Predictive maintenance for a device in the food industry using a digital twin and optimized production planning.

Description

[0001] The invention relates to a method for automatically detecting a malfunction of a device in the food or beverage industry, in particular for predictive maintenance according to the preamble of claim 1, a system for automatically detecting a malfunction according to the preamble of claim 9, a machine learning algorithm for automatically detecting a malfunction according to the preamble of claim 12, a method for determining an optimized production plan according to the preamble of claim 13 and a system for determining an optimized production plan according to the preamble of claim 15. Stand der Technik

[0002] Detecting an impending malfunction in a device is crucial for timely maintenance (such as replacing a worn part or cleaning a component) to prevent the malfunction (such as bearing failure). Often, the underlying cause (such as reduced lubrication or increased friction) has been present for some time, but the malfunction itself only manifests later. For example, a contaminated ball bearing in a device may continue to be used for a period before it is damaged and / or the device shuts down. Therefore, it is essential to identify the causal event (such as increased friction) that could lead to a future malfunction early on to prevent it from occurring.Detecting such a malfunction and carrying out timely maintenance measures as a result is also known as predictive maintenance.

[0003] Currently, predictive maintenance techniques rely on human experience (such as that of a plant operator), upon which predictive maintenance measures are based. For example, an experienced employee recognizes that certain output signals from a device, such as temperature or movement patterns, are unusual and determines a maintenance measure for the device based on their experience.

[0004] Alternatively, predictive maintenance can be performed based on operating data from a device (such as the output signals from the device's sensors). For this purpose, recorded operating data from previously occurring malfunctions of the device are used to train a machine learning algorithm. This trained machine learning algorithm then receives current operating data from the device during real-world operation and can thus be used to detect malfunctions.

[0005] To enable precise detection of all (or nearly all) potentially possible malfunctions of the device, these malfunctions must have previously occurred in the device so that the corresponding operational data (corresponding to each malfunction) can be used when training the machine learning algorithm. If there is no (or only limited) operational data for a specific malfunction, the machine learning algorithm cannot be trained (or can only be trained poorly) to detect that particular malfunction.

[0006] However, some malfunctions occur only very rarely. Furthermore, malfunctions often lead to damage to the affected device, and troubleshooting them is costly and time-consuming. While it is theoretically possible to deliberately cause malfunctions to generate relevant operational data (for training a machine learning algorithm), this is also very costly and time-consuming. This means that to accurately detect a malfunction, sufficient operational data for all possible malfunctions must first be recorded, but this data is difficult or very expensive to generate.

[0007] Another disadvantage of the prior art is that many additional sensors (besides those already present for the intended operation of the device) must be attached to the device in order to accurately detect an impending malfunction. These additional sensors collect additional operating data, which makes it possible to improve the detection of malfunctions, even if some malfunctions occur only rarely.

[0008] While the operational data generated by the additional sensors does improve the detection of malfunctions, as described above, this additional data must be transmitted, processed, and stored. This involves additional effort (for example, requiring additional storage and computing power) and can lead to a delay in malfunction detection. In particular, this delay (caused by the additional operational data) can prevent the necessary timely response to the malfunction. In any case, this delay is undesirable because it means that measures to prevent the malfunction can only be initiated with a delay.

[0009] Furthermore, the additional sensors may restrict the function or application of the device and require complex installation. Another problem is finding suitable locations for mounting the additional sensors so that they can collect the most informative operating data possible for detecting an impending malfunction (i.e., enabling the most precise possible detection of a malfunction).

[0010] NOVELS M ET AL: "Scheduling with simulation in the food & drinks industry", 1996 WINTER SIMULATION CONFERENCE PROCEEDINGS. CORONADO, CA, December 8-11, 1996; [WINTER SIMULATION CONFERENCE PROCEEDINGS], NEW YORK, IEEE, US, December 8, 1996 (1996-12-08), pages 1252-1256, describes the use of decision support tools for capacity planning and scheduling. The increasing use of decision support tools such as simulations is described to understand problems in the production environment and to create schedules that can be implemented on a daily basis. Production planning and scheduling have always been an important factor in maximizing plant efficiency. Inefficient planning and unclear objectives can lead to a significant increase in variable costs. Controlling variable costs is important for achieving healthy profit margins in any business.The costs of warehousing contribute significantly to variable costs, and case studies have shown a substantial reduction in inventory when production is planned taking limited resources into account. Aufgabe

[0011] The object of the invention is therefore to provide an improved method and a corresponding system for determining an optimized production plan for the production of one or more products using one or more production lines in the food industry or the beverage industry. Lösung

[0012] This problem is solved according to the invention by a method according to claim 1 and a system according to claim 9.

[0013] Preferred embodiments are described by the dependent claims.

[0014] One embodiment of the invention relates to a method for automatically detecting a malfunction of a device in the food or beverage industry, particularly for predictive maintenance. The method comprises generating simulation data using a digital twin of the device, wherein the simulation data is generated for a multitude of malfunctions; training a machine learning algorithm based on the simulation data and the multitude of malfunctions to generate a trained machine learning algorithm (60); and detecting a malfunction of the device by the trained machine learning algorithm based on operational data of the device.

[0015] The detected malfunction of the device can be an event, a fault, and / or a required maintenance measure. In particular, the event can include increased friction, increased vibration, imbalance, a wear parameter, or reduced lubrication; the fault can include motor damage, bearing damage, shaft breakage, a crack, a misplaced bottle, a short circuit, or a device stoppage; and the required maintenance measure can include cleaning, lubrication, oiling, fastening, adjusting, replacing, or repairing.

[0016] Furthermore, the detection of the malfunction can also include the detection of a component of the device in which the malfunction occurs.

[0017] Furthermore, the detection of the operational disruption can also include the detection of a likely time span until the occurrence of a malfunction caused by the operational disruption.

[0018] Furthermore, the method can be used to determine a production plan for a plant assigned to the device, based on the expected time span and / or duration of a maintenance measure to be carried out for the detected malfunction.

[0019] The operational data can include output signals from the device, and the simulation data can include output signals from the digital twin. Each output signal of the digital twin can correspond to a specific output signal of the device.

[0020] Furthermore, each of the device's output signals can be measured by a sensor of the device, which is used during the intended operation of the device.

[0021] Additionally, the device's operating data can include at least one additional output signal from the device, wherein the additional output signal is measured by an additional sensor of the device, which is required for the device to operate in a manner inconsistent with its intended purpose. Accordingly, the simulation data can include at least one corresponding additional output signal from the digital twin. Furthermore, the additional sensor of the device can be selected from a multitude of possible additional sensors using the digital twin in such a way that the output signal of the additional sensor optimizes the detection of the operational fault.

