JPH09155420A - Method for learning setup model of rolling mill - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately learn setup models such as a load calculation model using the data of actual rolling. SOLUTION: This method is constituted so that, at the time of setting up the rolling mill to improve the accuracy of thickness in the tip part of a material to be rolled, the data of actual rolling about the coil or the pass of rolling this time is applied to the load calculation model which is used for the load calculation of the rolling mill and learning is executed and the learning is reflected in the setup of the rolling mill for the next rolling. In such a case, when there are the load calculation model 1 requiring convergent calculation and the load calculation model 2 not requiring convergent calculation, first, learning calculation is executed by the model 1. Only when the calculation is possible, learning calculation is executed by the model 2, the model 2 is corrected based on that result, load calculation is executed using the model 2 after correcting. When learning calculation by the model 1 is impossible, setup calculation is executed based on the result of the last learning calculation.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for learning a setup model of a rolling mill, and more particularly, in a cold or hot rolling mill, when learning and modifying the setup model by using actual data obtained during rolling. The present invention relates to a learning method of a rolling mill setup model, which is suitable for application.

[0002]

2. Description of the Related Art One of the most important problems in a rolling mill is to keep the longitudinal thickness accuracy of a product obtained on the delivery side of the rolling mill within a predetermined range over the entire length.
Therefore, whether cold or hot, the batch type rolling mill is used for rolling, and the continuous rolling mill is used for rolling a running plate thickness change (Flying Gage Change; FGC) point like a welding point.
Setup is performed by setting and controlling the roll gap and roll speed of the rolling rolls to appropriate values so that a desired strip thickness can be obtained from the coil TOP portion (the tip end of the material to be rolled).

This setup is based on the thickness schedule,
This is performed by predicting the rolling state by calculating a mathematical model (setup model) using the tension schedule, rolling speed, etc. as input conditions. Of the mathematical models, the core of the mathematical model is the core of the coil TOP part. It is the rolling load calculation model that affects the strip thickness accuracy. Therefore, in the following, this model will be mainly described.

An approximate expression derived from the rolling theory is used for this rolling load calculation model. Such an approximate expression is described in detail in "Theory and practice of sheet rolling" (edited by Japan Iron and Steel Institute), pp. 33 to 37, in which several approximate expressions for cold rolling and hot rolling are illustrated. Has been done. These approximate equations can be roughly classified into those that include integral calculation in the model equation and those that do not.

However, when calculating the rolling load,
Since the flatness of the rolling roll must be taken into consideration, it is necessary to calculate the load equation and the roll flattening equation simultaneously, as shown in “Theory and practice of sheet rolling” on pages 39 to 41. At that time, as described above, the rolling load calculation model including the integral calculation in the model, when the rolling load is calculated by combining with the flat roll formula, it must be obtained by the convergence calculation, whereas the integral calculation is performed. In the case of a model that does not include it, calculation is possible only by expanding the model formula.

On the other hand, in the setup load calculation, in order to prevent the rolling mill from stopping due to a setup error, the calculation result must be output in advance.
This is because when the rolling mill is stopped, the work efficiency is reduced, which causes a significant increase in cost.

Thus, in the setup load calculation,
Since it is necessary to output the calculation result without fail, a model that does not require convergence calculation is generally used as a load calculation model because it may not converge depending on the input conditions in the model including convergence calculation as described above. Is. As an example of such a case where the convergence calculation is not required, there is a technique disclosed in Japanese Patent Publication No. 6-61566.

By the way, since the load calculation model includes many error factors such as mathematical approximation error, material of rolled material, and seasonal fluctuation of environment, it is intended to improve the accuracy of the model from the next time onward. A learning calculation is performed by grasping these errors and appropriately correcting the coefficients included in the load calculation model using the errors.

