KR950001935B1 - Process for improvement of rolling load in steel sheet - Google Patents

Abstract

The method improves the estimation of the plate roll load, a factor influencing a most sensitive effect on working conditions of FSU model of continuous hot rolling equipment, so that it can enhance a precision and a reliability of the estimation. The method comprises the steps of : obtaining an average learning coefficient of each stand from overall data; obtaining a calibration coefficient in order to make a group learning coefficient, obtained via classification of working conditions, approach the average value; obtaining a calibration function by use of the group learning coefficient; calibrating a transformation resistance model by use of the calibration model.

Description

Improvement method for predicting rolling load of sheet

1 is a distribution chart of learning coefficients according to classifications before operation.

2 is a distribution change chart of learning coefficients by operation classification after correction.

3 is a variation of learning coefficient distribution after stand correction.

4 is a bar graph showing the effect of reducing the model hit deviation.

The present invention relates to a method for improving the rolling load prediction accuracy of a plate having the most sensitive effect on the operating conditions of the FSU model of the filamentary rolling mill arranged in tandem method among continuous hot rolling mills of steel. will be. The hot rolling reduction setting model during hot rolling has a basic goal of predicting the change of shape, temperature, and material in the finishing rolling process from predictable process factors and controlling the operating conditions in order to operate according to each target value.

Conventionally, the rolling load was obtained through a test, and then the rolling load prediction equation was derived as a function of the test variable.

However, since each influencer influences the rolling load independently or through interaction, it is not easy to formulate the effect of each operation variable and accurately reflect it in the rolling load model. Therefore, the hit accuracy of the rolling load prediction model has a limit, and fine variables or error factors that are difficult to express in the model have been optimized by adjusting through learning. Japanese patent JP58084606: Kawasaki Steel Co., Ltd. has attempted to improve the rolling load by using a formula that takes into account the metallurgical change of the rolled material due to rolling on. This raises the question of improving the accuracy in practical application.

In European patents related to rolling loads (European Patent EP0289064: Fugo Vance Steel), the rolling load prediction was attempted to improve the prediction accuracy compared to the performance by using the multiplication factor multiplied by the deformation resistance. These effects do not show much difference from the existing model in that they can be expected only if they are accurately formulated through experiments.

The present invention improves the rolling load prediction accuracy by using the rolling results without going through the experiment, to reduce the time and cost burden through the experiment, as well as to provide a method of improving the prediction accuracy of the rolling load with much higher accuracy and reliability than the experiment There is a purpose.

EMBODIMENT OF THE INVENTION Hereinafter, this invention is demonstrated.

The present invention relates to a method for predicting the rolling load of a plate in a rolling reduction setting model of a finishing rolling mill arranged in a tandem method of a continuous hot rolling mill of steel, wherein the learning coefficient (α) value of each stand is determined from all the data. Obtaining an average value; Calculating a correction coefficient mi as shown in Equation (16) so that the learning coefficient for each group (2) obtained through the classification of the operating condition groups is close to the learning coefficient (2) value obtained from the entire data;

Where mi is the correction factor

αij: Operating conditions and learning coefficients for each stand

= Km _a / Km _c ] ij-1.0

α _T j: Learning coefficient obtained by each stand from all coils

Km _c ] ij: Operating conditions and calculated strain resistance for each stand

Km _c ^m ] ij: Km _c ] ij after correction

Obtaining a correction function (fi) by using the correction coefficient (mi) obtained by Equation (16) as a function of the classification condition for each group; Correcting with a deformation resistance model using the correction function fi as shown in Equation (3);

Km (new) = f ₁ , f ₂ ... fi. Km (old)

Where Km (old): strain resistance model before improvement, Km (new): strain resistance model after improvement, f ₁ : correction function

After correcting as described above, verifying the degree of reduction of the value of the learning coefficient by verifying the data using real data, and then selecting a condition giving the lowest deviation; The present invention relates to a method for improving the predicted rolling load of a plate comprising the steps of repeatedly performing the above steps.

That is, the present invention relates to a method for improving a rolling load prediction model of a plate to improve the accuracy of model hitting by using a correction function capable of compensating the deviation between the predicted rolling load and the measured rolling load for each operating condition and for each stand.

EMBODIMENT OF THE INVENTION Hereinafter, this invention is demonstrated in detail.

The present invention aims to improve the accuracy of rolling load prediction by removing predictive error factors for each operating condition based on statistical theory and regression analysis. To this end, the error factors of the used rolling load prediction model are analyzed and the existing model is modified by introducing a correction function that can reduce the prediction error. The correction function is a quantitative representation obtained to solve the problem from the error factors of the existing model. Here, the concept of the correction function will be explained against the deformation resistance prediction model, which is a model for determining the rolling load, in order to improve the accuracy of the rolling load.

