Euclidean distance as described in Patent Document 1, and can be applied as long as it is a measure for expressing the similarity of manufacturing conditions such as the Mahalanobis distance. In the following, the neighborhood teacher data extracted in response to the input of the manufacturing conditions of the material to be predicted will be referred to as "near teacher data extracted first".

Further, as described below, the data extraction unit 32 additionally extracts the neighborhood teacher data in response to the instruction from the neighborhood teacher data correction unit 33, and outputs the data to the neighborhood teacher data correction unit 33.

The proximity teacher data correction unit 33 determines whether or not there is a production condition outside the range of the production condition of the neighborhood teacher data extracted first for at least one of the production conditions of the plurality of types of the predicted target material (that is,). , Whether or not the manufacturing conditions of the material to be predicted are extrapolated with respect to the neighborhood teacher data extracted first). Then, the neighborhood teacher data correction unit 33 instructs the data extraction unit 32 to additionally extract the neighborhood teacher data when there is a manufacturing condition outside the range of the production conditions of the neighborhood teacher data extracted first. The neighborhood teacher data is modified so that it falls within the above range (that is, the manufacturing conditions of the material to be predicted are interpolated with respect to the neighborhood teacher data).

When the data extraction unit 32 additionally extracts the neighborhood teacher data in response to the instruction from the neighborhood teacher data correction unit 33, the distance function representing the similarity of the manufacturing conditions, the novelty of the actual data, and the predetermined conditions match. Additional neighborhood teacher data is extracted based on at least one of the things to be done.
When the distance function is used as a reference, the actual data that are considered to have the highest degree of similarity are preferentially extracted additionally. The prediction accuracy of the objective variable can be improved by using the actual data with as high a degree of similarity as the neighborhood teacher data. The distance function used here may be the same as or different from the distance function used when the neighborhood teacher data is first extracted.
In addition, when the newness of the actual data is used as a reference, the new actual data is preferentially additionally extracted in view of the time axis of the operation date and time. The equipment 1 changes over time, and it is possible to improve the prediction accuracy of the objective variable by using the new actual data as the neighborhood teacher data.
In addition, when the standard is that the predetermined conditions are met, for example, the actual data of the same steel type as the forecast target material is additionally extracted preferentially. By narrowing down the actual data of the same steel type to the neighborhood teacher data, it is possible to reduce the variation in the characteristics of the product and improve the prediction accuracy of the objective variable.
The criteria for additionally extracting the neighborhood teacher data may be any one of those listed here, a combination of a plurality of the criteria, or may be added. Further, the number of neighborhood teacher data to be additionally extracted may be, for example, a preset number, or may be determined each time by the data extraction unit 32 or the neighborhood teacher data correction unit 33 according to a preset algorithm. Good.

When additionally extracting the neighborhood teacher data in this way, at least a part of the neighborhood teacher data extracted first may be canceled (deleted).
The neighborhood teacher data correction unit 33 cancels at least a part of the neighborhood teacher data extracted first based on at least one of the distance function representing the similarity of the manufacturing conditions and the age of the neighborhood teacher data extracted first. .. In this case, the neighborhood teacher data correction unit 33 functions as a canceling means according to the present invention.
When the distance function is used as a reference, the neighborhood teacher data that is considered to have the lowest degree of similarity is canceled. Proximity teacher data for manufacturing conditions that are far from the manufacturing conditions for the material to be predicted are likely to have different data trends from the material to be predicted, and if a local regression model is generated based on actual data with such low similarity, the prediction error will occur. This is because there is a risk of producing. The distance function used here may be the same as or different from the distance function used when the neighborhood teacher data is first extracted and the distance function used when the neighborhood teacher data is additionally extracted.
In addition, when the age of the neighborhood teacher data extracted first is used as a reference, the old neighborhood teacher data is canceled in terms of the time axis of the operation date and time. This is because the equipment 1 changes over time, and if the neighborhood teacher data includes old actual data, it may cause a prediction error.
The criteria for canceling the neighborhood teacher data may be any one of those listed here, a combination of a plurality of the criteria, or other criteria may be added. Further, the number of neighborhood teacher data to be canceled may be, for example, a preset number, or may be determined each time by the data extraction unit 32 or the neighborhood teacher data correction unit 33 according to a preset algorithm.

FIG. 2 is a diagram showing an example of the relationship between the explanatory variable and the objective variable when the neighborhood teacher data is modified. FIG. 2 shows the additionally extracted neighborhood teacher data 301 and the canceled neighborhood teacher data 302. As shown in FIG. 2, as a result of modifying the neighborhood teacher data, the manufacturing conditions of the predicted target material (☆ in the figure) are included in the range R, that is, the manufacturing conditions of the predicted target material are relative to the neighborhood teacher data. It can be made into an interpolated state.

