MXPA99003811A - Optimizing the band width at the band ends on a mill train - Google Patents

Abstract

The width of bands to be laminated (4) on a mill train (3) is adjusted by means of vertical upsetting rollers (6), resulting, however, in a narrowing at the band ends due to the asymetric material flow there. In order to solve the problem, the upsetting rollers (6) are so designed as to move at the passage of the band ends in keeping with a curve (f) defined according to specified parameters (s). The parameters (s) are based on neuro-computer made predictions related to the milling process.

Description

PROCEDURE TO OPTIMIZE THE DISTRIBUTION OF BANDWIDTH AT THE END OF A BAND THAT GOES THROUGH A LAMINATION TRAIN FIELD OF THE INVENTION The invention relates to a method for optimizing the distribution of bandwidth at the ends of a band passing through a rolling mill.

BACKGROUND OF THE INVENTION One of the main problems in laminating bands, for example, strip steel, is to achieve a basic rectangular shape with a constant width along the band. To control the bandwidth, upset rollers are used in the rolling mill. If these rollers approach with a constant position, usually due to the flow of asymmetric material and other effects, the band is narrower at the ends of the same, that is, the beginning and the end of the band, which in the central part. To counteract the above, the position of the upsetting rollers is adjustable during the passage of the band, extending more, relatively to the central part, the position when passing the ends of the band, in the form of small deviations, called "short strokes" " This position correction at the beginning and at the end of the band is carried out in accordance with a movement curve ("short stroke control" (SCC) movement curve), which can be defined by predetermined parameters.

OBJECTIVES AND ADVANTAGES OF THE INVENTION The objective of the invention is to generate as best as possible a distribution of desired bandwidth at the ends of the band, by predetermining a movement curve for the position of the upsetting rollers. According to the invention, the objective is achieved by the methods indicated in independent claims 1, 2, 3 and 4. Advantageous developments of the methods according to the invention are provided in the sub-claims. Therefore, the determination of the parameters to form the movement curve, according to which the position of the upsetting rollers is adjusted during the passage of the ends of the bands, is carried out based on predictions about the process of lamination through neural networks, in which, through an online training of the same in the lamination process, the predictions can be continuously improved. Preferably, separate neural networks are used for the start and end of the band. For successive steps of the same band, that is, in the case of several upsets, separate neural networks can be used. If the number of upsetting is always fixed, a single neural network can also be used to determine the parameters for the movement curves of the upsetting rollers in successive upsets.

BRIEF DESCRIPTION OF THE DRAWINGS For a more detailed illustration of the invention, reference is made below to the figures of the drawings, which show: Figure 1, an example of the width distribution of a laminated strip and a derivative movement curve of the same for the upset rollers for the correction of the width distribution. Figure 2, an example of the fundamental control structure of a rolling mill with a unit for determining parameters for the definition of the movement curve. Figures 3 to 8, various embodiments of the unit to determine the parameters of the movement curve. And Figure 9, a detailed eraser based on the example according to Figure 8, for the determination of the parameters of the movement curve.

