JP6915428B2 - Manufacturing process control systems, methods and programs - Google Patents

Description

The present invention relates to a manufacturing process control system , method and program using a physical model that describes the behavior of the manufacturing process of a product.

In the manufacturing process of a product, a physical model that explains the behavior of the manufacturing process from a physical point of view is constructed, and the amount of operation in the manufacturing process is such that the product to be manufactured (hereinafter referred to as a target product) satisfies the manufacturing conditions. Control to set up (hereinafter referred to as setup control) is widely used.
However, the values of the parameters of the physical model are uncertain and may not be directly observed, so it is often difficult to construct a physical model that accurately explains the behavior of the manufacturing process.
For example, in plastic working of steel materials, it is necessary to control the temperature of the steel materials in order to obtain desired shapes and mechanical properties. However, there is no means to measure the temperature including the inside of the steel material, and when estimating by calculation, the unit amount of heat generated by the deformation of the material is required, but the knowledge so far derives this unit amount. There is no established way to do this.
Further, for example, in the downstream of the hot rolling process of the steel sheet, accelerated cooling is performed to cool the steel sheet to a predetermined temperature, thereby obtaining the mechanical properties required for the product. The heat transfer coefficient required to calculate this cooling temperature differs depending on how the cooling water comes into contact with the steel sheet and the surface properties of the steel sheet, but there is no means to measure the state of the cooling water, and the surface properties of the steel sheet are the heat transfer coefficient. The quantitative elucidation of the effect on the disease has not yet been made.

In view of these points, learning control that compensates for the uncertainty of the physical model is used from an empirical / statistical point of view rather than a theoretical or measurement technique.
For example, a learning control method that collects observable output actual values of the manufacturing process, identifies parameters so that the actual output values and the predicted output values calculated by the physical model match, and reflects them in the setup control from the next time onward. Is adopted. A specific method for identifying the parameters of the physical model is disclosed in, for example, Non-Patent Document 1.
In this case, even if an unknown parameter is uniquely identified over the entire range that the manufacturing condition can take (hereinafter referred to as the manufacturing condition space) and reflected in the control of the target product, the actual position in the manufacturing condition space is actually obtained. Since the true value of the parameter changes due to this, an error occurs and the control accuracy is often not maximized.
In the above-mentioned example of temperature calculation in plastic working of steel materials, the unit amount of heat generated by the deformation of the material changes depending on the steel type. Further, also in the above-mentioned example of accelerated cooling, the distance from the nozzle changes depending on the thickness of the steel sheet, so that the way the cooling water comes into contact with the steel sheet changes, and the surface texture also changes depending on the steel type, so that the heat transfer coefficient also changes.

Therefore, when manufacturing a product, the manufacturing conditions are classified by a certain value, and when the product is manufactured, the result of identifying the parameters is reflected in the classification of the manufacturing conditions to which the product belongs or the classification around it. , A stratified table learning method that applies the parameters of the division to which the product belongs to the physical model is often adopted.
FIG. 13 shows a configuration example of a control system for a manufacturing process that employs a stratified table learning method. The setup calculation unit 902 uses a physical model to determine the amount of operation of the manufacturing process 901 under the manufacturing conditions of the target product. At this time, the values obtained from the stratified table 904 are applied to the predetermined parameters of the physical model according to the manufacturing conditions of the target product. The parameter learning calculation unit 903 collects the observable output actual value of the manufacturing process 901, learns a predetermined parameter so that the output actual value and the output predicted value by the physical model match, and layers the result. It is reflected in the classification of the manufacturing conditions to which the target product belongs or the classification around it in the separate table 904.

As a technique related to the stratified table learning method, for example, in Patent Document 1, when correcting the learning coefficient corresponding to each material layer related to the processing in the process line, the error between the model predicted value and the actual value tends to be predetermined. It is described that the learning coefficient related to other strata having the above is also corrected.

When the non-linearity of the manufacturing process is strong and it is difficult to learn the parameters, a skilled operator may manually adjust the parameters of each category without learning the parameters by the parameter learning calculation 903.

Japanese Unexamined Patent Publication No. 6-259107 Japanese Patent No. 3351376

Compute Roll No. 2, Corona Publishing Co., Ltd., 1983, p49-59 Illustrated How to Learn Heat Transfer Engineering, Ohmsha, 1982

In the stratified table learning method, when the stratified table is divided, if the division is too large, the true value of the parameter within the same division is the same as the case where the parameter is uniquely identified over the entire manufacturing condition space as described above. Therefore, an error may occur and the control accuracy may not be maximized. In addition, if the categories are too small, there are few manufacturing records that fall into each category, and learning does not proceed in categories such as rare materials that are rarely manufactured, which causes an error. Patent Document 1 proposes a method of reflecting a value of a certain stratum on the value of another stratum as a countermeasure, but if the range of reflection is too wide, the classification will be classified. As with the case of being too large, the control accuracy is low.

