JPH10198402A - Method and device for identifying initial value of learning parameter - Google Patents

Abstract

PROBLEM TO BE SOLVED: To simply identify the initial value of a learning parameter for a physical model and to quickly and surely obtain a truely optimum initial value. SOLUTION: Experience data to be identified are entered from a data base (S3) and learning calculation for executing the calculation of a physical model, reading out the calculation result of the physical model and an experience data having the same manufacturing instruction as the calculation result from an indentifying experience data area, calculating an error or an error ratio between the calculation value and its corresponding experience value, and executing smoothing between the calculted error or error ratio and an accumulated error or error ratio is executed and stored in a learning parameter initial table prepared independently of a learning parameter table distinguished by material parameters and product parameters and allowed to be used for real operation and having the same structure as the learning parameter table to identify the initial value of a learning parameter for the physical model (S4).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for identifying initial values of learning parameters of a physical model.
The present invention is suitable for use in an identification method and an apparatus for automatically optimizing a learning parameter initial value of a physical model represented by a mathematical expression so as to approximate a product manufacturing process.

[0002]

2. Description of the Related Art In a manufacturing process in many industries including steel, a plurality of products are processed in various steps. For such a manufacturing process, when setting up equipment so that desired properties (for example, target values such as temperature, thickness, and hardness of steel materials) are obtained for a product, a physical model that approximates the processing process with a mathematical expression Is installed on a computer, the material parameters and product parameters of the target product are calculated as input conditions for this physical model, and the calculation results are set as equipment setting values in the actual processing process. Have gained.

However, there is no physical model in which the physical model completely matches the actual processing process, and the physical model is made compatible with a learning function described below.

[0004] This learning function installs a learning parameter table stratified by material parameters and product parameters correlated with a target physical model on a computer, and records actual data obtained by actually operating a processing process. The function of accumulating an error or error rate with the physical model calculation result as a learning parameter in the learning parameter table. By correcting the physical model calculation value with the learning parameter at the time of the next setting calculation, the defect of the physical model can be corrected. Is complementary.

However, in many industrial processes, equipment remodeling is performed and operating conditions are frequently changed. Therefore, the physical model itself has to be retuned each time to bring it closer to the actual processing process. In addition, since a tuned physical model cannot completely simulate an actual processing process, it is indispensable to prepare learning parameters according to a newly tuned physical model as a further complementing means.

For this reason, even if the physical model itself is tuned and applied to the actual operation, the learning parameters accumulated so far cannot be used, and the setting of the manufacturing equipment is calculated using the tuned physical model. At times, unless the learning parameter initial values suitable for the physical model are prepared, the target product characteristics cannot be stably obtained.

That is, it takes several months until all new learning parameters for all combinations of material parameters and product parameters to be manufactured are filled by the online learning calculation function, and the accuracy of the manufactured product during that time is as follows:
I had to be unstable.

Therefore, in order to correctly obtain an output result having desired characteristics after tuning the physical model, it is necessary to fill the learning parameter table with appropriate values before using it in the operation.

[0009] However, since this work takes an enormous amount of time, an off-line method has not been performed, and the only on-line method described below has been performed.

FIG. 4 is a diagram for explaining an example of a conventional online identification method of learning parameter initial values of a physical model. In the example of FIG.
In the learning calculation function B in the middle, in addition to the learning calculation function D using the existing physical model (physical model before tuning) and the existing learning parameter table F, which originally operate as functions, a new A learning calculation function E using a physical model and a new learning parameter table G are temporarily provided. Normal timing in actual operation, for example, in the case of a rolling process, timing at which actual data necessary for learning calculation is collected after rolling. Then, after performing the learning calculation using the existing physical model, storing the learning parameters in the existing learning parameter table, and subsequently performing the learning calculation using the tuned physical model, the learning parameters are stored in the new learning parameter table. To be stored.

In this way, the new learning parameter table is filled in accordance with the actual operation pitch, and the method is continued until the table is completely filled.

When all the operating conditions in the actual operation are completed, that is, when the new learning parameter table is completely filled, the learning parameters stored in the new learning parameter table are replaced with the learning parameters corresponding to the tuned physical model. As an initial value, it is transferred to the existing learning parameter table, and the learning calculation function B,
The model used for the setting calculation function H is also switched to the tuned physical model for use.

Thereafter, the learning calculation function E using the tuned physical model and the new learning parameter table G are removed from the learning function.

Until the new learning parameter table is completely filled and switched to the tuned physical model, the setting value calculation for the manufacturing equipment is started at a predetermined setting calculation timing by the equipment setting calculation function H being activated. ,
Calculate using the existing physical model before tuning, and set the target manufacturing equipment. The learning parameters used at this time use the learning parameters stored in the existing learning parameter table.

