JP2022170100A - Rolling load prediction method, rolling control method and rolling load prediction device - Google Patents

Abstract

To provide a rolling load prediction method and rolling load prediction device which can accurately predict a rolling load in a short time.SOLUTION: A rolling load prediction method according to the present invention comprises the steps of: calculating the similarity between object rolled material data and past data by using a prescribed operation factor as a condition item and comparing values for each condition item with each other; performing weighting by the similarity of the corresponding past data to the actual value of rolling load of each past data and calculating a learning coefficient applied to the rolling load by each condition item when rolling the object rolled material by using a weighted value; and acquiring the actual value of the prescribed operation factor after the previous pass rolling for the object rolled material, replacing the set value of the prescribed operation factor included in the object rolled material data with the acquired actual value and predicting the rolling load of the next pass on the basis of the calculated learning coefficient. The step of calculating the learning coefficient uses only the past data within a range of a prescribed Euclidean distance from a point in a data space corresponding to the object rolled material data as a processing target.SELECTED DRAWING: Figure 2

Description

The present invention relates to a rolling load prediction method, a rolling control method, and a rolling load prediction device.

In the process of manufacturing steel products such as thick steel plates, after heating a cast slab to a predetermined temperature in a heating furnace, the slab is rolled to the product thickness by rolling the slab multiple passes in a reversible rolling mill. . At this time, the rolling schedule, such as the amount of reduction in each pass, is automatically calculated and set by a computer. It is necessary to maximize the reduction amount.

Here, the reduction amount of each pass is restricted by the maximum normal load and maximum torque of the rolling mill. However, the reduction amount is restricted in the pass schedule. Therefore, if the prediction accuracy of the rolling load is poor, the constraint load within the pass schedule must be made smaller in consideration of the variation in prediction. Therefore, if the prediction accuracy of the rolling load (=actual rolling load/predicted rolling load) is improved, the constraint load can be increased to maximize the rolling efficiency.

Against this background, a method for reducing the prediction error of the rolling load by determining the learning value of the rolling load for this pass from the prediction accuracy of the rolling load for the previous pass and multiplying the predicted rolling load for this pass by the learning value. is proposed. However, in general, the steel plate temperature and the reduction ratio are different for each pass, and when the steel plate temperature is different, the deformation resistance, which is one of the explanatory variables of the rolling load, has an inflection point and becomes nonlinear with respect to the temperature near the transformation point. . Therefore, when the steel plate temperature straddles the transformation point between passes, simply reflecting the prediction accuracy of the rolling load of the previous pass in the current pass does not provide sufficient prediction accuracy of the rolling load. Therefore, in order to solve such problems, a method described in Japanese Patent Laid-Open Publication No. 2002-100001 has recently been proposed, which utilizes big data accumulated in a database.

JP 2015-66569 A

Theory and Practice of Plate Rolling / Edited by The Iron and Steel Institute of Japan, 1984

However, according to the method described in Patent Document 1, it is possible to explain the nonlinearity of the deformation resistance with respect to the steel plate temperature to some extent. is calculated and applied to learning, the calculation time in the database becomes enormous. Therefore, according to the method described in Patent Document 1, it is difficult to accurately predict the rolling load in a short period of time.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a rolling load prediction method and a rolling load prediction apparatus capable of accurately predicting a rolling load in a short period of time. Another object of the present invention is to provide a rolling control method capable of improving the rolling efficiency of a rolled material.

The rolling load prediction method according to the present invention is a rolling load prediction method for predicting the rolling load of each pass of a reversible rolling mill that rolls a target rolled material in multiple passes, and is a rolled material that is rolled in a past operation. a step of obtaining, as operation data, a plurality of past data for each operation including the set value or actual value of a predetermined operational factor and the actual value of the rolling load, and the value of the predetermined operational factor related to the target rolled material. a step of obtaining as rolled material data; a step of calculating the degree of similarity between the target rolled material data and the past data by using the predetermined operating factor as a condition item and comparing values for each of the condition items; The actual value of the rolling load of each past data is weighted by the similarity of the relevant past data, and the weighted value is used to apply the learning coefficient that each condition item exerts on the rolling load when rolling the target rolled material. obtaining the actual value of the predetermined operation factor after the previous pass rolling for the target rolled material, replacing the set value of the predetermined operation factor included in the target rolled material data with the obtained actual value, and predicting the rolling load of the next pass based on the calculated learning coefficient, wherein the step of calculating the learning coefficient is a range of a predetermined Euclidean distance from a point in the data space corresponding to the target rolled material data. It is characterized in that only the past data in the data is processed.