[0009] For example, the above-mentioned "Theory and Practice of Sheet Rolling" No. 2
As described on pages 91 to 292, in the learning calculation, the actual rolling data set in the rolling mill, such as the plate thickness and tension of the coil or the pass being rolled, is used as the input for the setup calculation. Calculate the load using the same model as the above, compare the calculated load obtained at that time with the actual rolling load, correct the coefficient of the load calculation model based on the error at that time, and use that coefficient for the next coil. Alternatively, it is reflected in the setup calculation of the next pass.

[0010]

When the above-described method of learning a setup model is applied to a rolling load model that does not require convergence calculation, there are the following problems.

As described above, when learning the load calculation model, the actual rolling data is used as the input condition for the load calculation.
For example, taking the actual plate thickness of a tandem rolling mill as an example, the plate thickness on the inlet side and the outlet side of each rolling stand is either measured by a thickness gauge or predicted by mass flow calculation or gauge meter calculation. Is.

However, the strip thickness actually measured by the strip thickness gauge may include a measurement error of the strip thickness gauge, an abnormal value due to an error, or an influence due to the distance between the strip thickness gauge installation point and the rolling stand. There is a variation in accuracy and error in the actual rolling data inevitably because the plate thickness obtained by mass flow calculation or gauge meter calculation includes a prediction error. Therefore, the learning coefficient calculated by the learning calculation may largely vary between the coils due to the variation in these errors. This is because the load calculation model uses a model that does not include the convergence calculation, and therefore, the learning calculation is performed and the learning coefficient is output even if any actual rolling data is input.

Therefore, in order to suppress this variation and improve the accuracy of the learning calculation, a guideline for evaluating the validity of the actual rolling data is set and whether or not the data should be used for the learning calculation based on the evaluation guideline. It is necessary to provide a means for determining whether or not, and executing the learning calculation only when it is determined that the data should be used for the learning calculation.

In the learning method described in Japanese Patent Publication No. 6-61566, the learning coefficient Z _PA is obtained by calculating the actual rolling load P _A and the load calculation model, as summarized in the flowchart of FIG. The ratio of the load P _CA is calculated (step S23), and when the average [Z _PA ] of the load ratio and the standard deviation σ are within a certain fixed range (ε1, ε2) as a filter process, It is supposed to learn (step S24).

However, in the above learning method, there is no theoretical basis for determining the reference range for judging whether or not to carry out the learning, and when this method is actually adopted, the learning accuracy is high. I didn't go up at all.

The present invention has been made to solve the above-mentioned conventional problems. A rolling mill setup model capable of accurately learning a setup model such as a rolling load calculation model using actual rolling data. The challenge is to provide learning methods.

[0017]

According to the present invention, when setting up a rolling mill to improve the plate thickness accuracy at the tip of a material to be rolled, the actual rolling data on the coil or the pass of the current rolling is used. In a learning method of a rolling mill setup model, which is applied to a setup model used for rolling mill setup calculation, learning of the model is performed, and the learning result is reflected in the rolling mill setup for rolling the next coil or the next pass. , If there is a first setup model that requires a convergence calculation and a second setup model that does not require a convergence calculation as the setup models, first, a learning calculation is performed using the first setup model. Only if the calculation is possible and the calculation by the first setup model is possible Learning calculation is performed by the second setup model, the second setup model is corrected based on the result of the learning calculation this time, and setup calculation is performed using the corrected second setup model to obtain the first setup. When the learning calculation by the model is impossible, the above problem is solved by performing the setup calculation based on the result of the previous learning calculation.

[0018]

BEST MODE FOR CARRYING OUT THE INVENTION The present inventors use actual rolling data as input and apply load to both a first load calculation model that requires convergence calculation and a second load calculation model that does not require convergence calculation. Was calculated and compared with the actual rolling load measured, and it was found that the load prediction accuracy by the second model was very high and the variation was very high only in the case of the actual rolling data of the coil where the calculation converged by the first model. Can be suppressed,
On the contrary, in the case of the actual rolling data of the coil whose calculation did not converge in the first model, it was found that the variation in the load prediction accuracy by the second model becomes extremely large.

