EP0994757B1 - Process and system for controlling or pre-setting a roll stand - Google Patents

Abstract

A process is disclosed for controlling and/or pre-setting a roll stand for rolling a strip. The roll stand is controlled and/or preset depending on at least one of the following values: rolling force, rolling torque and peripheral precession, which are calculated by means of a roll model. The calculation of the rolling force, rolling torque and peripheral precession by means of the roll model is carried out as a function of at least the hardness of the strip or the friction between the strip and the rolls of the roll stand, which are used as input values for the roll model. At least one of the input values of the roll model or at least one of the output values of the roll model is determined or corrected by at least one neural network.

Description

The invention relates to a method for controlling and / or Presetting of a roll stand for rolling a rolled strip According to the preamble of claim 1. It further relates to a corresponding device for carrying out the method.

A method according to the preamble of claim 1 and a Corresponding device are for example from the Fachaufsatz "Experiences with the use of neural networks in the Walzwerksautomatisierung "by D. Lindhoff et al., Published in Stahl und Eisen, Volume 114 (1994), No. 4, pages 49 to 53, known.

From DE 44 16 317 A1 a similar method is known. In this document, but no correction of the rolling force, the rolling moment and / or the lead by means of a further neural network.

From DE 195 27 521 C1 is a method of control and / or presetting a rolling stand for rolling a Metal band known in which the control and / or presetting of the roll stand in dependence on a rolling force, a rolling moment or an overfeed takes place, wherein the Rolling force, the rolling moment and / or the lead by means of a Roll model at least depending on the hardness of the Metal band and / or the friction between the metal band and the rolls of the rolling stand are calculated.

For presetting a rolling train or a roll stand before threading the rolled strip to be rolled or for control the rolling train or the roll stand after threading the rolled strip must be variables such as rolling force, rolling moment or overfeed or more of these sizes. It is possible, to determine these variables by means of a rolling model, in the as input variables the band height, the inlet of the rolled strip, the band width, the strip tension, the material hardness and / or the Friction between rollers and rolled strip. It has However, it has been shown that in such a process the quality requirements, especially for higher quality steels, often can not be met. It is accordingly Object of the invention, the quality of a rolled steel, in particular by compliance with thicknesses or hardness tolerances, to increase.

The object is achieved by a method with the Features of claim 1 and a device with the features of claim 6 solved.

FIG 1

die physikalischen Verhältnisse einer Walzstraße sowie den Zusammenhang zwischen Dickenreduktion und Materialhärte,

FIG 2

die physikalischen Verhältnisse in einem Walzspalt,

FIG 3

ein Vorgehen am Eingang des Walzmodells,

FIG 4

ein Vorgehen am Ausgang des Walzmodells,

FIG 5

ein erfindungsgemäßes Vorgehen am Ein- und Ausgang des Walzmodells,

FIG 6

ein Trainingsverfahren für neuronale Netze in einer besonders vorteilhaften Ausgestaltung,

FIG 7

ein alternatives Trainingsverfahren für ein neuronales Netz zur Bestimmung der Materialhärte und

Further advantageous and inventive details will become apparent from the following description of exemplary embodiments, with reference to the drawings and in conjunction with the dependent claims. In detail show:

FIG. 1

the physical conditions of a rolling mill and the relationship between thickness reduction and material hardness,

FIG. 2

the physical conditions in a nip,

FIG. 3

a procedure at the entrance of the rolling model,

FIG. 4

a procedure at the exit of the rolling model,

FIG. 5

an inventive procedure at the input and output of the rolling model,

FIG. 6

a training method for neural networks in a particularly advantageous embodiment,

FIG. 7

an alternative training method for a neural network for determining the material hardness and

1 shows the physical conditions of a rolling train and the relationship between thickness reduction and material hardness. In this case, reference symbols FS _i denote the lead on the ith frame, MR _i the rolling moment on the ith frame, FR _i the rolling force on the ith frame, FT _i the strip tension on the ith frame, eps _i the relative decrease in thickness on the i -th framework, MS the material hardness, ie the hardness of the rolled strip, V _i the belt speed after the ith frame and H _i the strip thickness after the ith frame. The relative decrease in thickness eps _i results from: eps i = H 0 - H i H 0 with H ₀ : strip thickness during unwinding and
H _i : band thickness after the i-th scaffold, i = 1,2,3,4,5 for 5 scaffolds.

The lead FS _i at the i-th frame is defined as FS i = V i V wi with V _wi = circumferential speed of the ith roller.

