Method and device for the optimized production of sheet metal parts

Abstract

Process for optimizing the production of sheet metal parts (B1, B2) with the following process steps: a) Cutting and separating the sheet metal parts (B1, B2); b) bending the sheet metal parts (B1, B2); the method comprising the following steps: A) training a neural network (24) executed on a Monte Carlo tree search framework (22) by means of supervised learning and self-play with reinforcement learning; B) detecting boundary conditions (10) of the sheet metal parts (B1, B2), wherein the boundary conditions (10) comprise geometric data of the sheet metal parts (B1, B2); C) creating an optimized production plan by the neural network (24), wherein the optimization includes both waste minimization and production time optimization; D) Issue of the production plan.

Description

Background of the invention

[0001] The invention relates to a method for optimizing the production of sheet metal parts. The invention further relates to a device for carrying out such a method.

[0002] Sheet metal parts are used in a wide variety of products in a wide variety of geometries. To manufacture products with sheet metal parts, the parts are cut from a large sheet of metal, separated, deburred, bent, joined, coated, and / or assembled.

[0003] The sheet metal parts are manufactured in so-called orders. An order includes (i) the manufacture of a cut, separated, bent and / or assembled sheet metal part or ii) the production of multiple cut, separated, bent and / or assembled sheet metal parts within a specified production deadline.

[0004] The individual sheet metal parts should be cut from the sheet metal panel in such a way that as little residual material (offcuts) as possible remains as waste. Since sheet metal parts from different orders can have different geometries, it can be advantageous to combine sheet metal parts from different orders on one sheet metal panel to save space.

[0005] However, the resulting temporal mixing of orders increases the complexity of production planning. In addition, the production of sheet metal parts can take place on several identical or similar production machines. For example, several identical or similar bending machines can be used to bend individual sheet metal parts. The production machines should be operated at the highest possible capacity.

[0006] Production planning, i.e. the planning of when which sheet metal part is processed on which production machine, becomes very complex due to the variables described, especially in the case of events such as production machine breakdowns, rush orders and / or freed-up production machine capacities.

[0007] Optimal production planning is referred to as solving a job shop scheduling problem (JSSP). Solutions and approaches to this problem can be found in the following publications: [1] F. Pfitzer, J. Provost, C. Mieth, and W. Liertz, “Event-driven production rescheduling in job shop environments,” in 2018 IEEE 14th International Conference on Automation Science and Engineering (CASE), IEEE, 2018, pp. 939-944; [2] M. Putz and A. Schlegel, “Simulation-based investigation of priority and picking rules for controlling material flow in the sheet metal industry”; [3] L. L. Li, C. B. Li, L. Li, Y. Tang, and Q. S. Yang, „An integrated approach for remanufacturing job shop scheduling with routing alternatives.“, Mathematical biosciences and engineering: MBE, vol. 16, no. 4, pp. 2063-2085, 2019; [4] M. Gondran, M.-J. Huguet, P. Lacomme, and N. Tchernev, „Comparison between two approaches to solve the job-shop scheduling problem with routing“, 2019; [5] J. J. van Hoorn, „The current state of bounds on benchmark instances of the job-shop scheduling problem“, Journal of Scheduling, vol. 21, no. 1, pp. 127-128, 2018; [6] S.-C. Lin, E. D. Goodman, and W. F. Punch III, „A genetic algorithm approach to dynamic job shop scheduling problem“, in ICGA, 1997, pp. 481-488; [7] T. Yamada and R. Nakano, „Scheduling by genetic local search with multistep crossover“, in International Conference on Parallel Problem Solving from Nature, Springer, 1996, pp. 960- 969; [8] B. M. Ombuki and M. Ventresca, „Local search genetic algorithms for the job shop scheduling problem“, Applied Intelligence, vol. 21, no. 1, pp. 99-109, 2004; [9] E. S. Nicoara, F. G. Filip, and N. Paraschiv, „Simulation-based optimization using genetic algorithms for multi-objective flexible jssp“, Studies in Informatics and Control, vol. 20, no. 4, pp. 333-344, 2011;

