Method for controlling a packaging machine, and packaging machine

Abstract

The present invention relates to a method for controlling a packaging machine for producing packages, the machine having a group of drives, in particular designed as servo drives, which each drive a machine element of the packaging machine and intersect in an overlapping region in which the machine elements could collide with one another, with products handled by the machine elements and / or with stationary machine parts, wherein the following measures are carried out: • a) inputting input data into a trained neural network; • b) calculating and outputting output data by means of the trained neural network, which data comprise data of individual movements for the drives, wherein the trained neural network, at least for each drive which is not in a predetermined end position, selects one or more individual movements from a group of individual movements assigned to this drive, each individual movement having a defined target position.

Description

[0001] METHOD FOR CONTROLLING A PACKAGING MACHINE AND PACKAGING MACHINE

[0002] Description

[0003] The present invention relates to a method for the (automatic) control of a packaging machine for producing packages, preferably for producing packaging for smokable products in the tobacco industry, comprising a group of drives, particularly designed as servo drives, each of which drives a machine element of the packaging machine and intersects in an intersection area where the machine elements could collide with each other, with products handled by the machine elements, and / or with stationary machine parts. Furthermore, the invention relates to a packaging machine for producing such packages, which can be controlled using the aforementioned method.

[0004] Packaging machines often have multiple groups of machine parts, especially in different areas of the machine, where, among other things, the movement paths of the members of the respective group can intersect in an overlapping area. On the other hand, in recent years, the mechanical (motion) coupling of individual moving machine parts, such as slides, conveyors, turrets, etc., to a rotating main or master shaft has increasingly been abandoned in packaging machines in favor of individual servo drives for these machine parts, electronically coupled to a machine cycle.

[0005] However, the use of separate servo drives for the individual moving machine parts naturally increases the risk of collisions between a first moving machine part or a product handled by such a part and a second moving machine part or a product handled by this part or a stationary machine part. Particular problems arise when synchronizing the individual drives and thus the machine parts they move or the products they handle. In particular, before the packaging machine is put into operation for the first time or is put back into operation, it can happen that the drives or the machine parts they drive are in undefined, non-synchronized positions relative to one another. To ensure subsequent collision-free (normal) operation of the machine, it is often necessary to align the machine parts beforehand and to move them into defined starting positions or return positions.Due to the fundamentally overlapping movement paths of the machine components, there is a significant risk of unwanted collisions between the machine components, so such alignment is often performed manually.

[0006] Until now, this drive synchronization has generally required individual planning and programming of the machine's control system. This involves planning and programming a multitude of conceivable motion sequences and initial position configurations separately in advance, taking into account potential collision situations in the overlap area, and storing them in the machine control system. This individual planning and programming is very time-consuming and complex. Furthermore, minor errors in the program code could go undetected and then lead to collisions in the machine during execution. Furthermore, with each additional servo drive, the programming effort increases exponentially if new or different overlaps arise in the movement paths of the machine components moved by the servo drives or the products they handle.

[0007] It is the object of the present invention to further develop the control method mentioned above and the packaging machine mentioned above.

[0008] This object is achieved by a control method having the features of claim 1 and a packaging machine having the features of claim 19.

[0009] Accordingly, the method according to the invention is characterized in that, in particular, but not only, for collision-free synchronization of the machine elements with one another, preferably during commissioning or maintenance of the packaging machine, the following measures are carried out: a) Input of input data into a trained (artificial) neural network, wherein the input data comprise at least position data representing the positions of the drives, in particular simulated actual positions or actual positions detected during operation of the packaging machine, and / or data indicating whether a hypothetical movement of the respective drive from a calculated or detected position of the drive into its assigned, predetermined end position or into at least one intermediate position deviating from the predetermined end position, in particular into the safe position, would lead to a collision or not,b) Calculating and outputting output data by the trained (artificial) neural network, which comprises data from (predefined) individual movements for the drives, wherein the trained neural network selects one or more individual movements from a group of individual movements assigned to this drive, each with a predefined target position, at least for each drive that is not in a predetermined end position, in particular in a synchronous position of the respective drive, in particular such that all of these drives can be transferred collision-free to their respective predetermined end position, preferably to their respective synchronous position, when the drives are moved according to the selected individual movements.

[0010] The starting point of the invention is therefore to provide and assign to each drive a defined number of individual movements, each with a fixed target position, which the drive would execute from its starting position to the target position if a control device or the control device of the packaging machine moves it in accordance with the respective individual movement.