[0022] According to the procedure described above, the trained machine learning algorithm for detecting the malfunction can determine a probability for each of the malfunctions based on the operating data (20A) of the device and recognize a malfunction of the malfunctions as the malfunction when the probability of the malfunction reaches a threshold.

[0023] In addition, the machine learning algorithm can be trained using recorded operating data from the device.

[0024] According to one embodiment, the device is a tripod used in the food industry.

[0025] Another embodiment of the invention corresponds to a system for automatically detecting malfunctions of equipment in the food or beverage industry, particularly for predictive maintenance, which comprises a digital twin of the equipment and a machine learning algorithm. The digital twin is configured to generate simulation data of the equipment, with the simulation data being generated for a multitude of malfunctions of the digital twin. The machine learning algorithm is trained based on the simulation data and the malfunctions of the digital twin and is configured to detect malfunctions of the equipment based on operational data of the equipment. Furthermore, the machine learning algorithm can be executed by a controller of the equipment, a processor connected to the equipment, or a cloud.

[0026] Another embodiment of the invention relates to a machine learning algorithm for automatically detecting a malfunction of a device in the food or beverage industry, in particular for predictive maintenance, wherein the machine learning algorithm is trained based on simulation data and malfunctions of a digital twin of the device, wherein the machine learning algorithm is set up to detect a malfunction of the device based on operating data of the device.

[0027] Another embodiment of the invention relates to a method for determining an optimized production plan for the production of one or more products using one or more production lines in the food or beverage industry. In a first step, the changeover time of the production plan is optimized. The changeover time includes product changes, bottle changes, variety changes, format changes, and / or packaging changes during product production. When optimizing the changeover time, the production plan is divided into a plurality of interconnected units. Each of these units corresponds to at least one product, one processing step, or at least one piece of equipment on a production line for producing one product of the production plan. Furthermore, when optimizing the changeover time, the units are arranged such that the changeover time has a minimum value.The next step involves optimizing the production costs of the production plan, which are the costs required to process the units. These production costs are determined by the setup times and production times for the units using the equipment, with the setup times and production times being defined as product- and equipment-dependent. Optimizing production costs includes assigning the units to specific production lines and determining the sequence in which the units are to be processed by these lines.

[0028] Furthermore, the changeover time and / or production costs can additionally include the duration of a maintenance measure for a detected malfunction of a device on a production line, whereby the malfunction is detected by a machine learning algorithm trained on simulation data and malfunctions of a digital twin of the device, or by the digital twin itself. The machine learning algorithm is configured to detect the malfunction of the device based on the device's operational data.

[0029] Another embodiment of the invention relates to a system that is configured to carry out the method described above for determining an optimized production plan.

[0030] Embodiments of the invention are described with reference to the following drawings. They show: Figur 1 an exemplary system for generating simulation data, training a machine learning algorithm, and predicting a malfunction based on current operating data of the device; Figur 2 a device that generates operating data and creates a digital twin of the device and simulation data for the device according to an exemplary embodiment of the invention; Figur 3A a digital twin of a device that generates simulation data for a multitude of operational malfunctions and for a multitude of simulation runs according to an exemplary embodiment of the invention; Figur 3B Exemplary simulation data of the digital twin for normal operation and various operational disruptions; Figur 4 the generation of a trained machine learning algorithm based on the simulation data and the associated operational disturbances according to an exemplary embodiment of the invention; Figur 5 the detection of a malfunction of a device by a machine learning algorithm trained on simulation data of the digital twin based on current operating data of the device according to an exemplary embodiment of the invention; Figur 6 a method according to the invention for predictive maintenance; Figur 7 an exemplary system for determining an optimized production plan for the production of one or more products using one or more production lines in the food or beverage industry; and Figur 8 an inventive method for production plan optimization.

[0031] The Fig. 1 Figure 1 shows an example of a system 100 according to the present invention. Here, system 100 is exemplarily divided into four components 200, 300, 400, and 500. First, the following is shown: Fig. 1 An overview was given before the four exemplary components 200, 300, 400, 500 were discussed in connection with the Fig. 2-5 will be described in more detail.

[0032] According to the invention, a digital twin 10B of a (real) device 10A is created. This is shown by way of example by component 200 and is described in more detail in connection with Fig. 2 . This digital twin 10B has the same physical properties as the real device 10A. Thus, the digital twin 10B provides simulation data 20B that corresponds one-to-one to the operating data 20A of the device 10A.

[0033] According to the invention, this digital twin 10B is used to generate simulation data 40 for possible operational malfunctions 30. This is shown by way of example by component 300 and is described in more detail in connection with the Fig. 3A and 3B .These simulation data 40 of the digital twin 10B correspond one-to-one to the operating data 20A of the real device. According to the invention, instead of inducing an operational malfunction in the real device 10A, the operational malfunction 30 is simulated using the digital twin 10B in order to obtain simulation data 40 that correspond one-to-one to the operating data of the device 10A when an operational malfunction occurs. An operational malfunction 30 here refers to events that, while not yet constituting a direct malfunction themselves, could lead to a future malfunction if no maintenance measures are taken, such as increased friction (e.g., due to contamination), increased vibration, imbalance, a wear parameter, or reduced lubrication.On the other hand, this also includes malfunctions that can occur as a result of this event, such as engine damage, bearing damage, axle breakage, a crack, a wrongly positioned bottle, a short circuit or a stopping of the device (because, for example, a critical value has been exceeded, which leads to a stopping of the device).

[0034] Thus, a multitude of operational disruptions 30 can be defined for the digital twin 10B, which can be based, for example, on empirical data, previously occurring operational disruptions, or randomly generated operational disruptions. These operational disruptions can also be defined and simulated for various components of the digital twin (such as a bearing, a chain, a gear, a motor, etc.). Subsequently, these operational disruptions 30 are simulated using the digital twin 10B, whereby a multitude of simulation data 40 can be generated for each of the operational disruptions 30. The simulation data generated in this way can then be used to train a machine learning algorithm 50 (such as a classifier).