Further, in the stratified table learning method, if the manufacturing conditions used for each stratum are increased, the classification of the stratified table becomes finer, which causes an error as described above. Therefore, the number of manufacturing conditions that can be used for each layer is limited, and even manufacturing conditions that are considered to affect the parameters of the physical model may not be included in the manufacturing conditions that are used for each layer.

The present invention has been made in view of the above points, and makes it possible to accurately identify predetermined parameters of a physical model that explains the behavior of a manufacturing process when manufacturing a specified target product, and to perform physics. The purpose is to improve the control accuracy of the control of the manufacturing process using the model.

The gist of the present invention for solving the above problems is as follows.
[1] A manufacturing process control system that controls the manufacturing process under the manufacturing conditions of the target product by using a physical model that explains the behavior of the manufacturing process of the product.
The manufacturing conditions of products manufactured in the past and the output actual value of the manufacturing process from a database that stores manufacturing performance data linked, manufacturing performance data based on the distance of the manufacturing conditions in the space between the production conditions of the target product Extraction means to extract
An identification calculation means that performs identification calculation of a predetermined parameter of the physical model using the manufacturing record data extracted by the extraction means, and an identification calculation means.
A regression calculation means for obtaining an error prediction value by predicting an error of an output predicted value by the physical model reflecting the predetermined parameter identified and calculated by the identification calculation means by a regression model.
A manufacturing process control system including a correction means for correcting an output predicted value by the physical model by using the error predicted value obtained by the regression calculation means when controlling the manufacturing process.
[2] The identification calculation means obtains the value of the predetermined parameter based on an objective function including an error between the output actual value in the production actual data extracted by the extraction means and the output predicted value by the physical model. The manufacturing process control system according to [1].
[3] The control system for a manufacturing process according to [2], which imposes a constraint condition that the value of the predetermined parameter is included in a predetermined range.
[4] In the objective function, in addition to the error between the output actual value in the manufacturing actual data extracted by the extraction means and the output predicted value by the physical model, the predetermined parameter value and the preset predetermined value are set. The control system for a manufacturing process according to [2] or [3], which further includes an error from a reference value of a parameter.
[5] to the physical model, and the predetermined parameters identified calculated by the identification calculation unit, wherein by inputting the manufacturing conditions of the products calculated output prediction value of the target product, the output predicted value the target A setup calculation means for determining the operation amount of the manufacturing process so as to substantially match the output target value under the manufacturing conditions of the product is provided.
The manufacturing process control system according to any one of [1] to [4], wherein the setup calculation means functions as the correction means and corrects an output predicted value of the target product.
[6] The setup calculation means determines the operation amount by performing a convergence calculation in which the operation amount is modified and the calculation of the physical model is repeated until the output prediction value substantially matches the output target value. The manufacturing process control system according to [5].
[7] the regression calculation means,
Using the physical model that reflects the predetermined parameters identified and calculated by the identification calculation means, the output predicted value for the manufacturing conditions in the manufacturing record data extracted by the extraction means is calculated.
An error between the output result value in the extracted manufacturing performance data by the extracting means and the output prediction value to generate the regression model by regressing the manufacturing conditions as explanatory variables,
Giving the products of manufacturing conditions on the regression model, the control of the manufacturing process according to any one of the characterized in that said asking you to the error prediction value for the target products [1] to [6] system.
[8] A method for controlling a manufacturing process that controls the manufacturing process under the manufacturing conditions of the target product by using a physical model that explains the behavior of the manufacturing process of the product.
Extraction means, from a database that stores manufacturing performance data and an output actual value correlated manufacturing conditions of products manufactured in the past and the manufacturing process, based on the distance of the manufacturing conditions in the space between the production conditions of the target product And the steps to extract manufacturing record data
A step in which the identification calculation means performs identification calculation of a predetermined parameter of the physical model using the extracted manufacturing record data.
A step in which the regression calculation means obtains an error predicted value by predicting an error of the output predicted value by the physical model reflecting the predetermined parameter identified and calculated by the regression model.
A method for controlling a manufacturing process, wherein the correction means includes a step of correcting an output predicted value by the physical model by using the error predicted value when controlling the manufacturing process .
[9] A program for controlling the manufacturing process under the manufacturing conditions of the target product by using a physical model that explains the behavior of the manufacturing process of the product.
The manufacturing conditions of products manufactured in the past and the output actual value of the manufacturing process from a database that stores manufacturing performance data linked, manufacturing performance data based on the distance of the manufacturing conditions in the space between the production conditions of the target product And the process of extracting
Using manufacturing performance data the extracted, the process for identifying calculation of certain parameters of the physical model,
A process of obtaining an error predicted value by predicting an error of an output predicted value by the physical model reflecting the predetermined parameter identified and calculated by a regression model.
A program for causing a computer to perform a process of correcting an output predicted value by the physical model using the error predicted value when controlling the manufacturing process.