[0011] If the control method according to the invention is to be used, for example, in accordance with the main application of the invention (although this is not the only possible one) for synchronizing the drives, a first individual movement could be provided which is assigned to a drive 1 and transfers this by rotation in a first, positive direction of rotation into its synchronous position - in this case forming its end position, and a second individual movement which transfers the drive 1 by rotation in a second, opposite, negative direction of rotation into its synchronous position. Furthermore, for example, a third and a fourth individual movement which transfer the drive 1 by rotation into positive and negative directions respectively.negative direction of rotation into a safe position, which is defined in such a way that, if the respective drive is in such a position, any other drive can assume any drive position within its respective set of possible drive positions without the machine parts colliding with each other or with the products they handle. In the same way as drive 1, several such individual movements could also be assigned to drive 2, drive 3, etc.

[0012] As is easily apparent, even with just two drives, starting from a multitude of different starting positions in which the drives can be located, for example, after a machine stop, there are a multitude of combinations of such individual movements which, when implemented, either successfully transfer the drives to the respective synchronous position without collision or, alternatively, cause the drives to collide (unintentionally).

[0013] The more drives there are in the respective constellation, the greater the number of possible combinations of individual movements starting from possible starting positions of the respective drives and the more complex the task would be to manually determine one or more combinations that lead to collision-free synchronization.

[0014] According to the invention, this task is performed by the appropriately trained (artificial) neural network. It is trained in such a way that it can select one or more of the individual movements assigned to the respective drive from the multitude of possible combinations, at least for the drives that are not yet in their synchronous position or, more generally, in their respective intended end position. Preferably, this can be done in such a way that all of these drives are transferred to their respective end or, if applicable, synchronous position without collision when the drives are moved according to the selected individual movements.In the further process, the drives can then be controlled accordingly, preferably by means of a control device of the packaging machine that has or is assigned to the trained neural network and processes the output data, based on the output data, and each can be moved to its predetermined end position, in particular its synchronous position. This is preferably achieved by moving the drives to their predetermined end position according to the selected individual movements, for example, in the temporal sequence determined by the neural network or in the derived or derivable chronological order of the individual movements.

[0015] With regard to the output data generated by the trained neural network, it can be provided that the trained neural network sequentially selects the individual movements of the groups assigned to the drives and outputs them as output data. In particular, by selecting at least one individual movement from at least one group of individual movements in successive calculation runs based on the input data for each calculation run. This is preferably done using input data updated before each subsequent calculation run, in particular on the basis of the output data of the previous calculation run; wherein it is at least conceivable to alternatively provide for the output of the selected individual movements as output data, without such updated input data being used between or for the individual subsequent calculation runs.

[0016] The input data for each subsequent calculation run can expediently comprise position data for at least one drive, which represent its calculated or detected position, which it would have or has if it were to be or is moved in accordance with the individual movement selected for it in the previous calculation run.

[0017] As for the above-mentioned end position of the respective drive, this is a (target) position in which the respective drive should be after completion of the respective control objective of the method according to the invention. If the control objective is to move the drives into their respective synchronous positions without collision, the end position of each drive would accordingly be the respective synchronous position of the respective drive. As for the above-mentioned defined target position of the respective individual movement, this can generally speaking be any (target) position into which the respective drive is transferred from its starting position by the control system when it is moved accordingly or in accordance with the individual movement. The target position can coincide with the end position, but this does not have to be the case. It can also be an intermediate position, for example.

[0018] For example, if the target position coincides with the synchronous position, the respective drive would be moved from its current starting position to this synchronous position if the controller is moved according to the single movement defined in this way. In this case, if the goal of the controller is to synchronize the drives, the target position of the drive would therefore coincide with the final position of the drive, which it should be in at the end of the process.

[0019] The defined target position of one or more individual movements of one or more groups of individual movements can, for example, also be the safe position mentioned above. If the safe position does not correspond to the end position, which is often the case, the respective drive would first be moved to the safe position if an individual movement with such a target position is selected and the controller implements this individual movement. This would then be followed, for example, by a single movement whose target position is the end position or, in this case, the synchronous position.

[0020] The trained neural network can also theoretically explicitly determine a temporal sequence for the individual movements of the drives, for example, in list form, in which these movements must be implemented sequentially by the drives in order to transfer these drives to their respective predetermined end positions, in particular to their synchronous positions, without collisions. Alternatively—and this is usually the case—it can also be provided that such a temporal sequence can be (indirectly) derived or is derived from the mere temporal sequence of the individual movements for the drives output by the trained neural network.As for the input data for the trained neural network, these may include data indicating, within the workspace in which the machine parts moved or can move by the drives, the presence or absence and / or the position and / or dimensions of products processed in the packaging machine and / or of, in particular, stationary machine components. In particular, data originating from one or more sensors that detect the presence or absence and / or the position and / or dimensions of the products or machine components.

