Determining target values in metal strip production
By combining recurrent neural networks and feedforward neural networks in a machine learning model, the production parameters of the rolling mill are adjusted, solving the problem that the existing technology failed to consider the equipment status and history, and achieving more accurate production parameter prediction and metal strip quality control.
Patent Information
- Application Number
- CN202580026267.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-04-30
- Filing Date
- 2025-04-09
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies fail to adequately consider the current equipment status and production history in rolling mills, resulting in inaccurate production parameter predictions. In particular, when equipment is changed or new materials are used, the average values of lookup tables and model outputs cannot adapt to real-time changes in the equipment.
A machine learning model combining recurrent neural networks and feedforward neural networks is adopted. Material parameters are adjusted by correction factors, and physical production parameter models are combined to determine the target values of production parameters for rolling equipment, taking into account the current status and historical information of the equipment.
It improves the accuracy of production parameter prediction, reduces average error, and can adapt to the current equipment status, ensuring that the quality of the metal strip is closer to the target specifications.
Smart Images

Figure CN123125553A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and apparatus for determining target values of production parameters for a rolling mill having at least one set of rolling mill stands for rolling metal products into metal strips, and a rolling mill having such apparatus, a computer program product, and a machine learning model. Background Technology
[0002] A rolling mill system comprises one or more mill stands, which are used to roll primary metal products, such as freshly cast billets, slabs, or pre-rolled metal strips, into final products. To produce new strips—strips with new characteristics (such as altered thickness, width, surface structure, and / or microstructure and / or similar characteristics compared to the previous strip)—the production or operating parameters of the equipment are adjusted during continuous operation. However, this first requires determining the corresponding target values for these parameters.
[0003] For example, in order to roll a new strip, it is first necessary to determine target values for important variables such as rolling force, torque, roll bending, and roll axial displacement. These predetermined values have a decisive impact on the subsequent product quality because the variables affect, for example, the thickness, profile, and / or straightness of the strip to be rolled. Furthermore, other operating parameters, such as mass flow rate settings, can be set based on these important variables.
[0004] To predetermine production parameters, lookup tables are typically used. In these tables, belt parameters characterizing the desired features of the belt to be produced are associated with corresponding production parameters, for example, those gathered through experience. Alternatively, physical models are also used, incorporating the physical relationships between belt parameters and production parameters.
[0005] Another approach is based on the use of so-called "artificial intelligence." The trained, machine-learned model is then able to output corresponding production parameters based on the provided parameters. Particularly advantageous is that this approach also incorporates the use of physical models. For example, in "Current and future development in neural computation in steel processing" by Schlang et al., *Control Engineering Practice*, Vol. 9 (2001), pp. 975-986, an architecture is described in which correction factors incorporated into the analytical model are calculated via two separate neural networks.
[0006] However, all these solutions to date have drawbacks: they typically do not consider, or only insufficiently consider, the current equipment condition, or even direct production or equipment history. When changes are made at the rolling mill or new materials are used, the lookup tables or models used can quickly become unavailable. Furthermore, such lookup tables and models can only output "average" target values for production parameters because equipment condition, i.e., production conditions, constantly changes slightly in practice, for example, through varying physical environmental conditions, wear, and / or similar factors.
[0007] For example, a so-called feedforward network can be used as a machine learning model, treating each new strip to be produced as independent of previously produced strips. Therefore, the predictions of such a network are always exclusively based on the average of previously collected experience, hence the term "long-term inheritance." In practice, however, in new strips to be rolled that are similar to the previous one, the predictions of this network are ignored, and instead, the target values of the production parameters corresponding to the previous strip are used. This is called "short-term inheritance." Summary of the Invention
[0008] The purpose of this invention is to further improve the determination of target values for production parameters of casting and rolling equipment, especially to consider the current equipment status in addition to the static relationship between the parameters and production parameters, and even the direct equipment history or production history when necessary.
[0009] The objective is achieved by the method and apparatus according to the independent claims, and rolling equipment having such apparatus, computer program products, and machine learning models.
[0010] The preferred embodiments are the subject of the dependent claims and the description below.
[0011] According to the first aspect, in a method, particularly a computer-implemented method, for determining target values of selected production parameters for a rolling mill having at least one set of rolling mill stands for rolling metal products into metal strips, values of material parameters for a predetermined reference material are provided. Furthermore, based on production parameters known in advance during metal strip production, a correction factor for the material parameters is determined by a machine learning model based on a recurrent neural network. Subsequently, target values of the selected production parameters for the rolling mill can be determined based on a physical production parameter model, incorporating the material parameter values adjusted by the correction factor and the target specifications for the metal strip to be rolled into the physical production parameter model.