[10] L. Asadzadeh, „A local search genetic algorithm for the job shop scheduling problem with intelligent agents“, Computers & Industrial Engineering, vol. 85, pp. 376-383, 2015;

[11] B. Waschneck, A. Reichstaller, L. Belzner, T. Altenmüller, T. Bauernhansl, A. Knapp, and Kyek, „Optimization of global production scheduling with deep reinforcement learning“, Procedia CIRP, vol. 72, pp. 1264-1269, 2018;

[12] M. Botvinick, S. Ritter, J. X. Wang, Z. Kurth-Nelson, C. Blundell, and D. Hassabis, „Reinforcement learning, fast and slow“, Trends in cognitive sciences, 2019.

[0008] From “Integrated sheetmetal production planning for laser cutting and bending, September 1, 2010 [<URL: https: / / halarchivesouvertes.fr / hal-00512900 / document> > a method for production planning is known.”

[0009] WO 2020 / 040763 A1 discloses the use of Monte Carlo Tree Search and Deep Reinforcement Learning in the creation of a production plan.

[0010] Furthermore, it has become known from WO 2017 / 157809 A1 to provide production planning with an optimization unit and a separate distribution unit.

[0011] Despite extensive efforts, satisfactory production planning has not yet been achieved due to the complexity of the task. Object of the invention

[0012] It is therefore an object of the invention to provide a method and a device for the optimized production of sheet metal parts. Description of the invention

[0013] This object is achieved according to the invention by a method according to claim 1 and a device according to claim 13. The subclaims give preferred developments.

[0014] The solution according to the invention thus comprises a method for optimizing the production of sheet metal parts. The method comprises at least the following process steps (one or more additional process steps may be provided before, after, and / or between the subsequent process steps): a) cutting and separating the sheet metal parts (in particular by punching or laser cutting); b) Bending the sheet metal parts.

[0015] The method comprises at least the following method steps (a further method step or several further method steps may be provided before, after and / or between the subsequent method steps): A) Training of a neural network based on a Monte Carlo tree search framework using supervised learning and self-play with reinforcement learning; B) Recording boundary conditions of the sheet metal parts, wherein the boundary conditions include at least geometric data of the sheet metal parts; C) Creating an optimized production plan through the neural network; D) Issue of the production plan.

[0016] According to the invention, optimization is thus provided using a neural network (NN). Neural networks are known to those skilled in the art, for example, from:

[13] Günter Daniel Rey, Karl F. Wender, “Neural Networks”, 2nd edition, 2010, Huber.

[0017] The neural network has decision nodes connected by edges. In this case, these are part of a Monte Carlo tree search (MCTS) framework, i.e., an algorithm with a decision tree. A promising path is selected in the decision tree (selection), the path is expanded (expansion), a simulation is performed based on the expanded path (simulation), and feedback, particularly in the form of strengthening or weakening, is provided to the decision tree based on the simulation result (backpropagation). Details on the implementation of an MCTS framework can be found in the following publication:

[14] G. Chaslot, S. Bakkes, I. Szita, and P. Spronck, “Monte-carlo tree search: A new framework for game ai,” in AIIDE, 2008.

[0018] In this case, the MCTS is performed by the neural network, which is pre-trained using supervised learning. Decision-making and further training are carried out using self-play and reinforcement learning.