[0021] As far as the technical basis of the trained neural network or the (artificial) neural network on which the trained neural network is based is concerned, it can comprise a multi-layer, in particular three-layer neural network with a layer of input neurons, possibly having several sublayers, a layer of hidden neurons, possibly having several sublayers, and a layer of output neurons, possibly having several sublayers.

[0022] In order to obtain the trained (artificial) neural network used according to the invention, the method according to the invention preferably comprises the following suitable steps for training an (artificial) neural network, in particular using the technique of deep reinforcement learning: a) inputting input data into an (untrained) or the neural network, wherein the input data comprise at least position data representing calculated or detected positions of the drives, and / or data indicating whether a hypothetical movement of the respective drive from a calculated or detected position of the drive to its assigned, predetermined end position or to at least one intermediate position deviating from the predetermined end position, in particular to the safe position, would lead to a collision or not, b) inputting output data into the neural network,which comprise, for each drive, data from a plurality of individual movements, each with a defined target position, which are each assigned to a group of individual movements assigned to the respective drive, in particular individual movements whose target position is a predetermined end position, preferably a synchronous position of the respective drive, and / or individual movements whose target position is a safe position defined in such a way that, in the event that the respective drive is in such a position, any other drive can assume any drive position within its respective set of possible drive positions without the machine elements colliding with each other or with products handled by them, c) training the neural network based on the input data and the output data,so that for each drive, it can select one or more individual movements from the group of individual movements assigned to it, each with a defined target position, with the proviso that all drives are transferred to their respective predetermined end position without collision when the drives are moved according to the selected individual movements.

[0023] The training may further comprise: a) calculating output data comprising one or more individual movements for each drive based on the input data and b) rewarding calculated output data if they comprise a number of individual movements with which all drives that are not in their predetermined end position can be transferred to their respective predetermined end position without collision.

[0024] In this case, it can further be provided within the scope of training that calculated output data are rewarded less during training if the number of individual movements includes an individual movement with which a drive is first moved into a safe position before it is moved into the predetermined end position, than calculated output data which do not include such an individual movement.

[0025] Furthermore, it can be provided that calculated output data are rewarded less during training if the number of individual movements includes one or more individual movements that lead to a collision when the drives are moved according to the selected individual movements than calculated output data that does not include one or more such individual movements.

[0026] The training of the neural network can be carried out entirely or partially with the help of a digital twin of the packaging machine that simulates at least the drives and the machine parts they move.

[0027] For training the neural network, positions can be automatically selected from a set of possible positions for the drives, in particular by means of a random generator, and made available to the neural network as input data.

[0028] Furthermore, it can be provided that, in the event that during operation of the packaging machine the drives are in actual positions for which the trained neural network cannot calculate and output output data comprising individual movements with which the or all drives can be transferred to their respective predetermined end position without collision, position data representing these actual positions are stored in a data memory, in particular in a data cloud.

[0029] The trained neural network can then be retrained on the basis of these stored position data, in particular with the help of the digital twin of the packaging machine, whereby the drives are then controlled during operation of the packaging machine on the basis of the output data of the retrained neural network and are each transferred to their predetermined end position.

[0030] The retrained neural network can then be transmitted back to the control device, so that the control device processes or can process the output data of the retrained neural network during operation of the packaging machine to control the drives and to transfer the drives to the respective predetermined end position. Further features of the present invention emerge from the appended patent claims, the following description of preferred embodiments of the invention and from the appended drawings.

[0031] It shows:

[0032] Fig. 1 shows a section of a packaging machine for producing packs for tobacco products, controlled according to the invention using a trained neural network, with several independent drives that move machine elements, wherein in a first overlapping area the movement paths of two machine elements overlap and in a second overlapping area the movement path of a machine element overlaps with the movement path of products that are moved by a machine element,

[0033] Fig. 2 a side view of the packaging machine along the

[0034] Viewing direction II in Fig. 1,

[0035] Fig. 3 is a schematic representation of the interaction of a

[0036] Computing device with digital twin of a packaging machine with a data cloud and the control device of the real packaging machine,

[0037] Fig. 4 exemplary groups of individual movements A1 to B4 for selected drives A and B of the packaging machine, from which the trained neural network selects individual movements to control the packaging machine,

[0038] Fig. 5 an exemplary list of possible input data EA-A1 to EA-B6 for the trained neural network,

[0039] Fig. 6 shows an overview of a training procedure for generating a trained neural network, Fig. 7 shows real data of hyperparameters optimized as part of a training procedure of a neural network,

[0040] Fig. 8 Information about the programs, libraries, formulas, etc. used to generate and / or train the neural network,

[0041] Fig. 9 Reward and success curves generated during training of the neural network.