If the roll is assumed to be circular, the peripheral speed of the i-th roll results according to: V wi = 2π • R AW • n i With
R _AW = roller radius of the i-th roller and
n _i = speed of the i-th roller.

Reference numerals 1, 2, 3, 4 and 5 denote rolling stands, reference numerals 6 a uncoiler, reference numeral 7 a rolled strip and 8, a coiler.

FIG. 1 also shows the relationship between material hardness MS and thickness decrease eps or thickness reduction. This relationship is particularly suitable by the function MS = MS (eps) = MSO + MSI • eps MSE described.

FIG. 2 illustrates the physical relationships in a roll gap, which are advantageously included in the modeling with a rolling model. The conditions in the nip are advantageously modeled by a strip model, where it is sufficient for reasons of symmetry to model only the upper or only the lower part of the roll stand, so that a boundary of the rolling model is the axis of symmetry 23 of the rolled strip 27. The strip 27 is split in the region of the contact surface strip - roller in strips 28 (due to the clarity, only one strip is provided with a reference numeral) perpendicular to the direction of movement of the strip 27. Within each strip 28, the material tension _forces F _{Q are} calculated in the horizontal and vertical directions and adjusted to each other via equilibrium conditions at the strip edges. In FIG. 2, some material tension forces F _{Q are entered} by way of example. The vertical material _tension forces F _Q lead to a flattening 26 of the roller 21. The calculation of the flattened roller radius R _B is carried out iteratively with the aid of the strip model and a model which describes the deformation of the roller.

The flow sheath 20 is the location where the material is straight with the peripheral speed of the roller 21 moves. In front the flow sheath 20, the material moves slower, behind the flow sheath 20 faster than the peripheral speed the roller 21. Except at the location of the flow sheath 20 occurs accordingly everywhere between work roll 21 and material one Relative movement 24, 25 on. This relative movement 24, 25 leads to considerable friction forces.

3 shows an improvement of the output variables of a rolling model 32 by changing input variables 32 of the rolling model 31 by means of a neural network based Information Processing 33. The procedure according to FIG. 3 as such is not the subject of the present invention. In this case, the rolling model 31 determines depending on the Input variables 30 and 32 Output variables 34. These output variables are rolling force, rolling moment and / or overfeed. The Inputs 32 are made by one on neural networks based information processing or a neural Network 33 as a function of input values 35 of the neural Network 33 is formed. The input variables 30 and 32 are z. As the tensile force in the rolled strip, the bandwidth, the inlet thickness of the rolled strip, the hardness of the rolled strip and / or the friction between roller and rolled strip. Of these, advantageously one, in particular the hardness of the rolled strip, through the neural network 33 detected. In this case, the input variables 35 of the neural network z. B. material-specific Data such as B. the alloying shares, the inlet thickness, the outlet thickness and characteristics of a previous processing such as B. Thickness reduction or temperature in the preceding Processing.

FIG. 4 shows another possibility for improving the output variables a rolling model 41 by correcting the output variables 47 of the rolling model 41. Also the procedure as shown in FIG 4 is not the subject of the present invention Invention. The rolling model 41 is determined as a function of Input variables 43, such as material hardness, friction between Rolls and rolled strip, tension, belt width or inlet thickness rolled strip, outputs 47. These outputs are rolling force, rolling moment and / or overfeed. The output variables 47 of the rolling model 41 are replaced by a correction block 45 corrected in response to correction parameters 44. Output variables of the correction block 45 are corresponding corrected values for rolling force FR, rolling moment MR and / or lead FS. It is particularly advantageous the output variables 47 of the rolling model 41 by multiplication with the correction parameters 44 to corrected values for rolling force FR, rolling moment MR or lead FS to link. The Correction parameters 44 are by means of a neural network 42 determined as a function of input variables 46.