According to the present invention, it is possible to accurately identify predetermined parameters of a physical model that explains the behavior of the manufacturing process when manufacturing a specified target product, and control accuracy of control of the manufacturing process using the physical model. Can be improved .

It is a figure which shows the structural example of the control system of the manufacturing process which concerns on 1st Embodiment. It is a figure which shows the structural example of the control system of the manufacturing process which concerns on 2nd Embodiment. It is a flowchart which shows the example of the convergence calculation by the setup calculation part in 2nd Embodiment. It is a figure which shows the outline of the steel pipe manufacturing process by a piercer. It is a characteristic figure which shows the machining end temperature predicted value and the machining end temperature actual value in the comparative example (1). It is a characteristic figure which shows the machining end temperature predicted value and the machining end temperature actual value in the comparative example (2). It is a characteristic figure which shows the machining end temperature predicted value and the machining end temperature actual value in Example (1). It is a characteristic figure which shows the machining end temperature predicted value and the machining end temperature actual value in Example (2). It is a characteristic figure which shows the value of the parameter obtained in Example (1). It is a characteristic figure which shows the value of the parameter obtained in Examples (3) and (4). It is a characteristic diagram which shows the reference value of a parameter. It is a characteristic figure which shows the machining end temperature predicted value and the machining end temperature actual value in Examples (3) and (4). It is a figure which shows the configuration example of the control system of the manufacturing process which adopted the stratified table learning method.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.
[First Embodiment]
FIG. 1 shows a configuration example of a control system for a manufacturing process according to the first embodiment.
The manufacturing process control system uses a physical model that explains the behavior of the manufacturing process 101 of the product to perform setup control for obtaining the operation amount of the manufacturing process 101 under the manufacturing conditions of the target product. When the manufacturing process 101 is plastic working of a steel material or accelerated cooling of a steel sheet, the manufacturing conditions are considered necessary to explain the behavior of the manufacturing process 101, for example, the size of the steel material, the component value of the steel material, the process start temperature, and the like. It becomes an item. The target product is a product to be manufactured from now on, and is a product to be identified as a parameter as described below.

The setup calculation unit 102 obtains the operation amount of the manufacturing process 101 under the manufacturing conditions of the target product by using the physical model. At this time, the value obtained by the parameter identification calculation unit 104 is applied to the predetermined parameter of the physical model. As a method for obtaining the manipulated variable of the manufacturing process, a convergent calculation method in which the manipulated variable is modified and the calculation is repeated until the output predicted value, which is a calculated value by the physical model, substantially matches the output target value, is generally used. It should be noted that the term "substantially matching" as used herein means that the condition that the predicted output value is within a predetermined range based on the output target value is satisfied. Further, the method is not limited to the convergence calculation method, and a method in which the physical model calculation is performed only once and the operation amount is corrected according to the difference between the output predicted value and the output target value may be used.

The database 105 stores and accumulates manufacturing record data in which the manufacturing conditions of the products manufactured in the past and the observable output actual values of the corresponding manufacturing process 101 are linked.

The extraction unit 103 extracts the manufacturing record data as similar data from the database 105 based on the distance in the manufacturing condition space from the manufacturing condition of the target product. When the manufacturing conditions of the target product are given, the extraction unit 103 calculates the distance between the manufacturing conditions of the target product and the manufacturing conditions in the manufacturing record data stored in the database 105 in the manufacturing condition space, and this distance is close. A predetermined number of manufacturing record data are extracted as similar data in order. As the distance in the manufacturing condition space, those known as the definition of the distance in the multidimensional space such as the Euclidean distance and the Mahalanobis distance can be applied. The manufacturing conditions that define the manufacturing condition space may be a part of the manufacturing conditions, not all of the manufacturing conditions.

The parameter identification calculation unit 104 performs the parameter identification calculation of the physical model using the manufacturing record data extracted from the database 105 by the extraction unit 103. The value of a predetermined parameter obtained by the parameter identification calculation unit 104 is given to the setup calculation unit 102.
Now, suppose that the physical model is represented by equation (1). y'is the output predicted value by the physical model, x is the manufacturing condition, and p is the parameter.