[0015]

However, in the case of the example of FIG. 4, a process for identifying a learning parameter initial value and a new learning parameter detable are provided in the learning calculation function of the original function. In addition to the processing time of the computer's CPU and the problem of extra consumption of shared resources such as memory, until the new learning parameter table is completely filled,
In other words, it is not possible to switch to a new physical model after tuning until production of all operating conditions (all steel types and all product sizes in the case of steel) has been completed. There is a problem that even if tuning to a physical model according to, it takes a lot of time to actually switch.

[0016]

According to the method of the present invention for identifying an initial value of a learning parameter of a physical model, an initial value of a learning parameter of a physical model for calculating a manufacturing parameter from a material parameter, a product parameter, and the like is calculated using a computer. In the identification method, material parameters, product parameters, and actual data for each target product are taken in for all target materials used in the identification process from a database that stores a plurality of operation data in the manufacturing process for each product. Using the material parameters, the product parameters, etc., execute the target physical model, calculate the error or error rate from the actual value corresponding to the calculated value, and calculate the error or error rate accumulated so far. Learning calculation to perform smoothing with The learning parameter table for use separately provided, by storing in the learning initial parameter table with the same structure as the table, and performs the identification of the learning parameter initial values of the physical model.

Another feature of the present invention is that the user can designate the operation data and the identification target of the product used for the identification process from the database.

The apparatus for identifying learning parameter initial values of a physical model of the present invention uses a computer to identify learning parameter initial values of a physical model for calculating manufacturing parameters from material parameters and product parameters. A database construction means for constructing a database by storing a plurality of operation data in the manufacturing process for each product, and specifying an operation data and an identification target of the product to be used for the identification process from the database and issuing an identification request. An identification requesting means to be performed, and a physical model calculation executing means for executing the calculation of the target physical model for all the products specified by the identification requesting means, and learning from the calculated value and the corresponding actual value. Learning calculation execution means to execute calculations, material parameters and product parameters The segmented, the learning parameter table used in actual operation is provided separately, comprising means for storing in the learning initial parameter table with the same structure as the table.

Another feature of the present invention is that the identification requesting means is further configured to be capable of designating operation data to be fetched as a product used in the identification processing and an identification target.

According to the present invention configured as described above, since the physical model itself and the database are provided in the device for performing identification, actual data used for the identification processing can be fetched by designating desired data. This makes it possible to automatically perform the learning without having to take a lot of time until the initial values of the learning parameters of the physical model for all possible operating conditions are obtained, as in the conventional online type shown in FIG. By specifying the past actual operation data to the minimum necessary, it becomes possible to identify in a short time.

[0021]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to the drawings.

[0022]

FIG. 1 is a flowchart showing a rough flow of a processing procedure of a method for identifying a learning parameter initial value of a physical model according to the present embodiment. Here, as an example of the processing process, a manufacturing process for manufacturing a product is considered, but the present invention can be applied to other processes.

In FIG. 1, in step S1, actual data of a product to be used for the identification process is designated from a predetermined database. The above-mentioned predetermined database stores a plurality of processing conditions and data of processing results corresponding to actual production of a product by actually processing materials in a processing process for each product. Each time a product is manufactured, it is accumulated sequentially.

In step S2, a physical model corresponding to a learning parameter initial value to be identified is specified.

When the user makes necessary designations in steps S1 and S2, the next step S3
After the result data in the specified range is retrieved from the database, in step S4, the result data of the first product is set from the retrieved result data.

Then, in step S5, the above step S
The calculation of the physical model is executed under the same processing conditions (included in the acquired result data) as when the product set in step 4 was actually processed in the processing process. In step S6, the following learning calculation is performed from the calculation result of the physical model in step S5 and the corresponding actual data, and the result is stored in the corresponding area of the learning parameter table stratified by the operating conditions at that time. Is done.

Β _{i + 1} = α · (A _actual / A _model ) + (1−α) · β _i β _i : Cumulative learning parameter value up to the i-th learning calculation target material of the stratified division α: Learning gain (Constant) A _model : Physical model calculation result under the relevant operating conditions A _actual : Actual value corresponding to A _model Then, in step S7, the above-described processing is completed for all the products specified in step S1. It is determined whether or not it has been done. At this stage, since only the first product has been processed, the process returns to step S4, and then returns to step S4.
Result data of the second product is set. And this 2
The processing of steps S4 to S7 is repeated for the second product.

In the same manner, step S1 is performed until the processing is completed for all the products specified in step S1.
4 to S7 are performed.

FIG. 2 is a diagram showing an example of the configuration of a system using the apparatus for identifying learning parameter initial values of a physical model according to the present invention.

In FIG. 2, reference numeral 1 denotes a process control computer for controlling the actual operation of a processing process (not shown). Here, as an example of the processing process, a manufacturing process (for example, a hot rolling process) for manufacturing a plurality of steel products in various steps is considered.