From these facts, in the learning calculation,
First, the load is calculated by the first model that requires the convergence calculation, and only when the calculation converges, the learning of the second model that does not require the convergence calculation is executed, and the result is used as the load calculation in the setup calculation. We have found that if this is reflected, learning can be performed with high accuracy and rolling accuracy can be improved.

The present invention was made based on the above findings. FIG. 1 is a flowchart showing the flow of a learning method of a load calculation model according to the embodiment of the present invention.

In the present embodiment, actual rolling data such as strip thickness, tension, roll speed, etc. are collected (step S1), and the load calculation is executed by the load calculation model 1 having a convergence calculation using these data as input (step S1). S2). Next, in step S3, it is determined whether or not the convergence calculation of the load calculation model 1 converges.
It is determined that the actual rolling data is not suitable for learning, learning calculation is not performed in the coil, and the previous (previous coil or previous pass) calculated value is directly output as the learning coefficient Z _PA (steps S4 and S7). ).

On the contrary, when the convergence calculation of the load calculation model 1 converges in the above step S3, it is judged that the actual rolling data is suitable for learning, and the load calculation model 2 having no convergence calculation is input with the data as input. The load calculation is executed (step S5) to obtain the load calculation result P _CA , and the learning coefficient Z _PA is calculated by comparing this with the actual rolling load P _A (step S6).

When the learning calculation is completed according to the above steps S1 to S6, the setup calculation is performed in the procedure of steps S8 to S12. In this setup calculation, the load is calculated by the same load calculation model 2 as that calculated in step S5 of the learning calculation by using the setting data such as the plate thickness / tension schedule and the mill speed which are scheduled to be set next. (Steps S8 and S9). Next, the calculated load P _C is corrected by the learning coefficient Z _PA obtained through the learning calculation input through step S7 (step S10), and the rolling position is corrected based on the corrected calculated load P _C ′. Is calculated (step S11) and is output to the controller to set up the next rolling (step S11).
12).

Next, the results when the method of the present invention and the conventional method are applied to the tandem rolling mill consisting of the first to fourth four stands (the second stand is not shown) shown in FIG. 2 are compared. And explain.

In the above tandem rolling mill, each stand is
The work rolls 1A to 1D and the backup rolls 2A to 2D that reinforce the work rolls are configured by a four-high rolling mill, and the work rolls can be adjusted by the reduction devices 3A to 3D. And in each of the 1st-4th rolling stands, the load at the time of rolling is 4A-4 load cells.
Detected by D, the plate thickness gauge 5 and the tensiometer 6 are installed on the first stand entrance side, and the board thickness gauges 5A to 5D and tensiometers 6A to 6D are installed on the stand exit sides, respectively. The plate thickness and the tension are detected on the side and the output side, respectively.

In this embodiment, in the tandem rolling mill, the load cells 4A to 4D and the plate thickness gauges 5 and 5A are used.
~ 5D, tensiometer 6, 6A ~ 6D, etc. actual rolling data is collected, the setup calculation device (not shown) learns the load calculation model according to the procedure of the flow chart shown in FIG. 1 and the learning result. A setup calculation based on is performed.

Here, as an example of the load calculation model 1 which requires the convergence calculation, the Bland & Ford equation (see "Theory and practice of strip rolling" on page 33) and the load calculation which does not require the convergence calculation are described. As an example of the model 2, Hill's formula (see "Theory and practice of strip rolling" on page 35) was used.

3 is set up by applying the load calculation model learned according to the flow chart of FIG. 1 to the tandem rolling mill, and the plate thickness is 4.0 to 3.0 mm at the mill entrance side, and is set out. The low-carbon steel plate is set to 300 by the plate thickness schedule such that the side becomes 2.0 to 0.5 mm.
It shows the change of the learning coefficient and the average of the off-gauge length which is ± 5% deviation of the coil TOP portion (the portion where the plate thickness deviation from the target value is ± 5% or more) when the coil is rolled. Further, FIG. 4 shows the results of rolling under the same conditions except that the load calculation model was learned by the conventional method shown in FIG. 5 for comparison with the method of the present invention. .