[0012] The rolling mill in the sense of this invention can be formed by at least one group of rolling mill stands having at least one mill stand. However, the rolling mill can also include other equipment components, such as heating devices for (re)heating the metal product to be rolled, descaling devices for descaling the metal product, vertical roll mills for profile shaping, cooling zones for cooling the rolled metal strip, one or more separating devices for separating the metal strip, one or more coilers for coiling the metal strip, and / or similar components. The rolling mill can be particularly part of a casting and rolling mill, but alternatively, it can be separate from the casting machine. The rolling mill can be designed for hot rolling or cold rolling.
[0013] In the context of this invention, determining the target value of the selected production parameter is preferably a pre-calculation or prediction (i.e., estimation) of the target value.
[0014] In the context of this invention, the material parameters are preferably material-related parameters, such as density, thermal conductivity, elasticity, yield stress, and / or similar parameters.
[0015] In the context of this invention, a machine learning model is preferably a statistical model generated by one or more algorithms based on training data. A machine learning model can also be referred to as a trained model. A machine learning model is suitably the result of machine learning. Such a machine learning model can also be commonly referred to as artificial intelligence, which has identified patterns and regularities in the training data and is thus able to evaluate or assess unknown data based on these patterns and regularities. The machine learning models described above and below are based on so-called "neural networks".
[0016] In the context of this invention, the prior known production parameters are preferably known values of production parameters relevant to metal strip production, particularly measurable or (pre-)set values of production parameters in metal strip production. For example, such prior known production parameters could be the cross-sectional dimensions of the metal product to be rolled, the temperature of the metal product to be rolled, the chemical composition of the metal product to be rolled, the roll diameter, the conveying speed of the metal product through or into the equipment, and / or similar parameters. Therefore, prior known production parameters can also be understood as prior known (e.g., set or measured) production parameter values.
[0017] In the context of this invention, the production parameter model is preferably a model of (particularly selected) production parameters. In the production parameter model, target specifications (i.e., desired parameters) are suitably associated with production parameters. For example, the production parameter model can be based on the physical relationship between the target specifications and the production parameters. Such a model can be analytically or numerically solvable.
[0018] One aspect of the invention is based on the following approach: using a combination of machine learning models and physical models to determine target values for selected production parameters of a rolling mill. Here, the starting point for data processing in this combination is preferably providing material parameter values for a predetermined reference material. These material parameter values can be determined or preset, for example, through simulation or calculation. However, the provided material parameter values cannot be directly used to calculate the target values. Instead, they must be adapted to the actual existing conditions. To this end, the machine learning model appropriately determines correction factors for the material parameters. Here, the machine learning model preferably starts from production parameters that are already known in advance, such as preset production parameters, such as the reduction rate determined according to the pass schedule, or the chemical composition of the metal product to be rolled, or measurable production parameters, such as the temperature of the metal product to be rolled. However, advantageously, the machine learning model also considers at least the current equipment state or operating state, or, if necessary, the production conditions of a predetermined number of previously produced strips. That is, when determining the correction factors, the machine learning model can at least include the current production conditions, which typically vary on a short timescale, and, if necessary, the production conditions of a predetermined number of previously produced strips.
[0019] The current operating state, equipment condition, or production conditions, at least in practice, are not directly and comprehensively measurable. In other words, the actual operating state, equipment condition, or current production conditions cannot be fully mapped through (adjustable or measurable) equipment or production parameters. This is because numerous variables influence the production of metal strip, and ultimately, not all of these variables can be controlled and / or detected. For example, environmental conditions such as ambient temperature, air humidity, and / or similar factors can play a role. Simultaneously, wear, such as wear on rolls, bearings, actuators, and / or similar components, can also have an impact.
[0020] However, the current operating state, equipment state, or current production conditions can be detected by a dedicated machine learning model or neural network. The model according to the invention can be, for example, adaptive, i.e., adapted to changing operating states. For this purpose, the machine learning model is suitably based on a recurrent neural network (RNN). This RNN possesses not only general, time-independent knowledge (i.e., so-called "long-term memory") about different materials and (physical) relationships during strip production in a casting and rolling mill, but also knowledge about the current equipment state (i.e., so-called "short-term memory"). With the aid of the RNN, the long-term and short-term memories can be combined in a very ingenious way. Thus, a near-optimal correction factor—and consequently a near-optimal target value for selected production parameters—can be determined, with at least a portion of the equipment history incorporated into the correction factor. Because of this machine learning model, when determining (preferably equipment-specific or even mill stand-specific) correction factors, it is possible to consider which correction factors were used to determine the target value of the prior strip without explicitly backtracking to earlier correction factors, e.g., without using these correction factors as is when producing similar strips.