The local regression model generation unit 34 generates a local regression model for obtaining the predicted value of the objective variable using the manufacturing conditions as explanatory variables based on the neighborhood teacher data. The neighborhood teacher data referred to here is the neighborhood teacher data itself extracted first unless the manufacturing conditions of the material to be predicted are extrapolated to the neighborhood teacher data extracted first. Further, if the manufacturing condition of the material to be predicted is extrapolated with respect to the neighborhood teacher data extracted first, the neighborhood teacher data is corrected by the neighborhood teacher data correction unit 33 so as to eliminate the extrapolation. Then, the local regression model generation unit 34 calculates the predicted value of the objective variable using the generated local regression model and the manufacturing conditions of the material to be predicted. Although the details are omitted for the generation of the local regression model, linear multiple regression or PLS as described in Patent Document 1 may be applied, and other known regression methods such as ridge regression and neural network may be used. Can be applied.

In the present embodiment, the neighborhood teacher data correction unit 33 instructs the data extraction unit 32 to additionally extract the neighborhood teacher data, but the data extraction unit 32 is not instructed to correct the neighborhood teacher data. The unit 33 may be configured to additionally extract the neighborhood teacher data directly from the database 31.

FIG. 3 is a flowchart showing a processing example by the neighborhood teacher data correction unit 33 of the computer 3. In the neighborhood teacher data correction unit 33, the manufacturing conditions x _{0, j} = (x _{o, 1} , ..., X _{0, m} ) of the material to be predicted and the neighborhood teacher data x _{i, j} extracted first are input. Will be done. i is a subscript indicating a steel sheet (i = 1 to n: n is the number of neighborhood teacher data), and j is a subscript indicating a manufacturing condition (j = 1 to m: m is the number of types of manufacturing conditions).
In step S1, the neighborhood teacher data correction unit 33 sets the manufacturing condition j = 1.

In step S2, the neighborhood teacher data correction unit 33 _{determines whether or not the manufacturing condition x 0, j} of the predicted target material is smaller than the minimum value of the neighborhood teacher data x _{1, j} , ···, x _{n, j.} Is determined. If it is smaller, the process proceeds to step S3, otherwise the process proceeds to step S4.

In step S3, the neighborhood teacher data correction unit 33 corrects the neighborhood teacher data. Specifically, the data extraction unit 32 is instructed to additionally extract the neighborhood teacher data. The data extraction unit 32 receives an instruction from the neighborhood teacher data correction unit 33, and among the actual data stored in the database 31 in which x _{i, j} <x _{0, j} , is represented by, for example, a distance function. Additional k neighborhood teacher data are extracted in descending order of similarity. Further, among the neighborhood teacher data extracted first, only one neighborhood teacher data are canceled in ascending order of similarity represented by the distance function. Both k and l are values smaller than the number n of neighboring teacher data and are set in advance.

In step S4, the neighborhood teacher data correction unit 33 _{determines whether or not the manufacturing condition x 0, j} of the predicted target material is larger than the maximum value of the neighborhood teacher data x _{1, j} , ..., X _{n, j.} Is determined. If it is larger, the process proceeds to step S5, otherwise the process proceeds to step S6.

In step S5, the neighborhood teacher data correction unit 33 corrects the neighborhood teacher data. Specifically, the data extraction unit 32 is instructed to additionally extract the neighborhood teacher data. The data extraction unit 32 receives an instruction from the neighborhood teacher data correction unit 33, and among the actual data stored in the database 31, x _{i, j} > x _{0, j} is represented by, for example, a distance function. Additional k neighborhood teacher data are extracted in descending order of similarity. Further, among the neighborhood teacher data extracted first, only one neighborhood teacher data are canceled in ascending order of similarity represented by the distance function. Both k and l are values smaller than the number n of neighboring teacher data and are set in advance.

In step S6, the neighborhood teacher data correction unit 33 determines whether or not j = m has been reached. As a result, if j = m has not been reached, j is incremented in step S7 and the process returns to step S2. On the other hand, when j = m is reached, the neighborhood teacher data (the neighborhood teacher data extracted first if there is no correction, and the neighborhood teacher data after correction if there is correction) is output to the local regression model generation unit 34.

