JP2018010521A - Product state prediction device, product state control device, product state prediction method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent occurrence of a deviation prediction value when predicting a model error of a product state such as cooling stop temperature of a steel plate.SOLUTION: A spot regression model generation unit 10 extracts achievement data having a manufacturing condition similar to a manufacturing condition of a prediction object material from a database 9 as neighbor teacher data, creates using the neighbor teacher data, a spot weighted regression model obtaining an error (referred as a model error prediction value) of a prediction value such as a cooling stop temperature of a prediction object material calculated based on the temperature prediction model, and calculates the model error prediction value by the spot weighted regression model. In this case, when a manufacturing condition of the prediction object material is in the extrapolation state where the manufacturing condition is out of the range R of the neighbor teacher data, the regression calculation is performed after excluding the manufacturing condition from the neighbor teacher data.SELECTED DRAWING: Figure 1

Description

  The present invention provides a product state predicting device, a product state control device, and a product state predicting method suitable for use in calculating a predicted value of a cooling stop temperature of a steel plate by a temperature prediction model, for example, in a steel plate cooling process. And the program.

In order to control the product state to a desired state in the product manufacturing process, the error between the actual value of the product state and the value calculated by the state prediction model, which is a physical model, is learned and used. Thus, a method is used in which an operation amount is determined so as to obtain a desired state by correcting a calculated value by the state prediction model.
For example, in the cooling step downstream of the hot rolling step of the steel plate, the mechanical properties necessary for the product are obtained by cooling the steel plate to a predetermined temperature.
Steel sheet cooling control is directly linked to the mechanical properties of the product and affects the yield, so that improvement in accuracy is required. In the cooling control, the temperature change of the steel sheet due to water cooling is estimated by heat transfer model calculation, and the amount of cooling water, the conveying speed of the steel sheet, and the like are determined so as to achieve a desired cooling stop temperature. The heat transfer model is created based on the knowledge of thermohydrodynamics and heat transfer engineering, but there are many factors that are difficult to model, such as the surface properties of the steel sheet and changes in equipment over time. It is difficult to accurately match the target value.

Therefore, a method of predicting the difference (model error) between the model calculated value of the cooling stop temperature and the actual value using a learning model and correcting the model calculated value of the cooling stop temperature is adopted.
Patent Document 1 discloses a cooling stop temperature of the thick steel plate in the cooling process, using a thick steel plate temperature prediction model for predicting the temperature change behavior of the thick steel plate in the cooling process. Prediction value calculation step for calculating a prediction value, extraction step for extracting past performance data of a steel plate having similar manufacturing conditions to the steel plate from the database in which past performance data is accumulated for each slab, extraction in the extraction step Based on the past actual data, a linear regression model equation is established, and an estimation step for estimating an error in a predicted value of the cooling stop temperature of the thick steel plate, cooling stop of the thick steel plate calculated in the predicted value calculation step Correction value calculation for calculating the correction value of the cooling stop temperature from the predicted value of temperature and the error of the prediction value of the cooling stop temperature of the steel plate estimated in the estimation step Step, it is disclosed that has a.

When extracting the performance data from the database, the similar determination of the manufacturing condition is performed based on, for example, the weighted Euclidean distance function of Expression (1). Here, x _i : = (x _{i, 1} ,..., X _{i, m} ) ^T is a manufacturing condition, q _j is a manufacturing condition coefficient, i is a subscript indicating a steel plate (manufacturing No) (i = 0: prediction) Target material). A predetermined number of performance data is extracted in ascending order of the Euclidean distance d _i .

A local regression model is generated using the performance data extracted in this way as neighborhood teacher data. For example, a locally weighted regression model represented by the equations (2) and (3) is generated. Here, β is a regression parameter, y is a model error actual value, y ^ (^ is attached on y) is a model error predicted value, and n is the number of neighboring teacher data. W _i is a weighting factor.

  As specific means, linear weighted linear multiple regression (Formula (4), Formula (5)) and local weighted PLS (Partial Least Squares) regression (Formula (6) to Formula (8)) have been proposed. .

The local weighted linear multiple regression is a method in which the model error predicted value y is expressed by a linear combination of the manufacturing conditions x (equation (4)) and a coefficient a _reg that minimizes the weighted error sum of squares is obtained. (Formula (5)). This method has a problem of multicollinearity in which the regression coefficient becomes unstable when there is a correlation between manufacturing conditions.
Therefore, in the local weighted PLS regression, a latent variable z that maximizes the covariance between variables is generated so as not to cause multicollinearity (formula (8)). C _tr is a matrix composed of eigenvectors of the matrix X ^T Wyy ^T WX. X: = (x ₁ ,..., X _n ) ^T , W: = diag (w ₁ ,..., W _n ), y: = (y ₁ ,..., Y _n ) ^T. Thereafter, linear multiple regression is performed on the latent variable z (Equation (6), Equation (7)) to avoid the problem of multicollinearity.