[0021] The target values for the selected production parameters can then be ultimately determined through production parameter modeling, where material parameter values corrected by correction factors are appropriately used as input variables for the production parameter model. Here, compared to conventional methods, such as predictions using lookup tables, the average error of the target value prediction is particularly small, and it automatically adapts to the current equipment state.
[0022] Preferred embodiments and improvements thereof are described below. Unless explicitly excluded, these embodiments can be arbitrarily combined with each other and with aspects of the invention described below.
[0023] The combination of two machine learning models enables exceptional flexibility in correcting material parameter values. To this end, an additional machine learning model based on a feedforward neural network determines further correction factors for the material parameters based on material feature data characterizing the state of the metal to be processed in the rolling mill. This additional machine learning model (hereafter occasionally referred to as the first machine learning model) thus suitably takes into account the material actually used in strip production and its characteristics, such as its chemical composition and / or temperature. It is suitably configured or trained specifically for the material. It can be used to consider or determine the hardness of the material currently in use. Conversely, the machine learning model initially discussed (hereafter occasionally referred to as the second machine learning model) takes into account the equipment configuration. It is suitably configured or trained specifically for the equipment.
[0024] The correction of the provided material parameter values is divided into material-specific correction and device-specific correction, allowing the use of dedicated machine learning models or neural networks. The first model can be, for example, static, while the second model—as already mentioned—is adaptive, i.e., capable of adapting to changing operating conditions. Therefore, the first machine learning model is suitably based on a feedforward neural network (FNN). Thus, for example, when a new device is put into operation, it is useful to use a different (first) machine learning model, because it no longer needs to be adapted to or trained on that device. Instead, the different machine learning model can be used across devices.
[0025] When using a machine learning model based on a recurrent neural network, earlier device states can influence the determination of correction factors over a longer period of time if necessary. Therefore, it is preferred that the (first) machine learning model is based on a recurrent neural network with controllable short-term memory. In the context of this invention, a recurrent neural network with controllable short-term memory is preferably a neural network with so-called LSTM modules or LSTM units (LSTM = "Long Short-Term Memory"). Therefore, a recurrent neural network with controllable short-term memory can also be called an LSTM network. Such a network has nodes or units in its hidden layers with input gates, forget gates, and output gates, which control the signal processing in their respective units and act as filters. These gates in the LSTM network control what is "remembered" and what is "forgotten".
[0026] When using machine learning models based on recurrent neural networks with controllable short-term memory, it is possible to limit the influence of earlier equipment states, especially very old ones, on the determination of the second correction factor. This allows for "resets," such as when changing rolls, during extended downtime of casting and rolling equipment, and / or similar situations.
[0027] One possibility for automatically and cost-effectively building a short-term memory for a second machine learning model lies in continuously determining correction factors during the sequential production of metal strip using the machine learning model. Here, continuous determination can be understood as using the second machine learning model during the sequential production of particularly different metal strips without resetting the model. Therefore, the short-term memory can be continuously and progressively built up during the regular operation of the rolling mill. Consequently, earlier operating states of the rolling mill can influence each subsequently determined correction factor—and thus, the determined target value.
[0028] Alternatively, it is advantageous to selectively activate or rebuild the short-term memory of the LSTM network for each strip to be produced. For this purpose, it is preferable to store pre-known production parameters or production parameter values for each hot-rolled strip produced during the operation of the rolling mill. The machine learning model can then iteratively determine correction factors for the provided material parameter values, starting from different production parameters in the stored pre-known production parameters, before a new strip begins production. Specifically, the correction factor can be iteratively determined before being incorporated into the physical model from the current pre-known production parameters. The iteratively determined correction factor is appropriately discarded here, i.e., it is not used directly as the basis for determining the target value under any circumstances. In this way, individual operating states can be selectively ignored. Therefore, this approach achieves enhanced control via the short-term memory of the LSTM network.
[0029] In the context of this invention, the commencement of production of a new metal strip is preferably the commencement of production of a metal strip with new target specifications and correspondingly changed features, i.e., changed strip parameters.