5 shows a particularly advantageous embodiment of the invention. Here, reference numeral 51 denotes a rolling model. Input variables 64 and MS of the rolling model 51 are the material hardness MS as indicated by reference numeral 64 rolling stock or scaffold-specific data such. B. friction between Rolls and rolled strip, tension, belt width and inlet thickness of the rolled strip. The material hardness MS is using a neural network, material network 50, depending on certain Input variables 60 are calculated. These input variables 60 can be: alloy parts, inlet thickness, outlet thickness, Temperature as well as information for the characterization of preprocessing such as B. previous degree of reduction or previous processing temperature. Outputs 65 of the Rolling model 51 are values for rolling force, rolling moment and / or Overfeed. These are in a correction block 53 in dependence corrected by correction parameters FRCP, MRCP, FSCP, by means of a neural network, scaffolding network 52, depending on of inputs 61. These input variables 61 are u. a. the tape thickness, the bandwidth as well roll-specific data. Output variables 66 of the correction element 53 are corrected values for rolling force, rolling moment and / or Overfeed. These are fed to a further correction element 55, this by means of the correction parameters FRCD, MRCD and FSCD further corrected. The correction parameters FRCD, MRCD, FSCD are using a neural network, daytime network 54, calculated as a function of input variables 62. These Input variables are u. a. Strip thickness, belt width and roll-specific Dates. Output variables 67 of the correction element 55 are corrected values for rolling force, rolling moment and overfeed, by means of another correction element in dependence of correction parameters FRCS, MRCS and FSCS on Getting corrected. The correction parameters FRCS, MRCS, FSCS be using a neural network, speed network 56, calculated as a function of input variables 63. The input variables 63 are the speed of the rolled strip as well u. a. Tape thickness, bandwidth and roll specific data. output 68 of the correction element 57 are corrected values for rolling force, rolling moment and overfeed, by means of a another correction element 59 as a function of a correction factor β for fine correction and adaptation to the current Rolled strip to be corrected. Output variables of the correction element 59 are corrected values for rolling force FR, rolling moment MR and lead FS. The correction members 53, 55, 57, 59 can z. B. multipliers. Basically, others come too Corrective strategies in question. Such correction strategies or joins of neural networks that are for the given Application can be used are known.

The material network 50 provides the material hardness MS z. In Form of the regression parameters MSI, MSO described in FIG and MSE. The skeleton network 52 provides framework-specific correction factors FRCP, MRCP and FSCP for rolling force, rolling moment and Overfeed. The material network 50 and the skeleton network 52 become advantageously trained with data, material and Represent the rolling stand over the life of the roll stand.

The daily network 54 supplies the correction factors FRCD, MRCD and FSCD for rolling force, rolling moment and overfeed, which is the relative small changes according to the daily form of the mill stand describe. Accordingly, the training of the day network is done 54 with young records, z. For example, records that are not older than three days.

The speed network 56 provides the speed-dependent Correction factors FRCS, MRCS and FSCS for rolling force, Rolling moment and advance. With the speed network 56 In particular, friction-specific deviations are compensated. The friction between roller and rolled strip is very strong from the belt speed. The friction is i. a. like this smaller, the higher the belt speed is, as between Rolled strip and rolls with increasing speed Lubricating film forms.

6 shows a training method for an inventive Structure according to FIG. 5. MSE, MSI represent this and MSO the material hardness according to MS in FIG 5. The meaning MSE, MSI and MSO is explained in FIG. FR, MR, FS, β, FRCL, MRCL, FSCL, FRCD, MRCD, FSCD, FRCS, MRCS and FSCS have the same meaning as in FIG 5. The input variables 86, 87, 88, 89 correspond to the input variables 60, 61, 62, 63 in FIG 5. Reference numerals 76, 77, 78 denote material networks with the associated training or learning algorithms. Reference numeral 81 denotes a scaffolding net with an associated one Learning or training algorithm, reference numeral 82 Day network with associated learning or training algorithm and Reference numeral 83 a speed network with associated Learning algorithm. Reference numeral 70 denotes a data memory or a database, in the data AC, FRA, MRA and FSA stored, the characteristics of a representative Cross-section of all rolled in the corresponding mill / rolling mill Form bands. FRA, MCA and FSA are the actual ones Values for rolling force, rolling moment and advance over one long period, e.g. over the life of the mill stand, considered. They are formed from the roll-specific data AC. Function block 80 denotes an inverted one Walzmodell and a regression model, whereby by means of the inverted Roll model from the data AC the actual material hardness determined on the individual stands of the rolling mill is and where by means of the regression model of the Values for material hardness of the individual scaffolds the actual Values for the parameters MSE, MSI and MSO are formed. By means of the values determined by the regression model 80 MSO, MSI and MSE the material networks 76, 77, 78 are trained. By means of the values MSE, MSI and MSO, from the material network 76, 77, 78 are output, as well as other input variables 90, a rolling model calculates 79 values for rolling force, Rolling moment and advance. The input quantities correspond 90 the input variables 64 of FIG. 5

The skeleton network 81 becomes dependent on the input variables 87 the data AC, FRA, MRA and FSA as well as the output quantities of the Roll model 79 trains. By means of a correction block 53 (see FIG 5) are the output variables of the roll model 79 with the correction parameters FRCL, MRCL and FSCL, which are the skeleton network 81 issues, corrected.