[0019] Reinforcement learning (RL) is a feedback-based learning process that specifically involves strengthening or weakening the decision tree of the MCTS framework. Reinforcement learning generally refers to a set of machine learning methods in which an agent independently learns a strategy to maximize received rewards. Instead of being shown which action is best in which situation, the agent receives a reward, which can also be negative, at specific times. Based on these rewards, the agent approximates a utility function that describes the value of a particular state or action. Details on the implementation can be found in the following publications:

[15] W. Zhang and TG Dietterich, “A reinforcement learning approach to job-shop scheduling,” in IJCAI, Citeseer, vol. 95, 1995, pp. 1114-1120;

[16] R. S. Sutton, A. G. Barto, et al., Introduction to reinforcement learning, 4. MIT press Cam- bridge, 1998, vol. 2;

[17] S. Mahadevan and G. Theocharous, „Optimizing production manufacturing using reinforcement learning.“, in FLAIRS Conference, 1998, pp. 372-377;

[18] S. J. Bradtke and M. O. Duff, „Reinforcement learning methods for continuous-time markov decision problems“, in Advances in neural information processing systems, 1995, pp. 393-400;

[19] S. Riedmiller and M. Riedmiller, „A neural reinforcement learning approach to learn local dispatching policies in production scheduling“, in IJCAI, vol. 2, 1999, pp. 764-771;

[20] C. D. Paternina-Arboleda and T. K. Das, „A multi-agent reinforcement learning approach to obtaining dynamic control policies for stochastic lot scheduling problem“, Simulation Modelling Practice and Theory, vol. 13, no. 5, pp. 389-406, 2005;

[21] T. Gabel and M. Riedmiller, „Scaling adaptive agent-based reactive job-shop scheduling to large-scale problems“, in 2007 IEEE Symposium on Computational Intelligence in Scheduling, IEEE, 2007, pp. 259-266;

[22] Y. C. F. Reyna, Y. M. Jim ' enez, J. M. B. Cabrera, and B. M. M. Hernändez, „A reinforcement learning approach for scheduling problems“, Investigación Operacional, vol. 36, no. 3, pp. 225-231, 2015;

[23] S. Qu, J. Wang, S. Govil, and J. O. Leckie, „Optimized adaptive scheduling of a manufacturing process system with multi-skill workforce and multiple machine types: An ontology-based, multi-agent reinforcement learning approach“, Procedia CIRP, vol. 57, pp. 55-60, 2016;

[24] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Ried- miller, „Playing atari with deep reinforcement learning“, arXiv preprint arXiv:1312.5602, 2013;

[25] A. Kuhnle, L. Schäfer, N. Stricker, and G. Lanza, „Design, implementation and evaluation of reinforcement learning for an adaptive order dispatching in job shop manufacturing systems“, Procedia CIRP, vol. 81, pp. 234-239, 2019;

[26] N. Stricker, A. Kuhnle, R. Sturm, and S. Friess, „Reinforcement learning for adaptive order dispatching in the semiconductor industry“, CIRP Annals, vol. 67, no. 1, pp. 511-514, 2018;

[27] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, „Trust region policy optimization“, in International conference on machine learning, 2015, pp. 1889-1897.

[0020] Supervised learning refers to training with predefined solutions. This type of supervised learning is generally a subfield of machine learning. Learning refers to the ability of artificial intelligence to replicate regularities. The results are known through natural laws or expert knowledge and are used to train the system. A learning algorithm attempts to find a hypothesis that makes the most accurate predictions possible. A hypothesis is a mapping that assigns each input value to the presumed output value. The method is therefore based on a predetermined output to be learned, the results of which are known. The results of the learning process can be compared with the known, correct results, i.e., "supervised." Details on the implementation can be found in the following publications:

[28] M. Gombolay, R. Jensen, J. Stigile, S.-H. Son, and J. Shah, “Apprenticeship scheduling: Learning to schedule from human experts,” AAAI Press / International Joint Conferences on Artificial Intelligence, 2016;

[29] H. Ingimundardottir and TP Runarsson, “Supervised learning linear priority dispatch rules for job-shop scheduling,” in International conference on learning and intelligent optimization, Springer, 2011, pp. 263-277.

[0021] The algorithm is preferably executed in the form of a single-player game.