In the rolling load prediction method according to the present invention, in the above invention, the step of calculating the learning coefficient includes selecting a predetermined number of past data in ascending order of distance from a point in the data space corresponding to the target rolling material data, It is characterized in that only selected past data is processed.

In the rolling force prediction method according to the present invention, in the above invention, the past data is divided into a plurality of databases according to the value range, and the step of calculating the learning coefficient includes a database corresponding to the value range of the target rolled material data Only the past data included in is to be processed.

In the rolling force prediction method according to the present invention, in the above invention, the step of calculating the learning coefficient includes expanding the predetermined Euclidean distance if the number of past data to be processed is less than a predetermined value, or It is characterized in that the past data included in the database adjacent to the database corresponding to the value range of the rolling material data is included in the processing target.

The rolling load prediction method according to the present invention is the above invention, wherein the step of calculating the learning coefficient includes a step of changing the number of past data to be processed in consideration of the degree of contribution to the calculation of the learning coefficient. Characterized by

A rolling control method according to the present invention includes a step of rolling a target strip to a target thickness by controlling a reversible rolling mill based on the rolling load predicted by the rolling load prediction method according to the present invention. characterized by comprising

A rolling load prediction device according to the present invention is a rolling load prediction device that predicts the rolling load of each pass of a reversible rolling mill that rolls a target rolling material in a plurality of passes, and is a rolling target rolled in past operations. Means for acquiring a plurality of past data for each operation including the set value or actual value of the predetermined operation factor and the actual value of the rolling load as operation data, and the value of the predetermined operation factor regarding the target rolled material means for obtaining rolling material data; means for calculating the degree of similarity between the target rolling material data and the past data by comparing the values for each of the condition items, with the predetermined operating factors as condition items; The actual value of the rolling load of each past data is weighted by the similarity of the relevant past data, and the weighted value is used to apply the learning coefficient that each condition item exerts on the rolling load when rolling the target rolled material. and acquiring actual values of predetermined operating factors after the previous pass rolling for the target rolled material, replacing the set values of the predetermined operating factors included in the target rolling material data with the acquired actual values, and means for predicting the rolling load of the next pass based on the calculated learning coefficient, wherein the means for calculating the learning coefficient is within a predetermined Euclidean distance range from a point in the data space corresponding to the target rolled material data. It is characterized in that only the past data in the data is processed.

According to the rolling force prediction method and the rolling force prediction device according to the present invention, the learning value of the rolling force can be calculated in a short time and the rolling force can be predicted with high accuracy. Further, according to the rolling control method of the present invention, it is possible to improve the rolling efficiency of the rolled material.

FIG. 1 is a schematic diagram showing the configuration of a rolling control device that is an embodiment of the present invention. FIG. 2 is a flow chart showing the flow of rolling force prediction processing, which is an embodiment of the present invention. FIG. 3 is a diagram for explaining an example of the processing of step S2 shown in FIG. FIG. 4 is a diagram for explaining another example of the process of step S2 shown in FIG. FIG. 5 is a diagram for explaining another example of the process of step S2 shown in FIG. FIG. 6 is a diagram for explaining another example of the process of step S2 shown in FIG. FIG. 7 is a diagram for explaining another example of the process of step S2 shown in FIG. FIG. 8 is a diagram for explaining another example of the process of step S2 shown in FIG.

DESCRIPTION OF THE PREFERRED EMBODIMENTS A rolling control apparatus, which is an embodiment of the present invention, will be described below with reference to the drawings.