[0030] The method according to the first aspect of the invention has particular advantage in determining the target values of rolling parameters for a group of mill stands or individual mill stands in a rolling mill. This is because, as already mentioned at the beginning, the quality of the produced metal strip is particularly relevant to these rolling parameters; the particularly precise determination of the required rolling force (where the static or only slowly varying relationship between strip parameters and rolling force can be taken into account, and the constantly occurring changes in operating conditions or equipment conditions (e.g., continuous wear of rolls, roll temperature fluctuations, and / or similar conditions)) is not only particularly important for achieving the desired metal strip thickness, but also for setting other parameters (such as mass flow rate) or calculating the profile and / or straightness.
[0031] Therefore, it is preferable that the selected production parameters are rolling parameters, particularly rolling force, roll torque, roll bending and / or roll axial displacement.
[0032] Since the thickness reduction of a metal (primary) product in one or more mill stands of a rolling mill is decisively related to the deformability of the metal being processed, it is further preferred that the material parameter whose value is provided is the yield stress. Here, the yield stress is the stress at which the material begins to deform plastically (rather than elastically). If the selected production parameters are rolling parameters, particularly the rolling force, then the yield stress can therefore be a material parameter that is particularly important for determining the target value.
[0033] Alternatively, this method can also be used to determine target values for other selected production parameters, such as profile, straightness, width, and / or compression in the roughing runner.
[0034] To further improve the long-term memory of the (second) machine learning model, after producing the metal strip using target values determined for the production parameters, it is preferable to determine the strip parameters of the produced metal strip. For example, these strip parameters can be detected by simulation or sensors. Suitably, the machine learning model is then adapted using the determined strip parameters. In other words, it is preferable that the recurrent network (with controllable short-term memory if necessary) is readjusted after each produced strip. By improving the model's long-term memory, a more reliable determination of the correction factors for the selected production parameters, and thus the target values, becomes possible. In normal operation, this allows the production of metal strips whose strip parameters are closer to the target specifications.
[0035] According to a second aspect of the invention, an apparatus for determining target values of selected production parameters for a rolling mill having at least one set of rolling mill stands for rolling metal products into metal strips includes a providing module designed to provide values of material parameters for a predetermined reference material. A second machine learning model based on a recurrent neural network is designed to determine correction factors for the material parameters based on pre-known production parameters during metal strip production. Finally, a calculation module is also provided, designed to determine the target values of the production parameters for the rolling mill based on a physical production parameter model, incorporating the material parameter values adjusted by the correction factors and the target specifications for the metal strip to be rolled into the physical production parameter model.
[0036] Because the (second) machine learning model is based on a recurrent neural network, it is able to specifically include the current operating state of the current production strip, and, if necessary, a predetermined number of directly prior operating states, in the production planning of the next strip. In other words, the (second) machine learning model thus possesses "short-term memory," which allows for significant improvements in the prediction of correction factors and, consequently, the prediction of target values.
[0037] The module in the sense of this invention can be configured in terms of hardware and / or software. The module can, in particular, have at least one processing unit preferably connected to a memory system and / or a bus system via data or signal connections. For example, the module can have a microprocessor unit (CPU) or a module thereof and / or one or more programs or program modules. The module can be configured to: process instructions implemented as a program stored in a memory system, detect input signals from a data bus, and / or output output signals to a data bus. The memory system can have one or more (particularly different) storage media, particularly optical, magnetic, solid-state, and / or other non-volatile media. The program can be constructed such that it embodies or at least a portion thereof of the methods described herein or is capable of performing or at least a portion thereof, enabling the apparatus according to the second aspect of the invention to perform the steps of this method, thereby particularly enabling the determination of target values for the production parameters of the rolling mill.
[0038] According to a third aspect of the invention, a rolling mill for producing metal strip comprises: at least one mill stand assembly for rolling metal products into metal strip; means according to a second aspect of the invention; and a control device designed to control the rolling mill based on target values of selected production parameters determined by the device. Using such a rolling mill, metal strip can be produced from metal (primary) products, such as, for example, freshly cast metal billets, slabs, or pre-rolled intermediate strips, the strip parameters of which are particularly close to target specifications. This is because, with the aid of a (second) machine learning model based on a recurrent neural network, in addition to the long-term stable relationship between strip parameters and production parameters, current or even other recently existing, continuously changing operating conditions in practice can be taken into account.
