METHOD FOR DETERMINING MECHANICAL PROPERTIES OF A ROLLED MATERIAL USING A HYBRID MODEL

Abstract

A method for determining mechanical properties of a first rolled material by a hybrid model that includes production datasets relating to further rolled materials, a physical production model and a statistical data model. The production dataset relating to the first rolled material is used to determine a first mechanical dataset, a further production dataset and a metallurgical dataset and also a second mechanical dataset. An averaged normalized distance value for production datasets relating to the further rolled materials is determined that is used to ascertain the mechanical properties of the rolled material as a weighted average from the first and second mechanical datasets. When creating the hybrid model, the physical production model is used to determine further production datasets relating to the further rolled goods for training the statistical data model.

Claims

1-15. (canceled)

16. A method for determining mechanical properties of a first rolled material, produced in a rolling mill, using a hybrid model, based on a production dataset of the first rolled material, the hybrid model comprising: production datasets of further rolled materials, produced in the rolling mill, from a first set that each have available corresponding sampling datasets, comprising the mechanical properties of the further rolled materials ascertained by a sampling device after production in the rolling mill; a physical production model comprising a set of metallurgical model parameters, for ascertaining: a first mechanical dataset, comprising the mechanical properties of the first rolled material after production in the rolling mill, a further production dataset, comprising further production data of the first rolled material during production in the rolling mill, and a metallurgical dataset, comprising metallurgical properties of the first rolled material during production in the rolling mill, and a trained and validated statistical data model for ascertaining a second mechanical dataset, comprising the mechanical properties of the first rolled material after production in the rolling mill; wherein the method comprises a first, a second, a third and a fourth step; wherein the first step involves ascertaining the first mechanical dataset, the further production dataset and the metallurgical dataset of the first rolled material from the production dataset of the first rolled material via the physical production model; wherein the second step involves ascertaining the second mechanical dataset of the first rolled material from the production dataset, the further production dataset and the metallurgical dataset of the first rolled material via the statistical data model; wherein the third step involves determining for the production dataset of the first rolled material an averaged normalized distance value in relation to a number of production datasets of the further rolled materials; and wherein the fourth step involves using the averaged normalized distance value to ascertain the mechanical properties of the first rolled material as a weighted average from the first and second mechanical datasets, the quotients of the mutually corresponding values of the second and first mechanical datasets being weighted in proportion to the averaged normalized distance value.

17. The method as claimed in claim 16, wherein the mechanical properties of the first rolled material comprise at least one of a tensile strength, a yield strength, and an elongation at break.

18. The method as claimed in claim 16, wherein: the averaged normalized distance value is formed from normalized distance values of the production dataset of the first rolled material in relation to a respective production dataset of one of the further rolled materials; and the number of production datasets comprises 0.1 to 0.5% of the training datasets of the statistical data model.

19. The method as claimed in claim 16, wherein the mechanical properties of the first rolled material are ascertained for at least one rolled material section of the first rolled material.

20. The method as claimed in claim 19, wherein: the rolling mill comprises roll stands, cooling devices, rolling mill sections, and at least one coiling device; the production datasets of the first rolled material and the further rolled materials each comprise at least one of the following manipulated or measured variables for at least one rolled material section of the first rolled material and the further rolled materials: one or more foreign element fractions in the rolled material section, at least one of a thickness and a width of the rolled material section before entering a first roll stand, at least one of a thickness and a width of the rolled material section after exiting a last roll stand, at least one temperature of the rolled material section recorded by a measuring arrangement, at least one roll gap value of one of the roll stands when the rolled material section passes through the applicable roll stand, at least one coolant stream emitted by one of the cooling devices when the rolled material section passes through the operating range of the applicable cooling device, or at least one speed of the rolled material section in one of the rolling mill sections.

21. The method as claimed in claim 19, wherein the further production dataset for at least one rolled material section of the first rolled material comprises at least one of: at least one modeled cooling rate of the rolled material section when passing through at least one of the rolling mill sections of the rolling mill; and at least one modeled temperature value of the rolled material section when passing through the at least one of the rolling mill sections of the rolling mill.

22. The method as claimed in claim 19, wherein the metallurgical properties of the metallurgical dataset comprise at least one of phase fractions and morphological characteristics of different metallurgical phases of at least one rolled material section of the first rolled material.