〔Constitution〕
First, referring to FIG. 1, the configuration of a rolling control device according to an embodiment of the present invention will be described. FIG. 1 is a schematic diagram showing the configuration of a rolling control device that is an embodiment of the present invention. As shown in FIG. 1, a rolling control apparatus 1 according to an embodiment of the present invention includes a four-high reversible rolling mill 10 having a pair of upper and lower work rolls 11a and 11b and a pair of upper and lower support rolls 12a and 12b. It is a device for controlling the rolling process for rolling the thickness of the rolled material S to a target thickness by controlling the operation data management device 2, the rolling process computer (rolling process computer) 3, the load detection device 4, and the thickness control device. (Automatic Gauge Control Programmable Logic Controller: AGC PLC) 5 and a reduction device 6 are provided.

The operation data management device 2 is configured by an information processing device such as a computer, and stores a plurality of past data for each operation regarding the rolled material S rolled in past operations as operation data. In this embodiment, the operation data management device 2 stores, as past data, a set of data of the value of the actual rolling load/predicted rolling load and the actual value or set value of the explanatory variable (predetermined operational factor) of the rolling load. is doing. Examples of explanatory variables of the rolling load include the rolling reduction, the average temperature in the thickness direction of the rolled material S, the carbon equivalent, the thickness before and after rolling, and the like. Note that the predicted rolling load can be calculated by a well-known method using, for example, formula (1) shown below. In Equation (1), P is the predicted rolling load, ld is the contact arc length, Qp is the rolling force function, km is the deformation resistance of the rolled material S, and W is the strip width of the rolled material S. See Non-Patent Document 1, for example, for a method of calculating each factor on the right side.

The rolling process computer 3 uses the operation data stored in the operation data management device 2 to determine the pass schedule of the rolling process to be controlled, such as the thickness change amount of each pass, the advance rate, the predicted rolling load, the initial reduction position, etc. Calculate

The load detection device 4 detects the rolling load of the pair of upper and lower work rolls 11a and 11b, and outputs an electric signal indicating the detected rolling load to the plate thickness control device 5. FIG.

The strip thickness control device 5 controls the reduction device 6 based on the rolling load detected by the pass schedule calculated by the rolling process computer 3 and the load detection device 4 .

The reduction device 6 controls the reduction amount of the rolled material S in each pass by controlling the gap (reduction position) between the pair of upper and lower work rolls 11a and 11b according to the control signal from the plate thickness control device 5. FIG.

In the rolling control apparatus 1 having such a configuration, the rolling process computer 3 executes the following rolling load prediction processing, thereby calculating the learned value of the rolling load in a short time and predicting the rolling load with high accuracy. making it possible. Hereinafter, with reference to FIG. 2, the flow of the rolling force prediction process, which is one embodiment of the present invention, will be described.

[Rolling load prediction processing]
FIG. 2 is a flow chart showing the flow of rolling force prediction processing, which is an embodiment of the present invention. The flowchart shown in FIG. 2 starts at the timing when the value of the explanatory variable of the rolling load (target rolling material data) for the rolled material S to be processed is input to the rolling process computer 3, and the rolling load prediction process is the process of step S1. proceed to

In the process of step S1, the rolling process computer 3 calculates the Euclidean distance from the point (data request point) in the operation data space corresponding to the target rolled material data for each past data stored in the operation data management device 2. calculate. Specifically, when the data request point and the past data point to be processed in the operation data space are points P and Q, respectively, and the number of explanatory variables of the rolling load is i (= 1 to n), the rolling process The computer 3 calculates the Euclidean distance d from the data request point P for the past data Q to be processed using the following formula (2). Here, in Expression (2), Q _i indicates the value of explanatory variable i in past data Q to be processed, and P _i indicates the value of explanatory variable i at data request point P. Thereby, the process of step S1 is completed, and the rolling force prediction process proceeds to the process of step S2.