[0039] The control device can be configured in hardware and / or software. Specifically, the control device can have at least one processing unit, preferably connected to a memory system and / or a bus system via data or signal connections. For example, the control device can have a microprocessor unit (CPU) or a module thereof and / or one or more programs or program modules. The control device can be configured to: process instructions executed as programs stored in a memory system, detect input signals from a data bus, and / or output output signals to a data bus.
[0040] According to a fourth aspect of the invention, a computer program product is configured to perform the method according to the first aspect of the invention. Such a computer program product can be executed, for example, within the framework of process automation in a rolling mill. The computer program product suitably includes instructions that, when executed by a computer, for example within the framework of process automation in a rolling mill, cause the computer to perform the method according to the first aspect of the invention.
[0041] According to a fifth aspect of the invention, the machine learning model is based on a recurrent neural network and provides correction factors designed to determine material parameters based on prior known production parameters during metal strip production. The machine learning model is suitably adapted for use as a (second) machine learning model in the method according to the first aspect of the invention and / or in the apparatus according to the second aspect of the invention. Preferably, the machine learning model can be obtained by training the recurrent neural network based on training data containing a large number of different prior known production parameters or production parameter values and correction factors associated with said production parameters or production parameter values.
[0042] Preferred embodiments and improvements thereof are described below. Unless explicitly excluded, these embodiments can be combined arbitrarily with each other and with aspects of the invention described below. Attached Figure Description
[0043] The above-described features, characteristics, and advantages of the present invention, and the ways in which they are implemented, will become clearer and more apparent from the following description of embodiments, which are explained in detail with reference to the accompanying drawings. Hereinafter: Figure 1 Examples of methods and apparatus for determining target values of selected production parameters for rolling mills are shown; Figure 2 Examples of recurrent neural networks are shown; and Figure 3 An example of rolling equipment is shown. Detailed Implementation
[0044] Where appropriate, the same reference numerals are used in the accompanying drawings for the same or corresponding elements of the invention.
[0045] Figure 1 An example of a method 100 and apparatus 10 for determining a target value S of a selected production parameter P of a rolling mill that can produce a metal strip 14 is shown.
[0046] In method step S1, a value W of material parameter M is provided for a predetermined reference material. The material parameter value W can be determined, for example, through simulation, analytical calculation, or numerical calculation. When determining the material parameter value W, other material properties or processing variables E of the reference material are suitably considered, such as its degree of deformation, its deformation rate, temperature, and / or similar variables, which are known in advance, for example, through measurement. For example, the material parameter value W can be provided by the providing module 20 of device 10.
[0047] Before the material parameter value W of the reference material can be used as the basis for determining the target value S, the material parameter value must be adjusted.
[0048] Therefore, a first correction factor c1 is determined in method step S2. The first correction factor c1 suitably corrects the material parameter M in terms of the material to be actually processed, particularly its hardness. For this purpose, the apparatus 10 has a first machine learning model 30 based on a feedforward neural network 32. The first machine learning model 30 is suitably designed to determine the first correction factor c1 of the material parameter M based on material feature data K characterizing the state of the metal to be processed in the rolling mill. The corresponding settings of the first machine learning model 30 are suitably performed via corresponding training of the feedforward neural network 32. Therefore, the first machine learning model 30 or the correspondingly trained feedforward neural network 32 is material-specific.
[0049] Material characteristic data K can, for example, characterize the chemical composition, temperature, and / or similar parameters of the metal to be processed. Suitably, the temperature to be used as input to the first machine learning model 30 is the temperature of the metal product 12 in the region where the equipment components controlled using the target value S also act. If the selected production parameter P for determining its target value S is the rolling force in the mill stand of the rolling equipment, as... Figure 1 As purely exemplarily shown in the example, it is preferable to use the temperature of, for example, the metal billet or the slab separated from it or the pre-rolled intermediate strip as it passes through the mill stand as the basis for determining the first correction factor c1.
[0050] However, the first correction factor c1 is insufficient to map the effect of the rolling mill, i.e., its components, on the processing of the metal product 12 in the rolling mill. Therefore, a second correction factor c2 is determined in a further method step S3. The second correction factor c2 suitably corrects the material parameter M in terms of the current operating state of the equipment, i.e., particularly in terms of the current operating parameters or production parameters. For this purpose, the device 10 has a second machine learning model 40 based on a recurrent neural network 42 (“recurrent neural network”). The second machine learning model 40 is suitably designed to be used for the pre-known production parameters P during metal strip production. The second correction factor c2 for the material parameter M is determined based on known values of some relevant production parameters in metal strip production. The corresponding settings of the second machine learning model 40 are suitably implemented via corresponding training of the recurrent neural network 42. Therefore, the second machine learning model 40 or the correspondingly trained recurrent neural network 42 is equipment-specific, particularly rack-specific.
