Ai model for controlling and / or regulating a forging process

Abstract

The invention relates to a method for generating and using an AI model (60) for a control device (50) for controlling and / or regulating and / or monitoring a forging process, preferably a free-form forging process, of a workpiece (2), wherein the method has the steps of: generating a database by running at least one computer simulation, preferably an FEM simulation, of at least one fictitious forging process on the basis of at least one set of fictitious process parameters; generating the AI model (60) by training a neural network using the generated database; and using the AI model (60) by inputting process parameters of the forging process of the workpiece (2) into the neural network, wherein the neural network of the AI model (60) generates an output field which is characteristic of the workpiece (2), said output field preferably comprising a temperature and / or shape-change distribution of the workpiece (2).

Description

[0001] Kl model for controlling and / or regulating a forging process

[0002] Technical area

[0003] The invention relates to a method for creating and applying a Kl model for a control device for controlling and / or regulating a forging process, in particular an open-die forging process, of a workpiece. The invention further relates to an arrangement comprising a forging device, preferably an open-die forging press, for forging a workpiece, a handling device for conveying and aligning the workpiece, and a control device for controlling and / or regulating the forging process.

[0004] Background of the invention

[0005] Forging is a non-cutting, forming process used primarily for steel, but also for some non-ferrous metals. The energy required for forming is transferred into the workpiece in the form of kinetic energy. If the forming movement is not mechanically limited by a fixed abutment or other means, it is called open-die forging.

[0006] A rough estimate of the geometry development of components during open-die forging is made, for example, based on analytical equations for creating pass schedules. A prediction of the strain distribution along the core fiber of the workpiece is made possible by semi-empirical models using analytical equations as well as material- and temperature-specific fitting parameters. Reference is made to the following references in this regard: Dominik Recker (2014), "Development of fast process models and optimization options for open-die forging," dissertation, IBF Aachen; Dirk Rosenstock (2018), "Fast process modeling, online visualization, and optimization in open-die forging," dissertation, IBF Aachen.

[0007] A 3D prediction of the strain distribution is currently only possible using finite element modeling (FEM). This can also be used to predict grain size distribution, pore closure behavior, and damage development in the material. For further information, see the following references: Wolfgarten, M., Rosenstock, D., Rudolph, F., & Hirt, G. (2019), "New approach for the optimization of pass-schedules in open-die forging," International Journal of Material Forming, 12, pages 973-983; Rudolph, F., Wolfgarten, M., Keray, V., & Hirt, G. (2021 ), “Optimization of Open-Die Forging Using Fast Models for Strain, Temperature, and Grain Size in the Context of an Assistance System”, Forming the Future: Proceedings of the 13th International Conference on the Technology of Plasticity, pages 1145-1159, Springer International Publishing.

[0008] While a rough estimate of the geometry development using simple analytical equations can be made very quickly, within a few seconds, it provides only insufficient predictive quality for calculating material behavior during the forging process. This also applies to the prediction of the strain distribution along the core fiber of the workpiece using semi-empirical models using analytical equations, although this information is particularly insufficient for estimating the overall component quality. The grain size distribution and damage development also significantly determine the performance of a component outside the core fiber.

[0009] Prediction using FEM, on the other hand, is very time-consuming. A simulation can take hours or even days, making it unsuitable for rapid production planning or optimizing the open-die forging process with regard to the resulting component quality. In practice, the prediction of component quality and the necessary pass schedules in open-die forging have so far been primarily based on experience. This experiential knowledge is not necessarily accessible in a structured manner within the company, but rather resides with the respective employee. This can lead to personnel-dependent fluctuations and a corresponding risk of failure.

[0010] A model- and computer-based prediction of the strain distribution as a basis for predicting component quality during open-die forging is currently only partially feasible, i.e., with insufficient accuracy, only for the core fiber, or with a significant time expenditure, as described above. Thus, it is not yet possible to design, control, or optimize the open-die forging process as a complex forming process with regard to those material properties that significantly determine component quality and performance. Process monitoring during open-die forging is also not yet feasible in 3D.