23. The method as claimed in claim 19, wherein a fifth step involves subjecting the ascertained mechanical properties of the first rolled material to a cooling correction in which carbide and nitride precipitations in the first rolled material are taken into account during cooling after production in the rolling mill.

24. A method for creating a hybrid model for determining mechanical properties of a first rolled material produced in a rolling mill, the hybrid model comprising: production datasets of further rolled materials, produced in the rolling mill, from a first set that each have available corresponding sampling datasets, comprising the mechanical properties of the further rolled materials ascertained by a sampling device after production in the rolling mill; a physical production model having a set of metallurgical model parameters for ascertaining: a first mechanical dataset, comprising the mechanical properties after production in the rolling mill, a further production dataset, comprising further production data during production in the rolling mill, and a metallurgical dataset, comprising metallurgical properties during production in the rolling mill, for each of the further rolled materials; and a statistical data model for ascertaining a second mechanical dataset, comprising the mechanical properties after production in the rolling mill, for each of the further rolled materials; wherein one or more consecutive iterations of a first step involve using the physical production model to ascertain for each of the production datasets of the further rolled materials the corresponding first mechanical dataset and further production and metallurgical datasets; wherein each iteration is followed by a first measurement function being used for component-by-component ascertainment, for all first mechanical datasets ascertained in the most recently performed iteration, of a measure of the deviation from the corresponding sampling datasets of the further rolled materials; wherein if the value of the first measurement function for at least one mechanical property in the first mechanical datasets of the further rolled materials is greater than a limit, the metallurgical model parameters are varied and, in a further iteration, the first mechanical datasets, the further production datasets and the metallurgical datasets of the further rolled materials are re-ascertained based on the physical production model with the modified metallurgical model parameters; wherein if the value of the first measurement function for all mechanical properties in the first mechanical datasets of the further rolled materials is less than or equal to the limit, the production datasets together with the further production and metallurgical datasets of the further rolled materials ascertained in the most recently performed iteration are transferred to the statistical data model as input variables and the corresponding sampling datasets are transferred to the statistical data model as target variables; wherein a second step involves training the statistical data model based on the transferred input and target variables using a first machine learning method; and wherein a third step involves validating the trained statistical data model based on production datasets and sampling datasets of rolled materials from a second set, which is disjunct from the first set, by supplying the further production datasets and metallurgical datasets ascertained for the rolled materials of the second set by the physical production model, together with the production datasets of the rolled materials of the second set, to the trained statistical data model as input variables and ascertaining corresponding second mechanical datasets by the trained statistical data model and ascertaining component-by-component-wise a measure of a deviation from the corresponding sampling datasets for said second mechanical datasets based on a second measurement function.

25. The method as claimed in claim 24, wherein the mechanical properties of the first rolled material and the sampling datasets, the first and the second mechanical datasets of the further rolled materials each comprise at least one of a tensile strength, a yield strength, and an elongation at break.

26. The method as claimed in claim 24, wherein the first step is preceded by the sampling dataset from each of the further rolled materials that corresponds to the production dataset being transformed to a normalized sample geometry according to Oliver's formula.

27. The method as claimed in claim 24, wherein the mechanical properties of the first rolled material and the first and second mechanical datasets of the further rolled materials are each ascertained for at least one rolled material section of the first rolled material and the further rolled materials.

28. The method as claimed in claim 27, wherein: the rolling mill has roll stands, cooling devices, rolling mill sections and at least one coiling device; and the production datasets of the first rolled material and the further rolled materials each comprise at least one of the following manipulated or measured variables for at least one rolled material section of the first rolled material and the further rolled materials: one or more foreign element fractions in the rolled material section, at least one of a thickness and a width of the rolled material section before entering a first roll stand of the rolling mill, at least one of a thickness and a width of the rolled material section after exiting a last roll stand of the rolling mill, at least one temperature of the rolled material section recorded by a measuring arrangement, at least one roll gap value of one of the roll stands when the rolled material section passes through the respective roll stand, at least one coolant stream emitted by one of the cooling devices when the rolled material section passes through the operating range of the respective cooling device, or at least one speed of the rolled material section in one of the rolling mill sections.

29. The method as claimed in claim 27, wherein the further production dataset for at least one rolled material section of each of the further rolled materials comprises at least one of: at least one modeled cooling rate of the rolled material section when passing through at least one of the rolling mill sections of the rolling mill; and at least one modeled temperature value of the rolled material section when passing through the at least one of the rolling mill sections of the rolling mill.

