WO2024079318A1

WO2024079318A1 - System and method for safety-related monitoring of an industrial plant

Info

Publication number: WO2024079318A1
Application number: PCT/EP2023/078490
Authority: WO
Inventors: Daniel Bakovic; Lukas ELWINGER
Original assignee: Pilz Gmbh & Co. Kg
Priority date: 2022-10-14
Filing date: 2023-10-13
Publication date: 2024-04-18

Abstract

The invention relates to a system (10, 1, 2, 3) for safety-related monitoring, in particular shut-down, of an industrial plant (12), said system comprising a controller (14) and a processing unit (16). The controller (14) is designed to use a safety-relevant input variable (18), which can be derived from at least one continuously recordable first measured value (20), to trigger a safety-related reaction based on the safety-relevant input variable in the event of a deviation from an expected input variable, said reaction causing the industrial plant (12) to switch to a safe state The processing unit (16) is designed to determine a virtual input variable (26) on the basis of continuously recordable second measured values (22), which do not fully correspond to the at least one first measured value (20), and to provide the controller (14) with said virtual input variable. The controller (14) triggers the safety-related reaction based on the safety-relevant input variable (18) if the virtual input variable (26) differs from the expected input variable.

Description

System and method for safety-related monitoring of a technical installation

[0001] The present invention relates to a system and a method for safety-related monitoring, in particular shutdown, of a technical system and a processing unit for use in such a system.

[0002] Safety-related monitoring of a technical system refers here to the safeguarding of hazardous areas on and around a technical system, in particular hazardous areas of a machine or industrial plant, by means of sensors and controls provided for this purpose. The sensors continuously record the state of the technical system and/or its surroundings and forward this information to the control system. The control system evaluates the information provided and transfers the technical system to a safe state if there is a danger to people or property. [0003] Sensors and controls are subject to special requirements in terms of function, availability and inherent fault tolerance, which are regulated by law and laid down in standards in most countries. In Europe, overarching standards include, for example, the standards on machine safety, in particular EN ISO 12100 (General principles for design - Risk assessment and risk reduction), EN ISO 13849-1/-2 (Safety-related parts of control systems) and EN IEC 62061 (Functional safety of safety-relevant control systems).

[0004] Safe drive systems can be used to safeguard hazardous areas. These are used to carry out a variety of safety functions in order to prevent dangerous situations when one and/or more operators intervene. Rotary encoders and/or integrated rotary encoders mounted on the electric drive are an essential component for safe position and speed determination. The normative basis for this is DIN EN 61800-5-2, which describes the requirements for the functional safety of electrical power drive systems with adjustable speed. Common safety functions described in the standard include safe torque off (STO), safe stop 1 (SS1), safe operating stop (SOS), safe stop 2 (SS2), safely limited speed (SLS), or safe direction of movement (SDI).

[0005] A signal chain for executing safety functions generally includes the safe detection of system variables by sensors, their safe processing and evaluation by a controller, and the safe control of actuators based on the processing and evaluation.

[0006] In the case of electric drives, rotary encoders can be built into the electric drive system internally or attached externally to safely detect the angular position, angular velocity and angular acceleration of a rotor. Depending on the risk reduction measures required, the safe rotary encoders can have a single-channel or two-channel redundant safety architecture up to a redundant diverse safety architecture consisting of two physical rotary encoders. Rotary encoders of this type as well as other safety-related sensors are complex systems. me and a decisive cost factor in the realization and implementation of safety-related monitoring of a technical system.

[0007] It is also known that an estimate of the rotor speed and/or rotor rpm is possible as a relevant system variable. Physical models can also be abstracted by mathematical differential equations in order to approximate relevant system variables by implicit system variables. In principle, initial approaches are also known to use artificial neural networks (ANN) for the approximation of important electrical and mechanical parameters in mechatronic systems. However, these approaches have not yet been taken into account in safety technology.

[0008] Against this background, it is an object to specify a system and method for the safety-related monitoring of a technical system that can be implemented cost-effectively and effectively and at the same time ensures standard-compliant and fail-safe monitoring.

[0009] According to one aspect of the present disclosure, this object is achieved by a system for safety-related monitoring, in particular shutdown, of a technical system, comprising: a controller that is set up to trigger a safety-related reaction related to the safety-relevant input variable based on a safety-relevant input variable that can be derived from at least one continuously detectable first measured value in the event of a deviation from an expected input variable, which causes the technical system to transition to a safe state, and a processing unit that is set up to determine a virtual input variable based on continuously detectable second measured values that do not completely correspond to the at least one first measured value and to provide it to the controller, wherein the controller is set up to trigger the safety-related reaction related to the safety-relevant input variable if the virtual input variable deviates from the expected input variable. [0010] According to a further aspect of the present disclosure, this object is achieved by a processing unit for use in a system for safety-related monitoring, in particular shutdown, of a technical installation, in which a controller, based on a safety-relevant input variable that can be derived from at least one continuously detectable first measured value, triggers a safety-related reaction related to the safety-relevant input variable in the event of a deviation from an expected input variable, which causes the technical installation to transition to a safe state, wherein the processing unit is configured to determine a virtual input variable based on continuously detectable second measured values that do not fully correspond to the at least one first measured value and to provide it to the controller, which triggers the safety-related reaction related to the safety-relevant input variable when the virtual input variable deviates from the expected input variable.

[0011] According to a further aspect of the present disclosure, this object is achieved by a method for safety-related monitoring, in particular shutdown, of a technical system, comprising: triggering a safety-related reaction related to a safety-relevant input variable by a controller, wherein the safety-related reaction causes a transition of the technical system to a safe state in the event of a deviation from an expected input variable and the safety-relevant input variable can be derived from at least one continuously detectable first measured value; determining a virtual input variable based on continuously detectable second measured values that do not completely correspond to the at least one first measured value; and providing the virtual input variable for the controller, wherein the controller triggers the safety-related reaction related to the safety-relevant input variable if the virtual input variable deviates from the expected input variable.

[0012] It is therefore an idea of the present invention to use a virtual sensor as an alternative or in addition to realize safety-related monitoring of a technical system and thereby save costs for expensive, complex and mechanically and/or electrically vulnerable sensors (primary sensors). The solution is based on using a conventional sensor, which is used to reliably provide a safety-relevant input The safety-relevant input variable is intended to be replaced by an alternative device which merely emulates the safety-relevant input variable. The safety-relevant input variable is a system variable which significantly influences the triggering of a safety-related reaction by a control system.

[0013] In contrast to a primary sensor, which provides a safety-relevant input variable based on a direct measurement of a measured value relevant to the safety-relevant input variable (first measured value), the alternative device is set up to determine a virtual input variable as the safety-relevant input variable based on other measured values (second measured values) that are only indirectly related to the safety-relevant input variable. The second measured values can represent, for example, current, temperature and/or vibration and can be recorded by the alternative device via secondary sensors. The secondary sensors can be designed to be simpler and more cost-effective than primary sensors and are present in many technical systems as part of the system or can be retrofitted at low cost.

[0014] The alternative device generates a virtual input variable as an equivalent for the safety-relevant input variable by continuously recording many different second measured values. For this purpose, the alternative device relates the second measured values to one another and weights them in order to emulate a representative safety-relevant input variable, which is ultimately used instead of or in addition to the safety-relevant input variable of a primary sensor to assess the hazardous situation and as a trigger for the safety-related reaction.

[0015] In this way, the number of complex primary sensors required to implement safety-related monitoring can be advantageously reduced, down to a system that relies exclusively on secondary sensors to ensure a safety level to be achieved via the alternative device.