[0011] Description of the invention

[0012] An object of the present invention is to improve the control and / or regulation and / or monitoring of a forging process, in particular an open-die forging process, of a workpiece.

[0013] The object is achieved by a method having the features of claim 1 and an arrangement having the features of claim 8. Advantageous further developments follow from the subclaims, the following presentation of the invention and the description of preferred embodiments.

[0014] The invention relates to the forging, in particular open-die forging, of a workpiece. The forging of the workpiece takes place in a forging device, preferably designed as an open-die forging press. Metallic materials, in particular steel, nickel, titanium, zirconium, and aluminum, are suitable as materials for the workpiece. Forging preferably takes place at temperatures above 800 °C. The geometries desired by forging include, for example, round, rectangular, square, and square.

[0015] The method according to the invention specifically serves to create and apply a Kl model for a control device for controlling and / or regulating and / or monitoring a forging process, preferably an open-die forging process, of a workpiece.

[0016] According to the method, a database is first created by executing at least one computer simulation, preferably a FEM simulation, of at least one fictitious forging process based on at least one set of fictitious process parameters. "Fictitious" here means that process parameters are used for the simulation(s) that do not necessarily have to be identical to the process parameters of the (real) forging process to be controlled, regulated, or monitored.

[0017] Based on the database generated in this way, a Kl model is created by training a neural network.

[0018] The neural network can be classified as a convolutional recurrent neural network, each comprising at least one convolution layer and one pooling layer for downsampling, at least one recurrent layer for describing the temporal progression, and optionally at least one transcription layer and one output layer. The exact number of layers and the number of neurons in each layer can be varied and determined depending on the application. The Kl model created in this way is applied within the (real) forging process by inputting process parameters of the forging process into the neural network, whereby the neural network of the Kl model generates an output field that is characteristic of the workpiece. The output field preferably comprises a temperature and / or deformation distribution of the workpiece.

[0019] The input of process parameters into the neural network can also be automated, for example by measuring instruments in the process such as a camera, distance measuring systems for the tool positions, timers, scales and the like.

[0020] The Kl model presented here, which uses at least one neural network, allows for a very accurate prediction of characteristic material properties, such as geometry development as well as temperature and strain distribution, for forging processes within a short time. Based on this, a pass schedule calculation can be performed.

[0021] The fast calculation time of, for example, a few (<10) seconds enables optimization of the component quality and performance with regard to selected target parameters, such as a homogeneous and small grain size, maximization of pore closure or minimization of induced damage.

[0022] Preferably, at least one material property of the workpiece is determined from the output field of the neural network of the Kl model, wherein the at least one material property comprises a local grain size distribution and / or pore closure behavior and / or damage development in the workpiece. The prediction system based on the neural network enables a combination with the prediction of the grain size distribution in the material, the pore closure behavior of casting cavities and similar defects in the material, and the damage development as well as the formation of possible component defects during the forging process. These properties can be predicted in the component, and the manufacturing process can be specifically designed with these criteria in mind. Any pass schedule calculation can thus take into account defined criteria regarding component quality and performance, allowing the process to be optimized accordingly.

[0023] Preferably, the output field of the neural network of the Kl model and / or the determined at least one material property of the workpiece is used to control and / or regulate and / or monitor the forging process. Such feedback to the plant control system enables the process to be carried out as closely as possible to the optimized process design. Process monitoring can provide the plant operator with additional information about the geometry development and enables warning messages to be issued if the process deviates from the pass schedule calculation, thus creating a risk that the component will not meet the required quality and performance.

[0024] Preferably, the database for the neural network is generated by finite element modeling by varying various fictitious process parameters and generating a multidimensional grid of the fictitious process parameters such that each point in the grid corresponds to a combination of the fictitious process parameters.

[0025] Preferably, the fictitious process parameters include workpiece dimensions and / or tool dimensions and / or temperatures and / or pause times and / or pass reductions.

[0026] Various parameter ranges for scanning by FEM are possible, concerning the geometry type, the material, the dimensions, process parameters, etc., which are given as examples in the description of embodiments below.