In the procedure of local weighted regression, for example, the form of Equation (9) is used for setting the weighting coefficient w _i . α is a parameter and is a positive real number.

Japanese Patent No. 5682484

Iron and steel, 15, 2509-2514, 1981

  As described above, when a locally weighted regression model is generated to obtain a model error predicted value, a value 801 that deviates significantly from the model error actual value may be output as a predicted value, as shown in FIG. (Hereinafter referred to as outlier predicted value). FIG. 8 is a diagram illustrating a relationship between the model error actual value and the model error predicted value. If the model calculation value of the cooling stop temperature is corrected while such a predicted deviation value 801 is generated, the cooling stop temperature may not be predicted with high accuracy.

  The present invention has been made in view of the above points, and an object of the present invention is to suppress occurrence of a predicted deviation when predicting a model error of a product state such as a cooling stop temperature of a steel plate. To do.

The gist of the present invention for solving the above problems is as follows.
[1] In a product manufacturing process, a product state prediction apparatus that calculates a predicted value of a product state by a state prediction model,
An extraction means for extracting neighborhood teacher data based on a distance function from a database in which performance data including manufacturing conditions is accumulated when calculating a predicted value of a state of a prediction target product;
When the manufacturing condition of the prediction target product is outside the range of the vicinity teacher data extracted by the extraction unit, at least one of the vicinity teacher data extracted by the extraction unit and the manufacturing condition of the prediction target product is Correction means to correct;
Regression model for obtaining an error in the predicted value of the state of the prediction target product calculated by the state prediction model using the vicinity teacher data extracted by the extraction unit or the corrected vicinity teacher data when corrected by the correction unit And calculating means for calculating the error according to the regression model and the manufacturing condition of the prediction target product or the corrected manufacturing condition when corrected by the correcting means,
Correction means for correcting the predicted value of the state of the prediction target product calculated by the state prediction model based on the error of the predicted value of the state of the prediction target product obtained by the calculation means; Product state prediction device.
[2] The correction unit excludes the manufacturing condition from the neighborhood teacher data extracted by the extraction unit when the production condition of the prediction target product is outside the range of the neighborhood teacher data extracted by the extraction unit. The product state prediction apparatus according to [1], characterized by:
[3] When the manufacturing condition of the prediction target product is larger than the maximum value of the manufacturing condition included in the proximity teacher data extracted by the extraction means, the correction unit sets the manufacturing condition of the prediction target product as the value. When it is smaller than the minimum value of the manufacturing conditions included in the proximity teacher data extracted by the extracting means, the manufacturing condition of the prediction target product is replaced with the minimum value of the proximity teacher data. The product state prediction apparatus according to [1], wherein
[4] The database stores, as actual data, the difference between the actual product state value and the predicted product state value calculated using the state prediction model in association with manufacturing conditions,
Any one of [1] to [3], wherein the extraction unit extracts performance data having manufacturing conditions similar to the manufacturing conditions of the prediction target product as neighborhood teacher data based on the distance function. The product state prediction apparatus according to one of the above.
[5] Any one of [1] to [4], wherein the product is a steel plate, the manufacturing process is a cooling step, the state of the product is a cooling stop temperature of the steel plate, and the state prediction model is a temperature prediction model. The product state prediction apparatus according to claim 1.
[6] The product state prediction apparatus according to any one of [1] to [4];
Control for controlling the operation amount of the manufacturing equipment used in the manufacturing process so that the predicted value of the state of the prediction target product corrected by the correction means matches the target value previously determined for each prediction target product And a product state control device.
[7] The product is a steel plate, the manufacturing process is a cooling step, the state of the product is a cooling stop temperature of the steel plate, the state prediction model is a temperature prediction model, and the operation amount is at least one of a cooling water amount and a steel plate conveyance speed. [6] The product state control device according to [6].
[8] A product state prediction method for calculating a predicted value of a product state using a state prediction model in a product manufacturing process,
When calculating the predicted value of the state of the prediction target product, an extraction step of extracting neighborhood teacher data based on a distance function from a database in which performance data including manufacturing conditions is accumulated;
When the manufacturing condition of the prediction target product is outside the range of the vicinity teacher data extracted in the extraction step, at least one of the vicinity teacher data extracted in the extraction step and the manufacturing condition of the prediction target product Correction steps to correct;
Regression model for obtaining an error in the predicted value of the state of the prediction target product calculated by the state prediction model using the vicinity teacher data extracted in the extraction step or the corrected vicinity teacher data when corrected in the correction step Generating the error and calculating the error according to the corrected manufacturing condition when the correction model is corrected in the manufacturing condition of the prediction target product or the correction step;
A correction step of correcting the predicted value of the state of the prediction target product calculated by the state prediction model based on the error of the predicted value of the state of the prediction target product obtained in the calculation step. Product state prediction method.
[9] A program for calculating a predicted value of a product state using a state prediction model in a product manufacturing process,
An extraction means for extracting neighborhood teacher data based on a distance function from a database in which performance data including manufacturing conditions is accumulated when calculating a predicted value of a state of a prediction target product;
When the manufacturing condition of the prediction target product is outside the range of the vicinity teacher data extracted by the extraction unit, at least one of the vicinity teacher data extracted by the extraction unit and the manufacturing condition of the prediction target product is Correction means to correct;
Regression model for obtaining an error in the predicted value of the state of the prediction target product calculated by the state prediction model using the vicinity teacher data extracted by the extraction unit or the corrected vicinity teacher data when corrected by the correction unit And calculating means for calculating the error according to the regression model and the manufacturing condition of the prediction target product or the corrected manufacturing condition when corrected by the correcting means,
A program for causing a computer to function as correction means for correcting the predicted value of the state of the prediction target product calculated by the state prediction model based on the error of the predicted value of the state of the prediction target product obtained by the calculation means .