With the output variables of the correction block 53, the input variables 88 as well as the data DC, FRD, MRD and FSD becomes the daily network 82 trained. Output variables of the daily network 82 are the correction parameters FRCD, MRCD and FSCD, the input variables in a correction block 55, by means of which the output variables of the correction block 53 are corrected. The parameters DC, FRD, MRD and FSD from the database 71 the data AC, FRA, MRA and FSA, in contrast to to the data AC, FRA, MRA and FSA only rolling belts of the last one Represent the day or the last days.

The outputs of the correction block 55, the input variables 89 and the ACC data are input variables in the speed network 83 and its learning algorithm. Furthermore, go in the speed network 83 or its learning algorithm correction parameters FRCS, MSCS and FSCS using a Speed correction element 85 are determined. there transforms the velocity correcting member 85 to a Standard speed normalized correction parameters FRC, FSC and MSC with respect to the actual speed of the rolled strip. The data ACC correspond to the data AC, but only represent the current rolled strip. Contains accordingly the database or the data memory 72 only the data for the current rolled strip. Output variables of the velocity network 83 are correction parameters FRCS, MRCS, FSCS, which are in enter a further correction block 57. The output of this Correction block enters a further correction block 59.

ermittelt, die mittels eines Walzmodells in Abhängigkeit der bekannten Datensätze berechnet werden.Also input to the correction block 59 is a parameter β which is stored in a memory 84. Output of the correction block 59 are corrected values for rolling force FR, rolling moment MR and lead FS. The adaptive values for rolling force FRA, rolling moment MRA, leading FSA used for training the neural networks and the correction values FRC, FSC and MSC for training of the neural networks for rolling force, overfeed and rolling moment become dependent on estimated values

determined, which are calculated by means of a rolling model depending on the known data sets.

The training of neural networks thus takes place in a long-term learning part 73, a day or short-term learning part 74 as well a speed learning part 75.

FIG. 7 shows an alternative training of the material network, wherein the material hardnesses are used for n rolling stands. In this case, reference numeral 70 denotes a database corresponding to FIG. 6, AC roll-specific data (see FIG. 6), reference numeral 100 a material network with a learning algorithm, 101 a regression model, and reference numeral 102 a rolling model. The material hardnesses MS _{1... N} at the individual stands and optionally the rolled _strip temperature T _strip and the total thickness reduction eps _{1... N} assigned to the individual rolling stands are output from the material network 100. Instead of the regression parameters MSU, MSI, MSE, the material network 100, which consists of one or more neural networks, outputs the material hardnesses MS _{1... N} , which are subsequently transformed by means of a regression model 101 into regression parameters MSO, MSI and MSE. The regression parameter MST is a parameter representing the temperature dependency, which can optionally be calculated if the temperature T _{strip of} the rolled _{strip also} enters the material network 100. This parameter is particularly advantageous if the method according to the invention is not used for cold rolling but for hot rolling.

,

und

, aus denen die Werte FRC, MRC und FSC, die zum Training der neuronalen Netze verwendet werden (vgl. FIG 6), berechnet werden, wird im folgenden verdeutlicht. Sind MSO, MSI und MSE Eingangsgrößen des Walzmodells, so gilt für das i-te Gerüst

The determination of the estimates

.

and

, from which the values FRC, MRC and FSC used to train the neural networks (see FIG. 6) are calculated, will be clarified below. If MSO, MSI and MSE are input variables of the rolling model, then the i-th framework is valid

und für die Voreilung

Analog applies to the rolling moment

and for the lead

. and denote the estimates of the respective models.

B denotes the band width, H _i-1 the band thickness before the ith frame, H _i the band thickness behind the ith frame, FT _i-1 the band tension before the ith frame, FT _i the band tension behind the i-th frame scaffold and v _Wi the peripheral speed of the work rolls in the i-th scaffold.

FRC, MRC and FSC are calculated FRC = FR is FR · · FRCS FRCD · FRCV MRC = MR is MR · · MRCS MRCD · MRCV FSC = FS is FS · · FSCS FSCD · FSCV where FR _is , MR _is and FS _is the current values for rolling force, rolling moment and overfeed.