[0035] Based on the operational faults 30 and the simulation data 40 of the digital twin 10B, the machine learning algorithm 50 can then be trained to generate a trained machine learning algorithm 60. The training of the machine learning algorithm 50 is carried out in such a way that the trained machine learning algorithm 60 is able to recognize a specific operational fault 30 based on the simulation data 40. This is exemplified by component 400 and is described in more detail in the context of Fig. 4 .The invention enables the generation of simulation data 40 for training the machine learning algorithm 50, which, since they were generated by the digital twin 10B of the real device 10A, correspond one-to-one to the operating data of the real device 10A. According to the present invention, it is no longer necessary to induce operational malfunctions of the real device 10A in order to generate corresponding operating data for these malfunctions; instead, simulation data 40 are generated for operational malfunctions 30 simulated on the digital twin 10B. This makes it possible to generate a large number of training data (namely the simulation data 40) for a large number of corresponding operational malfunctions 40 (including those malfunctions that only rarely occur during the operation of the real device).

[0036] The machine learning algorithm 60, trained on the simulation data 40 of the digital twin 10B, is then used to predict a malfunction of the (real) device 10A based on current operating data 20A of the real device 10A. This is exemplified by component 500 and is described in more detail in connection with Fig. 5 .The trained machine learning algorithm 60 can, for example, be translated into a generally executable code and ported to the controller of the (real) device 10A or an additional industrial PC (IPC). It is also possible to apply the trained machine learning algorithm 60 in a cloud environment. During operation of the (real) device 10A, the output signals of the (real) device 10A are continuously provided to the trained machine learning algorithm 60. The trained learning algorithm 60 then automatically determines (based on its trained knowledge and the output signals of the device 10A) the current state of the device 10A and detects whether an operational fault exists and / or is imminent.Furthermore, the machine learning algorithm 60 can identify the type of malfunction (which can be used to initiate appropriate maintenance measures to eliminate the malfunction and prevent a breakdown / damage), which component of the device is affected by the malfunction, and how long it will be before a breakdown occurs if no maintenance is performed. For example, in addition to recognizing that there is increased friction in a certain bearing, it can determine how long the device can continue to operate before the bearing fails, triggering an order for a new bearing and / or notifying a service technician.

[0037] A further advantage of the invention is that no (or only a few) additional sensors need to be attached to the device 10A, which (as described above) are used in the prior art to obtain additional operating data of the device in order to enable more precise detection of a malfunction. An additional sensor is understood here to be a sensor that is not already present in the device 10A for other reasons (not directly related to predictive maintenance), such as being necessary for the intended operation of the device, and is not used to improve predictive maintenance. A sensor that is already present in the device 10A (and used for the intended operation of the device) is referred to below as an operational sensor.

[0038] According to the present invention, it is possible to dispense with (or reduce the number of) additional sensors, since the digital twin 10B can generate a large number of simulation data 40 for all possible operational disturbances 30, thus eliminating the need for additional operational data from additional sensors when training the machine learning algorithm 50. This is particularly possible because the digital twin 10B generates the same output signals (as simulation data 20B) that are also measured in the real device 10A (as operational data 20A). That is, the digital twin 10B has the same operational sensors as the (real) device 10A and thus provides simulated output signals that correspond to the real operational data 20A of the real device 10A, in particular the real operational data provided by the operational sensors that are generated during the operation of the device 10A anyway.In other words, the digital twin 10B generates (for a variety of operational disturbances 30) a variety of simulation data 40 that correspond to the operational data 20A of the operational sensors of the real device 10A. With this simulation data 40 (i.e., the simulated output signals of the operational sensors of the digital twin, which correspond to the real output signals of the operational sensors of the real device), it is possible to precisely train the machine learning algorithm 50. Therefore, it is not necessary to attach a large number of additional sensors to the real device 10A, to conduct tests, or to monitor real operational disturbances.

[0039] Furthermore, with the help of the digital twin 10B, it is possible to select an additional sensor in the most optimal way if one is to be used. For example, the type of additional sensor (e.g., a vibration, voltage, or torque sensor) and its location on the device 10A can be determined using the digital twin 10B in such a way that the detection of an operational fault by a trained machine learning algorithm 60 is optimally improved. For this purpose, one or more additional sensors can be added to the digital twin 10B, and simulation data 40 can be generated for the operational sensors and the one or more additional sensors.Subsequently, a machine learning algorithm 50 is trained based solely on the simulation data of the operational sensors, and the same machine learning algorithm 50 is trained based solely on the simulation data of the operational sensors and one or more additional sensors. A comparison can then be made to determine whether the trained machine learning algorithm 60-2, which was additionally trained based on the simulation data of one or more additional sensors, can predict an operational fault better (e.g., earlier, or with higher probability) than the trained machine learning algorithm 60-1, which was trained solely on the simulation data of the operational sensors. This process can be repeated with one or more other additional sensors on the digital twin 10B, and a multitude of trained machine learning algorithms 60-1 ... 60-L can be generated using the respective simulation data.Finally, the influence of various additional sensors on the precision of the trained machine learning algorithms 60-1 to 60-L for detecting a malfunction can be compared. The additional sensor (or sensors) based on whose simulation data a trained machine learning algorithm with the best precision was generated can then be selected. This sensor, determined using the digital twin 10B, can then be attached to the real device 10A and used for predictive maintenance of the real device.

[0040] Fig. 1 Figure 1 shows an embodiment of the present invention for a device 10A in the food industry, particularly the beverage industry. In one example, the device 10A is a food industry tripod as known from DE 10 2013 208 082 A1. During operation, the device 10A generates operating data 20A, such as the time course of the output signals of the device 10A. Typically, the output signals include one or more of a torque, a velocity, an acceleration, a current, a voltage, a temperature, and a vibration, which are measured by corresponding sensors of the device 10A. The object of the invention is to automatically detect, based on the operating data 20A of the device 10A, whether (and / or when) the device needs maintenance in order to prevent a malfunction.Typical maintenance work includes, for example, cleaning individual components of the device, replacing wear parts, lubricating bearings, etc. Detecting a malfunction and carrying out maintenance measures in a timely manner is also known as predictive maintenance.

[0041] According to one embodiment of the invention, it is particularly possible to predict an operational malfunction only on the operating data of operational sensors of the device (i.e. sensors that are used and available for the operation of the device anyway) or on the operating data of the operational sensors and only a few selected additional sensors (which are used specifically for predictive maintenance).