DETAILED DESCRIPTION OF THE INVENTION The diagram in Figure 1 shows by way of example the distribution of width and of a band in its length 1 when going through a rolling mill, which comprises, in addition to horizontal flat rolling rollers for the regulation of the thickness of the band, also vertical upset rollers for the control of the width of the band. With a constant position of the upset rollers, due to asymmetric material flows, in the band there is a decrease in the width of the band at the ends, that is, the beginning and end of the band. To counteract the previous and obtain a basic rectangular shape of the band, when passing the ends of the band the position of the upsetting rollers is adjusted according to a movement curve f, which, in the example shown, consists of two straight sections adjustable, separated for the start and end of the band as well as for each step of the band, that is, each upset. The movement curve f is described by four parameters in the form of two position correction values ax and a2 and two coordinates of length l-_ and 12. The position correction values ax and a2 refer to the distance between rollers, so that the path of both upsetting rollers is half. Of course, it is possible to describe the movement curve in another way and with other parameters. The aim is to determine in such a way the parameters at r a2, lx and 12 of the movement curve f that an adjustment of the positions of the upset rollers according to the movement curve defined by the parameters leads to a predetermined theoretical distribution of the bandwidth at its ends, in this case, therefore, a basic rectangular shape. As illustrated below based on several examples, the foregoing is carried out with the aid of neural networks, although certain parameters can be predetermined, in this case, for example, the coordinates of length l-_ and 12, also as empirical values. Figure 2 shows the fundamental control structure for a rolling mill 3, in which an optimization of the actual bandwidth distribution yisfc of a band 4 passing through the rolling mill 3 takes place, according to a predetermined theoretical distribution of the bandwidth ysoll. The rolling mill 3 is a preliminary train having one or more horizontal structures with flat rolling rollers 5, being arranged before the last structure and, if necessary, also more horizontal structures, in this case the last two, respectively a vertical structure with upsetting rollers 6. Before each entry of a band 4 in the rolling mill 3, in a preliminary calculation unit 7, based on theoretical values SW and primary data PD and with the help of mathematical models 8 of the rolling process, process parameters x relevant to the rolling process are previously calculated and delivered to a basic automation 9, which carries out with them a pre-adjustment of the rolling mill 3. During the rolling process, by means of a unit of recording of measurement values 10 relevant dimensions of the rolling process are continuously recorded. The dimensions are fed to the basic automation 9 to comply with regulation functions, as well as to a further calculation unit 11. It uses the same mathematical models 8 as the preliminary calculation unit 7 and adapts the corresponding parameters of the model based on the dimensions that represent the actual course of the rolling process. In this way, the previous calculation for the band 4 respectively following to be laminated is continuously improved and adjusted to what happens in the actual process. To control the position of the upset rollers 6, the basic automation 9 contains a corresponding control device 12. This, depending on the parameters s it receives, for example s = (a1 to 2) according to Figure 1, generates a curve of movement f according to which the position of the upsetting rollers 6 is adjusted as the ends of the band pass. The determination of the parameters s for the motion curve f is carried out in a unit 13, first depending on the distribution of theoretical bandwidth ysou and the process parameters x calculated previously, and using at least one neural network 14, which provides a prediction about the processes of upsetting at the ends of the band. In order to improve and adjust the predictions of the neural network 14 to what actually happens in the process, at the exit of the rolling train 3, by means of a measuring device of widths 15, the actual distribution of yiat bandwidths and with it and the process parameters calculated later in the subsequent calculation unit 11, an adaptation of the neural network 14 is carried out. As regards the number of neural networks 14 used, preferably, for the start and end of the band separate neural networks are used. In addition, separate neural networks can be used for successive steps of the same band 4. If the number of upsets is variable band in band, then the use of separate neural networks is disadvantageous for the greater number of upsets, because in them fall less training data. Figures 3 and 4 show a first embodiment possibility of the unit 13 in two operating states.