In order to display the accuracy of the predicted (resisted) strain resistance and the performance strain resistance, the mathematical model is analyzed by designating the learning coefficient α as a parameter of the predicted accuracy.

α = (Km _a / Km _c -1.0). 100

Where α is the parameter (%)

Kma: Performance Strain Resistance (kg / mm ² )

Kmc: Calculated strain resistance (kg / mm ² )

In other words, if α is 0 (zero), the predicted value and the performance value are the same.If the value is positive, the forecast value is smaller than the performance value.If the value is negative, the predicted value is larger than the performance value, the absolute value The larger the value, the proportionately increase the difference between the forecast value and the performance value. Performance deformation resistance Kma is calculated | required from the rolling force P by stand obtained at the time of a rolling through the Sims formula as follows.

Km _a (kg / mm ² ) = P / (B _m L _d Q _p )

Where P: rolling load (kg)

Bm: Average plate width (mm)

Ld: Contact field (mm)

Qp: geometric factor

The calculated strain resistance (Kmc) is calculated using the Shida equation used in the existing model. The analysis of the hot deformation resistance mathematical model is to classify the process performance into major operating variables such as thickness, width, carbon equivalent and bender force, and then divide the classified data into groups of various variable ranges, and The characteristics of each classification group were identified by obtaining the mean and standard deviation of the values, and a stepwise correction was made with a mathematical model to minimize these values. By repeating this series of processes, the average value of α is approached to zero, thereby improving the accuracy of the model, reducing the learning dependence, and reducing the learning deviation, thereby improving the learning efficiency of the model. In other words, the strain resistance equation, Km (old) currently in use, is the basic form, and the correction function is introduced to directly correct the error factors generated by the current equation, thereby creating a new strain resistance equation, Km (new).

Km (new) = f ₁ , f ₂ ... f _i . Km (old) (3)

Where fi: correction function

The correction function fi is obtained through the following process.

Where i = operation condition classification number

j = stand number

ni = coil number per job classification

n = total number of coils (n _T = Σni)

αij, Kma] ij, Kmc] ij = Operating Condition and Learning Factor, Performance, and Calculation Strain Resistance by Stand

αTj, Kma] Tj, Kmc] Tj = learning coefficient, performance, and calculated strain resistance obtained by each stand from all coils

The correction coefficient is calculated so that the α value for each group obtained through the classification of the operating conditions is close to the α value obtained from the entire data.

(1 + αij) Kmc] ij = Kma] ij

In order for the learning coefficient value αij to be close to the learning coefficient αij obtained from the entire data, the right side term of the following equation should be obtained.

Thus in Kmc] ij After the correction is made, the learning coefficient value of each group obtained through the classification of the working condition group is as close to the learning coefficient obtained from the entire data as follows.

Where mi = correction factor

Where Kmc ^m ] ij: Kmc] ij after correction

That is, equations (7) and (8) have the same learning coefficient after correction.

The correction coefficient obtained through the above process is obtained by calculating the correction function (fi) as a function of the classification ship for each group and selecting the correction function that gives the lowest deviation by testing the obtained correction functions in actual operation.

Hereinafter, the present invention will be described in more detail with reference to Examples.

EXAMPLE

In the case of actual hot rolling, the operating variables are scattered over a wide range within the same stand, and there are also many differences in the prediction of rolling load for each stand. Therefore, it is not easy to improve the degree of hot deformation resistance by simply adjusting the coefficient of the model. Therefore, in order to investigate the cause of the difference, the work performance was classified by main condition survey as shown in Table 1, and the correction function was calculated for each operation condition and stand.

TABLE 1

Figure 1 shows the distribution of learning coefficients (α) by each classification before the correction, which shows that the thinner the thickness, the larger the carbon equivalent, and the higher the bending force, the lower the predictive strain resistance.

In order to reduce the variation of the learning coefficient indicated by the classification of each stand and the operating condition, the correction function for the thickness, carbon equivalent, and bender force except for the width of each classification among the four classifications of FIG. . From the learning coefficients obtained through equation (1) for each thickness group, the correction factor for model modification was obtained from equation (6), and the correction function was obtained by regression analysis. The correction function obtained through the regression analysis is as follows.

f (t) = 2.19-0.93.t + 0.24t ² -0.02.t ³

Here, f (t) = thickness correction function The correction function obtained after performing chemical composition and bender force correction in the same manner as the thickness classification before is as follows.

f (Ceq) = 1.14-7Ceq + 78.4.Ceq ²

f (P _B ) = 1.52-0.016.P _B + 8.38% · 10 ^-5 .P _B ²

Where f (Ceq) = carbon equivalent correction function

f (P _B ) = bender correction function

Three correction functions obtained through the above procedure cannot be used more than one at the same time.