[0042] According to the invention, a digital twin 10B of the device 10A is created. The digital twin 10B is a digital representation of the (real) device 10A, which replicates the real device 10A in a computer-executable component. Creating a digital twin is known in the prior art. The digital twin 10B of the device 10A can then be used to simulate all processes performed by the (real) device 10A. For example, the digital twin 10B can be used to simulate the temporal progression of the translational and rotational movement of the device 10A, as well as the control signals used and the resulting output signals of the device 10A. The digital twin 10B can thus generate simulation data 20B that correspond to the operating data 20A of the real device 10A.This means that while a process is simulated on the digital twin 10B, the digital twin 10B can generate simulation data 20B of the temporal output signals (such as one or more of a torque, velocity, acceleration, current, voltage, temperature, and vibration) measured by (simulated) sensors of the digital twin 10B. This simulation data 20B corresponds to the operational data 20A that the device 10A would output if the same process (which is simulated by the digital twin 10B and results in the simulation data 20B of the digital twin 10B) were carried out on the real device 10A.In particular, these simulation data 20B correspond to the temporal output signals (such as one or more of a torque, a velocity, an acceleration, a current, a voltage, a temperature and a vibration) that would be measured by the (real) sensors of the device 10A when performing the same process.

[0043] As described above, a problem in the prior art is that, typically, too little operational data 20A is available for the possible malfunctions of the device 10A, or operational data can only be generated with considerable effort. According to the invention, operational malfunctions of the device 10A are simulated by the digital twin 10B, which enables the generation of a large number of simulation data 40 for a large number of operational malfunctions 30. This demonstrates the Fig. 3A ,The following are illustrated as examples of K operational faults 30-1, 30-2 ... 30-K of the digital twin 10B. Each of these K operational faults can be a damage class or correspond to a damage class that can be predefined on the digital twin 10B. The digital twin 10B can perform a plurality (N) of simulation runs for each of the K operational faults. For each of the N simulation runs, the digital twin 10B can generate simulation data 40-1-1 ... 40-KN, each of which represents the time course of the M output signals of the digital twin 10B. In other words, N simulation runs are performed for each K operational fault, each containing M output signals. Thus, K x N x M time courses of individual output signals can be generated (where K, N, and M are integers).To have a comparison to normal operation (i.e., operation without operational disruption), simulation data 20B of the digital twin 10B can also be generated without operational disruption.

[0044] An operational disruption, as defined here, encompasses both events that do not themselves constitute a malfunction but could lead to one in the future (if no maintenance is performed), such as contamination, wear, imbalances, and / or increased friction, and the malfunctions themselves that can occur as a consequence of this event, such as motor damage, bearing damage, a crack in a component, incorrectly placed bottles, or even the device stopping (e.g., due to exceeding a critical value that triggers an (emergency) stop). The operational disruptions described above, 30-1, 30-2 ... 30-K, can be damage classes and / or defined as damage classes in the digital twin 10B and can be simulated using the digital twin 10B.For example, different classes of friction in one (or more) component(s) can be simulated as different operational disturbances using different coefficients of friction.

[0045] Possible events that could lead to a malfunction in the food industry tripod mentioned above include, for example, increased friction in a joint, reduced lubrication in a prismatic universal joint, or increased friction between the container and the mat chain due to contamination. Possible malfunctions that can occur in the food industry tripod mentioned above include, for example, bearing damage in a joint, loosening of the connection between the container and the mat chain, and motor failure.

[0046] As described above, for each of the K operational disruptions or damage classes (and, for example, also in the case where no operational disruption is present), N simulation runs of the digital twin 10B can be performed to generate N different simulation data for each operational disruption. For this purpose, N simulation runs of the digital twin can be performed, each subject to the same operational disruption but differing from the others. Here, N is any natural number greater than 1, preferably greater than 100. For example, random starting conditions for the simulation of the operational disruption can be chosen, or the parameters describing the operational disruption (such as the coefficient of friction) can be slightly varied.Furthermore, it is possible to subject the simulation runs to stochastic fluctuations, which leads to the individual simulation runs differing from one another despite the same (or very similar) operational disturbance. A simulation can then be stopped either after a predetermined time has elapsed or upon the occurrence of a specific event (in particular, the disturbance associated with an event). In this way, simulation data 40-1-1 ... 40-KN can be generated for each of the K operational disturbances 30-1, 30-2 ... 30-K. It was previously described that N simulation data are generated for each operational disturbance; however, this does not mean that exactly the same number of simulation data must be available for each operational disturbance 30-1, 30-2 ... 30-K. It is, of course, also possible that fewer or more than N simulation data are generated for one or more of the operational disturbances 30-1, 30-2 ... 30-K.

[0047] Thus, any number of simulation data sets 40-1-1 ... 40-KN for all possible operational faults can be generated by the digital twin 10B of the device 10A. This solves the problem described above in the prior art, namely that for some operational faults only a few recorded operational data sets 20A of the device 10A are available, or that these can only be generated with great effort.

[0048] An example of generating simulation data for a single one of the M output signals (e.g., for sensor 1 of the M sensors) for a single simulation run of the digital twin 10B for no operational fault 310 and for three exemplary operational faults 320, 330, 340 is shown below using the Fig. 3B shown. As mentioned above, N simulation runs with M output signals are typically performed for each of K operational faults. This is shown based on the Fig. 3B The described example of a simulation run and an output signal can easily be extended to this general case.

[0049] Figure 310 shows the simulated output signal of a sensor (in this example, Sensor 1) for a simulation run of digital twin 10B in the case of no operational faults. Figures 320, 330, and 340 further show the simulated output signal of a sensor (in this example, Sensor 1) for a simulation run of digital twin 10B in the case of various operational faults 30-1, 30-2, and 30-3. To simulate an operational fault, various components of digital twin 10B can be selected (such as a specific bearing, an axle, a motor, etc.), and different types of operational faults (such as increased friction, imbalance, etc.) can be simulated for each selected component. In Figures 320, 330, and 340, a specific bearing is selected as an example component of the digital twin.Furthermore, three different operational faults 30-1, 30-2, and 30-3 are simulated for this component as examples. Diagram 320 shows an example of an operational fault 30-1, which represents an event (in the sense described above), in this case, increased friction in the selected bearing. As described above, an event represents an operational fault that can be rectified by a maintenance measure, but which, if not rectified, can lead to a malfunction, such as damage to the device. In the example shown here, the operational fault of increased friction in the selected bearing is simulated for a specific period of time using the digital twin 10B and generates the simulation data 40-1-1 (here, for example, sensor 1). Diagram 330 shows an example of an operational fault 30-2, which represents a malfunction (in the sense described above), in this case, damage to the selected bearing.In the example shown here, the operational disruption "damage" of the selected bearing is simulated for a specific period of time using the digital twin 10B, generating the simulation data 40-2-1 (exemplified here for sensor 1). Furthermore, diagram 340 shows an example of an operational disruption 30-3, which includes both an event ("increased friction") and a malfunction ("bearing damage"), where the malfunction was caused by the event. As shown in diagram 340, an event is simulated initially, and a malfunction occurs during the simulation of the event. Typically, the malfunction can occur after a certain time (t = t1). In this way, a multitude of operational disruptions can be defined (e.g., by selecting a type of operational disruption and a location of the disruption), each of which can provide simulation data for M sensors and N simulation runs using the digital twin 10B.