[0042] The above-described relationships according to the invention will be explained below using a specific packaging machine (Figs. 1-2), namely a packaging machine 10 for packaging cigarettes 11 or for producing cigarette packs containing cigarettes 11, which is shown only in detail for this purpose. Packaging machines for cigarettes or other products are known to those skilled in the art and are therefore not described in detail here. It is understood that the control method according to the invention can also be applied to other types of packaging devices or packaging machines.

[0043] Furthermore, it is understood that the method according to the invention can be applied in all areas of a packaging machine in which two or more machine elements of the packaging machine driven by (electric motor) drives move on movement paths in a working space in which, in an overlapping area, collisions can occur between a moving machine element and another moving machine element or with products handled by the machine elements or with stationary machine parts.

[0044] Fig. 1 shows a part of the packaging machine 10, in which, in a first area, cigarettes 11 are pushed or slid in groups and in cycles from a cigarette magazine 13 into machine elements held in precise positions by machine elements designed as pushers 12, which are moved back and forth by a drive C in line with the machine cycle and are in this case designed as pockets 14 of a circulating pocket conveyor 15. The path of movement of the cigarette groups and the path of movement of the pockets 14 intersect in an overlap and transfer area 18, in which, for example, if the pockets 14 are incorrectly aligned in an inaccurate position, for example if they are laterally offset too far, the respective cigarette group may collide with one of the pocket walls of the pockets 14.

[0045] The pocket conveyor 15 then conveys the cigarette groups in cycles in the direction of a circulating carrier conveyor 16, which has individual machine elements moved by a drive B, in this case designed as carriers 17, which move on a conveyor track running transversely to the conveying plane of the pockets 14.

[0046] In a further overlap and transfer area 22, in which the conveying or movement path of the pockets 14 and the conveying or movement path of the drivers 17 intersect, an upper driver part 21 of one of the drivers 17, which is adapted to the inner contour of the respective pocket 14, is moved lengthwise through a respective pocket 14 that is held ready and in the process conveys the group of cigarettes located in the pocket 14 (taking the same along with it) transversely to the conveying plane of the pockets 14 out of the pocket 14 in the direction of a downstream wrapping station 19, in which an inner blank 20 is then placed on the group of cigarettes and folded around it. Collisions can also occur in the overlap and transfer area 22, for example a collision of a driver 17 with a held ready pocket 14 or a pocket wall thereof if the pockets 14 are incorrectly aligned in an inaccurate position relative to the drivers 17 orthe upper driver parts 21.

[0047] The drives A, B, C can be controlled individually and are each designed as servo drives with servo motor and corresponding position control.

[0048] The pushers 12, the pockets 14 (if necessary, also the entire pocket conveyor 15), and the carriers 17 (if necessary, also the entire carrier conveyor 16) can form the machine elements moved by the drives C, A, and B, respectively, which move along intersecting paths. Drives A, B form a first group of drives of the packaging machine 10, whose machine elements driven by them intersect in an overlapping area and which may need to be synchronized after a standstill of the machine 10, in particular before its initial start-up and / or restart. Drives A, C accordingly form a second such group.

[0049] Naturally, the invention is not limited to the specific machine components mentioned above. Rather, the invention encompasses all conceivable machine components moved by drives whose movement paths intersect in an overlapping area.

[0050] As mentioned, particular problems could arise, particularly but not exclusively, during initial commissioning or recommissioning of the packaging machine 10. For example, it may happen that the drives A, B, C and accordingly the machine elements 12, 14, 17 are in undefined, unsynchronized positions relative to one another. This makes it necessary to synchronize the machine elements 12, 14, 17 or the drives A, B, C that move them, before the machine starts normal operation, namely to move them into their predetermined synchronous positions in order to subsequently prevent collisions during normal operation of the machine and / or during commissioning. For subsequent collision-free (regular) operation of the packaging machine 10, it is therefore necessary in such a situation to align the machine elements 12, 14, 17 beforehand and to transfer them to defined starting positions or synchronous positions.

[0051] However, due to the fundamentally overlapping movement paths of the machine elements 12, 14 and 14, 17 in the respective group, there is a considerable risk of undesirable collisions between the machine elements 12, 14 and 14, 17 during such an alignment process.

[0052] In order to enable the most reliable transfer of the machine elements 12, 14 and 14, 17 into their synchronized positions, the packaging machine 10 is therefore controlled in a special manner in a separate, automatic synchronization mode using a trained artificial neural network prior to regular normal operation of the packaging machine 10. In this case, a or the corresponding control device of the packaging machine 10 comprises the trained neural network. This can be a multi-layered, in particular three-layered neural network in a known manner, with a layer of input neurons, possibly having several sublayers, a layer of hidden neurons, possibly having several sublayers, and a layer of output neurons, possibly having several sublayers.

[0053] This will be explained in more detail below using a selected group of drives as an example, namely drives A and B. It goes without saying that these explanations also apply to the other overlapping drive combinations.