[0022] The combination of a Monte Carlo tree search framework based neural network and training of this neural network using supervised learning and self-play with reinforcement learning leads to an optimization that significantly outperforms the known optimizations in sheet metal processing. Preferred embodiments

[0023] The output in process step D) can be sent to a manufacturing execution system (MES). This allows the production plan to be implemented directly on the production machines.

[0024] The method according to the invention may comprise, in addition to those already mentioned, one or more of the following process steps: c) Deburring of sheet metal parts; d) joining, in particular welding and / or soldering, the sheet metal parts; e) coating the sheet metal parts, in particular by painting and / or powder coating; f) Assembling the sheet metal parts.

[0025] Each of these process steps can be carried out by production machines and optimized by the method according to the invention.

[0026] In a preferred embodiment of the invention, the method according to the invention is implemented using the AlphaGo algorithm, and in a particularly preferred embodiment, using the AlphaGo Zero algorithm. In this case, the algorithm comprises the previously described Monte Carlo tree search framework with the neural network trained using supervised learning and self-play with reinforcement learning. AlphaGo or AlphaGo Zero has proven to be a very powerful algorithm for optimizing the production of sheet metal parts within the scope of the invention.

[0027] The AlphaGo Zero algorithm can be viewed on the following websites: • https: / / tmoer.github.io / AlphaZero / • https: / / towardsdatascience.com / alphazero-implementation-and-tutorial-f4324d65fdfc • https: / / medium.com / applied-data-science / how-to-build-your-own-alpha-zero-ai-using-python-and-keras-7f664945c188

[0028] AlphaGo bzw. AlphaGo Zero is preferably implemented in Python and / or Tensorflow. Further details on the implementation of AlphaGo bzw. AlphaGo Zero are eligible for the following promotions:

[30] D. Silver, A. Huang, CJ Maddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam, M. Lanctot, et al., „Mastering the game of go with deep neural networks and tree search“, nature, vol. 529, no. 7587, p. 484, 2016.

[31] G. Chaslot, S. Bakkes, I. Szita, and P. Spronck, „Monte-carlo tree search: A new framework for game ai.“, in AIIDE, 2008.

[32] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, et al., „Mastering the game of go without human knowledge“, Nature, vol. 550, no. 7676, p. 354, 2017.

[33] D. Silver, T. Hubert, J. Stiefwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, et al., “Mastering chess and shogi by self-play with a general reinforcement learning algorithm,” arXiv preprint arXiv:1712.01815, 2017.

[0029] The disclosure of all publications and websites cited herein is incorporated by reference in its entirety.

[0030] Further preferably, the training in process step A) is carried out using heuristically determined solutions of optimized production plans. This provides the neural network with a good starting point for its further optimization.

[0031] In particular, optimized production plans in the form of earliest-due-date (EDD) solutions can be used. These solutions have proven particularly advantageous, as rush orders often arise in practice, making prior production planning obsolete.

[0032] According to the invention, the optimization encompasses both waste minimization and production time optimization. This enables both fast and cost-effective, resource-efficient production. The goals of production time optimization are, in particular, the minimum overall delay and / or the minimum overall production time.

[0033] The constraints in process step B) can include the production deadlines for the sheet metal parts. Production time optimization can then take compliance with these deadlines into account. Compliance with these deadlines can be given higher priority than other objectives.

[0034] Alternatively or additionally, the boundary conditions in process step B) can include the values, i.e., the monetary values ​​or prices, of the sheet metal parts. This allows production to be optimized depending on the values ​​of the respective sheet metal parts. In general, this allows the value of a sheet metal part, for example, the price of its delayed production, to be qualified within the framework of the optimization according to the invention.

[0035] Further preferably, a waste score is assigned to the waste, and a production deadline score is assigned to the achievement of the production deadline, based on the value of the sheet metal parts, with the optimization minimizing both the waste score and the production deadline score. By assigning the scores, production time minimization can be treated or optimized on the same scale as waste minimization.

[0036] In this case, the decision node preferably stores the estimated maximum achievable total score value; the probability (= weighting) that the respective decision of the decision node is the best is preferably stored on the edges connecting the decision nodes.