Let the manufacturing record data (similar data) extracted from the database 105 be {x _i , y _i } (i = 1, ..., N). The subscript i is an index of similar data, and N is the number of similar data. x _i is the manufacturing condition of the i-th similar data, and y _i is the output actual value of the corresponding manufacturing process 101.
The parameter identification calculation unit 104 uses the physical model of the equation (1) _{to obtain the output predicted value y i} ′ of the i-th similar data by the equation (2), as in the equation (3) or the equation (4). , The value of the parameter p that minimizes the objective function J represented by the sum of squares or the sum of the absolute values of the error with the _{actual output value y i} of the similar data extracted from the database 105 is obtained.
A specific method for identifying the parameters of the physical model is disclosed in, for example, Non-Patent Document 1, and other known parameter identification methods can be applied.

As described above, instead of classifying the manufacturing conditions as in the stratified table learning method, the predetermined parameters of the physical model are organically determined based on the manufacturing performance data extracted based on the distance in the manufacturing condition space. Therefore, it is possible to accurately identify a predetermined parameter of the physical model that explains the behavior of the manufacturing process when manufacturing the specified target product.
That is, the manufacturing record data in which the distance in the manufacturing condition space is short is preferentially adopted, and the manufacturing record data in which the distance in the manufacturing condition space is long and the true value of the parameter changes is difficult to be reflected in the parameter identification calculation. Therefore, it is possible to eliminate the deterioration of the control accuracy that occurs when the division is too large as in the stratified table learning method.
In addition, for rare materials, there are not many manufacturing record data that are close to each other in the manufacturing condition space, but even when there is no manufacturing record data nearby, manufacturing record data with similar manufacturing conditions as much as possible. Can be extracted and reflected in the parameter identification calculation, so that it is possible to suppress the occurrence of an error due to the learning not progressing as in the stratified table learning method.
Further, in the stratified table learning method, the number of manufacturing conditions that can be used for each layer is limited, but in the parameter identification calculation to which the present invention is applied, the number of manufacturing conditions that can be used is theoretically infinite. be.

By the way, if the value of the parameter p that minimizes the evaluation function J is obtained by focusing only on the error between the _{actual output value y i} and the predicted output value y _{i ´, the value of the parameter p is excessively adjusted in some cases.} As a result, the value of the parameter p may be an unrealistic value. This occurs when the error between the actual output value y _i and the predicted output value y _i ′ is due to a factor other than the parameter p of the physical model. Even if the error between the actual output value y _i and the predicted output value y _i ´ becomes smaller in this way, if the value of the parameter p is unrealistic, the reliability of the predicted value of the physical model may decrease. There is.

Therefore, the value of the parameter p of the physical model may be limited by the following two methods. In addition, either of the first method and the second method described below may be used alone, or both of them may be used in combination.
The first method is a method of imposing a constraint condition that the value of the parameter p is included in a predetermined range.
_{Specifically, p Lj} ≤ p _j ≤ p _U j, where the subscript j is the index of the parameter p, p _{L j is} the lower limit of the j-th parameter p _{j, and p U j is} the upper limit of the j-th parameter p j. Set the range as follows. The lower limit value p _Lj and the upper limit value p _Uj shall be known in advance by physical consideration of the target manufacturing process.
If the value of the parameter p that minimizes the objective function J of the equation (3) or the equation (4) is obtained under this constraint condition, it is possible to prevent the value of the parameter p from becoming an unrealistic value. ..
Although an example of setting both the lower limit value p _Lj and the upper limit value p _Uj has been described, only one of them may be set.

In the second method, the objective function J further includes an error between the value of the parameter p and the preset reference value of the parameter p, in addition to the error between the _{actual output value y i} and the predicted output value y _{i ´.} This is the method.
Specifically, as in Eq. (3.1) or Eq. (4.1), the evaluation function J is added to the term representing the error between the _{actual output value y i} and the predicted output value y _{i ′, and a parameter.} A _{term representing the error between the value of p j} and its reference value p _0j (a term represented by the sum of squares of errors or the sum of absolute values) is included, and the evaluation function J is minimized by the parameter p. Find the value. p _0j is the reference value of the j-th parameter p _{j , K j} is the weighting coefficient for adjusting the magnitude of the deviation of the value of the j-th parameter p _j from the reference value, and M is the number of parameters of the physical model. The reference value p _j0 and the weighting coefficient K _j may be constant values or different values depending on the manufacturing conditions to be predicted.