[0016] The above-mentioned problem is thus completely solved. [0017] In a preferred embodiment, the system may further comprise a sensor which is configured to record the at least one first measured value, to generate the safety-relevant input variable from the at least one first measured value and to provide the safety-relevant input variable to the controller.

[0018] According to this embodiment, the system thus has a primary sensor, i.e. a conventional, real sensor that can provide the safety-relevant input variable in a known manner by continuously evaluating initial measured values. With the help of the sensor, a redundant safety architecture can be implemented in which a first channel is implemented by the sensor and a second, virtual channel is implemented by the processing unit. This makes it possible to implement safety-related monitoring in a simple and cost-effective manner even for systems that normatively require a two-channel safety architecture, but without necessarily having to use two physical sensors.

[0019] In a further embodiment, the processing unit can have an inference engine which is configured to generate the virtual input variable from the second measured values and a stored knowledge base.

[0020] An inference machine is generally referred to here as a data processing device that draws a conclusion from existing knowledge and available input data (second measured values), which in this case corresponds to the virtual input variable. The knowledge base can be, for example, a static or dynamic set of rules, a system model or a trained artificial neural network (ANN). The processing unit can draw conclusions from the second measured values and the knowledge base, for example as a rule interpreter, in order to determine the virtual input variable and to approximate the safety-relevant input variable. In the case of a trained artificial neural network, the processing unit can use the second measured values as input parameters for the trained artificial neural network, which returns the virtual input variable as an output. Alternatively or additionally, the use of Kalman filters to estimate the virtual input variable from the available second measured values is also conceivable. [0021] In a further embodiment, the processing unit can be configured to generate the knowledge base by machine learning, in particular by supervised learning and/or reinforcement learning

[0022] According to this embodiment, the processing unit can comprise, for example, an artificial intelligence system (AI system) that generates and/or expands the knowledge base through machine learning. The machine learning can comprise supervised learning, i.e. learning a function from given pairs of input and output values, and/or reinforcement learning, in which reinforcement learning algorithms extract relationships from data independently of an exact mathematical model by an agent independently learning a strategy through interaction with its environment in order to maximize the reward received. In addition, the AI system can also use artificial neural networks (ANN) with numerous intermediate layers between the input layer and the output layer in order to form a comprehensive internal structure that establishes a connection between the input data (second measured values) and the output data (virtual input variable for the control) (so-called deep learning or multi-layer learning). In this context, it is also conceivable that system relationships will be revealed that provide indications as to which additional parameters should possibly be recorded as further second measured values in order to improve the accuracy of the virtual input variable, or which and how many second measured values are required so that the given virtual input variable has sufficient significance.

[0023] In particular, the processing unit can be configured to perform the machine learning based on the second measured values and an output of the sensor.

[0024] According to this embodiment, the processing unit can learn a relationship between the second measured values and the safety-relevant input variable by observing the sensor and its output. The sensor can specify the correct output value for the safety-relevant input variable for different input states, which the processing unit relates to the recorded second measured values for the given input situation. In this way, the processing unit can build up a knowledge base in the sense of supervised learning. The design is particularly advantageous if, in a redundant safety structure, a primary sensor is only to be supplemented by the processing unit as a second channel.

[0025] Furthermore, the processing unit can be configured to perform the machine learning before the controller uses the virtual input variable.

[0026] The processing unit can thus learn the system relationships in a test phase in order to then apply them in operation. It is also conceivable that a test phase is carried out more frequently in order to continuously expand the knowledge base over time. In principle, it is also possible to continue the learning process during operation.

[0027] It is also conceivable to carry out machine learning on generically generated data by means of simulations before commissioning in order to create a knowledge base that maps the relationship between the second measured values and the safety-relevant system variable and/or to shorten the learning of the system relationships in a test phase.

[0028] In addition, the processing unit may have an interface to continuously expand the knowledge base.

[0029] According to this embodiment, the knowledge base can thus also be provided, expanded and/or updated externally. For example, updates based on manually recorded or automatically recognized new system relationships can be imported manually or automatically via the interface. It is also conceivable that the processing unit is connected to other processing units in a network or via a cloud service in order to exchange knowledge with them. Knowledge generation can thus advantageously also take place collectively.

[0030] In a further embodiment, the safety-relevant input variable can be the result of a direct signal processing of a first measured value or the result of a direct signal processing of the plurality of first measured values, wherein the virtual input variable is at least partially the result of an indirect signal processing of the second measured values, in particular an approximation and/or estimation.

[0031] The virtual input variable can thus be based in particular on implicit system variables that are not directly related to the safety-relevant input variable. In other words: the virtual input variable cannot be obtained solely by transforming measured values, but only by drawing conclusions based on other known or unknown system relationships. However, implicit measured values can advantageously be recorded more easily and at a lower cost than explicit system variables that are directly related to the safety-relevant input variable.

[0032] In a further embodiment, the controller can be set up to compare the safety-relevant input variable of a sensor with the virtual input variable and to trigger the safety-related reaction if the safety-relevant input variable and the virtual input variable differ from each other by a predefined amount.

[0033] According to this embodiment, a real and a virtual channel can thus advantageously be used as redundant processing channels.

[0034] In a further embodiment, the safety-relevant input variable and the virtual input variable can be identical in the way they represent a representative system variable of the technical installation, in particular in their unit and/or in their scope.

[0035] The virtual input variable and the safety-relevant input variable can therefore be identical in type, so that the controller cannot tell whether it is processing a virtual input variable or a safety-relevant input variable. In other words: The controller can read, process and evaluate the virtual input variable in the same way as the safety-relevant input variable that it receives from a private märsensor. An existing sensor can therefore be replaced without the need to adapt the control system.

[0036] In a further embodiment, the second measured values can be measured values inherent in the control of the technical system, at least in part.

[0037] Measured values inherent to the control are measured values that are relevant to the control of the technical system and are recorded in order to be able to control the technical system according to its function. The inherent measured values are usually recorded by simple sensors, also called standard sensors, and fed to a standard control system, e.g. a programmable logic controller (PLC), which uses these values as input to output certain control signals. Measured values from standard sensors can usually not be used or can only be used in addition to the implementation of a safety function, since the standard sensors that provide the measured values do not have the necessary inherent fault tolerance to ensure fail-safe recording and transmission. This is usually only possible with sensors that have appropriate safety-related devices. In a system according to the preferred embodiment, it is now possible to use the measured values of the standard sensors to implement a safety function if sufficient quality can be guaranteed by processing, e.g. using a Kl system. It is conceivable that the Kl system itself makes a statement about the quality of the virtual input variables provided and provides the virtual input variable depending on this. If the Kl system comes to the conclusion, for example, that the quality is insufficient, it can provide a corresponding virtual input variable or no input variable, whereupon the control system triggers the safety-related reaction.

[0038] The controller can in particular be a safety controller that has a multi-channel redundant structure to ensure a fail-safe execution of the safety function. The multi-channel, redundant structure can in particular provide two or more processing channels that process an input and provide an output separately and independently of one another. In addition, the channels can be set up in such a way that they monitor each other and deviating processing triggers the safety-related reaction. The person skilled in the art is familiar with corresponding safe control methods, in particular those in which a control system only allows the operation of a technical system if suitable input signals (safety-relevant input variables) are present and signal a safe state.

[0039] It is understood that the features mentioned above and those to be explained below can be used not only in the combination specified in each case, but also in other combinations or on their own, without departing from the scope of the present invention.

[0040] Embodiments of the invention are illustrated in the drawing and are explained in more detail in the following description.

Fig. 1 shows a schematic view of a system according to a first embodiment of the present disclosure.

Fig.2 shows a schematic view of an application scenario for a system according to an embodiment of the present disclosure.

Fig. 3 shows a detailed view of the system according to the first embodiment of the present disclosure.