A neuronal advancement model 140 is used as a neural network, which reproduces the stressing process in its natural cause / effect relationship. The input dimensions of the neural network 140 in its training phase (Figure 3) consist of the process parameters calculated later xnach and the sist parameters of the movement curve, which are determined by a conversion unit 16 from the measured movement curve fist, according to which the upsetting rollers 6 are moved in the rolling process. The neural network 140 provides a prediction of the distribution of bandwidths y, which is compared with the actual distribution of the yist bandwidths. Depending on the error? And determined in this way, an adaptation of the neural network 140 takes place, so that this, for predetermined parameters of the movement curve and the existing process parameters x, provides a prediction as accurate as possible of the distribution of bandwidths and obtained in this way. Figure 4 shows how the optimal parameters are determined sopt of a movement curve, with which a predetermined theoretical distribution of the ysou bandwidth distribution is achieved. For this purpose, a calculation unit 17 initially provides initial values sstart for the parameters s of the movement curve and is fed to the neural network 140 adapted together with the process parameters x previously calculated. The foregoing provides a prediction for the distribution of bandwidths y, which is compared with the theoretical distribution of bandwidths and are- If the variation between the predicted bandwidth distribution yy the theoretical distribution of bandwidths and ssoll exceeds a preset limit value, then the initial values sstart are modified by a value? s. The neural network 140 provides with the new parameters s = sstarC +? S a new prediction for the distribution of bandwidths y, which is again compared with the theoretical distribution of bandwidth ysoll. The parameters s for the movement curve are modified by steps in an amount? S until the variation between the predicted bandwidth distribution y and the theoretical distribution of bandwidth ysoll no longer exceeds the preset limit value. The parameters s determined in this way correspond to the optimal parameters sought after of the movement curve, with which the position of the upsetting rollers 6 is controlled. In the example shown in Figures 5 and 6 of an embodiment of unit 13, two neural networks 140 and 141 are used, of which the first neural network 140, as was already the case in the example of Figures 3 and 4 , is a neuronal advancement model, and the second neural network 141 is a neuronal regression model that describes the reversal of the natural cause / effect relationship. As shown in Figure 5, in a first operating state of unit 13, the first neural network 140 is trained in the same manner as described previously with the help of Figure 3. Upon completion of the training of the first network neural 140, according to Figure 6, by means of the second neural network 141, based on a predetermined theoretical distribution of the ysoll bandwidths and the process parameters x calculated previously, a prediction of the parameters s of the movement curve is generated , according to which the upsetting rollers 6 are adjusted as the band passes. The distribution of measured bandwidth ylst that results is compared to the theoretical distribution of the bandwidths and s? the error? and obtained by the first neural network 140 being propagated inversely and used for the adaptation of the variable weights of the network w (NN141) of the second neural network 141 according to the gradient descent process: d? y __ d? y d and d s d w [NNÍAÍ) d and d s d w (NN il) The exemplary embodiment for the unit 13 shown in Figure 7, contains as neuronal network a model of neuronal retreat 141 in accordance with that shown in "Figures 5 and 6. This neural network 141 provides in adapted state, depending on a predetermined theoretical distribution of bandwidths and ssoll and process parameters x previously calculated, a prediction of the parameters s of the movement curve, according to which the position of the upsetting rollers 6 is adjusted as the band passes in. To adapt the neural network 141 to the actual events of the process, after the passage of the band, the distribution of real bandwidths measured ylst and the process parameters further calculated naCh are fed to the neural network 141 as input dimensions, whose network response s is compared to the actual slst motion curve parameters determined to through the conversion unit 16 from the measured movement curve flst, depending on the error? s obtained in this way, an adaptive of the neural network 141. The embodiment of the unit 13 shown in Figure 8 is based on the condition that, in the case of an error in the distribution of widths, that is, in the case of a variation? and between the theoretical distribution of bandwidth ysoll and the distribution of real bandwidth measured yist, the flse movement curve for upsetting rollers 6 must be modified by the amount of this variation? and, to compensate for the error. Therefore, as a neural network 142, a neuronal regression model can be used. The theoretical distribution of bandwidth and sol? it is definitively fixed, for example, for a rectangular shape of the ends of the band, so that the neural network 142 has a reduced task approach with respect to the examples described above and for the prediction of the parameters s of the curve of only the process parameters x, or xnach, are fed to it as input dimensions. Before the entry of the band, the neural network 142, based on the process parameters x calculated previously, provides a prediction of the parameters s of the movement curve f, according to which a preliminary adjustment of the rolls of upsetting 6. After bandpass, the theoretical distribution of desired bandwidth and ssoll is compared to the distribution of real bandwidths and measured. With the variation? And obtained in