  According to the present invention, when predicting the model error of the product state such as the cooling stop temperature of the steel sheet, when the manufacturing condition of the prediction target product is outside the range of the vicinity teacher data, the vicinity teacher data and the prediction target product By correcting at least one of the manufacturing conditions, it is possible to suppress the occurrence of a predicted deviation value.

It is a figure which shows the structural example of a cooling control system. It is a block diagram which shows the function structure of a local regression model production | generation part. It is a figure which shows the relationship between manufacturing conditions and a model error actual value / predicted value. It is a flowchart which shows the correction process of the vicinity teacher data in 1st Embodiment. It is a flowchart which shows the correction process of the manufacturing conditions of the prediction object material in 2nd Embodiment. It is a figure which shows the function structure for a numerical experiment. It is a figure which shows the relationship between the model error track record value and model error prediction value in the invention method. It is a figure which shows the relationship between the model error actual value and model error prediction value in a comparison method.

Preferred embodiments of the present invention will be described below with reference to the accompanying drawings. In each embodiment, an example in which the present invention is applied to prediction and control of a cooling stop temperature of a steel sheet in a cooling process will be described.
(First embodiment)
FIG. 1 is a diagram illustrating a configuration example of a cooling control system including a cooling facility.
In the cooling facility, the steel plate 1 is transported by the transport table 3 and water is sprayed from the cooling nozzle 2 to cool the steel plate 1.
The cooling nozzle 2 controls the amount of cooling water based on the instruction value output from the cooling process computer 5. As a cooling water amount control method, a general cooling water amount control method such as flow rate valve opening adjustment or nozzle ON / OFF control can be applied.
Further, the transfer table 3 controls the transfer speed of the steel sheet based on the instruction value output from the cooling process computer 5. As a method for controlling the conveying speed of the steel plate, a general conveying speed control method such as control by a motor driver can be applied.
On the exit side of the cooling facility, the cooling stop temperature of the steel sheet is measured by the thermometer 4 and input to the cooling process computer 5. As a temperature measurement method, a general temperature measurement method such as a radiation thermometer or thermography can be applied.

  The cooling process computer 5 includes a cooling stop temperature calculation unit 7, a subtracter 8, a database 9, a local regression model generation unit 10, and a cooling setting calculation unit 11. The cooling process computer 5 receives the manufacturing conditions of the material to be predicted from the upper rolling process computer 6. Production conditions include, for example, steel plate size, steel type, target value of cooling stop temperature, and the like, but may include elements that are general as information affecting the cooling process, such as outside air temperature and rolling roll roughness. In the present embodiment, the cooling process computer 5 functions as a product state prediction device and a product state control device according to the present invention.

The cooling stop temperature calculation unit 7 calculates the cooling stop temperature calculation value of the steel sheet based on the temperature prediction model including the heat transfer model based on the manufacturing conditions of the steel sheet, the actual water amount value in the cooling facility and the actual transport speed value of the steel sheet (“ Calculated as “cooling stop temperature calculation value”).
The subtracter 8 takes the difference (referred to as “model error actual value”) between the actual cooling stop temperature value measured by the thermometer 4 and the calculated cooling stop temperature value output from the cooling stop temperature calculation unit 7.