[0050] As described above, the time courses of the translational and rotational movements of the device 10A, as well as its input, control, and output signals, can be simulated in each simulation run using the digital twin 10B. The time courses of the M output signals of the digital twin 10B can then be used as simulation data 20B, 40-1-1, ..., 40-KN. The digital twin 10B has the same physical properties as the (real) device 10A and provides the same M output signals (i.e., the operating data 20A) as the device 10A. This means that the simulation data 40-1-1 ... 40-KN, which are generated for the K operational disturbances using the digital twin 10B, correspond to the operating data 20A that the device 10A would output if it were subject to the respective operational disturbance. Thus, for each of the K operational disturbances 30-1 ...30-K simulation data of the digital twin 10B are generated, which correspond to the operating data of the real device 10A.

[0051] This simulation data from the digital twin 10B (corresponding to the operating data of the device 10A) can then be used to train a machine learning algorithm 50, such as a classifier. The machine learning algorithm 50 learns characteristic features that occur in the simulation data 20B, 40-1-1 ... 40-KN of the simulated malfunction 30-1 ... 30-K of the digital twin 10B. Subsequently, a machine learning algorithm 60 trained in this way can also find these characteristic features in the operating data 20A of the real device 10A and thus detect a malfunction of the real device 10A. This makes it possible not only to detect whether a malfunction exists, but also which malfunction exists, which component it affects, and / or how much time remains to perform maintenance before damage or a malfunction occurs.

[0052] The malfunctions 30-1, 30-2 ... 30-K, defined for the digital twin 10B and described above, can then be detected on the real device 10A using a trained machine learning algorithm, as described in more detail below. In particular, the trained machine learning algorithm can recognize both the type and the location of the malfunction, since a malfunction defined for the digital twin 10B can include both a type of malfunction (e.g., an event such as increased friction) and a location of the malfunction (e.g., which component of the digital twin 10B the malfunction occurs on).For example, if the operational fault of the digital twin is "increased friction in bearing X", the trained machine learning algorithm can recognize this two pieces of information based on the operating data of the device: the type of operational fault ("increased friction") and the location of the operational fault ("bearing X").

[0053] Furthermore, it is possible to determine the expected time interval until the occurrence of a malfunction associated with an event (such as the bearing damage caused by increased friction in Figure 340). According to one embodiment, this time interval can be determined directly based on the simulation data. One possibility, for example, is to calculate the mean of the time interval t1 (in Figure 340) until the occurrence of the malfunction associated with the event over all N simulation runs for that event. If the trained machine learning algorithm 60 then recognizes this event based on the operating data of the device 10A, the mean value calculated in this way can be output as additional information (in addition to the type and location of the event). In the example described in Figure 340, a corresponding notification could read: "Increased friction in bearing X, expected time interval until bearing damage occurs: t1".In an alternative embodiment, the machine learning algorithm could also be trained to predict the expected time until a malfunction occurs. The term "time" here can refer to a specific duration (such as 24 hours remaining) or a remaining production volume (such as 1000 units remaining) until the malfunction associated with the event occurs.

[0054] The Fig. 4 Figure 1 shows, by way of example, the training of a machine learning algorithm 50 based on the simulation data 20B, 40-1-1, ..., 40-KN and the operational disturbances 30-1 ... 30-K of the digital twin 10B. The machine learning algorithm 50 can be any suitable machine learning algorithm known in the prior art, such as a classifier or a regression method. Well-known examples include neural networks, support vector machines, boosting, naive Bayes, etc.

[0055] During training, the operational faults 30-1 ... 30-K (and the presence of no operational fault) or the damage classes can be used as ground truth, and the simulation data 20B, 40-1-1, ..., 40-KN as corresponding training data. The training determines the parameters of the machine learning algorithm 50, with which the trained machine learning algorithm 60 can optimally recognize the respective ground truth (e.g., whether and, if so, which of the operational faults 30-1 ... 30-K corresponding to the respective simulation data 20B, 40-1-1, ..., 40-KN) based on the training data (i.e., the simulation data 20B, 40-1-1, ..., 40-KN). The machine learning algorithm 50 with the optimal parameters determined in this way is the trained machine learning algorithm 60. Training machine learning algorithms (e.g.,The process of splitting the training data into a test set and a training set, optimizing an error term, and using cross-validation is known in the prior art. The trained machine learning algorithm 60 is then able, given a dataset of simulation data (e.g., 40-2-1), to accurately predict the operational fault that occurred during the simulation of that dataset (for 40-2-1, this is operational fault 30-2).

[0056] An additional or alternative learning objective (ground truth) for the machine learning algorithm can also be the predicted time until a malfunction occurs. Furthermore, multiple malfunctions or damage classes (such as several classes of increased friction) can be grouped together so that the trained machine learning algorithm only recognizes the aggregated malfunction (for example, only increased friction, and not the specific degree of increased friction in each individual case). Another possibility is to output the maintenance measure for the detected malfunction instead of, or in addition to, the learning objectives (ground truth) discussed above.

[0057] In one embodiment of the invention, in addition to the simulation data 20B, 40-1-1, ..., 40-KN of the digital twin 10B, historical operating data of the device 10A can also be used for training the machine learning algorithm. If historical operating data exists for malfunctions that are identical to one of the simulated malfunctions 30-1 ... 30-K, this historical operating data can be added to the simulation data for that malfunction, and the machine learning algorithm 50 can be trained based on the simulation data 20B, 40-1-1, ..., 40-KN and the historical operating data for that malfunction. Furthermore, it is possible to define additional malfunctions or damage classes that cannot be simulated with the digital twin 10B. An example of this is the intervention of an operator of the device 10A, where the operator, for example, takes over control of the device 10A (e.g.,(by overriding the device's automatic control). This is the case, for example, when the operator determines, based on their experience, that there is an irregularity in the operation of device 10A. Intervention by the operator of the device can thus be defined as a new operational fault (e.g., 30-K+1), and the operating data recorded before the intervention can be used as historical operating data for training the machine learning algorithm 50. Furthermore, if no operational fault is present, historical operating data of device 10A, recorded when no operational fault was present, can be used for training the machine learning algorithm 50 instead of (or in addition to) simulation data 20B (generated for normal operation without operational faults using the digital twin 10B).