[0027] Preferably, FEM simulations are carried out based on the defined grid, the results of which preferably include nodal displacements and / or temperatures and / or deformations, such as strains and shear stresses, and / or deformation increments and / or deformation rates. In addition to or as an alternative to the local variables in the material mentioned here, the results can also include global variables such as process forces, energy consumption, required process time, etc. In this respect, the system also offers the option of optimizing and controlling the pass schedules optimized according to these criteria. The neural network can also be trained with the results from different units (e.g., two hydraulic presses with 40 and 50 MN pressing force). This also makes it possible to output a preferred forming unit or to maintain a certain maximum force when generating a pass schedule.

[0028] The database for training the neural network therefore consists in particular of the multidimensional grid of process parameters including the corresponding FEM input parameters, the respective definition of the FEM model as well as the element- and node-based results of the simulation(s).

[0029] Preferably, the neural network of the Kl model, trained on the database, generates the output field for combinations of process parameters not considered when creating the database. In other words, by training on the database, the neural network is able to generate an output field in the form of a temperature and strain distribution for previously unconsidered combinations of process parameters or grid points within a short time, particularly within a few seconds (<10 seconds). The prediction accuracy is very high within the boundaries of the grid; outside the grid, the predictions are less reliable.

[0030] The above-mentioned object is further achieved by an arrangement comprising a forging device, preferably an open-die forging press, for forging a workpiece, a handling device for conveying and aligning the workpiece and a control device for controlling and / or regulating and / or monitoring a forging process which can be carried out by the forging device and handling device, wherein the control device comprises a Kl model with a neural network which is configured to generate an output field on the basis of process parameters of the forging process which is characteristic of the workpiece, preferably comprising a temperature and deformation distribution of the workpiece.

[0031] The features, technical effects, advantages and embodiments described with regard to the method apply analogously to the arrangement.

[0032] For the reasons stated above, the control device is preferably configured to determine at least one material property of the workpiece from the output field of the neural network of the Kl model, wherein the at least one material property comprises a local grain size distribution and / or a pore closure behavior and / or a damage development in the workpiece.

[0033] For the reasons stated above, the control device is preferably configured to use the output field of the neural network of the Kl model and / or the determined at least one material property of the workpiece to control and / or regulate and / or monitor the forging process. Further advantages and features of the present invention will become apparent from the following description of preferred embodiments. The features described therein can be implemented alone or in combination with one or more of the features set out above, provided the features do not contradict each other. The following description of preferred embodiments is made with reference to the accompanying drawings.

[0034] Short description of the characters

[0035] Preferred further embodiments of the invention are explained in more detail in the following description of the figures. In the figures:

[0036] Figure 1 shows schematically an arrangement with a forging device and a handling device; and

[0037] Figure 2 schematically shows a process for creating and applying a Kl model, based on a neural network, for a control device of the forging device.

[0038] Detailed description of preferred embodiments

[0039] Preferred embodiments are described below with reference to the figures.

[0040] Figure 1 schematically shows an arrangement 1 with a forging device 10 for forging a workpiece 2 and a handling device 20.

[0041] The forging device 10 is preferably designed for open-die forging, in particular as an open-die forging press. In this case, a piston force is exerted on the workpiece 2, which is arranged outside the forging device 10 in Figure 1, by means of a piston-cylinder system and a forging tool, in order to deform it in a desired manner.

[0042] Metallic materials, particularly steel, nickel, titanium, zirconium, and aluminum, are considered for workpiece 2. Forging preferably takes place at temperatures above 800 °C. The geometries desired by forging include, for example, round, rectangular, square, and rectangular.

[0043] The handling device 20 is configured to move the workpiece 2, in particular to a desired position and into a desired orientation. For this purpose, the handling device 20 comprises a transport section 22, for example, designed as a vehicle, in particular an autonomous vehicle, and a manipulator section 24 configured to hold, rotate, pivot, and the like the workpiece 2. The manipulator section 24 preferably comprises a pair of tongs for gripping the workpiece 2.