In the process of step S2, the rolling process computer 3 gives an index to the past data whose Euclidean distance from the data request point is within a predetermined range of Euclidean distance based on the result of the process of step S1. For example, when the operation data space is a two-dimensional space consisting of explanatory variables 1 and 2 as shown in FIG. An index is given to the past data B' included in a circular area R1 having a radius of an arbitrary Euclidean distance from the center. Thereby, the processing of step S2 is completed, and the rolling force prediction processing proceeds to the processing of step S3.

In the process of step S3, the rolling process computer 3 sets the explanatory variable of the rolling load as a condition item, and compares the values of the condition items to obtain the respective past data and the target rolled material data indexed in the process of step S2. Calculate the similarity with Specifically, the rolling process computer 3 calculates the degree of similarity w(d|h) with the target rolled material data using the following formula (3) for each past data indexed in the process of step S2: Calculate Here, in Expression (3), h is a parameter for adjusting the spread of the degree of similarity w(d|h), and the smaller the value, the similarity w( d|h) can be calculated. Also, d indicates the Euclidean distance of each past data calculated by the formula (2). Thereby, the processing of step S3 is completed, and the rolling force prediction processing proceeds to the processing of step S4.

In the process of step S4, the rolling process computer 3 uses the similarity w(d|h) calculated by the process of step S3 for the k pieces of past data indexed in the process of step S2 as follows: Regression coefficients β _i (i=1 to n) of i explanatory variables x _i composing the past data are calculated by the following formula (4). Thereby, the processing of step S4 is completed, and the rolling force prediction processing proceeds to the processing of step S5.

In the process of step S5, the rolling process computer 3 calculates the i explanatory variables x i of the target rolled material data and the regression coefficients β _i of the _{i explanatory variables x i calculated} in the process of step S4 as follows: By substituting in (5), the learning coefficient of the rolling load is calculated. Then, the rolling load process computer 3 calculates a value obtained by multiplying the predicted rolling load calculated using the above formula (1) by the learning coefficient y' as the predicted rolling load after learning correction. As a result, the processing of step S5 is completed, and the series of rolling load prediction processing ends.

As is clear from the above description, in the rolling load prediction process, which is one embodiment of the present invention, the rolling process computer 3 detects a point within a predetermined Euclidean distance from a point in the operation data space corresponding to the target rolled material data. Since the learning coefficient of the rolling load is calculated using only the past data in , the rolling load can be accurately predicted in a short time. Moreover, as a result, the rolling efficiency of the rolled material can be improved. In the rolling load prediction process, the number of past data is limited at the stage of calculating the degree of similarity, but the number of past data may be limited at the stage of calculating the learning coefficient of the rolling load. That is, by limiting the number of past data before calculating the learning coefficient of the rolling load, the rolling load can be accurately predicted in a short time.

The rolling process computer 3 selects a predetermined number of past data in ascending order of the Euclidean distance from the point in the operation data space corresponding to the target rolling material data, and uses only the selected past data to obtain the learning coefficient of the rolling load. may be calculated. According to such a configuration, regardless of whether the data point group is sparse or dense, a predetermined number of data points are used, so it is possible to achieve both a reduction in rolling load prediction time and prediction accuracy. .

Further, as shown in FIG. 4, the past data B is divided into a plurality of databases R2, R3, and R4 according to the range of values, and the rolling process computer 3 collects only the past data contained in the database including the data request point A (see FIG. 4 In the example shown in 1), the learning coefficient of the rolling load may be calculated using the database R2). Furthermore, if the number of past data to be processed is less than a predetermined value, the rolling process computer 3 expands the predetermined Euclidean distance, or is included in the database adjacent to the database corresponding to the value range of the target rolled material data Past data may be included in the processing targets. Specifically, in the example shown in FIG. 5, the past data to be processed is increased by enlarging the area R1 shown in FIG. 3 to the area R6. In the example shown in FIG. 6, the past data in the database R7, which is a combination of the database R3 adjacent to the database R2 shown in FIG. 4, is used as the past data to be processed.