In this case, the process control computer 1
For example, according to the conditions such as the thickness, hardness, and temperature of the coil to be manufactured, setting calculations for giving an instruction such as mill gap length information, load information, or heating information to a rolling mill, or actually execute a manufacturing process. A process of taking in actual data (information indicating how such a result was obtained under the processing conditions) obtained by operation is performed.

Reference numeral 2 denotes a computer for analysis (eg, an engineering workstation: EWS, hereinafter referred to as an analysis EW).
S) and constitutes a main part of the identification device for learning parameter initial values of a physical model according to the present invention. In the case of the present embodiment, the physical model and the learning calculation function are set on the EWS 2 for analysis, and the identification process of the learning parameter initial value of the physical model is mainly performed by the EWS for analysis.
2 is performed.

The analysis EWS 2 inputs the result data taken in by the process control computer 1 every time one product is manufactured, and sequentially accumulates the data on the manufacturing result data storage disk 3. The database formed in this way is configured by storing a plurality of processing condition data and the corresponding processing result data when a product is manufactured by actually processing a material in a manufacturing process for each product. When identifying the learning parameter initial value of the physical model, the actual data in this database is used. Reference numeral 4 denotes an interface terminal with an operator, and instructions such as a learning parameter initial value identification request to be described later are issued through these interface terminals 4. As shown, the process control computer 1 and the analysis EWS
2 and the interface terminal 4 are LAN (Local
The network is connected by a network such as Area Network, and is configured to mutually transmit and receive data.

FIG. 3 is a diagram showing a data flow and a control flow in the system shown in FIG. In FIG. 3, solid arrows indicate data flows,
The dotted arrow indicates the control flow. The numbers in parentheses indicate the order in which the processing is performed. This embodiment will be described in detail with reference to FIG.

As shown in FIG. 3, the process control computer 1 includes a production result data transmitting means 11. Each time the production of one product is completed (for example, rolling is completed in the case of a hot rolling process), the production result data transmission unit 11 outputs the production instruction data (for example,
Product target thickness) and rolling performance data (for example, rolling temperature,
EWS for analysis of production performance data for each product, such as values calculated by rolling load and material composition) and rolling performance data
2 (1).

In the analysis EWS 2, the production performance data receiving means 12 receives the production performance data for each product transmitted from the process control computer 1 and supplies it to the production performance data storage means 13 (2). The production result data storage means 13 stores the given production result data for each steel type (such as ordinary steel, electromagnetic steel, and alloy steel) and in time series in the production result data storage disk (production result data bank) 3.
(3).

As a data area for storing the production result data, for example, an index cyclic file having a capacity of 200,000 products is used. That is, it is possible to retrieve data for each product by using the coil number, steel type, etc. as a key, and when the data area is full, it is a file having a structure capable of deleting and storing the oldest data of the corresponding steel type. Conventionally, the database that stores the performance data is in the process control computer,
In consideration of the fact that the operation speed of the process control computer must be high, a computer having a small storage capacity had to be used. However, in the case of the present embodiment, since the database is in the analysis EWS 2, the database is large. Capacitive ones can be used.

When identifying the initial values of the learning parameters of the physical model, first, the operator operates the interface terminal 4.
An identification request is made using (4). The information to be input at this time is, for example, the following.

Physical model (eg, deformation resistance model, strip temperature prediction model, etc.) corresponding to the initial value of the optimization learning parameter, index of production performance data used for identification (eg, from what day to what day) Time specification indicating whether the data is the actual data of a product manufactured in the above, coil number specification, etc.).

When the identification management means 14 receives this identification request, it issues an instruction to prepare the use result data to the target result data preparation means 15 (5). That is, the identification management means 1
4 transfers the index of the usage result data designated by the interface terminal 4 as described above to the target result data preparation means 15 and waits for the preparation completion report. Upon receiving the use result data index, the target result data preparation means 15 searches the manufacturing result data bank 3 for the corresponding data one product at a time based on the index, and stores it in the identification result data area 16 (6). When the storing process is completed, a preparation completion report is made to the identification management means 14 (7).

As described above, according to the present embodiment, data to be used for the identification process can be arbitrarily specified from all the production result data stored in the database. As a result, the model to be subjected to the identification processing can be specified instead of being fixed, so that the operability can be further improved.

Next, the identification management means 14 having received the preparation completion report issues a calculation instruction to the model calculation means 18 (8),
The calculation of the target physical model specified above is executed as follows.

That is, among the input information necessary for the model calculation, the manufacturing command items and the temperature information are read from the manufacturing result data storage area in the identification result data area 16 for each product (9), and the model calculation is performed. Execute
Then, the calculation result is stored in a calculation result storage area in the identification result data area 16 (10).