That is, after one of the three correction functions is applied to the actual model, the two remaining correction functions are no longer valid. In order to select the correction function that is most effective in reducing the deviation, the above three correction functions were applied independently and then off-line test was performed using the field results. Table 2 shows the results of applying the three correction functions, the carbon equivalent correction formula was found to be the most effective in reducing the deviation of the learning coefficient.

TABLE 2

Therefore, the new model equation after the first correction is as follows.

Km ₁ (new) = f ₁ .Km (old)

f ₁ = f (Ceq) = 1.14-7Ceq + 78.4Ceq ²

The results obtained after the 2nd and 3rd corrections in the same way as the 1st correction are shown below.

Km (new) = f ₂ .Km ₁ (new) = f ₁ .f ₂ .Km (old)

P = Pc + f ₃ .xP _B

Where f ₁ = f (Ceq) = 1.14-7Ceq + 78.4Ceq ²

f ₂ = f (Ceq) = 1.09-1.445Ceq

f ₃ = f (P _B ) = 1.036-0.0015P _B + 9.555 · 10 ^-6 .P _B ²

Figure 2 shows the distribution of the learning condition coefficients obtained by applying the correction function and equation (13). Since the degree of dispersion was considerably reduced under the given conditions without considering the absolute distribution of the learning coefficients for each stand, the stand correction was performed to increase the predicted performance.

For each stand, the stand correction function was calculated in the same way as the previous correction so that the mean of the learning coefficient could be close to zero from the mean of the learning coefficient obtained through the correction of the previous section.

The correction function obtained through the regression analysis is as follows.

f (s) = 0.933-0.013.COS (s) + 0.105.SIN (s) -0.041.COS (2s) + 0.00237.SIN (2s)

Where f (s) = stand correction function

s = stand number

Therefore, the final correction equation is expressed as follows.

Km (new) = f ₁ .f ₂ .f ₄ .Km (old)

P = Pc + f ₃ .xP _B = Km.B.Ld.Qp + f ₃ .xP _B

Where P = rolling load, ton

P _B = ender force, _T on

Pc = calculated rolling load, _T on

x = bender effect function

f ₁ = f (Ceq) = 1.14-7Ceq + 78.4Ceq ²

f ₂ = f (Ceq) = 1.09-1.445.Ceq

f ₃ = f (P _B ) = 1.036-0.0015.P _B + 9.955 × 10 ^-6 .P _B ²

f ₄ = f (s) = 0.933-0.013Cos (s) + 0.105Sin (s) -0.041 Cos (2s) + 0.00237Sin (2s)

Figure 3 shows the distribution of learning coefficients after the correction of the stand. If the scale is greatly changed for each stand, it can be seen that the learning coefficient is very close to zero, which means that the prediction accuracy of the performance of each stand is greatly improved. In Figure 4, the reduction of the model-fitting deviation of the ultra low carbon steel obtained through the model modification is shown in FIG. 4, and the average deviation is reduced by 28.5%. This means that the prediction accuracy of the model has been greatly improved, and that the model's response to changes in operating conditions is more consistent with less variation.

Claims (1)

A method for estimating the rolling load of a sheet in a rolling reduction setting model of a filament rolling mill arranged in tandem method in a continuous hot rolling mill of steel, the method comprising: calculating average values of learning coefficients (2) for each stand from all data ; Calculating a correction coefficient mi as shown in Equation (16) so that the learning coefficient for each group (2) obtained through the classification of the operating condition groups is close to the learning coefficient (2) value obtained from the entire data; ···························· ... ············ 16 Where mi is the correction factor aij: operating conditions and learning coefficient for each stand = Kma / Kmc] ij-1.0 α _T j: Learning coefficient obtained for each stand from the whole coil Kmc] ij: Operating conditions and calculated strain resistance for each stand Kmc ^m ] ij: Kmc] ij after correction Obtaining a correction function (fi) by using the correction coefficient (mi) obtained by Equation (16) as a function of the classification condition for each group; Correcting the deformation resistance model using the correction function fi as shown in Equation (3); Km (new) = f ₁ .f ₂ ... fi.Km (old) Where Km (old): strain resistance model before improvement, Km (new): strain resistance model after improvement, fi: correction function After correcting as described above, obtaining a degree of reduction of the value of the learning coefficient by verifying the data using real data, and then selecting a condition giving the lowest deviation; And repeatedly performing the above steps.