Also, the past data may be selected in order from the one with the smallest Euclidean distance, and only the past data that sufficiently contributes to the calculation of the learning coefficient of the rolling load may be used. Specifically, in this case, as shown in FIG. 7, when the nearest neighbor point of the data request point A is X _a , its Euclidean distance (nearest neighbor distance) is da, and the similarity is Wa, the data request point _A The density ρ _da of past data inside a circle with a radius of d _a +Δd centered on is obtained by the following formula (6). Also, the degree of contribution Ca of the past data to the learning coefficient within this circle is obtained by the following _formula (7).

Next, as shown in FIG. 8, the contribution C of the past data to the learning coefficient in _a donut-shaped region excluding the circle R8 of the nearest neighbor distance da within the circle R9 obtained by multiplying the nearest neighbor distance _da by n _b is obtained by the following formula (8). However, the parameter k in Expression (8) indicates the number of past data in the doughnut-shaped area.

By adding the _donut - _shaped region shown in _FIG . 8 compared with the circle R8 shown in FIG. assumed). Therefore, when the radius of the donut-shaped region to be added first is r, the radius of the next donut-shaped region to be added is r ² , the radius of the next donut-shaped region to be added is r ³ , and so on. When the radius of the donut-shaped area is increased and the condition shown in the following formula (9) is satisfied, even if the learning coefficient is calculated using the past data of the area, the change in the result is small. means. In Equation (9), C _a is the degree of contribution to the learning coefficient of the past data included in the circular region and the donut-shaped region added so far, and C _b is the degree of contribution to the learning coefficient in the next donut-shaped region. Shows the contribution of the included past data to the learning coefficient. Also, in Expression (9), ε indicates a calculation end parameter. Further, when the number of data exceeds the threshold without satisfying the condition shown in formula (9), the learning coefficient is calculated using the past data up to that point.

In both cases, the total computational complexity is O(n ³ ), and the additional computational complexity is O(n logn ). For this reason, when the number of past data by this method is smaller than the number of past data in the case of a fixed number, even though the effect on accuracy is slight, speedup can be expected by reducing the amount of calculation. Note that O(n ³ ) and O(n·logn) are indices (order notation) that indicate calculation time estimates when the amount of data is sufficiently large. Specifically, O(n ³ ) indicates a calculation time proportional to the cube of the number of data, and O(n·logn) indicates a calculation time proportional to the first power×log of the number of data. In general, as the number of data increases, the part proportional to the cube of the number of data becomes dominant in the calculation time. Therefore, by reducing the number of data to be calculated in O(n ³ ), the calculation time can be greatly shortened. In this embodiment, the number of calculation data of O(n ³ ) can be reduced by the additional calculation amount O(n·logn). Since the amount of calculation is O(n ³ )>O(n·logn), the total calculation time can be shortened.

In order to evaluate the effect of the present invention, a comparison test was conducted on the rolling load prediction accuracy and calculation speed in the rolling of thick plates. In this example, a four-high reversible rolling mill with an average work roll diameter of 1200 mm and a backup roll diameter of 2200 mm was used. Comparative Example 1 shows the case where the rolling load is not applied using the learning coefficient. Example 1 for the case of using only the data, Example 2 for the case of using only a predetermined number of past data selected in ascending order of the Euclidean distance, only the past data contained in the database corresponding to the value range of the target rolled material data was used as Example 3, and the rolling load prediction accuracy (actual rolling load/predicted rolling load) was evaluated with 10000 points of data.

The results of Examples 1-3 and Comparative Examples 1 and 2 are shown in Table 1 below. First, in Comparative Example 1 that does not use data accumulated in the past, the standard deviation of the rolling load prediction accuracy was 6.3%, whereas in Comparative Example 2, the standard deviation of the rolling load prediction accuracy was 4. It was confirmed that it greatly improved to .0%. However, in Comparative Example 2, the calculation time of the learning coefficient was 6.0 seconds, which was long enough to be used in actual operation, and it was confirmed to be difficult to apply in terms of industrial production. On the other hand, in Examples 1 to 3, since the data used for calculating the learning coefficient was limited, the calculation time for the learning coefficient was significantly reduced to 0.3 seconds, which was a practically acceptable level. Although the standard deviation of the rolling load prediction accuracy is slightly worse than that of Comparative Example 2 using all data, it is difficult to apply Comparative Example 2 for industrial production, so Examples 1 to 3 are better. It can be said that this has a beneficial effect on industrial production.