As described above, for at least one of the manufacturing conditions of the plurality of types of the predicted target material, when there is a manufacturing condition outside the range of the manufacturing conditions of the neighborhood teacher data extracted first, the neighborhood teacher data from the database 31. By additionally extracting, the neighborhood teacher data is modified so that the manufacturing conditions of the material to be predicted are included in the range. As a result, it is possible to eliminate the case where the manufacturing conditions of the material to be predicted are extrapolated with respect to the neighborhood teacher data. In other words, as shown in FIG. 2, the manufacturing conditions (☆ in the figure) of the material to be predicted are interpolated with respect to the neighboring teacher data (☆ in the figure is included in the range R). Can be. In this case, even if noise such as measurement error or operation disturbance is added, the local regression model (solid line in the figure) can be prevented from being significantly deviated from the true model (dotted line in the figure). Therefore, the predicted value of the objective variable is unlikely to be far from the true model, the occurrence of the outlier predicted value can be suppressed, and the prediction accuracy of the state of the product can be improved. Then, if the prediction accuracy of the state of the product can be improved, it becomes possible to control the state of the product with high accuracy so as to be in a desired state.

In FIG. 1, one computer is shown to function as a database 31, a data extraction unit 32, a neighborhood teacher data correction unit 33, and a local regression model generation unit 34, but a plurality of computers are combined to form the database 31, data. It may function as an extraction unit 32, a neighborhood teacher data correction unit 33, and a local regression model generation unit 34, and input / output via a network.
Further, in FIG. 1, a configuration example in which the computer 3 is installed separately from the process computer 4 is used. This is because the state prediction device of the product requires a calculation to build a database and generate a local regression model, so that the computer load is expected to increase. As a matter of course, if the specifications of one computer (for example, the process computer 4) are sufficient, one computer may be made to function as a state prediction device and a setting calculation unit of the product.

The effect of the method to which the present invention was applied was verified by numerical experiments.
FIG. 4 shows a functional configuration for numerical experiments. The personal computer 100 reproduces the database 31, the data extraction unit 32, the neighborhood teacher data correction unit 33, and the local regression model generation unit 34, and performs a numerical experiment. The manufacturing conditions are input to the personal computer 100 based on the factory operation results. In the personal computer 100, the data extraction unit 32 extracts the neighborhood teacher data from the database 31, the neighborhood teacher data correction unit 33 corrects the neighborhood teacher data, and the local regression model generation unit 34 outputs the width reduction prediction value as the objective variable. To do.
Since the actual value of width shrinkage is stored in the factory operation record, the accuracy was evaluated by comparing it with the predicted value of width shrinkage by numerical experiments. In the numerical experiment, it is considered that the width shrinkage is predicted with high accuracy and the width control accuracy is also improved.

PLS is used in the local regression model generation unit 34, and the width is wide for each of the case where the method described in Patent Document 1 is used (comparative example) and the case where the method for modifying the neighborhood teacher data is used (example). The prediction accuracy of the amount of shrinkage was evaluated. In the embodiment, as described in the flowchart of FIG. 3, a method is adopted in which the neighborhood teacher data is additionally extracted and a part of the neighborhood teacher data extracted first is canceled.
FIG. 5 shows the results of numerical experiments. FIG. 5A is a diagram showing the relationship between the actual width shrinkage value and the predicted width shrinkage value in the examples. Further, FIG. 5B is a diagram showing the relationship between the actual width shrinkage value and the predicted width shrinkage value in the comparative example. In the comparative example, the mean square error RMSE was 1.66 mm, and the coefficient of determination (contribution rate) R2 was 0.834. On the other hand, in the example, the mean square error RMSE was 1.56 mm and the coefficient of determination R2 was 0.835, which were better results than those in the comparative example.

Although the present invention has been described above with one embodiment, the above-described embodiment is merely an example of embodiment of the present invention, and the technical scope of the present invention is limitedly interpreted by these. It should not be. That is, the present invention can be implemented in various forms without departing from its technical ideas or its main features.
The present invention is not only applied to cooling and width control of steel sheets as in the above embodiment, but is also universally applicable and has a wide effect in predicting the state of a product manufactured by a manufacturing process. Have.
The state prediction device of the product to which the present invention is applied is realized by, for example, a computer equipped with a CPU, a ROM, a RAM, or the like.
The present invention also provides software (programs) that realize the functions of the present invention to a system or device via a network or various storage media, and the computer of the system or device reads and executes the program. It is feasible.