FIGS. 3 and 4 show changes in the off-gauge length and the learning coefficient of the coil TOP part for a certain continuous 5 coils selected from the 300 coils actually rolled. In both cases, the coil after the third coil is selected when the required rolling load becomes low for some reason.

In the learning method of the prior art performed here, when the learning coefficient exceeds a certain value (1.2 in this example), it is considered as abnormal and learning is not performed. Learning is not reflected in the next fourth coil, and as a result, the off-gauge length is long also in the fourth and fifth coils as shown in FIG.

On the other hand, in the method according to the embodiment of the present invention, it is possible to predict the tendency that the required rolling load becomes low by learning with the third coil. Therefore, as shown in FIG. Appropriate setup calculation is done for the coil, and the off gauge length is shortened.

As described above, according to this embodiment, the off-gauge length can be shortened as compared with the conventional method.

As described above, by applying the embodiment of the present invention, the load calculation model can be learned only by the actual rolling data suitable for the learning, so that the accuracy of the load calculation model is improved. It is possible to significantly improve the product yield (the ratio of the plate length with the plate thickness deviation within the predetermined range with respect to the total length of the coil).

In the above description, the method of the present invention is described as
The case where the multi-stage rolling mill is applied to a tandem rolling mill in which five stands are arranged has been described, but the number of stands is not limited to this, and any number of stands such as a single stand (reversible type), four stands, six stands, and seven stands can be used. It may be applied to rolling mills, and the number of rolling stands is not limited to 4 and it can be 2
The number of steps, 6 steps, 12 steps, 20 steps, etc. may be used, and further, it may be applied to a cold rolling mill or a hot rolling mill.

A more specific example [embodiment] of the embodiment will be described in detail below.

[First Embodiment] In a cold rolling mill in which a four-high rolling mill similar to the tandem rolling mill shown in FIG. 2 is arranged in four stands, a load calculation model is calculated according to the flow chart shown in FIG. The low carbon steel sheet was rolled by performing the setup for the next rolling by the calculation of the model which was learned and corrected. In this example, the Bland & Ford equation is used as the load calculation model 1 that requires convergence calculation, and the H calculation is performed as the load calculation model 2 that does not require it.
The ill's formula was used.

Here, the plate thickness schedule of the material to be rolled is that the thickness of the mother plate on the mill entry side is 4.0 to 2.8 mm and that on the mill exit side is 2.0 mm to 0.5 mm. According to the schedule, a sheet having a plate width of 700 to 1800 mm was rolled. At that time, in the same manner as described above, the plate thickness of each stand entrance / exit side is measured with the plate thickness gauges 5 and 5A to 5D, and the tension is measured with the tension gauges 6 and 6.
Load cell 4A to 4D for each stand with A to 6D
Then, the load calculation model was learned and the next setup calculation was performed using these actual rolling data.

Here, while applying the method of the present invention,
For comparison, the conventional learning method performed according to the procedure shown in FIG. 5 is applied, and when each 100 coils are rolled, the on-gauge ratio with respect to the total coil length (the length within which the plate thickness deviation is within a predetermined allowable range) The ratio of the average on-gauge ratio of the conventional method is 83%.
However, the average on-gauge ratio of the method of the present invention was 9
It improved to 2%.

[Second Embodiment] In this embodiment, in a hot rolling mill (not shown) in which four-high rolling mills are arranged in seven stands, the load similarly learned and corrected according to the flow chart of FIG. Setup was performed using the calculation model, and austenitic stainless steel sheet (SUS30
4) was rolled. In this example, the formula of Tamano-Yanagimoto (see "Theory and practice of strip rolling" on page 37) is used as the first load calculation model, and the multiple regression formula of Yanamoto-Tamano is used as the second load calculation model (" The theory and practice of sheet rolling "(see page 38) was used.