30. The method as claimed in claim 27, wherein the datasets of the further rolled materials comprise at least one of phase fractions and morphological characteristics of different metallurgical phases of at least one rolled material section of each of the further rolled materials.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0128] The above-described properties, features and advantages of this invention, and the manner in which they are achieved, will become clearer and more clearly understandable in conjunction with the following description of an exemplary embodiment, which will be explained in more detail in conjunction with the figures, in which:

[0129] FIG. 1 schematically shows a hot rolling mill comprising a central computation unit having the hybrid model according to the invention,

[0130] FIG. 2 shows a flowchart for the method according to the invention for determining mechanical properties of a first rolled material on the basis of a hybrid model according to the invention,

[0131] FIG. 3 schematically shows the position of input and output data of the statistical data model, and

[0132] FIG. 4 shows a flowchart of the method according to the invention for creating a hybrid model.

DETAILED DESCRIPTION

[0133] Parts that correspond to one another are provided with the same reference signs in all of the figures.

[0134] FIG. 1 schematically shows a hot rolling mill 1. The hot rolling mill 1 comprises a heating furnace 3, a roughing line 4, a cropping shear 5, a finishing rolling line 7, a cooling section 8 and a coiling device 26. The finishing rolling line 7 comprises five roll stands R.sub.1, . . . R.sub.5; the cooling section 8 comprises four cooling devices Q.sub.1 . . . Q.sub.4. A rolled material 2, 2, 2 in the form of a slab is introduced into the heating furnace 3 and, after heating, heated to an increased temperature, for example to 1100? C. to 1200? C., and then transported in the production direction of the hot rolling mill 1 (shown in FIG. 1 by a horizontal arrow pointing to the right) in the direction of the roughing line 4.

[0135] In the roughing line 4, a first thickness reduction for the rolled material 2, 2, 2 is carried out, said rolled material being rolled from a slab format to a so-called roughed strip. Arranged downstream of the roughing line 4 when viewed in the production direction is a cropping shear 5, which can be used to cut off the strip head or the strip end of the roughed strip, since these sections of the roughed strip usually have to be scrapped. After passing through the cropping shear 5, the roughed strip enters the finishing rolling line 7, in which it is rolled to a specified final thickness. Furthermore, temperature measuring devices 6 in the form of a pyrometer arearranged immediately upstream of the first roll stand R.sub.1 and immediately downstream of the last roll stand R.sub.5in relation to the production direction of the hot rolling mill 1to contactlessly record a surface temperature of the rolled material 2, 2, 2.

[0136] After exiting the last roll stand R.sub.5, the rolled material 2, 2, 2 passes through a cooling section 8, in which it is cooled to a coiling temperature by applying coolant by way of the cooling devices Q.sub.1 . . . Q.sub.4. After leaving the operating range of the last cooling device Q.sub.4, the rolled material 2, 2, 2 is wound into a coil by a coiling device 26.

[0137] In line with the mentioned specialisms, the hot rolling mill 1 is divided into individual rolling mill sections A.sub.1 . . . A.sub.Z with Z=13 by data processing means. In FIG. 1, the rolling mill section A.sub.4, which comprises the area of the first roll stand R.sub.1 and the infeed-side pyrometer 6, is symbolically delimited by dash-dotted lines. The production process for the rolled material 2, 2, 2 in the hot rolling mill 1 is controlled or regulated on the basis of the respective production parameters P.sub.f, P.sub.i using an installation controller 27.

[0138] The installation controller 27 is connected to the specialisms of the individual rolling mill sections A.sub.1 . . . A.sub.13 by data processing means, this being symbolized in FIG. 1 by the bidirectional arrows. From the point of view of the installation controller 27, the production parameters P.sub.f, P.sub.i comprise both output data (e.g. default values for the roll gap of the individual roll stands R.sub.1 . . . R.sub.5), which are sent to the respective specialisms, and input data (such as a production speed in the respective rolling mill section or measured temperature values), which the installation controller 27 receives from the individual specialisms.