1: Equipment 2: High-level computer 3: Computer 4: Process computer, 31: Database, 32: Data extraction unit, 33: Neighborhood teacher data correction unit, 34: Local regression model generation unit

Claims (8)

A product state predictor that predicts the state of a product as an objective variable using manufacturing conditions in the manufacturing process.
In response to the input of multiple types of manufacturing conditions for the material to be predicted, the database that accumulates the actual data including the actual manufacturing conditions in actual operation and the actual values of the objective variable is converted into a distance function that expresses the similarity of the manufacturing conditions. Based on this, an extraction means for extracting actual data similar to the production conditions of a plurality of types of the predicted target material as neighborhood teacher data, and
When there is a manufacturing condition outside the range of the manufacturing condition of the neighborhood teacher data extracted by the extraction means for at least one of the manufacturing conditions of the plurality of types of the predicted target material, the neighborhood teacher data is obtained from the database. A correction means for modifying the neighborhood teacher data so that it is included in the above range by additionally extracting the data.
When it is extracted by the extraction means and the neighborhood teacher data is modified by the modification means, a regression model for obtaining a predicted value of the objective variable is generated based on the modified neighborhood teacher data, and the generated regression model and the prediction A product state prediction device including a calculation means for calculating a predicted value of the objective variable using the manufacturing conditions of the target material. When the neighborhood teacher data is additionally extracted, the correction means is based on at least one of a distance function representing the similarity of manufacturing conditions, the novelty of actual data, and the matching of predetermined conditions. The product state prediction device according to claim 1, wherein the teacher data is extracted. When the neighborhood teacher data is additionally extracted by the correction means, the canceling means for canceling at least a part of the neighborhood teacher data extracted by the extraction means in response to the input of the manufacturing conditions of a plurality of types of the predicted target material. The product state prediction device according to claim 1 or 2, wherein the product is provided with. The canceling means is at least one of the distance function representing the similarity of the manufacturing conditions and the age of the neighborhood teacher data extracted by the extracting means in response to the input of the manufacturing conditions of a plurality of types of the predicted target material. The third aspect of claim 3, wherein at least a part of the neighborhood teacher data extracted by the extraction means is canceled in response to the input of the production conditions of a plurality of types of the predicted target material based on the above. Product status predictor. The product state prediction device according to any one of claims 1 to 4, wherein the product is a steel plate. The product state prediction device according to any one of claims 1 to 5.
A manufacturing process control system including a control means for controlling equipment used in the manufacturing process based on a predicted value of an objective variable calculated by the calculation means. It is a product state prediction method that predicts the state of a product as an objective variable using the manufacturing conditions in the manufacturing process.
In response to the input of multiple types of manufacturing conditions for the material to be predicted, the database that accumulates the actual data including the actual manufacturing conditions in actual operation and the actual values of the objective variable is converted into a distance function that expresses the similarity of the manufacturing conditions. Based on this, an extraction step of extracting performance data similar to the production conditions of a plurality of types of the predicted target material as neighborhood teacher data, and
When there is a manufacturing condition outside the range of the manufacturing condition of the neighborhood teacher data extracted in the extraction step for at least one of the manufacturing conditions of the plurality of types of the predicted target material, the neighborhood teacher data is obtained from the database. A modification step that modifies the neighborhood teacher data so that it falls within the above range by additionally extracting,
When the data is extracted in the extraction step and the neighborhood teacher data is modified in the modification step, a regression model for obtaining the predicted value of the objective variable is generated based on the modified neighborhood teacher data, and the generated regression model and the prediction are generated. A method for predicting the state of a product, which comprises a calculation step of calculating a predicted value of the objective variable using the manufacturing conditions of the target material. A program for predicting the state of a product as an objective variable using the manufacturing conditions in the manufacturing process.
In response to the input of multiple types of manufacturing conditions for the material to be predicted, the database that accumulates the actual data including the actual manufacturing conditions in actual operation and the actual values of the objective variable is converted into a distance function that expresses the similarity of the manufacturing conditions. Based on this, an extraction means for extracting actual data similar to the production conditions of a plurality of types of the predicted target material as neighborhood teacher data, and
When there is a manufacturing condition outside the range of the manufacturing condition of the neighborhood teacher data extracted by the extraction means for at least one of the manufacturing conditions of the plurality of types of the predicted target material, the neighborhood teacher data is obtained from the database. A correction means for modifying the neighborhood teacher data so that it is included in the above range by additionally extracting the data.
When it is extracted by the extraction means and the neighborhood teacher data is modified by the modification means, a regression model for obtaining a predicted value of the objective variable is generated based on the modified neighborhood teacher data, and the generated regression model and the prediction are generated. A program for operating a computer as a calculation means for calculating a predicted value of the objective variable using the manufacturing conditions of the target material.

Images