[0051] Pre-known production parameters or production parameter values P It can be known in advance, for example, by measuring or presetting the operating parameters of the descaling machine, such as flushing pressure, flushing dosage, chemical flushing agent composition and / or similar parameters, or by measuring or presetting the operating parameters of each mill stand in the mill stand group, such as rolling force, roll torque, roll bending, roll axial displacement, roll temperature / roll temperature distribution and / or similar parameters.
[0052] Finally, the apparatus 10 has a calculation module 50. In method step S4, the calculation module 50 is used to determine the target value S of the selected production parameter P. Suitably, the material parameter value W, corrected using a first correction factor and second correction factors c1, c2, and the target specification Z for the metal strip 14 to be rolled are used as the basis for determination. The corrected material parameter value W and the target specification Z are preferably incorporated into the physical production parameter model 52. The physical production parameter model 52 can map the physical relationship between the target specification Z, i.e., the desired characteristics of the metal strip 14 to be produced, and the selected production parameter P, in this example, the rolling force to be applied. The calculation module 50 is accordingly designed to determine the target value S of the selected production parameter P based on the physical production parameter model 52.
[0053] Figure 2 An example of a recurrent neural network 42 that can serve as the basis for a second machine learning model is shown. Network 42 has an input layer 44 that holds X (e.g., measured or set) production parameter values p. i Where i = 1...X. Additionally, one or more hidden layers 46, also referred to as "hidden layers," are provided, each containing multiple nodes or units. Only one hidden layer 46 is shown at this time, comprising N nodes h. j , where j = 1...N. Finally, network 42 has an output layer 48 from which a second correction factor c2 can be extracted.
[0054] Regarding node h j In the calculation of the contribution of the second correction factor c2, as in the feedforward neural network, all values from the input layer 44 are first included. Therefore, the production parameters P are derived from the prior knowledge... The process begins by determining the second correction factor c2, as indicated by the solid black arrow. Here, through training network 42, at each node h... j The activation function of each individual is obtained, and the activation function of each individual is used to apply it to the corresponding node h. j The inputs ("inputs") are summed at each node h. jThe activation functions in the model can be distinguished from each other, for example, by the weights (formed during training) and the biases (formed during training).
[0055] At each node h j The activation function formed in the network corresponds to the establishment of a kind of "long-term memory" in network 42.
[0056] However, additionally, the corresponding node h from the previous calculation... j The output (“output”) is also included in this calculation. This is indicated by the dotted-line arrow. Furthermore, all other nodes h1...h from the previous calculations are also included. j-1 h j+1 ... h N The outputs are also included in this calculation, as indicated by the dashed arrow (for clarity, this is shown only for the first node h1).
[0057] The contribution to the second correction factor c2 is calculated based on the results previously calculated within hidden layer 46, thus achieving a kind of "short-term memory" for network 42. Therefore, the determined second correction factor c2 is related to the calculation of the previously determined second correction factor c2, and consequently, to the operating state of the device when the correction factor c2 was previously predicted. However, the operating state is already related to the previously determined correction factor or operating state, and so on. The influence of earlier operating states can therefore be persistent.
[0058] To counteract this, for example when the equipment state changes (such as when the rolls of a rolling mill stand are replaced), the recurrent neural network 42 can have nodes h with one input gate, one forget gate, and one output gate. j Or unit. The node h thus formed j Also known as LSTM cells (“long shortterm memory”). Signal processing in hidden layer 46 can be selectively controlled via gates. In particular, this allows for the selective avoidance or at least limitation, for example, temporal limitation, of the impact on the prediction of the correction factor c2 of previously produced bands.
[0059] Figure 3An example of a casting and rolling apparatus is shown. The casting and rolling apparatus includes: casting equipment 220 for casting metal billets 12, at least one mill stand group 230 for rolling the cast metal billets 12 or slabs separated therefrom into hot-rolled strips 14, a device 10 for determining target values S of the production parameters of the rolling equipment 200, and a control device 210, the control device being designed to control the rolling equipment 200, particularly at least one mill stand group 230, based on the target values S determined by the device 10.