[0139] The rolled material 2, 2, 2 is divided into a, preferably multiple, rolled material section(s) 20, 20, 20 of, for example, two to 5 meters in length by data processing means. The rolled material sections 20, 20, 20 are tracked during production in the rolling mill 1: this makes it possible to assign short-term fluctuations in the applicable process parameters P.sub.f, P.sub.i to individual rolled material sections 20, 20, 20 and to record said fluctuations over the length of the rolled material 2, 2, 2.

[0140] After the production process has concluded, material samples are taken from some of the rolled materials 2, 2 produced in the hot rolling mill 1 by means of a sampling device 9, the mechanical properties B.sub.i of which material samples are measured in a test laboratory, for example, and are transferred as sampling data B.sub.i together with the corresponding production data P.sub.i from the installation controller 27 to a central computation unit 28, which is embodied as a computer or programmable logic controller (PLC) and comprises the hybrid model 10 according to the invention. Production data P.sub.f, of rolled materials 2 that are not sampled, i.e. for which no sampling data are ascertained, are also transferred from the installation controller 27 to the central computation unit 28, this being symbolized by a bidirectional arrow between the installation controller 27 and the central computation unit 28. In FIG. 1, the hybrid model 10 according to the invention is represented by means of a rectangular symbol within the central computation unit 28. The sampling device 9 and the central computation unit 28 are not part of the hot rolling mill 1.

[0141] FIG. 2 shows a data flow diagram for the method according to the invention for determining mechanical properties J.sub.f of a first rolled material 2 produced in a hot rolling mill 1 with process steps S1, S2, S3 and S4 on the basis of a hybrid model 10 according to the invention. The hybrid model 10 comprises production datasets P.sub.i of further rolled materials 2 produced in the rolling mill 1, a physical production model 12 and a trained and validated statistical data model 14 as structural components and is shown in FIG. 2 using a broad line thickness. These structural components of the hybrid model 10 are each shown in FIG. 2 as rectangular symbols. Data of the first rolled material or the further rolled materials 2, 2 are identified by a circle symbol or a circle symbol with set braces. The sequentially performed steps S1 to S4 are shown in FIG. 3 by horizontal dashed lines. Arrows symbolize a data flow in which the data belonging to or coming from a source object are transmitted in the direction of the arrow to a target object.

[0142] Furthermore, the production datasets P.sub.i, the physical production model 12 and the data model 14 themselves can comprise data structures (for example a chronological sequence of individual data in the production datasets P.sub.i), internal computation rules (for example a heat conduction equation or the empirical metallurgical equations of the physical production model 12), and also relations between different data (for example links between data within the trained and validated statistical data model 14 when ascertaining a first mechanical dataset C.sub.f from the input variables of the data model 14). The production datasets P.sub.i, the physical production model 12 and the data model 14 can also each comprise further internal parameters: these include, for example, the metallurgical model parameters ? or internal parameters of the statistical data model 14, which govern the reaction thereof to input variables. The mentioned data structures, internal computation rules, relations between data and internal parameters are each realized in the present exemplary embodiment as data or programming structures in the central computation unit 28, meaning that the method according to the invention for determining mechanical properties J.sub.f of a first rolled material 2 can be carried out as a program by the central computation unit 28.

[0143] The first step S1 of the method according to the invention involves transferring a production dataset P.sub.f of a first rolled material 2 to the physical production model 12 as an input variable. The physical production model 12 and the metallurgical model parameters ? are used to ascertain a first mechanical dataset C.sub.f, a further production dataset F.sub.f and a metallurgical dataset K.sub.f for the first rolled material 2 from said input variable. For this purpose, the central computation unit 28 solves physical equations (for example comprising a thermal conduction equation and equations for the formation of metallurgical phases according to the aforementioned equations (1) to (4)) with applicable boundary conditions (for example according to equation (5)). According to the present exemplary embodiment, the further production dataset F.sub.f comprises temperatures and cooling rates of the first rolled material 2 during production thereof in the hot rolling mill 1, while the metallurgical dataset K.sub.f comprises phase fractions p.sub.k and/or morphological characteristics a.sub.k of the first rolled material 2 during production thereof in the hot rolling mill 1.

[0144] Furthermore, the central computation unit 28 ascertains the mechanical properties of the first rolled material 2 in the form of the first mechanical dataset C.sub.f from the production dataset P.sub.f of the first rolled material 2, from the metallurgical model parameters ? and by means of empirical metallurgical equations (for example according to the aforementioned equations (6) to (9)). According to the exemplary embodiment, the first mechanical dataset C.sub.f comprises a tensile strength Y.sub.s, a yield strength T.sub.s and an elongation at break A.sub.1 and is stored for the fourth step S4 of the method according to the invention, while the further production dataset F.sub.f and the metallurgical dataset K.sub.f are supplied to the statistical data model 14 as additional input variables.