In the physical model, the approximate value of the parameter p may be known from the physical knowledge of the target manufacturing process (in the case of physical property values, the reference value can be obtained from the science chronology, etc.) and the operator's past experience. , There is a case where it is desired to determine the value of the parameter p within a range that does not deviate significantly from the reference value. In such a case, if the value of the parameter p is determined based on the evaluation function including the error between the value of the parameter p and its reference value as in the equation (3.1) or the equation (4.1). , It is possible to prevent the value of the parameter p from becoming an unrealistic value.

In equation (3.1), both the first and second terms on the right side are quadratic equations, and in equation (4.1), both the first and second terms on the right side are linear equations. However, it is not necessary to make the order the same, and it can be selected according to the characteristics of the problem. For example, the first term on the right side may be a quadratic expression, the second term may be a linear expression, or vice versa. For example, when noise is included in the actual output value and an outlier exists, it is better to make the first term on the right side a linear expression to obtain a robust identification result that is less susceptible to noise. On the other hand, if the second term on the right side is made into a linear expression, it has the effect of reducing the number of parameters that change from the reference value, and when there are many parameters, which one is used to match the output predicted value with the actual output value. It makes it easier to see if the parameters need to be adjusted.

[Second Embodiment]
In the second embodiment, the regression calculation unit 106 obtains an error prediction value based on the control system of the manufacturing process according to the first embodiment, and the setup calculation unit 102 calculates the error prediction value using this error prediction value. An example of adding a configuration for correcting the output predicted value by the physical model of the product will be described.
FIG. 2 shows a configuration example of a control system for the manufacturing process according to the second embodiment. The elements common to the first embodiment are designated by the same reference numerals, the description thereof will be omitted, and the differences from the first embodiment will be mainly described.

The regression calculation unit 106 is given the manufacturing conditions of the target product, the manufacturing record data extracted from the database 105 by the extraction unit 103 as neighborhood teacher data, and the values of predetermined parameters obtained by the parameter identification calculation unit 104.
The regression calculation unit 106 inputs the predetermined parameters identified by the parameter identification calculation unit 104 and the production conditions in the production record data extracted from the database 105 by the extraction unit 103 into the equation (2), and the manufacturing conditions in the production record data are set. Calculate the output prediction value. Then, a regression model is generated in which the error between the output predicted value and the output actual value in the actual production data extracted from the database 105 by the extraction unit 103 is used as the objective variable and the manufacturing conditions are used as the explanatory variables. As the regression method, known regression methods such as linear multiple regression, PLS (Partial Least Squares) regression, and Ridge regression can be applied. Then, the manufacturing conditions of the target product are given to the generated regression model, and the error prediction value for the target product is obtained. The manufacturing conditions used as explanatory variables are not all of the manufacturing conditions, but may be a part of the manufacturing conditions (for example, the size of the steel material and the component value of the steel material).

Here, the manufacturing record data as similar data used by the parameter identification calculation unit 104 and the manufacturing record data as neighborhood teacher data used by the regression calculation unit 106 may be the same or different. Since the regression calculation and the parameter identification calculation predict different values, it is not necessary to make the manufacturing record data the same, and the manufacturing record data is extracted using the manufacturing conditions that are highly correlated with the predicted values. However, it is possible to calculate highly accurate coefficients (regression coefficients and parameter values). For example, as the manufacturing record data (similar data) used by the parameter identification calculation unit 104, while extracting the manufacturing record data having a short distance defined in the manufacturing condition space of the manufacturing condition A (for example, the size of the steel material and the process start temperature), As the manufacturing record data (neighborhood teacher data) used in the regression calculation unit 106, the manufacturing record data having a short distance defined in the manufacturing condition space of the manufacturing condition B (for example, the size of the steel material and the component value of the steel material) is extracted. You may. Although it is shown as one extraction unit 103 in FIG. 2, for example, it may be divided into an extraction unit for extracting similar data and an extraction unit for extracting neighborhood teacher data.

As described in the first embodiment, when the setup calculation unit 102 obtains the operation amount of the manufacturing process 101, the operation amount is calculated by modifying the operation amount until the output predicted value by the physical model substantially matches the output target value. A convergence calculation method that repeats the above is common.
FIG. 3 shows an example of convergence calculation by the setup calculation unit 102 in the second embodiment. The process of FIG. 3 is executed with a temporary operation amount of the manufacturing process set.
In step S301, the setup calculation unit 102 obtains an output predicted value under the manufacturing conditions of the target product using the physical model. At this time, the value obtained by the parameter identification calculation unit 104 is applied to the predetermined parameter of the physical model.
In step S302, the setup calculation unit 102 corrects the output predicted value obtained in step S301 using the error predicted value obtained by the regression calculation unit 106, and then determines whether or not it substantially matches the target value. .. As a result, if the output predicted value substantially matches the output target value (step S302: Y), the temporarily set temporary operation amount is adopted and output as the operation amount. On the other hand, if the output predicted value does not substantially match the output target value (step S302: N), the process proceeds to step S303, the currently set temporary operation amount is corrected, and the process returns to step S301.