The database 9 stores the model error actual value output from the subtracter 8 in association with the manufacturing condition as actual data. Further, the database 9 has a function of extracting performance data having manufacturing conditions similar to the manufacturing conditions as neighborhood teacher data and outputting the results to the local regression model generation unit 10 when the manufacturing conditions of the prediction target material are input. . In the present embodiment, the database 9 functions as the extraction means in the present invention.
When the performance data is extracted from the database 9, the similar determination of the manufacturing condition is performed based on, for example, the weighted Euclidean distance function of Expression (1). In the present embodiment, the Euclidean distance is taken as an example, but what is known as a definition of a multivariable system distance such as Mahalanobis distance is applicable.

  The local regression model generation unit 10 uses the neighborhood teacher data extracted from the database 9 to obtain a local error for obtaining a prediction value error (referred to as a model error prediction value) of the cooling stop temperature of the prediction target material calculated by the temperature prediction model. A weighted regression model is generated, and a model error prediction value is calculated using the local weighted regression model. In the present embodiment, the local regression model generation unit 10 functions as a correction unit and a calculation unit in the present invention.

  The cooling setting calculation unit 11 corrects the predicted value of the cooling stop temperature of the prediction target material calculated by the temperature prediction model based on the model error predicted value output from the local regression model generation unit 10. And the cooling setting calculation part 11 is at least among the amount of cooling water and the conveyance speed of a steel plate so that the predicted value of the cooling stop temperature of the corrected prediction target material matches the target value previously determined for each prediction target material. After correcting one of the indicated values, the indicated value of the cooling water amount and the conveying speed of the steel sheet is output to the cooling facility. In the present embodiment, the cooling setting calculation unit 11 functions as a correction unit and a control unit in the present invention. In this embodiment, the amount of cooling water and the conveying speed of the steel plate are used as the operation amount. However, other elements that are general as the operation amount of the cooling facility such as the height of the cooling nozzle 2 may be included.

Here, as shown in FIG. 8, the present inventor outputs the outlier predicted value 801 because the range of values that the neighborhood teacher data can take is narrow for one or a plurality of manufacturing conditions, and the prediction target material It has been found that the manufacturing conditions are in an extrapolated state that is outside this range.
3A to 3C show the relationship between certain manufacturing conditions and model error actual measurement / predicted values. 3A to 3C, the horizontal axis represents the manufacturing conditions as explanatory variables, and the vertical axis represents the model error actual value and the model error predicted value. A represents the maximum value and the minimum value that can be taken by the neighborhood teacher data for the manufacturing condition.

  As shown in FIG. 3 (a), when the range R of the neighborhood teacher data is wide for the manufacturing conditions, even if noise is added due to operation variations, measurement errors, etc., the model error prediction value obtained by the regression model is not so much. Does not fluctuate. Therefore, even if the manufacturing conditions of the material to be predicted (stars in the figure) are extrapolated, the model error predicted value is far from the range considered from the actual value (true model in the figure). Must not.

  On the other hand, as shown in FIG. 3B, when the range R of the neighborhood teacher data is narrow for the manufacturing condition, the model error predicted value obtained by the regression model varies greatly even with small noise. Still, if the manufacturing condition of the prediction target material is in an interpolated state, the model error predicted value is not so far from the range considered from the actual value. However, as shown in FIG. 3C, when the manufacturing condition of the prediction target material is in an extrapolated state when the range R of the neighborhood teacher data is narrow for the manufacturing condition, the model error prediction value is the actual value. It will be far from the range considered.

As described above, the out-of-range predicted value is output when the neighborhood teacher data range R is narrow for a certain manufacturing condition and is greatly affected by noise, and the manufacturing condition of the prediction target material is extrapolated. It can be said that the influence of the noise is amplified by being.
As a countermeasure to this problem, in the present embodiment, the regression calculation is performed after excluding the manufacturing conditions that are greatly affected by the noise as described above and amplifying the influence from the neighborhood teacher data.
In this case, it is conceivable to evaluate both the extent of the range R of the neighborhood teacher data and whether or not it is in an extrapolated state. However, the unit system of the range R of the neighborhood teacher data differs depending on the manufacturing conditions. Because the manufacturing range is different for each manufacturing location, it is difficult to make a unique determination. Therefore, it can be said that it is appropriate to evaluate whether or not the state is extrapolated.