[0145] The second step S2 involves ascertaining the second mechanical dataset S.sub.f of the first rolled material 2 from the production dataset P.sub.f, the further production dataset F.sub.f and the metallurgical dataset K.sub.f of the first rolled material 2 using the trained and validated statistical data model 14 and storing it for the fourth step S4 of the method according to the invention. In this case, according to the exemplary embodiment, the second mechanical dataset S.sub.f again comprises a tensile strength Y.sub.s, a yield strength T.sub.s and an elongation at break A.sub.1 of the first rolled material 2.

[0146] The third step S3 involves determining for the production dataset P.sub.f of the first rolled material 2 an averaged normalized distance value d.sub.f,m in relation to n production datasets P.sub.i (symbolized in FIG. 2 by set braces with index n) of the further rolled materials 2, this being shown schematically in FIG. 3.

[0147] The fourth step S4 involves using the averaged normalized distance value d.sub.f,m to ascertain the mechanical properties J.sub.f of the first rolled material 2 as a weighted average from the first and second mechanical datasets C.sub.f, S.sub.f of the first rolled material 2. The weighting is performed on a component-by-component basis, i.e. separately for each mechanical property of the first and second mechanical datasets C.sub.f and S.sub.f, the quotient C.sub.f,j/S.sub.f,j of the respective mechanical property of the second and first mechanical datasets S.sub.f, C.sub.f being weighted in proportion to the averaged normalized distance value d.sub.f,m. This means that the lower the value of d.sub.f,m, the stronger the weight of the relevant mechanical property from the second mechanical dataset S.sub.f in relation to that from the first mechanical dataset C.sub.f. Weighting is carried out on the basis of the aforementioned formula (10), a suitable weighting function w(d.sub.f,m) such as that of formula (11) being used.

[0148] FIG. 3 uses a simplified statistical data model 14, the input variables (corresponding to the production datasets P.sub.f and P.sub.i, the further production datasets F.sub.f and F.sub.i and the metallurgical datasets K.sub.f and K.sub.i of the first rolled material and the further rolled materials 2 and 2) of which comprise only two standardized production parameters z.sub.1 and z.sub.2 in total, to schematically show the position of a standardized production dataset P.sub.f of a first rolled material 2 in relation to four standardized production datasets P.sub.i of further rolled materials 2. In a realistic application, the number M of input variables (corresponding to the production datasets P.sub.f and P.sub.i, the further production datasets F.sub.f and F.sub.i and the metallurgical datasets K.sub.f and K.sub.i of the first rolled material and the further rolled materials 2 and 2) of the statistical data model 14 is greater than 2, for example 20 or 50. A representation of a realistic statistical data model 14 corresponding to FIG. 3 would accordingly comprise a mathematical map of an M-dimensional manifold to a curved M-dimensional manifold.

[0149] The mechanical property S.sub.f,j shown in FIG. 3 is a value (index j) from the second mechanical dataset S.sub.f. The production parameters of the first rolled material 2 and the further rolled materials 2 are converted to dimensionless and standardized values z.sub.1 and z.sub.2 in accordance with the aforementioned formulae (13a) and (13b), for which the smallest and largest absolute values x.sub.1,min, x.sub.2,min, x.sub.1,max, x.sub.2,max of each of the production parameters of all further rolled materials 2 are initially ascertained.

[0150] In FIG. 3, the two standardized production parameters z.sub.1 and z.sub.2 define a horizontal mathematical plane e.sub.1 in which the position of the standardized production parameters P.sub.f of the first rolled material (star) and of production parameters P.sub.i of four further rolled materials 2 (vacant circles) is identified. From the relative position of the standardized production parameters in the plane e.sub.1 in relation to each other it is possible for example to use formula (12) to ascertain a normalized distance value d.sub.f,i of the production dataset P.sub.f of the first rolled material 2 in relation to the respective production dataset P.sub.i of one of the further rolled materials 2, the value M=2 needing to be selected on the basis of the merely two input parameters z1 and z2 in formula (12) for the simplified example of FIG. 3. The connecting lines drawn in FIG. 3 between the standardized production parameters P.sub.f of the first rolled material 2 and the standardized production parameters P.sub.i of the further rolled materials 2 are an alternative to formula (12) in the form of the Euclidean norm within the plane e.sub.1 for computation of a normalized distance value d.sub.f,i of the production dataset P.sub.f of the first rolled material 2 in relation to the respective production dataset P.sub.i of a further rolled material 2.