[0054] First, in the synchronization operation according to the invention, the actual positions of the drives A and B are queried by the controller or control device (not shown) of the packaging machine 10, which has the trained neural network, namely in this case (by means of suitable rotary encoders) the angles of rotation of the servo motors of the individual drives A, B.

[0055] Based on the actual positions, it is possible to check whether both drives A and B are already in their respective assigned synchronous positions by comparing them with the synchronous positions stored in a data storage device 25 assigned to the controller (in this case, a cloud storage device, see Fig. 3). These positions represent the target positions or target angles of rotation for drives A and B. If this is the case, normal operation of the packaging machine 10 could begin immediately.

[0056] However, if at least one drive A or B is not yet in a synchronous position, this is not the case.

[0057] If, for example, both drives are not in their respective synchronous position, the control system uses the trained neural network to determine, in general terms, travel paths for the drives A, B, with the aid of which the drives A, B can be moved into their respective synchronous position without collision, without the machine elements 14 and 17 moved by them colliding with each other or with the products they handle.

[0058] For this purpose, in the present example, input data ED-A1 to ED-A7 or ED-B1 to ED-B7 are first transmitted to the trained neural network, see Fig. 5, which include, among other things, the current actual positions of the drives A and B, see ED-A1 or ED-B1.

[0059] Furthermore, status information is also transmitted to the trained neural network, such as information about whether the respective drive A or B is in its synchronous position, see ED-A2 or ED-B2. It would also be conceivable for one or both drives A and B to be in a predetermined safe position, see ED-A3 or ED-B3. By definition, a safe position of drive positions is any position (or any angle of rotation / any range of angles of rotation) for which the following applies: if the respective drive A or B is in such a position, the other drive A or B can assume any drive position within its respective set of possible drive positions without the machine elements 14 or 17 colliding with each other or with the products (cigarettes 11) handled by them.

[0060] The above information can be obtained by simply comparing the real actual position of drives A and B with the position data stored for the synchronous position or for the safe position.

[0061] The status information can also include information about whether, based on the current actual position of the respective drive, drive A or drive B can be moved to its respective safe or synchronous position when hypothetically moving it in the positive or negative direction of rotation or not (yes / no) without causing collisions. Such information can be obtained using conventional collision calculations, see ED-A4 to ED-A7 or ED-B4 to ED-B7.

[0062] Based on such input data, the trained neural network then selects at least one or more suitable individual movements for each drive from groups A' and B' of individual movements A1, A2, A3, A4 or B1, B2, B3 and B4, see Fig. 4, which are each assigned to drive A or drive B, if necessary sequentially in several calculation runs, so that as a result all drives A and B are transferred collision-free to their respective predetermined end position, preferably to their respective synchronous position, when the drives A, B are moved according to the selected individual movements.In the present case, the trained neural network could, for example, in a first calculation run for drive A, select individual movement A3 from the individual movements A1 to A4 assigned to it. This would result in drive A and with it the machine element 14 moved by it being moved from its initial position to its safe position, so that drive B can then, by definition, move freely together with its machine element 17 without any collisions with the machine element 14 or drive A. After selecting this individual movement A3, the control device will then generally move drive A immediately to the safe position before further individual movements are selected. However, this is not necessary.

[0063] In Fig. 1, for example, the pocket 14 a or the drive A is initially not in its synchronous position, since the pocket 14 a is arranged laterally offset by a (slight) distance to the path of movement of the driver 17 a, so that the driver 17 a would collide with the pocket 14 a if the drive B or the driver 17 a were to be moved accordingly to the pocket 14 a. A lateral movement of the pocket 14 a caused by the drive A, which reduces this lateral offset to zero, would then move the pocket 14 a or its drive A into the safe position so that the driver 17 a can move through the pocket 14 a without collision.

[0064] The new actual position of drive A after it has been moved to the safe position, as well as any other input data ED-AI to ED-B7 updated based on this new actual position, can then be fed back to the trained neural network, so that it performs another calculation run based on the updated input data ED-A1 to ED-B7 and selects a suitable individual movement from the individual movements B1 to B4 for drive B. Ideally, this could then be individual movement B1 or B2, for example, with which drive B is moved to its synchronous position.

[0065] In a further calculation run, the trained neural network would then, based on the newly updated input data ED-A1 to ED-B7 for drive A, select, for example, the individual movement A1, which would then transfer this drive A to its synchronous position.

[0066] As will be apparent to a person skilled in the art, depending on the multitude of input data ED-A1 to ED-B7 that depend on the respective starting positions, a wide variety of variants of combinations of individual movements A1 to B4 for the drives A and B are conceivable, which can guide the drives A and B into their end positions or synchronous positions without collision.