FIG. 2 is a block diagram illustrating a functional configuration of the local regression model generation unit 10.
Reference numeral 101 denotes an input unit that inputs the manufacturing conditions of the prediction target material and the neighborhood teacher data extracted from the database 9.

  Reference numeral 102 denotes a correction unit. When a manufacturing condition of a material to be predicted is in an extrapolation state that exists outside the range R of the neighborhood teacher data input by the input unit 101, an extrapolation state is set. Modify neighborhood teacher data to avoid it. Specifically, when a certain production condition for the material to be predicted exists outside the range R of the neighborhood teacher data, the production condition is excluded from the neighborhood teacher data.

  Reference numeral 103 denotes a local weighted regression calculation unit, which uses the neighborhood teacher data input by the input unit 101 or, if corrected by the correction unit 102, the locally weighted linear multiple regression (formula (4) ), Formula (5)), local weighted PLS regression (formulas (6) to (8)), and the like, to generate a local weighted regression model. When the neighborhood teacher data corrected by the correction unit 102 is used, the locally weighted regression model does not include the excluded manufacturing conditions as explanatory variables.

  A model error calculation unit 104 inputs the manufacturing condition of the prediction target material to the local weighted regression model generated by the local weighted regression calculation unit 103, and calculates a model error prediction value of the prediction target material. When using the neighborhood teacher data corrected by the correction unit 102 as described above, the locally weighted regression model does not include the excluded manufacturing condition as an explanatory variable, and thus the manufacturing condition of the prediction target material is not used. It will be.

  Reference numeral 105 denotes an output unit that outputs the predicted model error value of the prediction target material calculated by the model error calculation unit 104 to the cooling setting calculation unit 11.

FIG. 4 shows an example of neighborhood teacher data modification processing by the modification unit 102.
Assume that the manufacturing condition of the neighborhood teacher data is x _{i, j} and the manufacturing condition of the prediction target material is x _{0, j} . i is a subscript indicating a steel plate (i = 1 to n), and j is a subscript indicating a manufacturing condition (j = 1 to m).
In step S1, j = 1 is set.
In step S2, it is determined whether or not the manufacturing condition x _{0, j} of the prediction target material is in an extrapolated state, that is, whether or not the prediction target material exists outside the range R of the neighborhood teacher data. Specifically, the production condition x _{0, j of the} material to be predicted is smaller than the minimum value min (x _{1, j} ,..., X _{n, j} ) of the production condition x _{i, j} , or the production condition It is judged whether it is larger than the maximum value max (x _{1, j} ,..., x _{n, j} ) of x _{i, j} . As a result, if the manufacturing condition x _{0, j of} the prediction target material exists outside the range R of the vicinity teacher data, the process proceeds to step S3, and if it exists within the range R of the vicinity teacher data, the process proceeds to step S4.
In step S3, the manufacturing condition x _{i, j} is excluded from the neighborhood teacher data, and the process proceeds to step S4.
In step S4, it is determined whether j = m. If j = m has not been reached, j is incremented in step S5, and the process returns to step S2. If j = m has been reached, the corrected neighborhood teacher data is output, and this processing ends.

  As described above, when the production condition of the prediction target material exists outside the range R of the neighborhood teacher data, the production condition is excluded from the neighborhood teacher data. It is possible to eliminate the state. As a result, the influence of noise is not greatly affected or amplified, and the occurrence of outlier prediction values can be suppressed and the prediction accuracy of model error prediction values can be improved.

(Second Embodiment)
In the first embodiment, when a certain manufacturing condition of the prediction target material exists outside the range R of the vicinity teacher data, the vicinity teacher data is corrected. In the second embodiment, the prediction target material Modify manufacturing conditions. The configuration of the cooling control system including the local regression model generation unit 10 is the same as that of the first embodiment, and the following description will focus on differences from the first embodiment. A description of the common points is omitted.

  In the present embodiment, the correction unit 102 is in an extrapolation state when a certain manufacturing condition of the prediction target material is in an extrapolation state that exists outside the range R of the neighborhood teacher data input by the input unit 101. The manufacturing conditions of the material to be predicted are corrected so as to avoid this. Specifically, when a certain manufacturing condition of the prediction target material is larger than the maximum value of the manufacturing condition included in the vicinity teacher data, the manufacturing condition of the prediction target material is replaced with the maximum value of the vicinity teacher data, When the manufacturing condition is smaller than the minimum value of the manufacturing condition included in the vicinity teacher data, the manufacturing condition of the prediction target material is replaced with the minimum value of the vicinity teacher data.