Fig. 4 shows a schematic view of a system according to a second embodiment with a single-channel safety structure.

Fig. 5 shows a schematic view of a system according to a third embodiment with a two-channel security structure.

Fig. 6 shows a schematic view of a system according to a fourth embodiment with a two-channel security structure. Fig. 7 shows a schematic representation of a system with an information processing device according to another aspect of the present disclosure.

Fig. 8 shows a schematic representation of an architecture for using the information processing device according to Fig. 7.

Fig. 9 shows a simplified flow chart in a schematic representation.

Fig. 10 shows a schematic representation of an example of a neural network (autoencoder).

Fig. 11 shows a schematic representation of an example of another neural network.

[0041] Fig. 1 shows a schematic view of a system according to a first embodiment of the present disclosure. The system 10 is used for safety-related monitoring, in particular shutdown, of a technical system 12.

[0042] The technical system 12 can in particular be a robot or a machine in an industrial environment that can pose a danger to people and/or objects in its surroundings. In order to avoid or minimize injury to people or objects due to these dangers, the technical system 12 is monitored in a safety-related manner. Parts of the technical system 12 and/or its surroundings are monitored by suitable sensors and, if necessary, a safety-related reaction, in particular a shutdown of the technical system 12, is triggered on the basis of the monitoring. The detection of safety-relevant events, their evaluation and the subsequent triggering of a safety-related reaction are referred to below as a safety function. The system 10 represents a technical implementation of such a safety function. [0043] The system 10 has a controller 14 and at least one processing unit 16. The controller 14 is designed to trigger a safety-related reaction depending on a safety-relevant input variable 18. The safety-relevant input variable 18 is a signal that represents a safety-relevant state or a safety-relevant event in or on the technical system 12.

[0044] The safety-relevant input variable 18 can be a system variable that represents a state of parts of the technical system 12 or its environment. The controller 14 evaluates the safety-relevant input variable 18 and triggers the safety-related reaction depending on this. This means that the controller can trigger the safety-related reaction if the safety-relevant input variable 18 is either not present or does not correspond to a defined expectation. The safety-related reaction can in particular consist of switching off parts of the technical system 12 or transferring the technical system 12 to a safe state. The controller can also be set up to prevent the technical system 12 from starting up if the safety-relevant input variable does not correspond to the defined expectation.

[0045] The controller 16 can in particular be a fail-safe controller (FS controller) which, compared to a normal controller (standard controller), has special safety-related devices to ensure a fail-safe execution of the safety function. The special safety-related devices enable the fail-safe controller to ensure that the safety-related reaction is triggered when a safety-relevant event occurs. The FS controller can in particular switch off the technical device 12 in a fail-safe manner or prevent it from starting up in a fail-safe manner, in particular by means of a multi-channel, redundant structure of the internal signal processing.

[0046] The safety-related input variable 18 is determined from physical measured values 20. The measured values 20 can relate to chemical, physical or biological effects in or around the technical system 12. Usually, one or more measured values 20 are provided by a dedicated sensor. The safety-relevant input variable 18 can be determined by a (real) sensor 24. which directly accesses one or more measured values 20 (first measured values). The first measured values 20 are directly representative of the safety-relevant input variable 18, ie the first measured values 20 can be directly converted into the safety-relevant input variable 18 by processing, scaling and/or normalization. The first measured values 20 are therefore explicit measured values from which the safety-relevant input variable 18 can be directly deduced.

[0047] The processing unit 16 is set up to determine a virtual input variable 26 and to transfer it to the controller 14. The virtual input variable 26 can be equivalent in type and scope to the safety-relevant input variable 18, but can be based on other measured values 22 (second measured values). The second measured values 22 correspond to measured values that, in contrast to the first measured values 20, cannot be directly converted into the safety-relevant input variable 18 by processing, scaling and/or normalization. The second measured values 22 can therefore also be referred to as implicit measured values and the first measured values 20, in contrast, can also be referred to as explicit measured values.

[0048] The processing unit 16 has at least one device 28 that is able to derive the virtual input variable 26 from a large number of implicit second measured values. The device 28 can preferably be an artificial intelligence system (hereinafter referred to as AI system) that provides the virtual input variable 26 based on the second measured values and stored knowledge 30, which approximates the safety-relevant input variable 18. The processing unit 16 is thus set up to continuously provide a virtual input variable 26 as an approximated safety-relevant input variable from continuously recorded second measured values 22. The virtual input variable 26 can then be further processed by the controller 14 in accordance with the legal and normative requirements.

[0049] The Kl system can, for example, use a trained artificial neural network (ANN) as “knowledge” 30, to which the second measured values 22 are passed as input parameters and whose output corresponds to the virtual input variable 26. The training of the artificial neural network can, for example, be carried out in advance using the real sensor 24, as will be explained in more detail below. Alternatively, the trained artificial neural network can also be taken from another application of the technical system 12. In addition, the artificial neural network can also be continuously trained if the real sensor 24 is also present in a productive system. This can be the case in particular if the productive system has a two-channel security architecture in which a first channel is implemented by the real sensor 24 and a second channel by the processing unit 16.

[0050] It is understood that the device 28 can alternatively or additionally be implemented by an estimator, for example in the form of a Kalman filter, which estimates the safety-relevant input variable 18 on the basis of a system model and the second measured values 22. In comparison to a Kl system, however, this can lead to a loss of accuracy, since the Kl system is better able to recognize and take into account complex physical dependencies, especially when there is little explicit (systematic) knowledge.

[0051] Through reinforcement and/or supervised learning of the AI system, complex physical relationships can advantageously be automatically approximated. In comparison to classical methods based on observers and estimators, no laborious parameters have to be determined manually and optimized through iterative development loops in the system behavior. In principle, it is sufficient for an AI system to understand the system behavior of a technical system only to the extent that it is necessary to identify suitable system parameters that can be recorded as implicit measured values.

[0052] The virtual input variable 26 provided by the processing unit 16 can be further processed by the controller 14 in accordance with the legal and normative requirements, so that the use of a real sensor 24 can be dispensed with. Depending on the application, the processing unit 16 can also be used in addition to a real sensor 24, so that, for example, in a system with a two-channel safety architecture, one channel is implemented by the real sensor 24, while the real sensor 24 is omitted in the second channel. This can advantageously reduce the costs for the safety-related monitoring of a technical system 12.

[0053] In addition, a virtual channel also enables the upgrading of standard sensors for safety-critical applications through plausibility checks or tests.

[0054] The processing unit 16 can be based on a data processing device that has standard processing units (CPUs) or processing units adapted for the application (GPU, ASIC, etc.). It is also conceivable that the data processing of the processing unit 16 is outsourced to an external service, for example a cloud application. Likewise, the knowledge 30 can be stored in a memory of the processing unit 16 or made available via an interface to a data memory of the processing unit 16.

[0055] An application scenario for a system according to an embodiment of this disclosure is shown below with reference to Fig. 2. The same reference numerals are used for the same parts that are identical to those of the embodiment according to Fig. 1.

[0056] Fig. 2 shows a robot 32 as an example of a technical system 12. The robot 32 is arranged so that it can rotate on a base 34 and has a manipulator 36 with a tool 38 attached to the end. The robot 32 can be a commercially available industrial robot. The robot 32 has drives, gears and joints to move the tool 38 into working positions and to carry out activities. The motor 40 is shown here as an example of the drives. However, it goes without saying that the motor 40 can also be arranged inside the robot 32. The motor 40 drives an axis 41 and transmits a torque to the robot 32.