[Example A]
The effect of improving the control accuracy by applying the present invention was verified by numerical experiments. In the following, Example (1) corresponds to the first embodiment, and Example (2) corresponds to the second embodiment. In Example A, among the methods described in the first embodiment, a method that does not limit the value of the parameter is applied.
The manufacturing process targeted in Example A was a steel pipe manufacturing process using a piercer, and the processing end temperature by the piercer was set as an output to be predicted. FIG. 4 shows an outline of the steel pipe manufacturing process by Piasa. The billet 402 is a material to be processed, and the billet 402 is sandwiched while rotating the roll 401, which is a tool, and the plug 403 is supported by the bar 404 and pushed into the billet 402 while rotating to make a hole in the billet 402. Process into a tube with a diameter. The processing end temperature at the time of manufacturing a steel pipe by a piercer can be calculated by referring to Non-Patent Document 2 and Patent Document 2.

When the steel pipe is divided into small elements in the radial direction, the heat balance equation of the elements at the radial position r is expressed by Eq. (5).

In equation (5), c is the specific heat, ρ is the density, v _r is the volume of the element, and _Tr is the temperature of the element.
Q _next is the heat input due to heat transfer from the adjacent element, and is expressed by the equation (6). λ is the thermal conductivity and _Ar is the contact area with the adjacent element.
Q _fric is the frictional heat with the tool (roll) 401 and is expressed by the formula (7). η is the coefficient of conversion from work to heat, μ is the coefficient of friction, P is the contact surface pressure _{between the tool and the steel pipe, and ΔV s} is the speed difference between the tool and the steel pipe.
Q _form is the heat generated by the deformation of steel and is expressed by the formula (8). α is the conversion coefficient from deformation to heat, ε is strain, and K _f is the deformation resistance.

In the physical models of equations (5) to (8), most of the coefficients and variables are measurable, and theoretical equations and experimental equations are given, but the conversion coefficient α is formulated in the previous knowledge. However, since it differs for each manufacturing process and each manufacturing condition, it is a parameter that must be determined empirically, and it is a predetermined parameter that should be identified this time.

The numerical experiment was carried out by reproducing the setup calculation unit 102, the extraction unit 103, the parameter identification calculation unit 104, the database 105, and the regression calculation unit 106 shown in FIGS. 1 and 2 in the computer. The data for the numerical experiments used was from a database (called a factory database) that accumulates the operational results of steel pipe manufacturing plants.
However, since it is a numerical experiment, even if the operation amount of the manufacturing process is calculated as shown in FIGS. 1 and 2, the control accuracy cannot be compared by reflecting it in the actual operation. Therefore, the physical model calculation shown in FIG. 3 was performed using the parameters identified by applying the present invention and the actual results of the manipulated variable stored in the factory database, and the machining end temperature was predicted. This machining end temperature predicted value (output predicted value) was compared with the machining end temperature actual value (stored in the factory database), which is the result of operating the actual process using the actual operation amount. If this temperature prediction accuracy is improved, it is considered that the control accuracy when actually reflected in the operation is also improved. In the regression calculation in Example (2), the same manufacturing conditions and manufacturing record data as in the parameter identification calculation were used.

As a comparative example (1), as described in Non-Patent Document 1, the predicted machining end temperature value and the actual machining end temperature value using the parameters uniquely identified over the entire manufacturing condition space were compared. The result is shown in FIG. Further, as a comparative example (2), the machining end temperature predicted value using the parameters of the stratified table manually adjusted by the operator and the machining end temperature actual value were compared. The result is shown in FIG.
On the other hand, FIG. 7 shows the result of comparing the predicted machining end temperature value and the actual machining end temperature value by the numerical experiment of Example (1). Further, FIG. 8 shows the result of comparing the predicted machining end temperature value and the actual machining end temperature value by the numerical experiment of Example (2).
As shown in FIGS. 5 to 8, both the error average and the error σ are reduced in Example (1) as compared with Comparative Example (1) and Comparative Example (2), and further in Example (2). Both the error average and the error σ are reduced, and it can be confirmed that the temperature prediction accuracy is improved.