In this case, the local weighted regression calculation unit 103 uses the neighborhood teacher data input by the input unit 101 to perform local weighted linear multiple regression (equation (4), equation (5)) or local weighted PLS regression ( A locally weighted regression model is generated according to equations (6) to (8)).
Then, the model error calculation unit 104 inputs the manufacturing condition of the prediction target material (or the corrected manufacturing condition when corrected by the correction unit 102) to the local weighted regression model generated by the local weighted regression calculation unit 103. The model error prediction value of the prediction target material is calculated.

FIG. 5 shows an example of correction processing for the manufacturing conditions of the material to be predicted by the correction unit 102.
In step S11, j = 1 is set.
In steps S12 and S14, it is determined whether or not the manufacturing condition x _{0, j} of the prediction target material is in an extrapolated state, that is, whether or not it exists outside the range R of the neighborhood teacher data. Specifically, in step S12, is the production condition x _{0, j of the} material to be predicted smaller than the minimum value min (x _{1, j} ,..., X _{n, j} ) of the production condition x _{i, j} ? If it is determined that the value is smaller than the minimum value, the process proceeds to step S13. If the value is equal to or greater than the minimum value, the process proceeds to step S14. In step S14, it is determined whether the production condition x _{0, j of the} material to be predicted is larger than the maximum value max (x _{1, j} ,..., X _{n, j} ) of the production condition x _{i, j.} If greater, the process proceeds to step S15, and if less than the maximum value, the process proceeds to step S16.
In step S13, the production condition x0 _{, j} of the prediction target material is replaced with the minimum value min (x1 _{, j} ,..., _{Xn, j} ) of the neighborhood teacher data, and the process proceeds to step S16. In step S15, the production condition x0 _{, j} of the prediction target material is replaced with the maximum value max (x1 _{, j} ,..., _{Xn, j} ) of the neighborhood teacher data, and the process proceeds to step S16.
In step S16, it is determined whether j = m. If j = m has not been reached, j is incremented in step S17, and the process returns to step S12. If j = m has been reached, the manufacturing conditions of the corrected material to be predicted are output, and this process ends.

  As described above, when the manufacturing condition of the prediction target material exists outside the range R of the vicinity teacher data, the manufacturing condition of the prediction target material is adjusted to the maximum value or the minimum value of the manufacturing condition included in the vicinity teacher data. Therefore, it is possible to eliminate the extrapolation of the manufacturing condition of the prediction target material. As a result, the influence of noise is not greatly affected or amplified, and the occurrence of outlier prediction values can be suppressed and the prediction accuracy of model error prediction values can be improved.

The present invention can be applied not only to the prediction and control of the cooling stop temperature of the steel sheet in the cooling process, but also to a general application, and in a manufacturing process for manufacturing a product from raw materials at a certain manufacturing facility, a state prediction model is used. This is widely effective when a predicted value of a product state is calculated and manufacturing equipment is controlled based on the predicted value.
That is, the steel plate in the embodiment is an example of a product, the cooling process and equipment are examples of a manufacturing process and equipment, the cooling stop temperature of the steel sheet is an example of a product state, and the temperature prediction model is an example of a state prediction model. And when applying this invention to a general manufacturing process, what is necessary is just to make a state control system the same structure as FIG. 1, for example, the cooling stop temperature calculation part 7 in embodiment is made into a state calculation part, and a cooling setting calculation part 11 is set as a setting calculation unit, and the instruction value of the cooling water amount and the conveying speed of the steel sheet may be replaced with the instruction value of the operation amount of the manufacturing facility.
From the manufacturing condition and the actual value of the manipulated variable, the state calculation unit predicts the state of the product using the state prediction model, and the model error actual value that is an error between the predicted state value and the actual state value is the database 9. To accumulate. Then, when calculating the predicted value of the state of the prediction target product, the neighborhood teacher data is extracted from the database 9 storing the actual data based on the distance function, and the neighborhood teacher data is used for the local model generation unit 10. A local weighted regression model for obtaining the model error is generated, and the prediction calculated by the state prediction model is based on the predicted value of the model error of the prediction target product calculated by giving the manufacturing condition to the local weighted regression model. Correct the predicted value of the target product. By correcting in this way, the predicted value of the state can be brought close to the actual value. In addition, the setting calculation unit calculates the instruction value for the operation amount of the manufacturing facility so that the predicted value in the corrected state matches the target value determined for each prediction target product, and controls the manufacturing facility. As a result, the actual value of the product state can be brought closer to the target value.