[0044] The assembly 1 further comprises a control device 50, which communicates with the various devices, assemblies, and the like. The control device 50 is signal-connected to the components of the assembly 1 to be controlled or regulated and / or read out, thus in particular to the forging device 10 and the handling device 20.

[0045] Communication between the control device 50 and the components to be controlled or regulated and / or read out can be wired or wireless, digital or analog. The control device 50 can receive and / or send signals (control signals, data, etc.) accordingly, whereby both signal transport in one direction and in both directions falls under the term "communication" in this context. The control device 50 does not necessarily have to be implemented by a central computing device or electronic control system, but rather decentralized and / or multi-level systems, control networks, cloud systems, and the like are included. The controller can also be an integral component of a higher-level system control system or communicate with such a system. The control device 50 can also communicate with lower-level system controls, i.e., controls assigned to the corresponding devices.

[0046] The control device 50 comprises or has access to a Kl model 60 ("Kl" stands for artificial intelligence) based on one or more neural networks. The Kl model 60 allows the prediction of the geometric development of the workpiece 2 as well as the temperature and deformation distribution in the workpiece 2 during forging for defined forging sequences within a short time, in particular within a few seconds.

[0047] FEM simulations ("FEM" stands for finite element method) can be used as the database for the underlying neural network of the Kl model 60, in which the process parameters (workpiece and tool dimensions, temperatures, pause times, pass reduction, etc.) are varied. A database is generated for this purpose using a suitable number of support points (combinations with different process parameters), and the neural network is trained with it. At least within this selected range, the control device 50, with access to the Kl model 60, is then able to predict the temperature and deformation distribution of the workpiece 2. The range is selected according to the possible and targeted range of process parameters.

[0048] Based on the prediction of the temperature and deformation distribution of the workpiece 2, material properties that are crucial for the component quality and performance, such as the grain size distribution, the pore closure behavior and the damage development, can also be predicted.

[0049] This enables process design and pass schedule calculation within a few seconds from a start geometry to a final geometry by the Kl-Model 60 based on at least one neural network.

[0050] In addition, by predicting the temperature and strain distribution based on the Kl Model 60, the grain size distribution, pore closure behavior, and damage development can be predicted using conventional models. The process design can thus be optimized to ensure that defined criteria for component quality and performance are met or maintained.

[0051] The process design according to the above description is used to control and / or regulate and / or monitor the open-die forging process so that the component to be manufactured from workpiece 2 follows the process design as closely as possible.

[0052] Likewise, the existing temperature and deformation distribution in the workpiece 2 can be calculated almost in real time (e.g., with a delay of less than 2 seconds) during the process. Based on this, the current grain size distribution, pore closure, and damage development can be determined and, if necessary, displayed. Such process monitoring can also be used to generate a warning, for example, for the operator if the process is likely to result in the workpiece not achieving the defined component quality range. Figure 2 schematically shows an example process for creating and applying the Kl model 60 for the control device 50. The database required for training the neural network of the Kl model 60 is generated here using FEM simulations.

[0053] In a first step S1, a multidimensional grid is defined for the FEM simulation(s), whereby different parameter ranges for scanning, ie the grid points of the multidimensional grid to be simulated, are possible.

[0054] With regard to the geometry type, for example, straight components with round or square cross-sections, multi-stage shafts, stepped shafts, curved components, twisted components, etc. as well as combinations of geometry variants can be recorded.

[0055] With regard to the material of workpiece 2, iron-based materials and steels as well as materials based on aluminum, titanium, nickel, copper, zinc and alternative formable materials with workpiece temperatures from -273 °C to 2800 °C can be considered.

[0056] In terms of mass and dimensions of the component, masses from 0.1 kg to 5000 t operating weight with lengths from 10 mm to 120 m and edge lengths (or diameters for round cross-sections) from 10 mm to 25 m are included.