[0057] The motor 40 is coupled to a power supply 42 which supplies the motor 40 with energy. In the power supply 44 from the power supply 42 to the motor 40, Contactors 46 are arranged, which must be actively energized (energized) so that energy is transferred to the motor 40. The contactors 46 are redundantly coupled to a fail-safe controller 14 via lines 48. When the controller 14 generates an output signal on the lines 48, the contactors 46 are energized and the motor 40 is energized.

[0058] The provision of the output signal by the controller 14 depends on an input signal 52 applied to an input module 50. The input signal 52 here comes from a fail-safe rotary encoder 54 which is coupled to the rotary axis 41 driven by the motor 40. The rotary encoder 54 corresponds to the real sensor 24 (primary sensor) described with reference to Fig. 1 and has one or more measuring sensors which record first measured values 20 and provide them as an electrical signal. The rotary encoder 54 converts the signals from the measuring sensors into relevant system variables such as an angular position, an angular velocity or an angular acceleration of the rotary axis. A safety-relevant event can be derived from the relevant system variable, as a result of which a safety-related reaction is to be triggered. The relevant system variable can thus be a safety-relevant input variable 18 in the sense of the present disclosure and can be fed to the controller 14. The controller 14 can compare the safety-relevant input variable 18 with an expected input variable and, if necessary, trigger a safety-related reaction via the connected contactors 46 and disconnect the motor 14 from the power supply 42.

[0059] An additional module 56 is also coupled to the controller 14, which has a processing unit 16 according to the present disclosure. The additional module 56 is coupled to a plurality of second measuring sensors 58, 60, 62, 64, 65, which provide second measured values 22 to the processing unit 16. The second measured values determined by the second measuring sensors 58, 60, 62, 64, 65 cannot be directly converted into the safety-relevant system variable by processing, scaling and/or normalization, or are not sufficient in terms of their accuracy or reliability to indicate the safety-relevant system variable alone within a defined tolerance range. [0060] The second measuring sensors can be, for example, an ammeter 58 in the power supply 44 to the motor 40, a temperature sensor 60 on the robot 32 or a vibration sensor 62 on the robot 32. The second measuring sensors can also be designed as an interface 64 to a standard control, via which the control parameters of the robot 32 can be determined as second measured values. Finally, a simple rotary encoder 65 can also serve as the second measuring sensor, the measured values of which alone are not sufficient to explicitly specify the safety-relevant system variable with a corresponding level of reliability or accuracy.

[0061] The second measuring sensors 58, 60, 62, 64, 65 are in particular measuring sensors that are basically present for the operation of the technical system 12. Thus, the second measured values 22 recorded by the second measuring sensors 58, 60, 62, 64, 65 can be measured values that must necessarily be recorded for the control of the technical system 12. Furthermore, the second measuring sensors 58, 60, 62, 64, 65 can be simple sensors that directly determine a physical parameter.

[0062] Of course, the second measuring sensors 58, 60, 62, 64, 65 are not limited to the measuring sensors shown here. Preferably, all parameters recorded on the technical system 12 are fed to the processing unit 16, including those that are not explicitly linked to the safety-relevant input variable via a model of the technical system 12. Rather, it can be deliberately left to the Kl system to recognize the relevant system dependencies itself. Accordingly, it is conceivable to also provide the processing unit 16 with measured values as second measured values that are not recorded directly in relation to the technical system 12, but can be assigned to the environment of the technical system 12 and/or other actors in the environment of the technical system 12.

[0063] It is understood that the number of second measured values 22 in productive operation may differ from the number of second measured values 22 during a training phase. For example, it is conceivable that the Kl system initially determines from a large number of second measured values 22 those measured values that are required for productive operation in order to achieve a sufficient approximation of the safety-relevant inputs. input variable 18. In the productive system, only the second measured values 22 relevant for the evaluation are then fed to the processing unit 16.

This makes it possible to realize efficient evaluation even on cost-effective hardware.

[0064] The second measuring sensors 58, 60, 62, 64, 65 are coupled to the further module 56 in order to provide the second measured values 22 to the processing unit 16. The processing unit 16 processes the second measured values 22, in particular by means of the previously described Kl system, and generates a virtual input variable 26 which approximates the safety-relevant input variable 18. The virtual input variable 26 is fed to the controller 14, which can trigger the safety-relevant reaction based on this. In the present case, the controller 14 can, for example, only provide the output signal for the contactors 46 if the virtual input variable 26 provided by the processing unit 16 is present and is within a predetermined value range.

[0065] In the embodiment shown in Fig. 2, the two security measures complement each other to form a redundant safety architecture. The input module 50, together with the fail-safe sensor 54, forms a first, classic channel of the redundant safety architecture. The further module 56 with the processing unit 16 and the second measuring sensors 58, 60, 62, 64, 65 forms a second, virtual channel. By cross-matching the two channels, the controller 14 can ensure sufficient safety even if one of the two channels is faulty.

[0066] By designing the second channel as a virtual channel, real, particularly complex, sensors such as another secure rotary encoder 54 can advantageously be saved. In addition, the virtual channel can easily be adapted to other conditions or technical systems 12, preferably without having to make significant hardware adjustments. The claimed system can therefore be used flexibly for various applications. [0067] By designing the second channel as a virtual channel, real, particularly complex sensors, such as another secure rotary encoder 54, can advantageously be saved. In addition, the virtual channel can easily be adapted to other conditions or technical systems 12, preferably without having to make significant hardware adjustments. The claimed system can therefore be used flexibly for various applications.

[0068] Fig. 3 shows the system from Fig. 1 in a detailed view. The same reference numerals here also designate the same parts as previously in relation to Fig. 1.

[0069] In Fig. 3, the real sensor 24 and the processing unit 16 are shown in a detailed view. The real sensor 24 therefore has an input unit 66, an evaluation unit 68 and an output unit 70. Similarly, the processing unit 16 has an input unit 72, a Kl system 74 and an output unit 76.

[0070] The input unit 66 is designed to record at least a first measured value 20 based on chemical, physical or biological effects and to convert it into a corresponding electronic signal 78. The evaluation unit 68 processes the electronic signal 78 and forwards the processed electronic signal 80 to the output unit 70. The processing by the evaluation unit 68 can include processing, scaling and/or normalizing the electronic signal 78. For this purpose, sensor characteristics, calibration data, compensation equations and other processing algorithms can be stored in the evaluation unit 68. The output unit 70 has one or more interfaces to make the processed electronic signal 80 available to a downstream controller 14 in analog and/or digital form as a safety-relevant input variable 18.

[0071] The input unit 72 of the processing unit 16 can record second measured values 22 that differ from the first measured values 20, analogously to the input unit 66. In addition, the input unit 72 can also record parameters that are not attributable to chemical, physical or biological effects. For example, the input unit 72 can also record setting parameters of the technical System 12 or other data describing the technical system 12 or its environment. The recorded second measured values 22 are transferred to the Kl system 74 as input 82 for evaluation. As part of a stationary and/or dynamic inference, the Kl system 74 processes the input 82 and generates an output 84, which is provided by the output unit 76 as a virtual input variable 26 in a manner analogous to the output unit 70 of the real sensor 24. The virtual input variable 26 corresponds in type and scope to the safety-relevant input variable 18, so that the virtual input variable 26 can be processed by the controller 14 in the same way, preferably redundantly to the safety-relevant input variable 18 of the real sensor 24.

[0072] The Kl system 74 processes the input 82, whereby the function for approximation and the relevant knowledge can be generated by supervised and/or reinforcement learning. The same Kl system 74 can be used for this, whereby it operates in a training environment or is supplied with additional training data. In particular, the real sensor 24 can act as a "teacher" in the training environment. However, the productive system does not require a real sensor 24. Rather, in a single-channel system, only the Kl system 74 can be used, which generates the virtual input variable 26, which is used exclusively by a downstream controller 14. A corresponding system is shown in Fig. 4 and is explained below.