[0067] In order to ensure collision-free transfer of drives A and B to their respective end positions or synchronous positions by means of the possibly several individual movements, the order in which the individual movements are carried out usually also plays a role.

[0068] This sequence can, for example, be indirectly derived from the (temporal) sequence in which the trained neural network outputs the individual movements A1 to B4 one after the other after each calculation run. For example, if the trained neural network selects individual movement A3 in a first run (for drive A) as described above, then individual movement B1 for drive B in a second calculation run, and finally individual movement A1 for drive A in a third calculation run, the sequence A3 - B1 - A1, in which drives A and B are moved one after the other, can result directly.

[0069] Theoretically, however, it is also conceivable, for example, that the trained neural network immediately outputs all individual movements A3, B1, A1 after only one run and then specifies the sequence in which the individual movements A3, B1, A1 are to be executed, for example separately.

[0070] It is naturally particularly important to train the trained neural network before use in a control operation as described above in such a way that it learns, based on or depending on the possible combinations of input data ED-A1 to ED-B7, to select from the predefined individual movements A1 to A4 or B1 or B4 of the individual drives A or B, with which these are moved either to their safe position or to their synchronous position, that combination of individual movements A1 to B4 with which this is achieved - if necessary in a predetermined sequence - without collision.

[0071] For this purpose, according to the invention, a so-called deep reinforcement learning of the still untrained or possibly still incompletely or insufficiently trained neural network can be carried out.

[0072] As is known to those skilled in the art, during such training, the neural network interacts with its environment by repeatedly feeding it data (in iterations) and receives rewards and / or punishments based on the actions it selects based on the data. Through these repetitions, the neural network learns to classify and associate the data it receives with the actions it selects more precisely each time.

[0073] This works by continually adjusting the weighting of the individual connections between the neural network's neuron layers. The model generated in the learning runs can then be regularly applied to data that the neural network hasn't yet encountered during training.

[0074] In this case, input data ED-A1 to ED-B7 are repeatedly fed to the neural network. Positions selected from a set of possible positions for the drives, for example, using a random generator, can then be repeatedly fed to the neural network. This and the other input data can also be generated using a digital twin 24 of the packaging machine 10 that simulates at least the drives A and B and the machine parts they move.

[0075] The respective selection of individual movements A1 to A4 or B1 to B4 from the respective group of individual movements A1-A4 or B1-B4 is then evaluated either by the computing device 23 based on predefined evaluation criteria or by a user to determine whether or not they lead to collision-free travel paths of the drives A, B into their respective synchronous positions. According to refined evaluation criteria, travel paths that require no or few intermediate movements of the drives A, B into safe positions can be rewarded more highly than travel paths that require such intermediate movements or that require more such intermediate movements.

[0076] The trained neural network could also be retrained in certain configurations. For example, if, during operation of the packaging machine 10, the drives A, B are in actual positions for which the trained neural network cannot calculate and output output data, the position data representing these actual positions could be stored in the data memory 25.

[0077] The trained neural network can then be retrained on the basis of these stored position data, in particular with the aid of the digital twin 24 of the packaging machine 10. The retrained neural network can then later be transmitted back to the control device, so that the control device processes or can process the output data of the retrained neural network during operation of the packaging machine 10 for controlling the drives A, B and for transferring the drives A, B to the respective predetermined end position.

[0078] The drives A, B can then be controlled accordingly during operation of the packaging machine 10 on the basis of the output data of the retrained neural network.

[0079] Only as a supplement and without claim to completeness, Fig. 6 shows an example of a process diagram from which further information can be obtained with regard to possible training processes, whereby an expert in the field of artificial intelligence or artificial neural networks will understand that the values ​​and processes mentioned there can naturally also be selected differently.

[0080] Figures 7 to 9 also show examples of real data used in practice to generate such a trained neural network, only for supplementary purposes and without any claim to completeness. Figure 7 shows so-called hyperparameters, which were used in this example to control the training algorithm.

[0081] Fig. 8 shows the programs, libraries, reward functions, layers and number of neurons used.

[0082] Fig. 9 shows reward and success curves depending on the number of iterations during training.