The plate thickness schedule of the material to be rolled is that the base plate thickness is 5.0 to 3.0 mm and the mill exit side plate thickness is 2.5 mm to 1.0 mm. Rolls of 1000 to 1300 mm were rolled. At that time, the plate thickness on the inlet side and the plate side on the outlet side of the hot tandem mill were measured with corresponding plate thickness meters, the tension between each stand was measured with a tensiometer, and the load of each stand was measured with a load cell. And about the 1st-6th stand delivery side board | plate thickness, it estimated using the gauge meter type | formula and used the estimated value for learning.

Again, applying the method of the present invention,
For comparison, the conventional learning method performed in the procedure shown in FIG. 5 was applied to obtain the on-gauge ratio for the total coil length when 100 coils were rolled, and the respective averages were compared. Was 80%, the on-gauge ratio of the method of the present invention was improved to 91%.

As described above in detail, according to the present invention, the learning accuracy of the load calculation model is improved, the plate thickness accuracy of the coil TOP portion is improved, and the reduction of the yield can be remarkably improved.

In the above concrete example, the number of stands, the number of steps, the plate thickness schedule, the plate width, the material, etc. are different.
Although the examples are shown, the items are not limited to these.

The present invention has been specifically described above, but the present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the invention.

For example, in the above embodiment, the load calculation model is shown as the setup model, but the present invention is not limited to this, and a crown / shape calculation model or a temperature calculation model may be used.

[0046]

As described above, according to the present invention,
A setup model such as a load calculation model can be accurately learned using the actual rolling data. Therefore,
The accuracy of the setup model is improved, the control accuracy of the plate thickness and the like in the coil TOP is improved, and the product yield can be significantly improved.

[Brief description of the drawings]

FIG. 1 is a flowchart showing an example of learning calculation and a procedure of setup calculation according to the present invention.

FIG. 2 is an explanatory diagram showing an outline of a tandem rolling mill to which an example of learning calculation according to the present invention is applied.

FIG. 3 is an explanatory diagram showing a result of applying the learning method of the present invention.

FIG. 4 is an explanatory diagram showing a result of applying a conventional learning method.

FIG. 5 is a flowchart showing a procedure of a conventional learning method.

[Explanation of symbols]

 1A to 1D ... Work roll 2A to 2D ... Backup roll 3A to 3D ... Rolling down device 4A to 4D ... Load cell 5, 5A to 5D ... Plate thickness gauge 6, 6A to 6D ... Tensiometer

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Yasushi Sunamori Inventor, 1 Kawasaki-cho, Chuo-ku, Chiba-shi, Chiba Inside the Chiba Works, Kawasaki Steel Co., Ltd. (72) Moriyuki Miyahara 1 Kawasaki-cho, Chuo-ku, Chiba-shi, Chiba Address Inside Kawasaki Steel Co., Ltd. Chiba Works

Claims (2)

[Claims] Claim: What is claimed is: 1. When setting up a rolling mill to improve the plate thickness accuracy at the tip of the material to be rolled, the actual rolling data for the coil or the pass of the current rolling is
In the learning method of the rolling mill setup model, which is applied to the setup model used for the setup calculation of the rolling mill to learn the model, and the learning result is reflected in the setup of the rolling mill rolling the next coil or the next path, When there are a first setup model that requires a convergence calculation and a second setup model that does not require a convergence calculation as the setup models, first, a learning calculation is performed by the first setup model, and this calculation is performed. It is determined whether or not is possible, and the learning calculation is performed by the second setup model only when the calculation by the first setup model is possible, and the second calculation is performed based on the result of this learning calculation. Modify the setup model, perform the setup calculation using the modified second setup model, A learning method of a rolling mill setup model, characterized in that, when the learning calculation by the first setup model is impossible, the setup calculation is performed based on the result of the previous learning calculation. 2. The method for learning a rolling mill setup model according to claim 1, wherein the setup model is a rolling load calculation model.