[0073] Fig. 4 shows a schematic view of a system 1 according to a second embodiment with a single-channel security structure. Here too, the same reference numerals correspond to the same parts as in Fig. 1.

[0074] The single-channel system 1 has a processing unit 16 and a controller 14. The processing unit 16 is constructed as explained in more detail with reference to Fig. 1 and Fig. 3 and generates a virtual input variable 26 from second measured values 22, which can be processed by the controller 14 like a safety-relevant input variable 18. The single-channel system 1 can thus completely dispense with a real sensor that explicitly records measured values (first measured values) with regard to the safety-relevant input variable 18. In the single-channel system 1, the controller 14 thus evaluates only an emulated safety-relevant input variable based on implicit second measured values 22.

[0075] Here too, the device 28 can be a Kl system, which in this case was previously trained with a real sensor, whereby the real sensor is no longer present in the later production system. Alternatively, the Kl system can also use pre-generated knowledge 30, e.g. an artificial neural network trained for the given application, which was trained on a comparable technical system 12 and/or in a simulation on a digital twin of the real production system. Since the knowledge 30 is thus essentially stored in the form of data, the processing unit 16 can be easily and flexibly adapted to changed conditions without hardware having to be replaced.

[0076] Fig. 5 shows a schematic view of a system 2 according to a third embodiment with a two-channel security structure.

[0077] The two-channel system 2 essentially corresponds to the first embodiment, whereby the real sensor 24 is not optional here. The two-channel system 2 thus has a real sensor 24 in a first processing channel 86 and a processing unit 16 with a Kl system in a second processing channel 88. This embodiment has the advantage that, compared to a classic two-channel system in which two real sensors are used, one real sensor can be saved. In addition, this embodiment increases the diversity of the system, since the two channels are based on different functional principles and thus common cause failures (CCF) can be effectively excluded.

[0078] Alternatively, the two-channel system can also be formed from a first processing channel and a second processing channel, each of which has a processing unit 16. A corresponding embodiment is shown in Fig. 6.

[0079] The two-channel system 3 according to Fig. 6 therefore has two (or more) processing units 16A, 16B which exclusively process virtual input variables 26A, 26B for the downstream controller 14. The processing unit 16A realizes a first processing channel 90 and the processing unit 16B realizes a second processing channel 92.

[0080] The embodiment according to Fig. 6 thus implements a two-channel system 3 that does not require any complex real sensors and therefore does not record any explicit measured values in relation to the safety-relevant input variable. In this way, a two-channel system can be implemented particularly cost-effectively and efficiently. It goes without saying that in this case the processing units 16A, 16B can be designed differently, for example by using different knowledge to approximate the safety-relevant input variable. It is also conceivable that the processing units 16A, 16B record different second measured values 22 in order to feed them to the Kl system as input.

[0081] It goes without saying that the invention is not limited to one- or two-channel embodiments, but by adding channels with additional real sensors and/or AI-based processing units, systems with more channels can also be realized.

[0082] Finally, it should be noted that elements of the disclosed devices and systems can be implemented by suitable hardware and/or software elements, e.g. suitable circuits. A circuit is a structural arrangement of electronic components, including conventional circuit elements, integrated circuits, including application-specific integrated circuits, standard integrated circuits, application-specific standard products and field-programmable gate arrays. In addition, a circuit can include central processing units, graphics processors and microprocessors that are programmed or configured according to a software code. A circuit is not pure software, even if it contains the hardware described above that executes software. [0083] A further aspect of the present disclosure relates to a device, in particular an information processing device, and/or a method for KI-based analysis, in particular in electric drives and devices driven thereby.

[0084] Analyses based on current and/or vibration analyses have been used for some time to detect, report and monitor motors and drive units. The previous approaches are based on statically or adaptively defined threshold values that are determined during the development or commissioning of the motor or drive unit. However, this does not allow dynamic adaptation to specific dependencies during a process (e.g. during the start-up process). Parameters, patterns and weightings must be detected and defined manually, often using only a few parameters (e.g. current and vibration) because humans can only see a certain amount of information and recognize patterns and relationships in it. Much other system information that is in principle available in the form of parameters such as temperature, speed, axis angle, humidity, friction, torsion and strain is not used or is used insufficiently in the evaluations. This means that system information that could be relevant for better behavior prediction is lost or not used.

[0085] If system information is intelligently evaluated in isolated applications, this is done centrally on dedicated computer systems. This means that the information must be transported away from the location where it was collected. In the case of a computer system outside the company (e.g. cloud service), this leads to a loss of data integrity and data sovereignty. In addition, transporting the data to a (central) separate evaluation means increased energy consumption. Data communication must also be reliable and stable in order to avoid failures and reduced availability. Such issues are particularly problematic in mobile applications such as driverless transport vehicles (Automatic Guided Vehicles (AGV)), autonomous mobile robots (AMR) and drones.

[0086] One task may therefore be to provide improved analysis options. In particular, one task may be to analyze static and dynamic behaviors by evaluating (many) different parameters without having to manually determine the relationships between these and the system behavior beforehand during development. Another task can be to specify an analysis option that avoids a complex transfer of data and ensures data integrity and data sovereignty. At the same time, it can be a task to provide an analysis option that can be implemented efficiently and effectively.

[0087] One approach is an information processing device and/or a method for anomaly detection and/or classification of driving profiles and movement patterns and/or interpolation of downtimes and/or autotuning of control parameters of individual and/or linked axes via real and/or simulated and/or recorded, as well as real and/or simulated and/or synthetic sensor data, by means of sensor data-supported (electrical power, motor currents, torques, mass moments of inertia, electrical voltage, vibrations, motor speeds, axle angle values, temperatures, time, date, strains/deformations, air humidity, aerosol composition, LiDAR, radar, etc., in particular) Kl algorithms, i.e. neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM, ...).

[0088] It is therefore conceivable that a control logic contains an intelligent evaluation algorithm (Kl) which, based on its feature extractions learned in a learning process, generates the best possible estimates, so-called output values, in inference mode for new input variables. These can be used to implement wear detection and/or the prediction of failures and/or the detection of impermissible operating states of one or more electric drives and the movement axes mechanically coupled to them. Supervised, unsupervised or reinforcement learning processes can be used as learning processes. In a separate process, these learn to independently approximate the basic target function of the system from existing data. In order to be able to continuously improve a system, data must be collected, archived and played back in inference mode. This process should be carried out periodically. [0089] By continuously collecting information about the position, state and context of the axes in a drive and a stored learning model, statements about wear and expected damage to driven machine parts can be predicted (Physics-aware AI). The inference can be carried out directly on embedded hardware and no longer on a dedicated local computer or local and external data centers.

[0090] The sensor data recorded for controlling the mechanics, e.g. power, motor currents, electrical voltage, vibrations, motor speeds, axis angle values, temperatures, strains/deformations, humidity, aerosol composition, LiDAR or radar, are recorded in the drive and can be trained to the course pattern valid during operation.

[0091] By adding application-specific background information as coded domain knowledge, such as path radii, driving speed, current load (transport load), temperature (engine, environment, ...), energy consumption, battery charge level, service calendar, statements about the utilization of the machine parts can be made with the help of appropriate algorithms for anomaly detection.

[0092] This knowledge can be used to predict overloads and wear of the mechanics and thus provide a diagnostic option for defects, errors and undesirable behavior, wear, increased power loss, incorrect parameterization of the drive train and its individual components. This can be used for preventive maintenance routines on the machine or to reduce and optimize energy consumption.