[0151] The computation of the normalized distance values d.sub.f,i is followed by arithmetic averaging to ascertain the averaged normalized distance value d.sub.f,m of the production dataset P.sub.f of the first rolled material 2 in relation to the production datasets P.sub.i of the further rolled materials 2 from the n smallest normalized distance values d.sub.f,i, the number n comprising 0.1 to 0.5 percent of the total number of further rolled materials 2.

[0152] FIG. 3 furthermore shows, along the vertical axis, which runs normally to the plane e.sub.1, the respective value, for each of the four sampling datasets P.sub.i shown for the further rolled materials 2, of a mechanical property with index j from the associated sampling dataset B.sub.i as a solid circle. These data points from the sampling datasets B.sub.i approximate a further, generally curved plane e.sub.2, which is indicated in FIG. 3 by curved lines. In this context, approximate means that the data points from the sampling datasets B.sub.i are not necessarily situated exactly on the plane e.sub.2, but may also be positioned at a short local normal distance therefrom, because the mathematical plane e.sub.2 in the specific example of FIG. 3 is a regression plane. Within the framework of the predictive accuracy of the statistical data model 14, the regression plane e.sub.2 represents an approximation to the data points of the sampling datasets B.sub.i of the further rolled materials 2, which have been used to train the statistical data model 14.

[0153] Along the vertical axis, FIG. 3 shows the value S.sub.f,j predicted by the statistical data model 14 for the mechanical property from the first mechanical dataset S.sub.f of the first rolled material 2 above the standardized production dataset P.sub.f of the first rolled material 2: this computed value replaces a physical sampling process for the first rolled material 2 and is usually close to the plane e.sub.2this means that if the statistical data model 14 has a one hundred percent predictive accuracy, the data point S.sub.f,j is also an element of the plane e.sub.2. However, since only a finite number of data of further rolled materials 2 are available for training the statistical data model 14 in the considered range of the production parameters z.sub.1 and z.sub.2, the value S.sub.f,j predicted by the statistical data model isdepending on the averaged normalized distance value d.sub.f,mlocated at a specific local normal distance from the plane e.sub.2. In other words: in heavily populated ranges of the training data of the statistical data model 14 (resulting in a small value of d.sub.f,m for a first rolled material 2), in which a high number of training datasets is present, the value S.sub.f,j predicted by the statistical data model 14 is usually closer to the plane e.sub.2 approximated by the body of the sampling data B.sub.i than in sparsely populated ranges of training data (large value of d.sub.f,m).

[0154] FIG. 4 shows a data flow diagram for the method according to the invention for creating a hybrid model 10. The method according to the invention comprises three sequential steps S1 to S3i.e. steps to be carried out one after the other. Analogously to FIG. 2, the hybrid model 10 is shown using a broad line thickness and comprises production datasets P.sub.i of further rolled materials 2, a physical production model 12 with a set of metallurgical model parameters ? and a statistical data model 14. The further rolled materials 2 and rolled materials 2 form a first set I.sub.1 and a second set I.sub.2, respectively, and are each shown in FIG. 4 by a circle symbol with set braces. As in FIG. 2, the sequential steps of the method are represented by horizontal dashed lines: solid arrows again symbolize applicable data flows.

[0155] The rolled materials of the first and second quantities I.sub.1, I.sub.2 are each produced in the hot rolling mill 1, for which the hybrid model 10 is created, and then physically sampled, with the result that corresponding production data P.sub.i and sampling data B.sub.i are available. These data have been transmitted, for example as shown in FIG. 1, to the central evaluation unit 28, in which the hybrid model 10 is also created.

[0156] Also as in FIG. 2, in the present exemplary embodiment for creating a hybrid model 10 internal data structures, internal computation rules, relations between data and internal parameters are each realized as data or programming structures in the central computation unit 28, and so the method according to the invention for creating a hybrid model 10 can again be carried out as a program by the central computation unit 28.