[Example B]
As Example B, the result of applying the method of limiting the value of the parameter among the methods described in the first embodiment to the problem of predicting the machining end temperature by the same piercer as in Example A is shown.
FIG. 9 shows a method that does not limit the value of the parameter, that is, the value of the parameter obtained in Example (1) of Example A. The horizontal axis shows the rolling number by the piercer, and the vertical axis shows the parameter value.
Here, as the parameters of the physical model, as described in Example A, the conversion coefficient from deformation to heat (also called the processing heat generation coefficient) α and the correction coefficient of the deformation resistance due to the heating furnace temperature shown in the equation (9) are used. Identify β. Although the description is omitted in Example A, the correction coefficient β is also used as the parameter to be identified in Example A.

In the method that does not limit the value of the parameter, the parameter is identified on the assumption that the cause of the error between the actual output value and the predicted output value is all the parameter value. Therefore, when the cause of the error is an unknown factor other than the parameter, the value of the parameter may change significantly for each material. The correction coefficient β does not physically take a negative value, but in the example of FIG. 9, it has become a negative value in many materials.

FIG. 10 shows the parameter values obtained by the method of limiting the parameter values. As in FIG. 9, the horizontal axis shows the rolling number by the piercer, and the vertical axis shows the parameter value.
FIG. 10A shows the result of using the first method described above, and sets the upper and lower limit values of the parameters (referred to as Example (3)). The upper and lower limits were set within a range assumed to be correct as a physical phenomenon, the conversion coefficient α was limited to 0.5 to 1.5, and the correction coefficient β was limited to 0 to 0.5. By imposing such a constraint condition, the correction coefficient β does not become a negative value, and the identification result is within a predetermined range.

FIG. 10B shows the result of using the second method described above, and uses an evaluation function including an error between the parameter value and the reference value thereof (referred to as Example (4)). The reference value of the conversion coefficient α and the reference value of the correction coefficient β are set to values that the operator considers to be approximately correct from past experience. FIG. 11 shows a reference value of the conversion coefficient α and a reference value of the correction coefficient β. In Example (4), the upper and lower limit values are not set for the conversion coefficient α, but the lower limit value is set to zero for the correction coefficient β. The evaluation function used the equation (3.1) and set the weighting coefficient K _j = 1. The effect differs depending on the magnitude of the weighting coefficient K _j , but the parameters are adjusted so that the physical model has high accuracy within a range that does not deviate significantly from the reference value.
Compared with the result of FIG. 9, the variation of the parameter was small. Regarding the conversion coefficient α and the correction coefficient β, in the case of FIG. 9, the RMSE with the reference value was 0.186 and 0.086, respectively, whereas in the case of FIG. 10B, the RMSE with the reference value was 0.186 and 0.086, respectively. The values were 0.181 and 0.030, respectively, and the parameter values were determined so that the RMSE with the reference value would be small. In particular, with the correction coefficient β, the difference from the reference value became significantly smaller, and excessive fluctuations in the parameters for matching with the actual data were suppressed.

As described above, FIG. 7 shows the result of comparing the predicted machining end temperature value and the actual machining end temperature value according to the first embodiment. The error in this case was 6.81 ° C.
On the other hand, FIG. 12A corresponds to FIG. 10A and shows the result of comparing the machining end temperature predicted value and the machining end temperature actual value according to the embodiment (3). The error in this case was 6.99 ° C., and the prediction accuracy of the physical model was the same as that of Example (1) in which the value of the parameter was not limited, and the prediction accuracy was good.
Further, FIG. 12 (b) corresponds to FIG. 10 (b), and shows the result of comparing the machining end temperature predicted value and the machining end temperature actual value according to the embodiment (4). The error in this case was 6.76 ° C., and the prediction accuracy of the physical model was the same as that of Example (1) in which the value of the parameter was not limited, and the prediction accuracy was good.
In this way, by setting the upper and lower limit values for the parameters of the physical model and evaluating the error from the reference value, excessive fluctuations in the parameters were suppressed, and highly reliable identification results were obtained. ..

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 the technical idea or its main features.
For example, the setup calculation unit 102, the extraction unit 103, the parameter identification calculation unit 104, and the regression calculation unit 106 are realized by a computer device including, for example, a CPU, a ROM, a RAM, and the like. In this case, each part may be configured by one hardware, or each part may be configured by a plurality of hardware.
Further, in the above embodiment, the extraction unit 103 functions as the extraction means according to the present invention, and the parameter identification calculation unit 104 functions as the calculation means according to the present invention to form the parameter identification device. In this case, the extraction unit 103 and the parameter identification calculation unit 104 are not limited to being composed of one hardware, but may be composed of a plurality of hardware.
The present invention can also be realized by supplying software (program) 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.