this way, the measured movement curve flst is corrected to a theoretical movement curve fsoll, whose corresponding ssoll parameters are determined by a conversion unit 18. The neural network 142, based on the parameters of process further calculated xnach that are now fed, provides a prediction of the parameters of the movement curve s, which are compared with the ssoll parameters of the theoretical motion curve fsoll, using the error? s so obtained, for adaptation of the neural network 142. Of course, the conversion interface between the motion curve f and its parameters 18 can also be arranged in another way, by converting the parameters s predicted by the neural network 142 into a predicted motion curve f and comparing the curve of predicted movement f with the theoretical movement curve fso ?? - The above results from the following example. Figure 9 shows a detailed eraser, based on the example of Figure 8, to determine the movement curve f for upsetting rollers 6. As shown in Figure 1, for each of the total three upset i ( = 1,2,3), the movement curve f (i) must consist respectively of two straight sections, which are described in total by the four parameters a ± (±), a2 (i), l (i) and l2 (i) • The process parameters x, or xnach relevant for the determination of the motion curve f, include the bandwidth b (i), the band thickness d (i) and the band temperature T (i) after each upset i, the width reduction? B (i) and the thickness reduction? D (i) of the band 4 after each upset i, as well as a coefficient OI as a measure of the hardness of the material (resistance to deformation) of the band 4. The prediction of the parameters of the movement curve a -_ (i) and a2 (i), that is, of the position correction values are carried out for the three upsets through a neural network 142, which has correspondingly six network outputs ok (= 0_ 5). The position correction values ax (i) and a2 (i) are calculated as a product between the network outputs ok that are between -1 and +1 and the respective width reduction? B (i) of the band 4. The foregoing has the consequence that none of the position correction values ax (i) and a2 (i) can be greater than the respective width reduction? B (i). The length coordinates l -_ (i) and l2 (i) are preset by a device 19 as empirical values. The length coordinate 12 (1), which corresponds to the length of the area of action of the upsetting rollers 6 at the first upset on the band 4, is set for the beginning of the band, for example, as 3 times and for the end of the band, like 2 times the roughing width. For the following upsetting, the length of the area of action of the upsetting rollers 6 is divided by two each time, so that it is valid 12 (2) = 1/2 12 (1) and 12 (3) = 1/4 12 (1). The other length coordinates are established with lx (i) = 1/3 l2 (i). The values determined in this way for the length coordinates refer to the band 4 after leaving the rolling mill 3, when measuring the distribution of bandwidths. For this reason, in order to control the upsetting rollers 6 in each upsetting i, these values, due to the stretching of the strip caused in each step of the strip by the rolling train 3, must be converted to the strip length before the respective upsetting i in proportion to the length of the band 4 after leaving the rolling mill 3. This conversion is carried out based on the temperature T (i), the width b (i) and the thickness d (i) of the band 4 before the respective upsetting i, the temperature, width and thickness of the strip 4 after leaving the rolling mill 3 and the expansion coefficient a. The movement curve parameters a- i) and a2 (i) predicted by the neural network 142 based on the previously calculated process parameters x and the movement curve parameters lx (i) and l2 (i) preset through of the unit 19, are transmitted to the basic automation 9 for the adjustment of the rolling mill 3. By laminating the strip 4 in the rolling mill 3, by means of the measuring value recording device 10 and the measuring device of width 15, the yist width distribution and the fist movement curve of the upsetting rollers 6 are measured by points. In a unit 20, the error? And between the predetermined theoretical distribution of the Ysoii Y ^ bandwidths is first computed for distribution of real bandwidths measured yist and, then, the theoretical movement curve fsoll from the measured movement curve fist and the error? and in, in this case, for example, seven pre-established support points j (= 0. ..6). The calculation of the support values fsoll / j of the theoretical movement curve is in sum for all the upsets i, that is, the theoretical movement curve fsoll is the sum of the theoretical movement curves fso ?? (i) of each upset i. The neural network 142, based on the subsequently calculated process parameters xnach fed to it after the passage of the web 4 through the rolling mill 3, provides predictions about the position correction values! (i) and a2 (i), from which, in a unit 21 at the support points j, support values fi of the predicted summed movement curve f resulting from the predicted position correction values are calculated. ax (i) and a2 (i) in sum for all the upsets i. In a unit 22, by 'comparing the values of support fso ??, j of the theoretical movement curve summed fsoll with the support values fj of the summed motion curve predicted f, the error is determined? f = fsoi?, j ~ fj • A From the error? fj, in another unit 23 , you get the square error summed of all the points of support which is used for the adaptation of the neural network 142 according to the gradient descent process. As already mentioned before, the rolling mill 3 of the exemplary embodiment shown in Figure 1 is a preliminary train. To also take into account the influence of subsequent treatment lines, such as finishing trains and cooling paths, on the distribution of bandwidths, this is measured at the end of the cooling path and fed to unit 13 for the determination of the parameters s of the movement curve.