[0157] The first step S1 involves at least one pass through the physical production model 12 taking place: this results in the first mechanical dataset, the further production dataset and the metallurgical dataset C.sub.i, F.sub.i, and K.sub.i being ascertained for each of the production datasets P.sub.i of the further rolled materials 2.

[0158] Each ascertainment of these datasets is followed by a first measurement function ?.sub.1 being used for component-by-component ascertainment, i.e. ascertainment for all individual mechanical properties in the further mechanical datasets C.sub.i, of a measure of the deviation of the ascertained first mechanical datasets C.sub.i from the sampling datasets B.sub.i of the further rolled materials 2, the first measurement function ?.sub.1 used, according to the present exemplary embodiment, being the expression of the aforementioned formulae (16) or (17). If the value of the first measurement function (?.sub.1) ascertained in this process for at least one mechanical property in the first mechanical datasets C.sub.i of the further rolled materials 2 is greater than a specified limit ?.sub.max, the physical production model 12 is adapted by varying its metallurgical model parameters ?: this is shown in FIG. 4 by the dashed arrow of the first program branch identified by 1, the arrow tip of which protrudes into the rectangular symbol of the physical production model 12.

[0159] Another pass through the physical production model 12 then takes place, this also being referred to as a further iteration. The first mechanical, further production and metallurgical datasets C.sub.i, F.sub.i and K.sub.i are re-ascertained by solving the applicable equations of the physical production model 12but in this case with the modified metallurgical model parameters ?and then the first measurement function ?.sub.1 is again used to perform a check on the deviation of the values of the newly ascertained first mechanical datasets C.sub.i from the corresponding values of the sampling datasets B.sub.i.

[0160] If, after a pass through the physical production model 12, the value of the first measurement function ?.sub.1 for all mechanical properties C.sub.i,j (j denotes a single value within a mechanical dataset Ci) in the first mechanical datasets C.sub.i of the further rolled materials 2 is less than or equal to the limit ?.sub.max, the production datasets P.sub.i of the further rolled materials 2, together with the corresponding further production and metallurgical datasets F.sub.i, K.sub.i that were ascertained in the most recently performed iteration, are transferred to the statistical data model 14 as input variables and the corresponding sampling datasets B.sub.i are transferred to the statistical data model 14 as target variables. This case is represented in FIG. 4 by the arrows of the first program branch identified by 0.

[0161] The second step S2 involves training the statistical data model (14) on the basis of the transferred input and target variables using a first machine learning method V.sub.1 (shown in FIG. 4 by a square with rounded corners). Machine learning methods are known to those skilled in the art. Internal parameters of the statistical data model 14 are defined as appropriate: the statistical data model trained in this way is able to take a set of input parameters (for example the production parameters and further production parameters of a rolled material) and predict the corresponding mechanical properties without a sampling process. In the analogous example of FIG. 3, this means that the trained statistical data model 14 is able to map the plane e.sub.1 (corresponding to the standardized production parameters z.sub.1, z.sub.2) to thegenerally curvedplane e.sub.2 for each mechanical property.

[0162] The subsequent third step S3 involves validating the trained statistical data model 14 on the basis of production datasets P.sub.i and sampling datasets B.sub.i of rolled materials 2 from a second set I.sub.2. The second set I.sub.2 is disjunct from the first set I.sub.1, which means that none of the rolled materials whose production and sampling data P.sub.i, B.sub.i have been used to train the statistical data model 14 are used for validating the trained statistical data model 14.

[0163] In the present exemplary embodiment of the method according to the invention, for validation, the production datasets P.sub.i of the rolled materials of the second set I.sub.2 are supplied to the physical production model 12 with the metallurgical model parameters y that have been used in the most recently performed iteration: this is shown in FIG. 4 by the dashed arrow to the physical production model 12, which is also shown in dashed form, in the third method step S3. A single pass through the physical production model 12 is used to ascertain the further production datasets F.sub.i and the metallurgical datasets K.sub.i for the rolled materials of the second set I.sub.2, and these are then supplied together with the corresponding production datasets P.sub.i to the trained statistical data model 14 as input variables. Subsequently, the statistical data model 14 ascertains corresponding second mechanical datasets S.sub.i for the rolled materials of the second set I.sub.2 from said input variables.