[0058] The machine learning algorithm 60 described above, trained with the aid of the digital twin 10B, is then used according to the invention to detect an operational malfunction of the (real) device 10A, as shown in Fig. 5 This is shown by way of example. For this purpose, the (current) operating data 20A of the device 10A are provided to the trained machine learning algorithm 60, which can detect an operational malfunction of the device 10A based on the operating data 20A. In particular, the trained machine learning algorithm is configured to recognize which operational malfunction is present (e.g., an event such as increased friction), where this operational malfunction is present (e.g., increased friction in the universal joint), and / or to determine a probable time interval after which a malfunction will occur (e.g., damage to the universal joint in 12 hours). According to one embodiment of the invention, the machine learning algorithm determines a probability P_k for each of the operational malfunctions (and also for no operational malfunction) based on the current operating data 10A (where k = 0 for no operational malfunction, and k = 1 ... K for the operational malfunctions 30-1 ...30-K), which indicates the probability of the malfunction k occurring. If the trained machine learning algorithm determines a probability P_k for a malfunction k that is greater than a threshold P_G, a notification 70 can be generated indicating a malfunction. Notification 70 can, for example, contain the detected malfunction k (e.g., its location on the device and the type of malfunction) and / or an estimated time (e.g., how many hours remain or how many more units can be produced) until a malfunction of device 10A associated with the malfunction occurs. Furthermore, the notification can also include a maintenance measure to be performed to resolve the malfunction and / or the duration of this maintenance measure.For example, notification 70 can be sent to an operator of the device and / or to maintenance personnel. The trained machine learning algorithm 60 can be ported to the controller of the device 10A or to an additional industrial PC (IPC). Alternatively, it is also possible to apply the trained machine learning algorithm 60 in a cloud environment.

[0059] As previously described, the trained machine learning algorithm 60, which was trained on the simulation data 20B, 40-1-1, ... 40-KN of the digital twin 10B, is able to detect an operational malfunction of the device 10A. This has the advantage that the operational malfunctions (and any associated incidents) do not have to occur on the real device 10A (in order to obtain the operational data associated with the malfunctions for training the machine learning algorithm), but can be simulated using the digital twin.

[0060] Additionally, it is possible to generate a large number of simulation data for each of the operational faults 30-1 ... 30-K, which improves the precision of the trained machine learning algorithm in detecting an operational fault. This also makes it possible to reduce the number of additional sensors, as used in state-of-the-art systems to improve predictive maintenance, or even to eliminate the need for additional sensors altogether.

[0061] According to a further embodiment of the invention, the digital twin 10B is used to identify one (or more) additional sensors that optimally improve the prediction of a malfunction (according to the present invention). For this purpose, various additional sensors (such as a vibration, current, torque, temperature, or acceleration sensor) can be added to the digital twin 10B at different locations. The influence of these additional sensors on the detection of a malfunction (as described above) can then be compared, and one or more of the additional sensors can be selected that optimally improve the prediction of the malfunction. These selected sensors of the digital twin 10B can then also be attached to the (real) device 10A to improve the detection of a malfunction of the device 10A.In other words, various additional sensors can be simulated using the digital twin 10B and compared with regard to their information content for detecting the operational disturbances 30-1 ... 30-K simulated on the digital twin. In this way, different sensors, such as a vibration, current, torque, temperature, or acceleration sensor, can be applied at different locations on the digital twin 10B (which is a one-to-one replica of the device 10A) and tested (by means of simulation by the digital twin) to determine which sensor at which location on the device provides an output signal that optimally improves the precision of the machine learning algorithm trained on this simulation data (compared to other additional sensors of the digital twin and compared to using no additional sensors).

[0062] For example, if Z additional sensors are attached (at Z locations on the digital twin) and the digital twin already has X operational sensors (i.e., sensors that are already present and not additional sensors), then the digital twin provides Z+X output signals as simulation data for each simulation run. The digital twin (with the Z additional sensors) can now be used as described in the context of... Fig. 3 The simulation is described for the K fault cases 30-1 ... 30-K, each with N simulation runs. This yields simulation data with K x N x (X+Z) individual output signals, similar to the data described above. Subsequently, Z+1 different trained machine learning algorithms can be generated by training one machine learning algorithm on the simulation data without the output signals from the additional sensors, and by training a machine learning algorithm for each of the Z additional sensors with the respective output signal of one of the Z additional sensors. The precision (such as the optimized error term of the trained machine learning algorithm or the Bayesian information criterion) of the Z+1 trained machine learning algorithms can then be compared. The additional sensor that corresponds to the trained machine learning algorithm with the best precision (e.g.,the sensor that has resulted in the lowest error term (or the n additional sensors that have resulted in the n best trained learning algorithms) can then be selected and the same sensor can then also be attached to the real device at the same location (as on the digital twin) and used to detect an operational malfunction of the device according to the invention.

[0063] The Fig. 6 Disclosing a method 600 according to the invention for automatically detecting a malfunction of a device (10A) in the food or beverage industry, particularly for predictive maintenance. In a first step 610, simulation data 20B, 40-1-1, 40-2-1, ..., 40-KN are generated using the digital twin 10B of the device 10A. The simulation data for a plurality of malfunctions 30-1-1, 30-2-1, ..., 30-KN of the digital twin 10B are generated. In a second step 620, the machine learning algorithm 50 is trained based on the simulation data 20B, 40-1-1, 40-2-1, ..., 40-KN of the digital twin 10B and the multitude of simulated operational disturbances 30-1-1, 30-2-1, ..., 30-KN to generate a trained machine learning algorithm 60.This trained machine learning algorithm 60 is then used to detect the operational fault of the (real) device 10A based on operating data 20A of the (real) device.

[0064] According to a further embodiment of the invention (which can be combined with all the features described above), predicting a malfunction can also be used to optimize the production plan of a plant with one or more production lines (e.g., in the food industry, and especially the beverage industry). In this case, information about the impending maintenance of a device on a line of the plant, the nature of the maintenance (i.e., its duration), and / or the timeframe within which the maintenance must be performed (i.e., how much time remains or how many products can still be produced before the maintenance must be performed at the latest to prevent a malfunction) can be used to optimize the plant's production plan.This information, which can be made available through the detection of an operational disruption according to the invention, can thus be incorporated into the automatic optimization of the plant's production plan.