1 rolling control device 2 operation data management device 3 rolling process computer (rolling process computer)
4 Load detector 5 Plate thickness controller (AGC PLC)
6 reduction device 10 reversible rolling mill

Claims (7)

A rolling load prediction method for predicting the rolling load of each pass of a reversible rolling mill that rolls a target rolling material in multiple passes,
a step of acquiring, as operation data, a plurality of pieces of past data for each operation, including the set value or actual value of a predetermined operation factor and the actual value of the rolling load regarding the rolled material to be rolled in the past operation;
a step of acquiring the value of the predetermined operation factor related to the target rolled material as target rolled material data;
a step of calculating the degree of similarity between the target rolled material data and the past data by using the predetermined operating factors as condition items and comparing values for each of the condition items;
The actual value of the rolling load of each past data is weighted by the similarity of the relevant past data, and the weighted value is used to apply the learning coefficient that each condition item exerts on the rolling load when rolling the target rolled material. a step of calculating
Acquire the actual value of the predetermined operation factor after the previous pass rolling for the target rolled material, replace the set value of the predetermined operation factor included in the target rolled material data with the obtained actual value, and use the calculated learning coefficient predicting the rolling load for the next pass based on
including
The rolling load prediction method, wherein the step of calculating the learning coefficient processes only past data within a range of a predetermined Euclidean distance from a point in the data space corresponding to the target rolled material data. The step of calculating the learning coefficient is characterized by selecting a predetermined number of past data in ascending order of distance from a point in the data space corresponding to the target rolled material data, and processing only the selected past data. The rolling load prediction method according to claim 1. The past data is classified into a plurality of databases according to the value range, and the step of calculating the learning coefficient processes only the past data included in the database corresponding to the value range of the target rolled material data. The rolling load prediction method according to claim 1 or 2. In the step of calculating the learning coefficient, when the number of past data to be processed is less than a predetermined value, the predetermined Euclidean distance is expanded, or a database adjacent to the database corresponding to the value range of the target rolled material data 4. The rolling force prediction method according to any one of claims 1 to 3, wherein included past data is included in the object to be processed. 5. Any one of claims 1 to 4, wherein the step of calculating the learning coefficient includes a step of changing the number of past data to be processed in consideration of the degree of contribution to the calculation of the learning coefficient. The rolling load prediction method described in the item. By controlling the reversible rolling mill based on the rolling load predicted by the rolling load prediction method according to any one of claims 1 to 5, the thickness of the target rolled material is rolled to the target thickness. A rolling control method, comprising: A rolling load prediction device that predicts the rolling load of each pass of a reversible rolling mill that rolls a target rolling material in multiple passes,
Means for acquiring, as operation data, a plurality of pieces of past data for each operation, including set values or actual values of predetermined operation factors and actual values of the rolling load, regarding the rolled material to be rolled in past operations;
means for acquiring the value of the predetermined operation factor related to the target rolled material as target rolled material data;
means for calculating the degree of similarity between the target rolled material data and the past data by comparing the values for each of the condition items, with the predetermined operating factor as a condition item;
The actual value of the rolling load of each past data is weighted by the similarity of the relevant past data, and the weighted value is used to apply the learning coefficient that each condition item exerts on the rolling load when rolling the target rolled material. a means for calculating
Acquire the actual value of the predetermined operation factor after the previous pass rolling for the target rolled material, replace the set value of the predetermined operation factor included in the target rolled material data with the obtained actual value, and use the calculated learning coefficient means for predicting the rolling load for the next pass based on
with
The rolling load prediction device, wherein the means for calculating the learning coefficient processes only past data within a range of a predetermined Euclidean distance from a point in the data space corresponding to the target rolled material data.

Images