[0037] For example, the waste score and the production deadline score can be used as a price. Then the price of waste material can be weighed against the price of a sheet metal part produced late.

[0038] The process can be optimized according to the following function: rabs:=−c(W)+∑i=110vI−λ max{0,Ti}

[0039] Where c(W) represents the value for the total material used (including offcuts, i.e. waste), T i and v irepresent the delay and the value of order part i, respectively. λ is a parameter that penalizes delay. r abs reflects the sum of the sheet metal parts, each reduced proportionally to production deadlines, minus the total material costs. Using this formula, a reward of the neural network can be generated, in particular scaled to [0, 1], where the maximum possible score r max (without delay and without waste).

[0040] The process steps B) to D) can be triggered as required by the occurrence of an event, whereby the event is read in via an event interface.

[0041] Preferably, the event takes the form of a request for further processing of a sheet metal part, in the form of freed-up production machine capacity, in the form of a production machine breakdown and / or in the form of an urgent order.

[0042] The event can be triggered automatically and read in via the event interface. The event is particularly preferably triggered by a production machine, an indoor localization system, and / or a manufacturing execution system and read in via the event interface. In the case of an indoor localization system, planning can be further optimized automatically using events transmitted by the indoor localization system's tags.

[0043] To further improve the neural network, a user evaluation of the production plan output in process step D) can be read in a process step E).

[0044] The invention further relates to a method for producing sheet metal parts, in which a previously mentioned method is carried out and then the process steps a) and b) are carried out on the basis of the optimized production plan.

[0045] In the process for manufacturing sheet metal parts, after process steps a) and b), process steps c), d), e) and / or f) can be carried out on the basis of the optimized production plan.

[0046] The object of the invention is further achieved by a device for carrying out a method described here, wherein the device has a computer for storing and executing the neural network, a boundary condition interface for reading in the boundary conditions and a production plan interface for outputting the production plan.

[0047] A user rating interface can be provided for reading user ratings. The neural network can be cloud-based to facilitate training with user ratings, especially anonymized ones.

[0048] The device according to the invention can have the event interface and further comprise a production machine, an indoor localization system (with multiple tags that transmit events), and / or a manufacturing execution system, wherein an event triggered by the production machine, the indoor localization system, and / or the manufacturing execution system can be read in via the event interface. In this case, the device can be optimized in an automated or semi-automated manner.

[0049] Further advantages of the invention will become apparent from the description and the drawings. Likewise, the above-mentioned and further-described features can be used individually or in combination in any desired manner. The embodiments shown and described are not intended to be exhaustive, but rather are exemplary in nature for describing the invention. Detailed description of the invention and drawing Fig. 1 shows a schematic of the production process for sheet metal parts. Fig. Figure 2 shows schematically the optimization of the production process.

[0050] Fig. 1 shows a schematic of the production of various orders. Fig. 1 are exemplary orders A 01 to A 10 shown. The orders A 01 -A 10 include the manufacture of products P 01 to P 10 which are made of several, especially different, sheet metal parts with their respective geometric data. For reasons of clarity, Fig. 1 only the sheet metal parts B1 and B2 are provided with a reference symbol.

[0051] As shown by clock symbols in Fig. As indicated in Figure 1, the individual sheet metal parts B1 and B2 have different production times. Furthermore, orders A 01 to A10 different production deadlines F 01 to F 10 Piggy banks indicate that sheet metal parts B1 and B2 have different (monetary) values. The described specifications represent boundary conditions 10 for sheet metal parts B1 and B2.