[0065] In principle, according to the embodiment of the invention described here, capacity utilization forecasts of the plant are used for automatic production planning (which can be supplemented with information about predictive maintenance). The automatic determination of an optimal production plan involves dividing the production plan into individual units and determining an optimal sequence for the units of the production plan. For example, when determining the optimal production plan, factors such as production time, setup time, and changeover time (which includes, for example, product changes, bottle changes, and / or packaging changes) are taken into account. Based on this information, an optimal production plan is automatically generated that minimizes the time required to produce all planned products (each in a specific quantity).An optimal production plan ensures that the plant is only down for the shortest possible time, thus increasing the plant's production volume.

[0066] The first step involves optimizing the changeover time of the production plan with the goal of minimizing the longest changeovers. Changeover time can encompass the duration of all activities required for a changeover, such as product changes, bottle changes, variety changes, format changes, and / or packaging changes. The production plan is broken down into small, coherent units, and these are arranged to minimize the overall changeover time. Each unit can correspond to at least one product, one processing step, or at least one piece of equipment on a production line for producing one product in the production plan. Optionally, the maintenance time, determined by the trained machine learning algorithm described above, can also be taken into account.In particular, information about the duration of maintenance and / or the timeframe in which it must be performed can be used to optimize the changeover time of the production schedule. For example, it is possible to schedule the maintenance time of a piece of equipment so that it coincides with the changeover time of the system to which that piece of equipment belongs. This allows maintenance to be carried out at the same time as the product changeover.

[0067] In a second step, the production costs of the production plan can be optimized. Based on the first step, the optimization of the changeover time can be extended to include the cost-optimized allocation of the units of the production plan to the individual production lines of the plant. This involves optimizing the production plan to minimize the production costs K required to process the units of the production plan. K = ∑ i = 1 n H i ∗ t ges , max where n is the number of lines, H i the machine hourly rate of line i in euros per hour and t ges,max The maximum production time on a line is in hours.

[0068] The production time on line i t i,ges is calculated from: t i , ges = t Nach , i , ges + t Pro , i , ges where t Nach,i,ges specifies the retrofit time of line i in hours and t Pro,i,ges specifies the production time of line i in hours.

[0069] The retrofit time of line i t Nach,i,ges This is calculated, for example, as follows: t Nach , i , ges = sume M 1 , i . ∗ S i + sume M 2 , i . ∗ S i + sume M 3 , i . ∗ S i where M 1,i , M 2,i , M 3,i the retrofit time matrices for line i are (in this example there are 3 retrofit time matrices; however, more or less than 3 matrices are also possible), S i is the state change matrix for line i and sum means the sum over all elements of the matrix.

[0070] The production time of line i t Pro,i,ges In this example, it is calculated as follows: t Pro , i , ges = sum t Pro , i . ∗ z akt , i where t Pro , i = t Pro , i , 1 t Pro , i , 2 t Pro , i , 3 with t Pro,i,j as the production time of a state j on line i in hours (in this example there are 3 states j = 1, 2, 3; however, more or fewer states are also possible; in particular, different lines i may have different numbers of states), z akt,i a vector containing all selected states on line i is with z akt,i = ( sum ( S i , 2) + sum ( S i , 1) T< ). / ( sum ( S i , 2) + sum ( S i , 1) T< , and sum specifies the calculation of the sum of the elements of a vector (e.g., in a computational tool like Matlab).

[0071] As already mentioned in detail, information about how long one (or more) devices on one of the lines i can still be operated before a malfunction occurs (predicted by the trained machine learning algorithm 60). For example, information that a device can only process a certain number of product units (or can only run for a certain duration) before the malfunction occurs can be used as an additional boundary condition during the optimization of the production plan.

[0072] The optimization process results in an allocation of units (from the production plan) to the individual lines, as well as a sequence in which the units of the production plan should be processed on the individual lines to achieve maximum production efficiency for the entire plant and thus minimize costs. The setup time and production time can be defined as product- and line-dependent, so that the plant specifications are taken into account when determining the optimized production plan. Ultimately, an optimal production plan can be determined that minimizes the time required to produce all planned products. This enables optimal plant design.

[0073] This is exemplified in the Fig. 7 shown where the changeover time, the production times of each state j of each line i t Pro,i,j , the retrofit time matrices M j,i and the machine hour rate H i The first input 710 and the plant's line layout are provided as the second input 720 to a computing tool 730. As described above, the first input 710 can optionally also include the maintenance time, which is determined using the machine learning algorithm 60 trained on simulation data from the digital twin 10B, and / or the time until this maintenance measure must be carried out. Based on the first input 710 and the second input 720, the computing tool 730 then determines an optimal production plan 740. The optimal production plan specifies on which line and in which sequence the individual units of the production plan should be run, so that the production costs K are minimized.

[0074] The optimization process described above incorporates the changeover time as well as the production and retooling times for the units (of the production plan). As described above, information about upcoming maintenance measures (i.e., whether and what maintenance is required for a specific piece of equipment, how long this maintenance will take, and / or how much time remains to perform the maintenance in order to avoid downtime) can also be included. According to the embodiment presented here, it is possible to perform predictive maintenance on a piece of equipment when the system is already shut down (e.g., during a changeover). This allows the duration of a maintenance measure to be factored into the optimization of the production plan.

[0075] The Fig. 8Figure 800 shows a method according to an embodiment of the invention for determining an optimized production plan 740 for the production of one or more products using one or more production lines 721-1, ..., 721-N in the food or beverage industry. In a first step 810, the changeover time of the production plan is optimized. This first step 810 comprises dividing the production plan 820 into a plurality of related units and arranging the units 830 such that the changeover time is minimal. As described above, this first step 810 can additionally be performed based on the duration of a maintenance measure for a detected malfunction of a device 10A of a production line of the plurality of production lines 721-1, ..., 721-N.

[0076] Subsequently, in a second step 840, the production costs to be incurred for processing the units are optimized. This second step 840 includes assigning 850 the units to a product line and determining 860 a sequence in which the units are to be processed on the production lines. As described above, this second step 840 can additionally be performed based on the expected duration until the predicted malfunction occurs (which was determined using the trained machine learning algorithm 60).

[0077] It is further noted that the features mentioned in the embodiments described above are not limited to these specific combinations, but are also possible in any other combinations as further embodiments.