[0083] List of reference symbols

[0084] 10 packaging machines

[0085] 11 cigarettes

[0086] 12 kickers

[0087] 13 cigarette magazine

[0088] 14 pockets

[0089] 14 a bag

[0090] 15 pocket conveyors

[0091] 16 flight conveyors

[0092] 17 drivers

[0093] 17 a driver

[0094] 18 Overlap and takeover area

[0095] 19 changing station

[0096] 20 Internal cutting

[0097] 21 Driving part

[0098] 22 Overlap and takeover area

[0099] 23 Computing device

[0100] 24 digital twin

[0101] 25 data storage

[0102] A - C drives

Claims

Patent claims 1. A method for (automatic) controlling a packaging machine (10) for producing packages, in particular for producing packaging for smokable products in the packaging industry, comprising a group of drives (A, B, C), in particular designed as servo drives, each driving a machine element of the packaging machine (10) and intersecting in an intersection area (18, 22) in which the machine elements could collide with each other, with products (11) handled by the machine elements, and / or with stationary machine parts, wherein, in particular for collision-free synchronization of the machine elements with each other, preferably during commissioning or maintenance of the packaging machine (10), the following measures are carried out: a) inputting input data into a trained neural network, wherein the input data comprise at least position data representing the positions of the drives (A, B, C),in particular simulated actual positions or actual positions detected during operation of the packaging machine (10), and / or data which indicate whether a hypothetical movement of the respective drive (A, B, C) from a calculated or detected position of the drive (A, B, C) into the predetermined end position assigned to it or into at least one intermediate position deviating from the predetermined end position, in particular into the safe position, would lead to a collision or not, b) calculating and outputting output data by the trained neural network, which comprise data of individual movements for the drives (A, B, C), wherein the trained neural network calculates, at least for each drive (A, B, C) which is not in a predetermined end position, in particular in a synchronous position of the respective drive (A, B, C), from a position assigned to this drive (A, B,C) assigned group of individual movements, each with a defined target position, selects one or more individual movements, in particular in such a way that all of these drives (A, B, C) are moved collision-free into their respective predetermined end position, preferably into their respective synchronous position, can be transferred if the drives (A, B, C) are moved according to the selected individual movements.

2. Method according to claim 1, characterized in that the defined target position is a position into which the respective drive (A, B, C) is transferred when it is moved according to the individual movement.

3. Method according to claim 1 or 2, characterized in that the defined target position of one or more individual movements of one or more groups of individual movements is a safe position which is defined in such a way that, in the event that the respective drive (A, B, C) is in such a position, any other drive (A, B, C) can assume any drive position within its respective set of possible drive positions without the machine elements colliding with one another or with products handled by them, and / or that the defined target position of one or more individual movements of one or more groups of individual movements is the predetermined end position, in particular the synchronous position of the drive (A, B, C) assigned to the group in each case.

4. Method according to one or more of the preceding claims, characterized in that the trained neural network selects the individual movements of the groups sequentially and outputs them as output data, in particular by selecting at least one individual movement from at least one group of individual movements in successive calculation runs on the basis of the input data in each calculation run, preferably using input data updated before each subsequent calculation run, in particular on the basis of the output data of the previous calculation run.

5. Method according to claim 4, characterized in that the input data for the subsequent calculation run comprise position data for at least one drive (A, B, C) which represent its calculated or detected position which it would have or has if it were to be or is moved in accordance with the individual movement selected for it in the previous calculation run.

6. Method according to one or more of the preceding claims, characterized in that the trained neural network determines a chronological sequence for the individual movements for the drives (A, B, C), in which these must be implemented one after the other by the drives (A, B, C) in order to be able to transfer these drives (A, B, C) collision-free into their respective predetermined end position, in particular into their synchronous position, or in that a chronological sequence can be or is derived from the chronological sequence of the individual movements for the drives (A, B, C) output by the trained neural network, in which these must be implemented one after the other by the drives (A, B, C) in order to be able to transfer these drives (A, B, C) collision-free into their respective predetermined end position, in particular into their synchronous position.

7. Method according to one or more of the preceding claims, characterized in that the drives (A, B, C), preferably by means of a control device of the packaging machine (10) which has the trained neural network or is assigned thereto and processes the output data, are controlled on the basis of the output data and are each transferred into their predetermined end position, in particular their synchronous position, in particular by the drives (A, B, C) being moved into their predetermined end position in accordance with the selected individual movements, preferably in the temporal order determined by the neural network or in the derived or derivable temporal order of the individual movements.

8. Method according to one or more of the preceding claims, characterized in that the input data comprise data which, within the working space in which the machine elements moved by the drives (A, B, C) move or can move, characterize the presence or absence and / or the position and / or dimensions of products processed in the packaging machine (10) and / or of in particular stationary machine components, in particular data which originate from one or more sensors which detect the presence or absence and / or the position and / or dimensions of the products or the machine components.

9. Method according to one or more of the preceding claims, characterized in that the trained neural network comprises a multi-layer, in particular three-layer neural network with a layer of input neurons, optionally having a plurality of sublayers, a layer of hidden neurons, optionally having a plurality of sublayers, and a layer of output neurons, optionally having a plurality of sublayers.