[0164] A second measurement function ?.sub.2, which is provided in the present exemplary embodiment by the aforementioned formula (18), is used to ascertain a measure of a deviation of the individual variables in the mechanical datasets S.sub.i from the respective corresponding variables in the sampling datasets B.sub.i of the rolled materials of the second set I.sub.2. If each value ascertained on the basis of the second measurement function ?.sub.2 is less than or equal to a specified limit ?.sub.max, the statistical data model 14 with the internal parameters determined during training is transferred to the hybrid model 10 (shown in FIG. 4 by the arrow of the second program branch identified by 0). The limit ?.sub.max may be the same value as was used for adapting the physical production model in the first method step S1.

[0165] If one of the values ascertained on the basis of the second measurement function ?.sub.2 is greater than the specified limit ?.sub.max, the statistical data model is adapted further by, in the present exemplary embodiment of the method according to the invention, training the statistical data model 14 (or each partial data model) on the basis of a second machine learning method V.sub.2, which is different from the first machine learning method V.sub.1, and the second method step S2 is repeated: this case is indicated in FIG. 4 by the dashed arrow of the second program branch identified by 1. In contrast to the aforementioned first machine learning method, the invention involves the second machine learning method being employed only in the event of unsuccessful validation of the statistical data model 14 (or one of the partial data models).

[0166] Alternatively or in addition, in the event of unsuccessful validation (i.e. ?.sub.2??.sub.max), the first and second quantities I.sub.1, I.sub.2 of rolled materials whose production and sampling data P.sub.i, B.sub.i are used to train and validate the statistical data model 14 can also be modified (not shown in FIG. 4). In turn, one and the same second machine learning method V.sub.2 can be used for validating each of the individual partial data models, said second machine learning method merely being different from the first machine learning method V.sub.1 of the respective partial data model. Alternatively, the second machine learning methods V.sub.2 for validating the individual partial data models may also differ from each other.

LIST OF REFERENCE SIGNS

[0167] 1 rolling mill, hot rolling mill [0168] 2, 2, 2 rolled material [0169] 3 heating furnace [0170] 4 roughing line [0171] 5 cropping shear [0172] 6 temperature measuring device, pyrometer [0173] 7 finishing rolling line [0174] 8 cooling section [0175] 9 sampling device [0176] 10 hybrid model [0177] 12 physical production model [0178] 14 statistical data model [0179] 20, 20, 20 rolled material section [0180] 26 coiling device [0181] 27 installation controller [0182] 28 central evaluation unit [0183] a.sub.k morphological characteristic [0184] A.sub.1 . . . A.sub.13 rolling mill section [0185] B.sub.i sampling dataset [0186] B.sub.i,j value from the sampling dataset [0187] C.sub.f, C.sub.i first mechanical dataset [0188] C.sub.f,j, C.sub.i,j individual value in first mechanical dataset [0189] d.sub.f,i normalized distance value [0190] d.sub.f,m averaged normalized distance value [0191] ?.sub.j function parameters [0192] e.sub.1, e.sub.2 plane [0193] ?.sub.1, ?.sub.2 measurement function [0194] ?.sub.max limit [0195] F.sub.f, F.sub.i further production dataset [0196] ? metallurgical model parameters [0197] I.sub.1, I.sub.2 set [0198] J.sub.f mechanical properties [0199] J.sub.f,j single mechanical property [0200] K.sub.f, K.sub.i metallurgical dataset [0201] ?.sub.j function parameters [0202] p.sub.k phase fraction [0203] P.sub.f, P.sub.i production dataset [0204] q.sub.z cooling rate [0205] Q.sub.1 . . . Q.sub.4 cooling device [0206] R.sub.1 . . . R.sub.5 roll stand [0207] S.sub.1 . . . S.sub.5 method step [0208] S1. . . S3 method step [0209] S.sub.f, S.sub.i second mechanical dataset [0210] S.sub.f,j value from the second mechanical dataset [0211] T.sub.z temperature value [0212] V.sub.1, V.sub.2 machine learning method [0213] w.sub.i weighting factor [0214] w.sub.j(d.sub.f,m) weighting function [0215] x.sub.k,f, x.sub.k,i component in production dataset [0216] x.sub.k,min minimum value for component k [0217] x.sub.k,max maximum value for component [0218] z.sub.k,f, z.sub.k,i standardized value in production dataset