  For example, when the present invention is applied to sheet width prediction and control in hot rolling, the cooling facility in the first embodiment is a sheet width processing facility such as an edger or a sizing press. In addition, the manufacturing conditions are factors that are considered to affect the plate width, such as the steel plate size, the steel type, and the apparatus entry-side steel plate temperature. The cooling nozzle 2 and the transfer table 3 serve as hydraulic pressure reduction devices if they are edgers, and actuate actuators for plate width control such as mechanical cranks if they are sizing presses. The thermometer 4 is a plate width measuring device such as a line sensor. The cooling process computer 5 is a rolling process computer, and the upper rolling computer 6 is a heating process computer or a business computer. Moreover, the cooling stop temperature calculation part 7 becomes a board width calculation part which calculates an apparatus delivery side board width based on the model of a nonpatent literature 1, for example. Further, the cooling setting calculation unit 11 is based on a model described in Non-Patent Document 1, for example, and a width reduction setting calculation unit that sets the opening degree and speed of the edger and the sizing press so that the apparatus delivery side plate width becomes a desired value. Become.

The effect of the method to which the present invention is applied was verified by numerical experiments.
FIG. 6 shows a functional configuration for a numerical experiment. A numerical experiment is performed by reproducing the database 9 and the local regression model generation unit 10 with a personal computer. Manufacturing conditions are input to a personal computer from a factory database that has accumulated operational results in the factory, and model error prediction values are output by the database 9 and the local regression model generation unit 10.
Since the model error actual value is also stored in the factory database, it was compared with the model error predicted value by the database 9 and the local regression model generation unit 10 to evaluate its accuracy. Unlike FIG. 1, the model error prediction value is not reflected in actual operation, but it is considered that the accuracy of the cooling stop temperature control is improved by accurately predicting the model error.

  On the basis of the contents described in Patent Document 1, the local regression model generation unit 10 uses local weighted PLS regression, and when the neighborhood teacher data is not corrected (comparison method), the neighborhood teacher data is obtained according to the first embodiment. The prediction accuracy of the model error was evaluated for each of the cases where the correction was made (Invention Method 1) and where the manufacturing conditions of the material to be predicted were corrected according to the second embodiment (Invention Method 2).

FIG. 8 is a diagram illustrating the relationship between the model error actual value and the model error predicted value in the comparison method. As described above, the predicted deviation value 801 is generated under the influence of noise. As an evaluation index, the mean square error RMSE was 16.11 ° C., and the coefficient of determination (contribution rate) R ² was 0.088.

FIG. 7A is a diagram showing the relationship between the model error actual value and the model error predicted value in Invention Method 1, and FIG. 7B is the relationship between the model error actual value and the model error predicted value in Invention Method 2. FIG. In the invention methods 1 and 2, it is possible to suppress the occurrence of outlier predicted values. In invention method 1, the mean square error RMSE was 10.82 ° C., and the coefficient of determination (contribution rate) R ² was 0.588, which is a better result than the comparison method. In Invention Method 2, the mean square error RMSE was 10.81 ° C. and the coefficient of determination R ² was 0.589, which was a better result than the comparative method.

A product state prediction apparatus, a steel sheet cooling stop temperature prediction apparatus, and a steel sheet cooling control apparatus to which the present invention is applied are realized by, for example, a computer apparatus including a CPU, a ROM, a RAM, and the like. In FIG. 1, an example in which the cooling stop temperature calculation unit 7, the local regression model generation unit 10, and the cooling setting calculation unit 11 operate on one process computer (cooling process computer 5) has been described. It is possible to operate in such a manner that input / output is performed via a network. Moreover, although the manufacturing conditions are input from the rolling process computer 6, the present invention is applied to a configuration that is common as a form of accelerated cooling control in hot rolling, such as input from a higher-level business computer. Applicable.
The present invention also provides software (program) that implements the functions of the present invention to a system or apparatus via a network or various storage media, and the system or apparatus computer reads out and executes the program. It is feasible.

  1: Steel plate, 2: Cooling nozzle, 3: Transfer table, 4: Thermometer, 5: Cooling process computer, 7: Cooling stop temperature calculation unit, 8: Subtractor, 9: Database, 10: Local regression model generation unit, 11: Cooling setting calculation unit, 101: Input unit, 102: Correction unit, 103: Regression calculation unit with local weight, 104: Model error calculation unit, 105: Output unit

Claims (9)