[0057] The forging device 10 can comprise, as a forming unit, one or more hammers with impact energies from 1 Nm to 125,000 kNm, screw presses with forming energies from 1 J to 250,000 kJ, as well as hydraulic, servo-hydraulic, or servo-mechanical presses with maximum forces from 1 kN to 10,000 MN, crank-lever presses, etc. Process parameters include relative pass reductions of 0.1 - 90% per stroke, pass widths from 1 mm to 10 m, if applicable, rotation angles between each stroke or pass of 0.1 - 180°, and curvature or twist angles per stroke of 0.1 - 75°.

[0058] Furthermore, various boundary conditions can be used for scanning to create a FEM simulation model, including ambient temperatures from -100 to +900 °C and tool temperatures from -100 to +1400 °C.

[0059] The database required for training the neural network of the Kl model 60 is then generated in a second step S2 by FEM simulations.

[0060] For this purpose, at least one FEM simulation model is created that corresponds to the aforementioned parameters of the grid points of the multidimensional grid to be simulated. Each point in this grid corresponds to a combination of process parameters, which are then calculated for at least one pass using a suitable finite element program (e.g., Simufact, Forge NxT, Deform, Q-Form, or similar). A simulation can encompass a single or multiple strokes, one or more passes, or even an entire forging sequence, including intermediate heat treatments.

[0061] The workpiece 2 under consideration is meshed in 2D with preferably at least 4 nodes or in 3D with preferably at least 8 nodes. A maximum of 10,000,000 or 1,000,000 nodes, respectively, that comprise the FEM mesh, are preferably used in each simulation.

[0062] The FEM simulation is started, and after the calculation is complete, the simulation results are saved. The saved data includes, in particular, nodal displacements, temperatures, strains (strains and shear stresses), strain increments, and strain rates. The simulation results thus generated are exported, preferably automatically, as a database.

[0063] The sum of all input data paired with the stored simulation results forms the data basis for the neural network.

[0064] It is helpful if the individual FEM simulations are defined in such a way that the number of elements, and thus the number of support points in the model, is as equal as possible for all simulations in a database. However, this is not mandatory.

[0065] The database for training the neural network thus consists of the multidimensional grid of process parameters including the corresponding FEM input parameters, the respective definition of the FEM model as well as the element- and node-based results of the simulations.

[0066] In a third step S3, the neural network of the Kl model 60 is trained. This is done in a suitable programming environment on the generated database.

[0067] The neural network can, for example, be classified as a convolutional recurrent neural network, consisting of at least one convolution layer and one pooling layer for downsampling, at least one recurrent layer to describe the temporal course, and possibly at least one transcription layer and one output layer. The exact number of layers and the number of neurons in each layer are determined by the size of the database and, in particular, the definition of the finite element model, i.e., the number of elements and nodes. When applying the neural network trained in this way, step S4, an input point within the multidimensional grid can now be used, which does not have to correspond to a grid point of the grid. It should be noted that the database should be chosen such that no points outside the grid are necessary for future predictions.Although results are also generated for such points, the deviation from reality increases with the distance from the grid.

[0068] For each input point, the neural network then generates a prediction for the selected simulation results, ie at least nodal displacements, temperatures, strains (strains and thrusts), strain increments and strain rates.

[0069] By training on the database, the neural network is thus able to generate an output field in the form of a temperature and strain distribution for previously unconsidered combinations of process parameters (or grid points) within a short time, for example, within a few seconds (<10). The accuracy of the prediction is very high within the boundaries of the grid; outside the grid, the predictions are less reliable.

[0070] The output field can be used to predict material properties that are crucial for component quality and performance using suitable models based on temperature and strain distribution. These include, among other things, the local grain size distribution, pore closure behavior, and / or damage development within the component.

[0071] The generated result variables can, for example, be coupled with conventional models for predicting grain sizes, damage values, and / or pore closure in order to predict the performance of components and optimize them manually or automatically. The optimized process design is used to control and / or regulate and / or monitor the open-die forging process (press and manipulator movements) by the control device 50. In the event of deviations from the optimized process, a recalculation and, if necessary and possible, a course correction of the further process can be performed.

[0072] If the process deviates so significantly from the specification that the required component quality can no longer be achieved, or if reheating becomes necessary, the control device 50 can issue a corresponding warning message to which an operator can then react.