FRA, MRA and FSA are calculated from: FRA = FR is FR MRA = MR is MR FSA = FS is FS

Except for the parameters MSO, MSI and MSE and the friction μ Actual values are available for all input variables. The coefficient of friction μ However, z. B. tabular deposited. It is also possible, the friction μ with a neural network in analog How to determine the material hardness.

With the method according to the invention, either the material hardness or the friction can be determined. It is further conceivable to determine both quantities by means of a neural network. However, it has been shown that it is usually is sufficient, only one of the two unknowns, material hardness or friction, by means of a neural network. If the material hardness z. B. according to the invention by means of a neural network and for friction (rough) Estimates are used, the material network is able to the errors with respect to the rolling force, the rolling moment or the lead caused by inaccurate knowledge of the friction between Rolled strip and roller arise, correct. tries have shown that with bad estimates of friction and using the method according to the invention the neuronal Net a poor estimate of material hardness determines that this deviation from the actual material hardness However, the error compensated for the coefficient of friction. On This way is indeed by the inventive method a suboptimal value for the material hardness obtained, however a particularly precise value for rolling force, rolling moment and Overfeed. For a control or a presetting, at the expected values for rolling force, rolling moment and Overfeed, but not the actual value of the material hardness, is relevant, it is sufficient in most cases, only a material network, but not a friction network to use.

Claims (6)

1. Process for controlling and/or presetting a roll stand (1 - 5) for rolling a metal strip (7),

   the control and/or presetting of the roll stand (1 - 5) taking place in dependence on a rolling force (FRi), a rolling torque (MRi) and/or a forward slip (FSi),

   the rolling force (Fri), the rolling torque (MRi) and/or the forward slip (FSi) being calculated by means of a rolling model (51) at least in dependence on the hardness (MS) of the metal strip (7) and/or the friction between the metal strip (7) and the rolls of the roll stand (1 - 5),

   at least one of the input variables (FRi, MRi, FSi, FT0 - FT5, H0 - H4, MS) of the rolling model (51) being determined or corrected by means of a neural network (50),

   the rolling force (FRi), the rolling torque (MRi) and/or the forward slip (FSi) being corrected by means of a further neural network (52, 54, 56),

   characterized

   in that the further neural network (52, 54, 56) comprises a stand network (52), which is trained with roll-specific data which form an average value over the lifetime of the roll stand (1 - 5), and

   in that the further neural network (52, 54, 56) comprises at least one supplementary neural network (54, 56), which takes into account influences with time constants in the range from a day to several days.
2. Process according to Claim 1, characterized in that the input variable (FRi, MRi, FSi, FT0 - FT5, H0 - H4, MS) determined or corrected by means of the neural network (51) is the hardness (MS) of the metal strip (7) and/or the friction between the metal strip (7) and the rolls of the roll stand (1 - 5).
3. Process according to Claim 1 or 2, characterized in that the neural network (50), and possibly also the further neural network (52, 54, 56), is or are retrained.
4. Process according to one of Claims 1 to 3, characterized in that the supplementary neural network (54, 56) also comprises a speed network (56), which takes into account the influence of the speed (v0 - v5) of the metal strip (7) on the rolling force (FRi), the rolling torque (MRi) and/or the forward slip (FSi).
5. Process according to Claim 4, characterized in that the speed network (56) is trained with data of the particular metal strip (7) at a given time.
6. System for carrying out a process according to one of Claims 1 to 5, with a rolling model (51), a neural network (50) arranged upstream of the rolling model (51) and a further neural network (52, 54, 56) arranged downstream of the rolling model (51),

   a rolling force (FRi), a rolling torque (MRi) and/or a forward slip (FSi) being able to be calculated by means of the rolling model (51) at least in dependence on the hardness (MS) of a metal strip (7) and/or the friction between the metal strip (7) and the rolls of a roll stand (1 - 5),

   at least one of the input variables (FRi, MRi, FSi, FT0 - FT5, H0 - H4, MS) of the rolling model (51) being able to be determined or corrected by means of the upstream neural network (50),

   the rolling force (FRi), the rolling torque (MRi) and/or the forward slip (FSi) being able to be corrected by means of the further neural network (52, 54, 56),

   the further neural network (52, 54, 56) comprising a stand network (52), which is trained with roll-specific data which form an average value over the lifetime of the roll stand (1 - 5),

   the further neural network (52, 54, 56) comprising at least one supplementary neural network (54, 56), which takes into account influences with time constants in the range from a day to several days, and

   the roll stand (1 - 5) being able to be controlled or preset on the basis of the rolling force (FRi), the rolling torque (MRi) and/or the forward slip (FSi).

Images