[0060] In addition to the casting equipment 220 and the mill stand group 230, the casting equipment 220 suitably includes a mold 222 and a billet guide 224 extending in an arc and including a secondary cooling system 226. The mill stand group 230 has multiple, in this example, four mill stands 231, 232, 233, and 234. The casting and rolling equipment also suitably includes additional equipment components. For example, upstream of the mill stand group 230, a first separation device 240 (e.g., a swing shear) is arranged for separating the slab from the cast metal billet 12, and a descaling device 250. Downstream of the first mill stand group 230, a cooling zone 260 for cooling the hot-rolled strip 14 to the coiling temperature, multiple (in this example, two) coilers 270 for coiling the cooled hot-rolled strip 14, and at least one second separation device 280 for separating the coiled hot-rolled strip from the subsequent hot-rolled strip 14 can be provided.
[0061] In this example, the rolling mill 200 comprises a descaling device 250, mill stands 231-234 of the mill stand group 230, a cooling zone 260, a coiler 270, and a second separation device 280. The rolling mill may also have other components, such as a heating device, like a tunnel furnace or induction heater, arranged upstream of the mill stand group 230, for heating the metal slab 12 or separating the slab from it. The rolling mill 200 does not necessarily have to be combined with the casting and rolling equipment; conversely, the rolling mill 200 may also be separate from the casting equipment 220.
[0062] If a new hot-rolled strip 14, i.e., a hot-rolled strip 14 with changed characteristics (e.g., different strip thickness) compared to the currently produced hot-rolled strip 14, is produced during continuous operation, the selected production parameters for the rolling mill 200, such as the corresponding target value S of the rolling force in one of the mill stands 231-234, can be determined by means of the apparatus 10. For this purpose, the apparatus 10 appropriately processes the production parameters or production parameter values P that are known in advance for the production of the new hot-rolled strip 14. For example, the cross-sectional dimensions of the metal billet 12, the required thickness reduction, the current rolling force, current roll torque, current rolling temperature and / or similar parameters in each of the stands 231-234, the flushing agent pressure and / or flushing agent flow rate in the descaling unit 250, and the cooling capacity and / or similar parameters in the cooling zone 260. Similarly, the apparatus 10 suitably processes the target specification Z characterizing the desired features of the hot-rolled strip 14 to be produced.
[0063] Based on the defined target value S, the control device 210 can output a corresponding control signal I to the equipment components, which embody the selected production parameters or can set the defined target value S. For example, it can output the control signal I to the mill stands 231-234 of the mill stand group 230 to set the rolling force to achieve the target specification Z.
[0064] Since device 10 relies on a (second) machine learning model based on a recurrent neural network when determining the target value S, it can take into account the temporal evolution of the equipment state up to the current equipment state. Therefore, through the network's short-term memory, the determination of the target value S can be implicitly based on an "instantaneous record" of the equipment state, even if the equipment state as a whole (at least in practice) cannot be characterized by parameter values. Furthermore, the effects of the metal strip 14 produced immediately afterward can also be incorporated into the target value determination, such as the resulting potentially uneven roll wear or its temperature rise. As described above... Figure 2 This effect is based on the architecture of recurrent neural networks, therefore it does not require targeted manual adaptation of the (second) machine learning model, i.e., continuous training. Instead, the machine learning model adapts automatically.
[0065] Although the invention has been described and illustrated in detail with reference to preferred embodiments, the invention is not limited to the disclosed examples, and those skilled in the art can derive other variations therefrom without departing from the scope of protection of the invention.
[0066] List of reference numerals
[0067] 10 devices
[0068] 12 Metal billets and metal products
[0069] 14. Metal strip, hot-rolled strip
[0070] 20 Provide Modules
[0071] 30 Other (first) machine learning models
[0072] 32 Feedforward Neural Networks
[0073] 40 (Second) Machine Learning Model
[0074] 42 Recurrent Neural Networks
[0075] 44 Input Layer
[0076] 46 Hidden Layers
[0077] 48 Output Layer
[0078] 50 Calculation Modules
[0079] 52 Physical Production Parameter Model
[0080] 100 methods
[0081] S1 provides material parameter values
[0082] S2 determines an additional (first) correction factor.