101: Manufacturing process 102: Setup calculation unit 103: Extraction unit 104: Parameter identification calculation unit 105: Database 106: Regression calculation unit

Claims (9)

A manufacturing process control system that controls the manufacturing process under the manufacturing conditions of the target product using a physical model that describes the behavior of the manufacturing process of the product.
The manufacturing conditions of products manufactured in the past and the output actual value of the manufacturing process from a database that stores manufacturing performance data linked, manufacturing performance data based on the distance of the manufacturing conditions in the space between the production conditions of the target product Extraction means to extract
An identification calculation means that performs identification calculation of a predetermined parameter of the physical model using the manufacturing record data extracted by the extraction means, and an identification calculation means.
A regression calculation means for obtaining an error prediction value by predicting an error of an output predicted value by the physical model reflecting the predetermined parameter identified and calculated by the identification calculation means by a regression model.
A manufacturing process control system including a correction means for correcting an output predicted value by the physical model by using the error predicted value obtained by the regression calculation means when controlling the manufacturing process. The identification calculation means is characterized in that the value of the predetermined parameter is obtained based on an objective function including an error between the output actual value in the manufacturing actual data extracted by the extraction means and the output predicted value by the physical model. The control system for the manufacturing process according to claim 1. The control system for a manufacturing process according to claim 2, wherein a constraint condition is imposed such that the value of the predetermined parameter is included in the predetermined range. The objective function is, in addition to the error between the output actual value in the manufacturing actual data extracted by the extraction means and the output predicted value by the physical model, the value of the predetermined parameter and the preset reference of the predetermined parameter. The control system for a manufacturing process according to claim 2 or 3, further comprising an error from the value. The predetermined parameter identified and calculated by the identification calculation means and the manufacturing conditions of the target product are input to the physical model to obtain the output predicted value of the target product, and the output predicted value is the manufacturing of the target product. A setup calculation means for determining the operation amount of the manufacturing process so as to substantially match the output target value under the conditions is provided.
The control system for a manufacturing process according to any one of claims 1 to 4, wherein the setup calculation means functions as the correction means and corrects an output predicted value of the target product. The setup calculation means is characterized in that the operation amount is determined by performing a convergence calculation in which the operation amount is modified and the calculation of the physical model is repeated until the output prediction value substantially matches the output target value. The control system for the manufacturing process according to claim 5. The regression calculation means
Using the physical model that reflects the predetermined parameters identified and calculated by the identification calculation means, the output predicted value for the manufacturing conditions in the manufacturing record data extracted by the extraction means is calculated.
An error between the output result value in the extracted manufacturing performance data by the extracting means and the output prediction value to generate the regression model by regressing the manufacturing conditions as explanatory variables,
The given conditions for producing the target product, the control system of the manufacturing process according to any one of claims 1 to 6, characterized in that asking you to the error prediction value for said products to the regression model. A method for controlling a manufacturing process that controls the manufacturing process under the manufacturing conditions of the target product by using a physical model that explains the behavior of the manufacturing process of the product.
Extraction means, from a database that stores manufacturing performance data and an output actual value correlated manufacturing conditions of products manufactured in the past and the manufacturing process, based on the distance of the manufacturing conditions in the space between the production conditions of the target product And the steps to extract manufacturing record data
A step in which the identification calculation means performs identification calculation of a predetermined parameter of the physical model using the extracted manufacturing record data.
A step in which the regression calculation means obtains an error predicted value by predicting an error of the output predicted value by the physical model reflecting the predetermined parameter identified and calculated by the regression model.
A method for controlling a manufacturing process, wherein the correction means includes a step of correcting an output predicted value by the physical model by using the error predicted value when controlling the manufacturing process . A program for controlling the manufacturing process under the manufacturing conditions of the target product by using a physical model that explains the behavior of the manufacturing process of the product.
The manufacturing conditions of products manufactured in the past and the output actual value of the manufacturing process from a database that stores manufacturing performance data linked, manufacturing performance data based on the distance of the manufacturing conditions in the space between the production conditions of the target product And the process of extracting
Using manufacturing performance data the extracted, the process for identifying calculation of certain parameters of the physical model,
A process of obtaining an error predicted value by predicting an error of an output predicted value by the physical model reflecting the predetermined parameter identified and calculated by a regression model.
A program for causing a computer to perform a process of correcting an output predicted value by the physical model using the error predicted value when controlling the manufacturing process.

Images