[0093] An embodiment of the present disclosure may comprise an information processing device and/or a method which, in an evaluation unit, by means of an intelligent evaluation algorithm on one and/or more electric drives and, if applicable, mechanical movement axes coupled thereto, by recording measured values by one and/or more sensors, by analyzing and evaluating these measured values, enables anomaly detection and/or early prediction of wear and/or early prediction of possible component failures and/or prediction of quality characteristics of the manufacturing process in which the intelligently monitored electrical drives and/or mechanical motion axes are involved.

[0094] In one embodiment, it can be provided that the measured values of different sensors, which represent any physical quantity in an interacting mechatronic system to be monitored, can be read in analog and/or digitally via one and/or more interfaces.

[0095] In one embodiment, it can be provided that these interfaces are hardware-based and/or wireless communication paths.

[0096] In one embodiment, it can be provided that the evaluation unit is a physical information processing device. However, it is also conceivable that the evaluation takes place by means of the intelligent evaluation algorithm in a cloud-based runtime environment. In this case, the physical information processing device can function as a sensor data transmitter.

[0097] In one embodiment, it can be provided that the information processing device has communication paths to local and/or cloud-based application programs.

[0098] In one embodiment, it can be provided that the information processing device has an integrated control unit and offers a readout of the predictions and/or estimates made in the evaluation algorithm via one and/or several diverse interfaces. Further manually configurable binary switching outputs can be integrated in the control unit.

[0099] In one embodiment, it can be provided that the control unit is formed by an evaluation unit physically contained in the information processing device. Alternatively or additionally, the control unit can also be formed by an evaluation unit physically contained in the An evaluation algorithm implemented in the cloud must be set up, which can control the control unit.

[0100] In one embodiment, it can be provided that by functionally different sensors and the associated diverse measurement value acquisition in the interacting mechatronic system to be monitored, the smallest changes can be detected by linking a large number of physical variables that are unique in the system.

[0101] In one embodiment, it can be provided that the evaluation algorithm comprises one and/or more algorithms.

[0102] In one embodiment, it can be provided that the evaluation algorithm comprises generalized feature extractions which are learned in a supervised and/or unsupervised learning process using real and/or synthetically generated data for anomaly detection and/or for early wear prediction and/or for early prediction of possible component failures and/or for prediction of quality characteristics via the manufacturing process in which the intelligently monitored electric drives and/or mechanical motion axes are involved.

[0103] In one embodiment, it can be provided that the approximated feature extractions are stored as one or more models in a program memory.

[0104] In one embodiment, it can be provided that the evaluation unit comprises an interface to a local user program and/or a cloud-based web service in which configurations and data can be managed and the intelligent evaluation algorithm can be trained for new situations and system-changing interventions, modifications and changes. [0105] In one embodiment, it can be provided that machine learning methods are used for the intelligent evaluation algorithm and an inference can be carried out in real time.

[0106] In one embodiment, it can be provided that the evaluation algorithm continuously improves anomaly detection and/or an early wear prediction and/or an early prediction of possible component failures and/or a prediction of quality characteristics of the manufacturing process in which the intelligently monitored electric drives and/or mechanical motion axes are involved, through new training cycles with new training data locally or cloud-based.

[0107] In one embodiment, it can be provided that the measured values are physical quantities that are measured by direct and/or indirect measuring devices, and that the measuring devices are arranged at different locations of the mechatronic system to be monitored.

[0108] In one embodiment, it can be provided that the measured values to be recorded and evaluated are one or more of the following physical quantities and/or system quantities: temperature, electric current, electric voltage, electric field strength, inductance, electric power, torque, motor speed, mechanical stresses, force, mechanical energy, angular momentum, air pressure, air humidity, mechanical vibrations, mechanical strains, rotational position, translational position, angular velocity, angular acceleration, angular momentum, speed, acceleration, impulse, jerk, frequency, time and/or angular frequency.

[0109] In one embodiment, it can be provided that the accuracy of the statement and estimation of the evaluation algorithms for anomaly detection and/or an early prediction of wear and/or the early prediction of possible component failures and/or the prediction of quality characteristics over the production process can be improved by application-related process information and/or by coded domain knowledge. process involving the intelligently monitored electrical drives and/or mechanical motion axes.

[0110] In one embodiment, the application-related process information may include path radii, travel speeds, current load, energy consumption, battery charge status, service and maintenance calendars.

[0111] By collecting and evaluating a large number of measured values that represent physical quantities in a complex interacting mechatronic system, the accuracy can be increased enormously compared to processing by manually programmed algorithms, since deep neural networks perform feature extraction on unstructured data.

[0112] Through unsupervised and/or supervised learning, the feature extractions can be used for anomaly detection and/or for early wear prediction and/or for early prediction of possible component failures and/or for prediction of quality characteristics of the manufacturing process in which the intelligently monitored electrical drives and/or mechanical motion axes are involved, both in stationary machine operation and in dynamic machine operation. With conventional methods, only the stationary operating case could be monitored sufficiently well so far.

[0113] Further aspects and features of the technology according to the present disclosure can be found in Figs. 7 to 11 and the description below.

[0114] Fig. 7 shows a schematic representation of a system with an information processing device. The information processing device is designated here by the reference number 200.

[0115] A system 100 here has at least one mechatronic system with an electric drive 110. The electric drive can have mechanically coupled axes of movement. One or more sensors on the electric drive provide one or more measured values 111. In addition, system parameters 112 of the mechatronic system can be provided.

[0116] The measured values/system parameters 111, 112 are fed to an information processing device 200. For this purpose, the information processing device 200 can be coupled to the mechatronic system via an interface 210. The interface 210 can be wired and/or wireless, analog and/or digital.

[0117] The information processing device 200 can centrally have one or more evaluation units 220a, 220b, which execute one or more evaluation algorithms alone or jointly. An evaluation unit 220a can be physically integrated into the information processing device 200. Alternatively or additionally, an evaluation unit 220b can be implemented virtually on an external device 400, for example as a local or cloud-based application program.

[0118] The information processing device 200 can also have a program memory (among other things for storing feature extracts) and a control unit 230. The control unit 230 can have at least one interface via which the information processing device 200 is coupled to the outside world. The information processing device 200 can output control signals via the interface, for example. The control signals can be analog or digital control signals, in particular binary control signals 300. Furthermore, the control signals can be control values 301, 302, e.g. for predicting an anomaly or a wear prediction.

[0119] Fig. 8 schematically shows an architecture in which an information processing device 200 can be used.

[0120] In the system shown, a control logic contains an intelligent evaluation algorithm (Kl) which, on the basis of its feature extractions learned in a learning process, generates the best possible estimates of new input variables in an inference operation, which are referred to as output values. These can be used to implement wear detection and/or the prediction of failures and/or the detection of impermissible operating states of one or more electric drives and the movement axes mechanically coupled to them. Supervised, unsupervised or reinforcement learning methods can be used as learning methods. In a separate process, these learn to independently approximate the basic target function of the system from existing data. In order to be able to continuously improve a system, data must be recorded, archived and fed back into the inference process. This process must be carried out periodically.

[0121] Fig. 9 shows schematically a simplified flow diagram of a process briefly explained below.

[0122] In the first step, the sensor data is read in. This can include reading in the data from connected sensors or reading from a sensor database (archived sensor data).

[0123] In the next step, the sensor data is processed. In this process, the respective parameter values can be normalized to values from 0 to 1. Incorrect data can also be filtered and removed in this step. If necessary and possible, the sensor data is also synthetically multiplied in this step.

[0124] The processed data are then fed into an intelligent evaluation algorithm. The evaluation algorithm can be roughly divided into the following sub-processes: creating a CL model, training the CL model, deploying the CL model, archiving inference and sensor data.