[0073] Below, a concrete example is presented to illustrate the above process for creating and applying the Kl Model 60:

[0074] Forged, cuboid-shaped workpieces 2 with unit weights of 100 kg to 25 t made of structural steel, a starting geometry in the range of 320 mm x 320 mm x 125 mm and 1265 mm x 1265 mm x 2,000 mm and finished dimensions of 160 mm x 160 mm x 500 mm to 632.5 mm x 632.5 mm x 8,000 mm are assumed, which are manufactured at a starting temperature of 1250 °C.

[0075] If a workpiece 2 cools below 900 °C during production, reheating is necessary. For open-die forging, a hydraulic press with a maximum press force of 25 MN is used as the forging device 10. The relative pass reductions range from 3 to 15%, and the bite width is 50 to 250 mm. The workpieces 2 are rotated 90° between each pass. The ambient temperature is approximately 35 °C, and the tool temperature is approximately 300 °C. To generate the database according to step S2, these ranges of dimensions and process parameters as well as the fixed boundary conditions are simulated. For this purpose, the boundary points and 18 additional linear points within the parameter spectrum are simulated. With 20 simulations per variable and the variables “start geometry”, “end geometry”, “relative pass reduction” and “bite width” this results in 20x20x20x20 = 16,000 simulations.By physically reasonable limitations of the start and end geometry (the smallest end geometry cannot be created from the largest start geometry) and the bite width (which cannot be longer than the total component length), this number can be reduced to 10,000 simulations.

[0076] An entire process simulation, for example, takes approximately three days, so the initial effort is very substantial. However, since many forging passes are very similar, the processes can be broken down into their individual passes, with each process consisting of four to forty passes. If thermally and geometrically similar passes are combined, a total of 200,000 simulations are obtained, each requiring approximately 30 minutes of simulation time. With a total simulation time of 100,000 hours (instead of 720,000 hours for the original ten).000 simulations), whereby in principle any number of these simulations can be carried out and evaluated in parallel, all simulations can be started, calculated and evaluated automatically within one week using a high-performance cluster with, for example, 700 computing cores, including saving the simulation results together with the input parameters in a database in such a way that this can be used for the subsequent training of the neural network.

[0077] The generated database is then used, for example, in the C# programming language to train a neural network. Within this parameter field, this network generates an output within a few seconds that corresponds to the result of an FEM simulation. The results generated in this way, like the results of the FEM simulations, can be used to evaluate the performance of the produced components. For this purpose, existing models for predicting grain size, damage, and / or pore closure behavior are used. The short computing time of just a few seconds to generate a result is used to optimize the process sequences, particularly the forging strategy involving press and manipulator movements, with a view to maximizing pore closure behavior.

[0078] The forging strategies optimized in this way for the entire product spectrum are used in subsequent production to control the forging device 10 and the handling device 20. In the event of occasional deviations from the process strategy, an adapted forging strategy is calculated within a few seconds, and the remaining process sequence is replaced by the newly calculated sequence. If the calculation shows that the required component quality can no longer be achieved at this temperature, the system operator is notified that reheating is necessary. An adapted forging strategy is then calculated for the reheated component, enabling the required component quality to be achieved, and this strategy is implemented.

[0079] If deviations from the process strategy are detected by the existing measuring instruments, the forging strategy can also be adjusted automatically. In addition to the tool positions, measuring instruments can also include a timer (for deviations from pause or non-productive times), a camera for position and geometry detection, and / or a scale if material loss becomes excessive.

[0080] In a further development of the concept presented here, software for the automatic generation of the database (creation, calculation and evaluation of the simulations) can be developed in a suitable simulation program.

[0081] Furthermore, a system can be developed for the simplified generation / prediction and, if necessary, automatic optimization of the strain distributions by the neural network (if necessary with regard to a goal, for example, with regard to a defined grain size range, maximization of pore closure or minimization of damage).