[0083] S3 Determine the (second) correction factor
[0084] S4 Determine the target value
[0085] 200 Rolling Mill Equipment
[0086] 210 Control device
[0087] 220 casting equipment
[0088] 222 Casting mold
[0089] 224 Cast billet guide component
[0090] 226 Secondary Cooling System
[0091] 230 Rolling Mill Stands
[0092] 231-234 Rolling Mill Stands
[0093] 240 First Separation Unit
[0094] 250 Descaling Device
[0095] 260 Cooling Zone
[0096] 270 winding machine
[0097] 280 Second Separation Unit
[0098] W value
[0099] M Material Parameters
[0100] c1 First correction factor
[0101] c2 Second Correction Factor
[0102] E. Material characteristics or processing variables
[0103] K Material Characteristic Data
[0104] S target value
[0105] Z Target Specification
[0106] P Selected production parameters
[0107] P Pre-known production parameters
[0108] I control signal
[0109] p i Production parameter values
[0110] h i node.
Claims
1. A method (100) for determining a target value (S) of a selected production parameter (P) for a rolling mill (200) for rolling a metal product (12) into a metal strip (14), wherein, -Provide the value (W) of material parameter (M) for the predetermined reference material (S1). - A machine learning model (40) based on a recurrent neural network (42) is used in metal strip production with pre-known production parameters (P). Based on this, a correction factor (c2) for the material parameter (M) is determined (S3), and - Based on the physical production parameter model (52), determine (S4) the target value (S) of the selected production parameter (P) of the rolling equipment (200), and incorporate the material parameter value (W) adjusted by the correction factor (c2) and the target specification (Z) of the metal strip (14) to be rolled into the physical production parameter model.
2. The method (100) according to claim 1, wherein, Another machine learning model (30) based on a feedforward neural network (32) determines (S2) an additional correction factor (c1) for the material parameter (M) based on material feature data (K) characterizing the state of the metal to be processed next in the rolling mill.
3. The method (100) according to claim 1 or 2, wherein, The machine learning model (40) is based on a recurrent neural network (42) with controllable short-term memory.
4. The method (100) according to any one of the preceding claims, wherein, The machine learning model (40) continuously determines the correction factor (c2) during the sequential production of the metal strip (14).
5. The method (100) according to any one of claims 1 to 3, wherein, - During the operation of the rolling mill (200), the pre-known production parameters (P) are stored for each produced metal strip (14). ),as well as - Before rolling a new metal strip (14), the machine learning model (40) reads the production parameters (P) from different stored pre-known production parameters (P). The correction factor (c2) used for the material parameter (M) is determined repeatedly in the process.
6. The method (100) according to any one of the preceding claims, wherein, The selected production parameters (P) are rolling parameters, particularly rolling force, roll torque, roll bending and / or roll axial displacement.
7. The method (100) according to any one of the preceding claims, wherein, The material parameter (M) with a specified value (W) is the yield stress.
8. The method (100) according to any one of the preceding claims, wherein, After producing the metal strip (14) using the target value (S) determined for the selected production parameter (P), the strip parameters of the produced metal strip (14) are determined, and the machine learning model (40) is tuned using the determined strip parameters.
9. An apparatus (10) for determining a target value (S) of a selected production parameter (P) for a rolling mill (200) for rolling a metal product (12) into a metal strip (14), the rolling mill having at least one mill stand group (230), the apparatus having: - A providing module (20) is configured to provide a value (W) of a material parameter (M) for a predetermined reference material. - A machine learning model (40), which is based on a recurrent neural network (42) and is designed for metal strip production from pre-known production parameters (P). The correction factor (c2) for the material parameter (M) is determined in the following: - Calculation module (50), which is configured to determine the target value (S) of the selected production parameter (P) of the rolling mill (200) based on the physical production parameter model (52), and incorporate the material parameter value (W) adjusted by the correction factor (c2) and the target specification (Z) of the metal strip (14) to be rolled into the physical production parameter model.
10. A rolling mill (200) for producing metal strip (14), said rolling mill having: - At least one mill stand assembly (230) for rolling metal products (12) into metal strips (14). - The apparatus (10) according to claim 9, and - Control device (210), which is configured to control the rolling mill (200) based on a target value (S) determined by the device (10) for a selected production parameter (P) for the rolling mill (200).
11. A computer program product for performing the method (100) according to any one of claims 1 to 8.
12. A machine learning model (40) for use in the method according to any one of claims 1 to 8, said machine learning model being based on a recurrent neural network (42) and configured to be trained in metal strip production from a previously known production parameter (P) using training data comprising a plurality of different pre-known production parameter values and correction factors associated with said production parameter values. The correction factor (c2) used for the material parameter (M) is determined in ).