When the model calculation is completed, a model calculation completion report is sent to the model learning calculation means 19 (1).
1). The model learning calculation means 19 receiving this report reads the calculation result of the target physical model 18 and the actual value having the same manufacturing instruction item from the identification actual data area 16 (12), and performs the model learning calculation as follows. And store it in the corresponding area of the learning parameter initial value accumulation area stratified by operating conditions (13).

Β _{i + 1} = α · (A _actual / A _model ) + (1−α) · β _i β _i : Cumulative learning parameter value up to the i-th learning calculation target material of the stratified division α: Learning gain (Constant) A _model : physical model calculation result under the relevant operating conditions A _actual : _actual value corresponding to A _{model When the} model learning calculation is completed for one product in this way, the model learning calculation means 19 Then, a learning calculation completion report is made to (14). In response to this, the identification management means 14 checks whether or not the processing has been performed for all the products specified as described above, and if not completed, returns to the processing of (8) again.

The learning parameter initial value identification calculation of the model as described above is executed for all the products specified as described above.

As described above, the production result data collection is also performed by the production result data transmitting means 11 and the production result data receiving means 1.
2 and is automatically performed by the production result data accumulating means 13, and the acquisition of the result data and the like used for the identification process is performed as follows.
It is done automatically just by specifying the desired data,
The number of manual operations can be significantly reduced.

Further, as shown in FIG. 4, the identification of the learning parameter initial value depends on the production schedule of the actual operation.
Obtain reliable initial values of the optimal learning parameters without inconvenience, such as waiting for the operation of a product with a low frequency of operation, until all the learning parameters are obtained. Can be.

[0049]

As described above, according to the present invention, the result data of the product used for the identification process is designated from the database storing the result data, and the result data used for the identification process is fetched. The identification process of the learning parameter initial value including the above can be performed very easily. Therefore, the identification of the initial values of the physical model parameters can be easily performed by reducing the manual operation, and the identification of the initial values of the learning parameters depends on the production schedule of the actual operation, and it takes a long time to obtain all the learning parameters. It is possible to obtain a reliable optimal learning parameter initial value in a short time without inconvenience such as spending.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a rough flow of a processing procedure of a method for identifying a learning parameter initial value of a physical model according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration example of a system using an apparatus for identifying a learning parameter initial value of a physical model according to the present invention.

FIG. 3 is a block diagram showing a data flow and a control flow in the system shown in FIG. 2;

FIG. 4 is a block diagram showing an example of a computer function for executing a conventional method (online method) of identifying learning parameter initial values of a physical model.

[Explanation of symbols]

1: Computer for process control 2: EWS for analysis 3: Manufacturing result data storage disk (manufacturing result data bank) 4: Terminal for interface 11: Manufacturing result data transmitting means 12: Manufacturing result data receiving means 13: Manufacturing result data storing means 14: identification management means 15: target result data preparation means 16: identification result data area 18: target model calculation means 19: model learning calculation means 20: learning parameter initial value accumulation area

Claims (4)

[Claims] An identification method for optimizing a learning parameter initial value of a physical model for calculating a manufacturing parameter from a material parameter, a product parameter, and the like using a computer, wherein a plurality of operation data in a manufacturing process is stored for each product. For all target materials used in the identification process from the database made up of these, the material parameters, product parameters, and actual data are fetched for each product, and the material parameters,
Using the product parameters, etc., execute the target physical model, calculate the error or error rate with the actual value corresponding to the calculated value, and smooth the error or error rate accumulated so far. Is executed, and stored in a learning parameter initial table having the same structure as that of the learning parameter table provided separately from the learning parameter table used in the actual operation and divided by the material parameter and the product parameter. A method of identifying a learning parameter initial value, characterized in that the value of is set as an optimal value of the learning parameter initial value. 2. The identification of learning parameter initial values according to claim 1, wherein the operation data of the product used for the identification process and the identification target can be specified by the user from the database. Method. 3. An identification apparatus for optimizing a learning parameter initial value of a physical model for calculating a manufacturing parameter from a material parameter and a product parameter using a computer, wherein a plurality of operation data in the manufacturing process are stored for each product. Database construction means for constructing a database in this manner, identification request means for designating the operation data of the product to be used for the identification process and the identification target from the database, and an identification request, and the identification request means specified by the identification request means. Physical model calculation execution means for executing the calculation of the target physical model for all products, and the error or error rate between the actual value corresponding to the calculated value and the error, and the error accumulated so far. Or a learning calculation execution means for smoothing with the error rate, and material parameters or product parameters Min was, the learning parameter table used in actual operation is provided separately, characterized in that comprises means for storing in the learning initial parameter table with the same structure as the table, the identification device of the learning parameter initial values. 4. The learning parameter according to claim 3, wherein the learning parameter initial value identification device is further configured to be able to designate operation data to be taken in as a product used in an identification process and an identification target. Initial value identification device.