A product state prediction device for calculating a predicted value of a product state by a state prediction model in a product manufacturing process,
An extraction means for extracting neighborhood teacher data based on a distance function from a database in which performance data including manufacturing conditions is accumulated when calculating a predicted value of a state of a prediction target product;
When the manufacturing condition of the prediction target product is outside the range of the vicinity teacher data extracted by the extraction unit, at least one of the vicinity teacher data extracted by the extraction unit and the manufacturing condition of the prediction target product is Correction means to correct;
Regression model for obtaining an error in the predicted value of the state of the prediction target product calculated by the state prediction model using the vicinity teacher data extracted by the extraction unit or the corrected vicinity teacher data when corrected by the correction unit And calculating means for calculating the error according to the regression model and the manufacturing condition of the prediction target product or the corrected manufacturing condition when corrected by the correcting means,
Correction means for correcting the predicted value of the state of the prediction target product calculated by the state prediction model based on the error of the predicted value of the state of the prediction target product obtained by the calculation means; Product state prediction device.   The correction means excludes the manufacturing conditions from the vicinity teacher data extracted by the extraction means when the manufacturing conditions of the prediction target product are outside the range of the vicinity teacher data extracted by the extraction means. The product state prediction apparatus according to claim 1.   When the manufacturing condition of the prediction target product is larger than the maximum value of the manufacturing condition included in the proximity teacher data extracted by the extraction means, the correction means sets the manufacturing condition of the prediction target product as the vicinity teacher data. And the manufacturing condition of the prediction target product is replaced with the minimum value of the vicinity teacher data when the manufacturing condition is smaller than the minimum value of the manufacturing condition included in the vicinity teacher data extracted by the extraction means. The product state prediction apparatus according to claim 1. The database stores, as actual data, the difference between the actual product state value and the predicted product state value calculated using the state prediction model in association with manufacturing conditions,
4. The extraction unit according to claim 1, wherein the extraction unit extracts performance data having manufacturing conditions similar to the manufacturing conditions of the prediction target product as neighborhood teacher data based on the distance function. 5. The state prediction apparatus of the product described in 1.   5. The apparatus according to claim 1, wherein the product is a steel plate, the manufacturing process is a cooling step, the state of the product is a cooling stop temperature of the steel plate, and the state prediction model is a temperature prediction model. Product state prediction device. The product state prediction apparatus according to any one of claims 1 to 4,
Control for controlling the operation amount of the manufacturing equipment used in the manufacturing process so that the predicted value of the state of the prediction target product corrected by the correction means matches the target value previously determined for each prediction target product And a product state control device.   The product is a steel plate, the manufacturing process is a cooling step, the product state is a cooling stop temperature of the steel plate, the state prediction model is a temperature prediction model, and the operation amount is at least one of a cooling water amount and a steel plate conveyance speed. The product state control device according to claim 6. A product state prediction method for calculating a predicted value of a product state using a state prediction model in a product manufacturing process,
When calculating the predicted value of the state of the prediction target product, an extraction step of extracting neighborhood teacher data based on a distance function from a database in which performance data including manufacturing conditions is accumulated;
When the manufacturing condition of the prediction target product is outside the range of the vicinity teacher data extracted in the extraction step, at least one of the vicinity teacher data extracted in the extraction step and the manufacturing condition of the prediction target product Correction steps to correct;
Regression model for obtaining an error in the predicted value of the state of the prediction target product calculated by the state prediction model using the vicinity teacher data extracted in the extraction step or the corrected vicinity teacher data when corrected in the correction step Generating the error and calculating the error according to the corrected manufacturing condition when the correction model is corrected in the manufacturing condition of the prediction target product or the correction step;
A correction step of correcting the predicted value of the state of the prediction target product calculated by the state prediction model based on the error of the predicted value of the state of the prediction target product obtained in the calculation step. Product state prediction method. A program for calculating a predicted value of a product state using a state prediction model in a product manufacturing process,
An extraction means for extracting neighborhood teacher data based on a distance function from a database in which performance data including manufacturing conditions is accumulated when calculating a predicted value of a state of a prediction target product;
When the manufacturing condition of the prediction target product is outside the range of the vicinity teacher data extracted by the extraction unit, at least one of the vicinity teacher data extracted by the extraction unit and the manufacturing condition of the prediction target product is Correction means to correct;
Regression model for obtaining an error in the predicted value of the state of the prediction target product calculated by the state prediction model using the vicinity teacher data extracted by the extraction unit or the corrected vicinity teacher data when corrected by the correction unit And calculating means for calculating the error according to the regression model and the manufacturing condition of the prediction target product or the corrected manufacturing condition when corrected by the correcting means,
A program for causing a computer to function as correction means for correcting the predicted value of the state of the prediction target product calculated by the state prediction model based on the error of the predicted value of the state of the prediction target product obtained by the calculation means .

Images