[0125] The creation of the Kl model can include the selection of a suitable architecture (autoencoder, "normal" neural network, KNN, SVM, ...) and a suitable model (number of layers and parameters, ...). Examples of a suitable architecture are shown in Figs. 10 and 11. The selection can be made in the time interval in which the archived sensor data is played back. [0126] When training the Kl model, the number of epochs and batch sizes can be specified, as well as the accuracy and loss at which the training should be considered complete. The training can take place in the time interval in which the archived sensor data is played back and/or a new Kl model is created.

[0127] Deployment refers to the transfer and loading of the code representing the AI model onto the information processing device. This can also take place in the time interval in which the archived sensor data is played back and/or a new AI model is created and/or trained.

[0128] The inference is then carried out on the information processing device using the processed sensor data. This can be done sequentially in the process with other tasks or in parallel to them in a separate task/thread.

[0129] The sensor data are then archived and the raw data from the input sensors are archived or logged locally or on servers. If archiving is done locally, the data can also be backed up on a higher-level system at a defined time interval, e.g. periodically on a weekly basis.

[0130] After archiving, the process starts again.

[0131] A variant/extension level represents an adapted system that continues to actively learn and adapts its behavior dynamically and independently. This leads to a constant improvement in the understanding of the system and thus to an improved output or instruction.

[0132] A further variant/expansion stage is an adapted system that, in addition to inference, also enables data processing and training in the device.

[0133] It is conceivable that a type of control function (open control loop) with inference is initially implemented on a fixed, previously determined database that is not updated during operation. The variants mentioned are then to be understood as a type of control function, in which the database or the model adapts by updating the database and reinforcement learning.

[0134] Various application examples are specified below. In particular, the technology can be used in stationary single- or multi-axis coupled motion applications, e.g. for anomaly detection in the drive train (electrical, pneumatic, hydraulic, ...), classification of movement patterns and driving profiles, interpolation of downtimes of the individual or linked axes or KL-supported autotuning for controller parameterization.

[0135] In addition, the technology can be used advantageously in mobile, single- or multi-axis platforms, e.g. for anomaly detection in drive trains of all kinds (wheel, chain, running, flight, underwater drives, ...), a classification of driving profiles and movement patterns of all kinds, an interpolation of downtimes of drive trains of all kinds or an AI-supported autotuning for controller parameterization.

[0136] The technology can also be used advantageously in single and multi-machine systems. Here, an edge device can take over the entire task chain (sensor data acquisition, processing, inference, action instructions) for individual machines or a network of several similar machines.

[0137] A concrete application example is the use of the disclosed technology in driverless transport vehicles (Automatic Guided Vehicles (AGV)) or autonomous mobile robots (AMR).

[0138] For example, anomaly detection in the drive train of mobile robotic platforms can be carried out with an information processing device and/or a method for anomaly detection in the drive train of mobile robotic platforms (AGV and AMR) via, directly and/or recorded as well as real and/or simulated and/or synthetic sensor data, as well as real and/or simulated and/or synthetic sensor data, by means of sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axis angle values, temperatures, deformation ations/deformations, air humidity, aerosol composition, LiDAR, radar etc. va) Kl algorithms, ie neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM ...), for the (early) detection and reporting of defects, errors and malfunctions, for the detection and reporting of wear, for the detection and reporting of increased power loss, ie for the reduction and optimization of energy consumption and/or for the detection and reporting of incorrect parameterization of the drive train and its individual components.

[0139] It is conceivable that it can be designed as a separate information processing device or as an integrable module or similar in existing products and/or controls (drive controllers, automation systems, ...) and/or data centers (internal or external to the company).

[0140] The use of AI-based analysis has the advantage in this context that much more complex (more parameters) and therefore more realistic models can be created. The inference can be carried out on embedded hardware and thus very locally at the location where the data is created. In addition, no transport of the data beyond the pure function is necessary. The data remains where it is needed. A higher level of data integrity can be guaranteed. At the same time, reliability can be increased because there is no communication with a central evaluation unit. Processing directly on the device, i.e. on the periphery, can also increase energy efficiency and speed up response.

[0141] In a further application example, the disclosed technology can be used to classify driving profiles and movement patterns of mobile robot platforms, e.g. by an information processing device and/or a method for classifying driving profiles and movement patterns of mobile robot platforms (AGV and AMR) by means of direct and/or recorded, as well as real and/or simulated and/or synthetic sensor data, by means of sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axis angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar etc.) Kl algorithms, ie neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM ...), for the recognition Detection and reporting of the direction of travel (straight ahead and cornering, forwards and backwards), detection and reporting of orientation/alignment in space, monitoring of predefined driving profiles and movement patterns and/or detection and reporting of ascent and descent in space (drone).

[0142] In addition, the disclosed technology can be used for service life interpolation, e.g. by an information processing device and/or a method for interpolating drive train service lives in mobile robotic platforms (AGV and AMR) via direct and/or recorded, as well as real and/or simulated and/or synthetic sensor data, by means of sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axis angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar, etc.) Kl algorithms, i.e. neural networks - deep learning (e.g. autoencoders) and/or machine learning (e.g. KNN, SVM ...), to calculate and report the remaining drive train service life up to a certain degree of wear.

[0143] An application example is the use of the disclosed technology in moving, single or linked axes.

[0144] An example of this is an information processing device and/or a method for detecting anomalies in the drive train with moving axles by means of direct and/or recorded, as well as real and/or simulated and/or synthetic sensor data, by means of sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axle angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar etc. h. Neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM ...), for the (early) detection and reporting of defects, errors and undesirable behavior, for the detection and reporting of wear, for the detection and reporting of increased power loss, i.e. for the reduction and optimization of energy consumption, and/or for the detection and reporting of incorrect parameterization of the drive train and its individual components. [0145] Another example is an information processing device and/or a method for classifying driving profiles and movement patterns of moving axes via direct and/or recorded, as well as real and/or simulated and/or synthetic sensor data using sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axle angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar etc.) algorithms, i.e. neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM ...), for detecting and reporting the direction of rotation (straight ahead and cornering, forwards and backwards), for detecting and reporting the orientation/alignment in space, for monitoring predetermined driving profiles and movement patterns and/or for detecting and reporting ascending and descending in space (drone).

[0146] Another example is an information processing device and/or a method for interpolating drive train service lives for moving axles via direct and/or recorded, as well as real and/or simulated and/or synthetic sensor data using sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axle angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar, etc.) CL algorithms, i.e. neural networks - deep learning (e.g. autoencoders) and/or machine learning (e.g. KNN, SVM ...), for calculating and reporting the remaining drive train service life up to a certain degree of wear.

[0147] Another application example is the use of the disclosed technology in presses.

[0148] For example, the anomaly detection on presses can be carried out by an information processing device and/or a method for anomaly detection on presses using direct and/or recorded as well as real and/or simulated and/or synthetic sensor data, using sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axis angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar etc.) algorithms, i.e. neural networks - deep learning (e.g. autoencoders) and/or machine learning (e.g. KNN, SVM ...), for the (early) detection and reporting of defects, errors and undesirable behavior, for Detection and reporting of wear, detection and reporting of tool breakage, detection and reporting of increased power consumption, i.e. to reduce and optimize energy consumption, detection and reporting of incorrect parameterization of the drive train and its individual components, detection and reporting of the pressing force and/or detection and reporting of good or bad parts of the pressing process.

[0149] Another example is an information processing device and/or a method for classifying pressing processes by means of direct and/or recorded, as well as real and/or simulated and/or synthetic sensor data, by means of sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axis angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar, etc.) Kl algorithms, i.e. neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM ...), for detecting and reporting different pressing tools, for detecting and reporting different pressed parts, for detecting and reporting different materials of the blank and/or for detecting and reporting different material thicknesses of the blank.