[0082] The Kl-Model 60 presented here, using at least one neural network, allows for a very accurate prediction of the geometry development as well as temperature and strain distribution for open-die forging processes within a few seconds. A pass schedule calculation can be performed, which also provides a basis for predicting component quality and performance.

[0083] Based on this, the grain size distribution, pore closure behavior, and / or damage development in the component can be predicted, and the process can be specifically designed with these criteria in mind. The pass schedule calculation can thus take into account defined criteria regarding component quality and performance, and the process can be optimized accordingly.

[0084] A feedback to the plant control system enables the process to be carried out as close as possible to the optimized process design.

[0085] Process monitoring can provide the plant operator with additional information about the geometry development and enables warning messages to be issued if the process deviates from the pass schedule calculation, thus posing a risk that the component will not meet the required quality and performance. Where applicable, all individual features presented in the exemplary embodiments can be combined and / or interchanged without departing from the scope of the invention.

[0086] List of reference symbols

[0087] 1 Arrangement with forging device and handling device

[0088] 2 Workpiece 10 Forging device

[0089] 20 Handling device

[0090] 22 Transport section

[0091] 24 Manipulator section 50 Control device

[0092] 60 Kl model

Claims

Patent claims 1. A method for creating and applying a Kl model (60) for a control device (50) for controlling and / or regulating and / or monitoring a forging process, preferably an open-die forging process, of a workpiece (2), the method comprising: Creating a database by executing at least one computer simulation, preferably FEM simulation, of at least one fictitious forging process based on at least one set of fictitious process parameters; Creating the Kl model (60) by training a neural network using the generated database; and Applying the Kl model (60) by inputting process parameters of the forging process of the workpiece (2) into the neural network, wherein the neural network of the Kl model (60) generates an output field that is characteristic of the workpiece (2), preferably comprising a temperature and / or deformation distribution of the workpiece (2).

2. Method according to claim 1, characterized in that at least one material property of the workpiece (2) is determined from the output field of the neural network of the Kl model (60), wherein the at least one material property comprises a local grain size distribution and / or a pore closure behavior and / or a damage development in the workpiece (2).

3. Method according to claim 1 or 2, characterized in that the output field of the neural network of the Kl model (60) and / or the determined at least one material property of the workpiece (2) is used to control and / or regulate and / or monitor the forging process.

4. Method according to one of the preceding claims, characterized in that the database for the neural network is generated by finite element modeling by varying various fictitious process parameters and generating a multidimensional grid of the fictitious process parameters, so that each point in the grid corresponds to a combination of the fictitious process parameters.

5. Method according to claim 4, characterized in that the fictitious Process parameters workpiece dimensions and / or Tool dimensions and / or temperatures and / or pause times and / or pass reductions.

6. Method according to claim 4 or 5, characterized in that on the Based on the defined grid, FEM simulations are carried out, the results of which preferably include node displacements and / or Temperatures and / or changes in shape and / or strain increments and / or strain rates.

7. Method according to one of the preceding claims, characterized in that the neural network of the Kl model (60) trained on the database generates the output field for combinations of process parameters not taken into account when creating the database.

8. Arrangement (1) with a forging device (10), preferably an open-die forging press, for forging a workpiece (2), a handling device (20) for conveying and aligning the Workpiece (2) and a control device (50) for controlling and / or regulating and / or monitoring a forging process which can be carried out by the forging device (10) and handling device (20), wherein the control device (50) comprises a Kl model (60) with a neural network which is set up to generate an output field on the basis of process parameters of the forging process which is characteristic of the workpiece (2), preferably comprising a temperature and / or deformation distribution of the workpiece (2).

9. Arrangement (1) according to claim 8, characterized in that the control device (50) is set up to determine at least one material property of the workpiece (2) from the output field of the neural network of the Kl model (60), wherein the at least one material property comprises a local grain size distribution and / or a pore closure behavior and / or a damage development in the workpiece (2).

10. Arrangement (1) according to claim 8 or 9, characterized in that the control device (50) is set up to use the output field of the neural network of the Kl model (60) and / or the determined at least one material property of the workpiece (2) for controlling and / or regulating and / or monitoring the forging process.

Images