[0052] The sheet metal parts B1, B2 are arranged on a sheet metal panel 12 in such a way that the waste is minimal. Fig. 1, this can lead to the mixing of sheet metal parts B1, B2 of different orders A 01 -A 10 The sheet metal parts B1, B2 are machined on 14 production machines, of which Fig. 1 Production machines c1, c2 (cut) for cutting and separating, production machines b1, b2 (bend) for bending and production machines a1, a2 (assemble) for assembling the sheet metal parts B1, B2 are shown. In addition, further Fig. 1, production machines 14 for processing the sheet metal parts B1, B2, for example for deburring, joining and / or coating the sheet metal parts B1, B2, may be provided. The finished products comprising the sheet metal parts B1, B2 are in Fig. 1 at reference number 16.

[0053] The distribution of the sheet metal parts B1, B2 to the production machines 14 represents a highly complex problem given the different boundary conditions 10 of the sheet metal parts B1, B2. This is particularly the case since the individual process steps take different lengths of time, production machines 14 may fail, and / or rush orders may be received.

[0054] The inventive optimization of the production process is in Fig. 2 shown. Fig. 2 shows a device 18 for optimized production or optimized production planning of the sheet metal parts B1, B2 from Fig. 1. For this purpose, an algorithm 20 is provided. The algorithm 20 is preferably available as AlphaGo or AlphaGo Zero. The algorithm 20 includes a Monte Carlo tree search framework 22. The Monte Carlo tree search framework 22 is modified by a neural network 24. Supervised learning is initially performed, i.e., training based on heuristically determined problem solutions.

[0055] This is followed by self-play with reinforcement learning as a single-player game. This is Fig. 2 in steps 26 (selection), 28 (expansion), 30 (simulation), and 32 (backpropagation). In step 26, a decision path is selected via specific decision nodes; in step 28, the decision tree is randomly expanded with the decision nodes; the resulting result is simulated in step 30; and the decision nodes are reweighted (strengthened or weakened) based on this simulation result in step 32. Steps 26 to 32 are repeated several times.

[0056] The determination of the most optimal distribution of production steps is preferably carried out with regard to both waste minimization (nesting) and production time optimization (scheduling). This process can be described as optimization by a nesting agent and a scheduling agent, in which the agents make decisions in a simulation environment and receive a reward based on the quality of their decisions. The simulation is a representation of sheet metal production.

[0057] The optimized production plan is output via a production plan interface 34, in particular to a manufacturing execution system 36. The manufacturing execution system 36 controls the production machines 14, i.e., the actual sheet metal production, with the optimized production plan.

[0058] The boundary conditions 10 are fed to the algorithm 20 via a boundary condition interface 38. User ratings 40 can be fed to the algorithm 20 via a user rating interface 42.

[0059] Alternatively or additionally, an event interface 44 can be provided, via which an event 46 can be read. The event 46 can be triggered by the manufacturing execution system 36, one or more production machines 14 and / or an indoor localization system 48. The event 46 can include, for example, a failure of a production machine 14, freed-up capacity of a production machine 14, errors in production, new orders and / or order changes. In particular, the event 46 includes the further production planning for a sheet metal part B1, B2 (see Fig. 1) that has just completed a production step in a production machine 14.

[0060] The algorithm 20 is executed on a computer 50. The computer 50 may be cloud-based to facilitate the use of user ratings 40 from different users. The manufacturing execution system 36 may (as indicated) be executed on the same computer or a different computer.

[0061] Taking a summary of all figures of the drawing, the invention relates in summary to a method for optimizing the production of sheet metal parts B1, B2. The method optimizes the allocation of sheet metal parts B1, B2 for processing on various production machines 14 and outputs an optimized production plan. For this purpose, an algorithm 20 is provided which has a decision tree in the form of a Monte Carlo tree search framework 22 and a neural network 24. The algorithm 20 is trained with each new query through self-play and reinforcement learning. Pre-training of the algorithm 20 is achieved through supervised learning. The algorithm 20 preferably optimizes the production plan primarily with regard to minimally delayed production deadlines F 01 to F 10of the sheet metal parts B1, B2, and secondarily with regard to minimal waste. By assigning scores, both objectives can be evaluated jointly. The method can include the receipt of query-triggering events 46 and / or the operation of production machines 14 according to the production plan. The invention further relates to a device 18 for carrying out the method. List of reference symbols A 01 to A 10 Orders P 01 to P 10 Products B1, B2 sheet metal parts F 01 to F 10 Production deadlines c1, c2 cutting production machines b1, b2 bending production machines a1, a2 assembly production machines 10 boundary conditions 12 metal plate 14 production machines 16 products 18 Device 20 Algorithm 22 Monte Carlo tree search framework 24 neural network 26 Step - selection 28 Step - expansion 30 step simulation 32 step - backpropagation 34 Production plan interface 36 manufacturing execution system 38 Boundary condition interface 40 user reviews 42 User rating interface 44 Event interface 46 Event 48 Indoor Localization System 50 computers