[0150] Another example is an information processing device and/or a method for interpolating tool life in presses using direct and/or recorded, as well as real and/or simulated and/or synthetic sensor data, using sensor data-based (power, motor currents, electrical voltage, vibrations, motor speeds, axis angle values, temperatures, strains/deformations, air humidity, aerosol composition, LiDAR, radar, etc.) Kl algorithms, i.e. neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM ...), for calculating and reporting the remaining tool life up to certain threshold values of the pressed part quality.

[0151] Another application example is the use of the disclosed technology in drones, in access control or in workspace monitoring. [0152] In the case of drones, in particular unmanned aerial vehicles (UAVs), the aspects of anomaly detection, classification of movement profiles and interpolation of downtimes described in connection with the above application examples can be applied in the same way.

[0153] An example of the application of the described technology in the field of access control is an information processing device or method for classifying people for access authorization to systems using sensor data-based (aerosol sensor or similar) algorithms, i.e. neural networks - deep learning (e.g. autoencoder) and/or machine learning (e.g. KNN, SVM ...) for evaluating the individual aerosol print analogous to access control using a password, RFID or fingerprint. Such a method is superior to existing methods because a key, i.e. the aerosol print, cannot be passed on and the input is still contactless.

[0154] In the field of workspace monitoring, the disclosed technology can be used, for example, as a device with an information processing system and an interface to one or more acoustic sensors and/or an interface to an external data source in order to detect (emergency) signals from people and process them in a computing module using an algorithm. When a specific (emergency) signal is detected, the computing module can control an output module that generates a shutdown signal that can be read out via one and/or more interfaces. Such a device could also be integrated into existing control systems.

[0155] With the aid of the device described, it would be possible to switch off a system or to bring it into a defined state by issuing an (emergency) signal in the form of a call for help, a verbal return command or a verbal position indication even if the worker is trapped in the setup mode or cannot operate an emergency shutdown device for another reason. [0156] In workspace monitoring, the disclosed technology can also be used to detect or classify living beings.

[0157] The scope of the present invention is determined by the following claims and is not limited by the features explained in the description or shown in the figures.

[0158] List of reference symbols

10, 1 , 2, 3 systems

12 technical system

14 Control

16 Processing unit

18 Input size

20 measured values (first measured values)

22 measured values (second measured values)

24 Sensor

26 virtual input variable

28 Facility

30 Knowledge

32 robots

34 Base

36 Manipulator

38 Tools

40 Engine

41 Axis

42 Power supply

44 Power supply

46 shooters

48 lines

50 Input module

52 Input signal

54 encoders

56 Additional module , 60, 62, 64, 65 sensor (second sensor) input unit evaluation unit output unit input unit Kl system output unit , 80 electronic signal input output , 90 processing channel (first processing channel) , 92 processing channel (second processing channel) 0 mechatronic system 0 electric drive (with or without mechanically coupled motion axes)1 measured value(s) of a sensor 2 system parameters of the control (optional) 0 information processing device 0 interface (wired/wireless; analog/digital) 0a evaluation unit (physically integrated) 1 one or more evaluation algorithms 2 program memory (e.g. for storing feature extractions) 0 control unit 0 control signals (e.g. binary) 1 control value (prediction of an anomaly) 2 control value (wear prediction) 0 external device 0b evaluation unit (virtual on an external device)

Claims

Patent claims

1. System (10; 1; 2; 3) for safety-related monitoring, in particular shutdown, of a technical system (12), comprising: a controller (14) which is set up to trigger a safety-related reaction related to the safety-relevant input variable on the basis of a safety-relevant input variable (18) which can be derived from at least one continuously detectable first measured value (20) in the event of a deviation from an expected input variable, which causes the technical system (12) to transition to a safe state, and a processing unit (16) which is set up to determine a virtual input variable (26) based on continuously detectable second measured values (22) which do not completely correspond to the at least one first measured value (20) and to provide it to the controller (14), wherein the controller (14) is set up to trigger the safety-related reaction related to the safety-relevant input variable (18) if the virtual input variable (26) deviates from the expected input variable.

2. System according to claim 1, further comprising: a sensor (24) which is configured to record the at least one first measured value (20), to generate the safety-relevant input variable (18) from the at least one first measured value (20), and to provide the safety-relevant input variable (18) to the controller (14).

3. System according to claim 1 or 2, wherein the processing unit (16) comprises an inference engine which is configured to generate the virtual input variable (26) from the second measured values (22) and a stored knowledge base (30).

4. System according to claim 3, wherein the processing unit (16) is configured to generate the knowledge base (30) by machine learning, in particular by supervised learning and/or reinforcement learning

5. System according to claim 2 and 4, wherein the processing unit (16) is configured to perform the machine learning based on the second measured values (22) and an output of the sensor (24).

6. System according to claim 4 or 5, wherein the processing unit (16) is arranged to carry out the machine learning before use of the virtual input variable (26) by the controller (14).

7. System according to claim 6, wherein the processing unit (16) is configured to perform machine learning on generically generated data by simulations prior to commissioning.

8. System according to one of claims 3 to 7, wherein the processing unit (16) has an interface to continuously expand the knowledge base (30).

9. System according to one of claims 1 to 8, wherein the safety-relevant input variable (18) is the result of direct signal processing of the one first measured value (20) or the result of direct signal processing of the plurality of first measured values (20), and wherein the virtual input variable (26) is at least partially the result of indirect signal processing of the second measured values (22), in particular an approximation and/or estimation.

10. System according to one of claims 2 to 9, wherein the controller (14) is configured to compare the safety-relevant input variable (18) of the sensor (24) with the virtual input variable (26) and to trigger the safety-related reaction if the safety-relevant input variable (18) and the virtual input variable (26) differ from one another by a predefined amount.

11. System according to one of claims 1 to 10, wherein the safety-relevant input variable (18) and the virtual input variable (26) are identical in their way of representing a representative system variable of the technical installation (12), in particular in their unit and/or in their scope.

12. System according to one of claims 1 to 11, wherein the second measured values (22) are at least partially measured values inherent to the control (14) of the technical system (12).

13. System according to one of claims 1 to 12, wherein the controller (14) is a safety controller with a multi-channel, redundant structure.

14. Processing unit (16) for use in a system for safety-related monitoring, in particular shutdown, of a technical installation (12), in which a controller (14) uses a safety-relevant input variable (18) that can be derived from at least one continuously detectable first measured value (20) to trigger a safety-related reaction related to the safety-relevant input variable (18) in the event of a deviation from an expected input variable, which causes the technical installation (12) to transition to a safe state, wherein the processing unit (16) is set up to determine a virtual input variable (26) based on continuously detectable second measured values (22) that do not fully correspond to the at least one first measured value (20) and to provide it to the controller (14), which triggers the safety-related reaction related to the safety-relevant input variable (18) when the virtual input variable (26) deviates from the expected input variable.

15. Method for safety-related monitoring, in particular shutdown, of a technical system (12), comprising:

- triggering, by a controller (14), a safety-related reaction related to a safety-relevant input variable (18), which causes a transition of the technical system (12) to a safe state in the event of a deviation from an expected input variable, wherein the safety-relevant input variable (18) can be derived from at least one continuously detectable first measured value (20);

- determining, by a processing unit (16), a virtual input variable (26) based on continuously detectable second measured values (22) which do not completely correspond to the at least one first measured value (20); and

- Providing the virtual input variable (26) for the controller (14), wherein the controller (14) triggers the safety-related reaction related to the safety-relevant input variable if the virtual input variable (26) deviates from the expected input variable.