Claims (9)

NOVELTY OF THE INVENTION Having described the above invention, it is considered as a novelty, and therefore, the content of the following is claimed as property: CLAIMS

1. A method to optimize the distribution of bandwidths at the ends of a band passing through a rolling mill, adjusted the position of the upsetting rollers as the ends of the band pass, according to a movement curve described by preset parameters, where, based on the process parameters of the rolling process and based on the parameters of the movement curve, a prediction on the distribution of bandwidths and the error between the distribution of bandwidths is provided by a neural network predicted and the distribution of measured real bandwidths is used for the adaptation of the neural network in the sense of a reduction of the error, and where, with the adapted neural network, optimal values are iteratively determined for the parameters of the movement, for which the predicted bandwidth distribution varies minimally from a distribution of theoretical bandwidths to predetermined and according to which the position of the upsetting rollers is subsequently adjusted.

2. A method to optimize the distribution of bandwidths at the ends of a band passing through a rolling mill, adjusting the position of the upsetting rollers when passing the ends of the band, according to a movement curve described by parameters pre-established, where based on the process parameters of the rolling process and based on the distribution of measured real bandwidths, through a neural network is provided a prediction on the parameters of the movement curve and the error between the parameters predicted of the movement curve and the actual parameters of the movement curve, according to which the position of the upsetting rollers was adjusted, is used for the adaptation of the neural network in the sense of a reduction of the error, and where , with the adapted neural network, based on the process parameters of the lamination process and a predetermined theoretical bandwidth distribution, it is det They eliminate optimum values for the parameters of the movement curve, according to which the position of the upsetting rollers is adjusted.

3. A procedure to optimize the distribution of bandwidths at the ends of a band passing through a rolling mill, adjusting the position of the upsetting rollers when passing the ends of the band, according to a movement curve described by parameters pre-established, where based on the process parameters of the rolling process and based on the actual parameters of the movement curve, according to which the position of the upset rolls was adjusted, a prediction is provided by a first neural network on the distribution of bandwidths and the error between the distribution of predicted bandwidths and the distribution of real measured bandwidths is used for the adaptation of the first neural network in the sense of a reduction of the error, where, in addition, , based on the process parameters of the rolling process and a predetermined theoretical bandwidth distribution, through a second network ne A prediction is provided on optimal values for the parameters of the movement curve, according to which the position of the upset rolls is adjusted, and where the error between the distribution of theoretical bandwidths and the distribution of bandwidths real measure, it propagates back through the first adapted neural network and is used for the adaptation of the second neural network in the sense of a reduction of the error.

4. A procedure to optimize the distribution of bandwidths at the ends of a band passing through a rolling mill, adjusting the position of the upsetting rollers as the ends of the band pass, according to a movement curve described by pre-established parameters, where based on the process parameters of the rolling process, through a neural network a prediction is provided on optimal values for the parameters of the movement curve, according to which it is adjusted the position of the upsetting rollers, where from a measured movement curve and the error between a predicted bandwidth distribution and the actual measured bandwidth distribution, a theoretical movement curve is determined, and where the error between the predicted optimal values for the parameters of the movement curve and the corresponding parameters determined from the theoretical movement curve or the error between the predicted movement curve from the predicted optimal values for the parameters of the curve of movement and the theoretical movement curve, is used for the adaptation of the neural network in the sense of a reduction of this error.

5. A method according to claim 1, characterized in that separate neural networks are used for the start of the band and for the end thereof.

6. A method according to claim 1, characterized in that the movement curve for adjusting the position of the upsetting rollers, at each step of one end of a strip, consists of two straight sections, which are defined by two »values of correction of position and two coordinates of length, preset these as empirical values and forming the values of correction of position the parameters of the curve of movement.

7. A method according to claim 1, characterized in that separate neural networks are used for consecutive steps of the same band.

8. A method according to claim 1 of claim 1, characterized in that one and the same neural network is used for a fixed predetermined number of consecutive steps of the same band.

9. A method according to claim one of the preceding claims, characterized in that the bandwidth, band thickness and band temperature after each step, width reduction and reduction are used as process parameters. of the thickness of the band in each step and a measure for the resistance to the deformation of the material of the band.