Claims

[1] Procedure for optimising the production of sheet metal parts (B1, B2) with the process steps: a) Cutting and separating the sheet metal parts (B1, B2); b) bending the sheet metal parts (B1, B2); the method comprising the following steps: A) training a neural network (24) executed on a Monte Carlo tree search framework (22) by means of supervised learning and self-play with reinforcement learning; B) detecting boundary conditions (10) of the sheet metal parts (B1, B2), wherein the boundary conditions (10) comprise geometric data of the sheet metal parts (B1, B2); C) creating an optimized production plan by the neural network (24), wherein the optimization includes both waste minimization and production time optimization; D) Issue of the production plan. [2] A method according to claim 1, wherein the method comprises one or more of the following process steps: c) deburring the sheet metal parts (B1, B2); d) joining the sheet metal parts (B1, B2); e) coating the sheet metal parts (B1, B2); f) Assemble the sheet metal parts (B1, B2). [3] Method according to claim 1 or 2, wherein the method steps A) to D) are carried out with an algorithm (20), wherein the algorithm (20) is based on AlphaGo or AlphaGo Zero and wherein the algorithm (20) comprises the neural network. [4] Method according to one of the preceding claims, in which the training in method step A) is carried out with heuristically determined solutions of optimized production plans. [5] Method according to claim 4, in which optimized production plans are used in the form of earliest-due-date solutions. [6] Method according to one of the preceding claims, in which the boundary conditions (10) in method step B) additionally comprise the production deadlines of the sheet metal parts (B1, B2). [7] Method according to claim 6, wherein the boundary conditions (10) in method step B) additionally comprise the values ​​of the sheet metal parts (B1, B2). [8] Method according to claim 7, wherein the waste is assigned a waste score and the achievement of the production deadline is assigned a production deadline score based on the value of the sheet metal parts (B1, B2), wherein the optimization minimizes both the waste score and the production deadline score. [9] Method according to one of the preceding claims, in which the method steps B) to D) are carried out in an event-triggered manner, the reading of the event (46) being carried out via an event interface (44). [10] Method according to claim 9, wherein the event (46) is in the form of a request for further processing of a sheet metal part (B1, B2), in the form of freed-up production machine capacity, in the form of a production machine failure and / or in the form of a rush order. [11] Method according to claim 9 or 10, wherein the event (46) is triggered by a production machine (14), an indoor localization system (48) and / or a manufacturing execution system (36) and is read in via the event interface (44). [12] Method according to one of the preceding claims, in which in a method step E) a user evaluation (40) of the production plan output in method step D) is read in and the neural network (24) is further trained with the user evaluation (40). [13] Device (18) for carrying out a method according to one of the preceding claims, wherein the device (18) comprises a computer (50) for storing and executing the neural network (24), a boundary condition interface (38) for reading in the boundary conditions (10) and a production plan interface (34) for outputting the production plan. [14] Device according to claim 13 in conjunction with claim 11, wherein the device (18) has the event interface (44) and the device (18) further comprises a production machine (14), an indoor localization system (48) and / or a manufacturing execution system (36), wherein an event (46) triggered by the production machine (14), the indoor localization system (48) and / or the manufacturing execution system (36) can be read in via the event interface (44).

Images