Claims

1. A method (800) for determining an optimized production plan (740) for producing one or more products by means of one or more production lines (721-1 ... 721-N) in the food industry or the beverage industry, the method comprising: optimizing (810) a changeover time of the production plan, wherein the changeover time comprises a product changeover, bottle changeover, product variant changeover, format changeover and / or packaging changeover during production of the products, wherein optimizing the changeover time comprises: • breaking down (820) the production plan into a plurality of interconnected units, wherein each of the units corresponds to at least one product, a processing step or at least one device of a production line for producing a product of the production plan; and • arranging (830) the units such that the changeover time assumes a minimum value; and optimizing (840) production costs of the production plan required for processing the units, wherein the production costs are determined by retrofitting times and production times for the units by means of the devices, wherein the retrofitting times and production times are defined as a function of the product and the device, and wherein optimizing the production costs comprises: • assigning (850) each of the units to a respective production line; and • determining (860) an order in which the units are to be processed by the production lines; wherein the changeover time and / or the production costs additionally comprise a duration of a maintenance operation for a detected operating fault of a device (10A) of a production line, wherein the operating fault is detected by a machine learning algorithm (60), wherein the machine learning algorithm is trained based on simulation data (20B; 40-1-1, 40-2-1, ..., 40-K-N) and operating faults (30-1, 30-2, ..., 30-K) of a digital twin (10B) of the device (10A), and wherein the machine learning algorithm is configured to detect the operating fault of the device (10A) based on operating data (20A) of the device.

2. Method (600) according to claim 1, wherein the detected operating fault of the device (10A) is an event, a failure event, and / or a maintenance operation to be performed, wherein, in particular: the event comprises increased friction, increased vibration, an imbalance, a wear parameter, or reduced lubrication; the failure event comprises motor damage, bearing damage, axle breakage, a crack, a mispositioned bottle, a short circuit, stoppage of the device, or any other type of mechanical damage; and the maintenance operation to be performed comprises cleaning, lubricating, oiling, fastening, adjusting, replacing, or repairing; and / or wherein detecting (630) the operating fault further comprises detecting a component of the device at which the operating fault occurs.

3. Method (600) according to claim 1 or 2, wherein detecting (630) the operating fault further comprises detecting an estimated time period until occurrence of a failure event caused by the operating fault.

4. Method (600) according to claim 3, wherein the production plan (740) is determined based on the estimated time period and / or on a duration of a maintenance operation to be performed for the detected operating fault.

5. Method (600) according to claim 4, wherein the operating data (20A) comprise output signals of the device (10A), wherein the simulation data (20B, 40-1-1, 40-2-1, ..., 40-K-N) comprise output signals of the digital twin (10B), wherein each output signal of the digital twin corresponds to a respective output signal of the device; and wherein each output signal of the device (10A) is measured by a sensor of the device that is used during intended operation of the device.

6. Method (600) according to claim 5, wherein the operating data (20A) of the device (10A) comprise at least one additional output signal of the device, the additional output signal being measured by an additional sensor of the device that is not required for intended operation of the device; wherein the simulation data (20B; 40-1-1, 40-2-1, ..., 40-K-N) comprise at least one corresponding additional output signal of the digital twin (10B); and wherein the additional sensor of the device (10A) is selected, by means of the digital twin (10B), from a plurality of possible additional sensors such that the output signal of the additional sensor optimizes detection (630) of the operating fault.

7. Method (600) according to any one of claims 1 to 6, wherein the trained machine learning algorithm (60) is configured, for detecting the operating fault, to determine a probability for each of the operating faults (30-1, 30-2, ..., 30-K) based on the operating data (20A) of the device (10A); and wherein the trained machine learning algorithm (60) detects (630) one of the operating faults (30-1, 30-2, ..., 30-K) as the operating fault when the probability of the operating fault reaches a threshold value.

8. Method (600) according to any one of claims 1 to 7, wherein the machine learning algorithm (60) is additionally trained using recorded operating data of the device (10A); and / or wherein the device (10A) is a tripod of the food industry.

9. A system (700) for determining an optimized production plan (740) for producing one or more products by means of one or more production lines (721-1 ... 721-N) in the food industry or the beverage industry, the system being configured to: optimize (810) a changeover time of the production plan, wherein the changeover time comprises a product changeover, bottle changeover, product variant changeover, format changeover and / or packaging changeover during production of the products, wherein optimizing the changeover time comprises: • breaking down (820) the production plan into a plurality of interconnected units, wherein each of the units corresponds to at least one processing step or at least one device of a production line for producing a product of the production plan; and • arranging (830) the units such that the changeover time assumes a minimum value; and optimize (840) production costs of the production plan required for processing the units, wherein the production costs are determined by retrofitting times and production times for the units by means of the devices, wherein the retrofitting times and production times are defined as a function of the product and the device, and wherein optimizing the production costs comprises: • assigning (850) each of the units to a respective production line; and • determining (860) an order in which the units are to be processed by the production lines; wherein the changeover time and / or the production costs additionally comprise a duration of a maintenance operation for a detected operating fault of a device (10A) of a production line, wherein the operating fault is detected by a machine learning algorithm (60), wherein the machine learning algorithm is trained based on simulation data (20B; 40-1-1, 40-2-1, ..., 40-K-N) and operating faults (30-1, 30-2, ..., 30-K) of a digital twin (10B) of the device (10A), and wherein the machine learning algorithm is configured to detect the operating fault of the device (10A) based on operating data (20A) of the device.

10. System (700) according to claim 9, wherein the detected operating fault of the device (10A) is an event, a failure event, and / or a maintenance operation to be performed, wherein, in particular: the event comprises increased friction, increased vibration, an imbalance, a wear parameter, or reduced lubrication; the failure event comprises motor damage, bearing damage, axle breakage, a crack, a mispositioned bottle, a short circuit, or stoppage of the device; and the maintenance operation to be performed comprises cleaning, lubricating, oiling, fastening, adjusting, replacing, or repairing; and / or wherein detecting (630) the operating fault further comprises detecting a component of the device at which the operating fault occurs.

11. System (700) according to claim 9 or 10, wherein detecting (630) the operating fault further comprises detecting an estimated time period until occurrence of a failure event caused by the operating fault.

12. System (700) according to any one of claims 9 to 11, wherein the production plan (740) is determined based on a duration of a maintenance operation to be performed for the predicted operating fault and on the estimated time period until occurrence of the failure event caused by the operating fault.

13. System (700) according to any one of claims 9 to 12, wherein the operating data (20A) comprise output signals of the device (10A), wherein the simulation data (20B, 40-1-1, 40-2-1, ..., 40-K-N) comprise output signals of the digital twin (10B), wherein each output signal of the digital twin corresponds to a respective output signal of the device; and wherein each output signal of the device (10A) is measured by a sensor of the device that is used during intended operation of the device.