10. Method according to one or more of the preceding claims, characterized in that the method comprises steps for training a neural network in order to obtain the trained neural network, in particular using the technique of deep reinforcement learning: a) inputting input data into the neural network, wherein the input data comprise at least position data representing calculated or detected positions of the drives (A, B, C), and / or data indicating whether a hypothetical movement of the respective drive (A, B, C) from a calculated or detected position of the drive (A, B, C) into its assigned, predetermined end position or into at least one intermediate position deviating from the predetermined end position, in particular into the safe position, would lead to a collision or not, b) inputting output data into the neural network which, for each drive (A, B,C) comprise data from several individual movements, each with a defined target position, which are each assigned to a group of individual movements assigned to the respective drive (A, B, C), in particular individual movements whose target position is a predetermined end position, in particular a synchronous position of the respective drive (A, B, C), and / or individual movements whose target position is a safe position defined in such a way that, in the event that the respective drive (A, B, C) is in such a position, any other drive (A, B, C) can assume any drive position within its respective set of possible drive positions without the machine elements colliding with each other or with products handled by them, c) Training the neural network based on the input data and the output data so that it can select one or more individual movements for each drive (A, B, C) from the group of individual movements assigned to it, each with a defined target position, with the proviso that all drives (A, B, C) are transferred to their respective predetermined end position without collision when the drives (A, B, C) are moved according to the selected individual movements.

11. The method according to claim 10, characterized in that the training further comprises: a) calculating output data comprising one or more individual movements for each drive (A, B, C) based on the input data and b) rewarding calculated output data if they comprise a number of individual movements with which all drives (A, B, C) that are not in their predetermined end position can be transferred to their respective predetermined end position without collision.

12. Method according to claim 11, characterized in that calculated output data are rewarded less during training, the number of individual movements of which includes an individual movement with which a drive (A, B, C) is first moved into a safe position before being moved into the predetermined end position, than calculated output data which do not include such an individual movement.

13. Method according to claim 11 or 12, characterized in that calculated output data are rewarded less during training, the number of individual movements of which includes one or more individual movements that lead to a collision when the drives (A, B, C) are moved in accordance with the selected individual movements, than calculated output data that do not include one or more such individual movements.

14. Method according to one or more of the preceding claims 10 - 13, characterized in that the training of the neural network is carried out entirely or partly with the aid of a digital twin of the packaging machine (10) simulating at least the drives (A, B, C) and the machine parts moved by them.

15. Method according to one or more of the preceding claims 10 - 14, characterized in that for the training of the neural network, in particular by means of a random generator, positions are selected several times, preferably automatically, from a set of possible positions for the drives (A, B, C) and made available to the neural network as input data.

16. Method according to one or more of the preceding claims, characterized in that the trained neural network, in the event that during operation of the packaging machine (10) the drives (A, B, C) are in actual positions for which the trained neural network cannot calculate and output output data, comprising individual movements with which the or all drives (A, B, C) can be transferred collision-free into their respective predetermined end position, position data representing these actual positions are stored in a data memory, in particular in a data cloud.

17. Method according to one or more of the preceding claims, at least according to claim 7 and claim 16, characterized in that the trained neural network is retrained on the basis of these stored position data, in particular with the aid of the digital twin of the packaging machine (10), and that the drives (A, B, C) are then controlled during operation of the packaging machine (10) on the basis of the output data of the retrained neural network and are each transferred to their predetermined end position.

18. Method according to claim 17, characterized in that the retrained neural network is transmitted to the control device, so that the control device processes or can process the output data of the retrained neural network during operation of the packaging machine (10) for controlling the drives (A, B, C) and for transferring the drives (A, B, C) to the respective predetermined end position.

19. A packaging machine for producing packages, in particular for producing packaging for smokable products in the packaging industry, comprising a group of drives (A, B, C), in particular designed as servo drives, each driving a machine element of the packaging machine (10) and intersecting in an overlap region in which the machine elements could collide with each other or with products handled by the machine elements, wherein the packaging machine (10) has a control device, in particular for carrying out the method according to one or more of the preceding claims, which comprises a trained neural network, preferably for collision-free synchronization of the machine elements with each other, preferably during commissioning or maintenance of the packaging machine (10), wherein input data comprising at least position data can be input into the trained neural network,represent the actual positions in which the drives (A, B, C) are currently located, and wherein the trained neural network calculates and outputs output data comprising data of individual movements for the drives (A, B, C), and at least for each drive (A, B, C) that is not in a predetermined end position, in particular a synchronous position of the respective drive (A, B, C), selects one or more individual movements from a group of individual movements assigned to this drive (A, B, C), each with a defined target position, in particular such that all of these drives (A, B, C) can be transferred collision-free to their respective predetermined end position, preferably to their respective synchronous position, when the drives (A, B, C) are moved according to the selected individual movements.

20. Packaging machine according to claim 19, characterized in that the packaging machine has one or more further features of claims 1-18. *****

Images