Method for an ML- and / or ai-assisted self-monitoring of functions of a sensor, and sensor equipped with such a monitoring function

Abstract

The invention relates to a computer-implemented method for the automatable monitoring of at least one function of a sensor, wherein a monitoring function is provided which is based on machine learning (ML) or artificial intelligence (AI) and is set up in the sensor. The method has the following steps in particular: providing at least one technical behavior model trained by means of machine learning on an existing sensor of a specified sensor type, said behavior model comprising internal technical properties and possible external operating conditions of the sensor and being adapted to a specified intended technical use or use scenario of the sensor; setting up the monitoring function in the sensor on the basis of the technical behavior model so as to detect a possible deviation of at least one of the internal technical properties of the sensor during the operation of the sensor and so as to detect an anomaly of the at least one technical property of the sensor using the detected deviation of the at least one technical property of the sensor, the monitoring function detecting an anomaly on the basis of the at least one technical behavior model adapted to the respective intended use or use scenario; and outputting, by means of the sensor, an indication or warning signal with respect to a sensor malfunction relating to a detected anomaly.

Description

[0001] METHOD FOR ML- OR AI-BASED SELF-MOUNTING OF FUNCTIONS OF A SENSOR AND A SENSOR EQUIPPED WITH SUCH A MONITORING FUNCTION

[0002] The invention relates to the automated or automatable technical monitoring or remote monitoring of the functionality of a sensor or at least one sensor function of a sensor.

[0003] State of the art

[0004] Sensors or sensor arrangements for monitoring the condition of technical equipment are known in the prior art. For example, the "condition monitoring sensors" developed by the present applicant detect various physical quantities, such as vibration, temperature, humidity, and / or ambient pressure of a technical device. This allows the technical condition of, for example, a machine, an industrial plant, and its associated components to be reliably monitored. The condition monitoring sensors thus enable the efficient and trouble-free operation of such a machine or plant and thereby significantly increase its effectiveness.

[0005] The physical quantities recorded by the condition monitoring sensors are pre-processed and interpreted, for example, by integrated evaluation electronics, and thus deliver corresponding results via, for example, an IO-Link connection to a higher-level system of the respective machine or plant, to a higher-level evaluation system, or to a programmable logic controller (PLC). For example, German patent DE 10 2021 133 338 A1 describes a method for monitoring an IO-Link device using machine learning, in which the current, voltage, or electrical power at a port of an IO-Link master is recorded. Using a model for the current, voltage, or electrical power previously trained through machine learning, any malfunctions are monitored.This also allows for the detection of anomalies, errors, deviations or maintenance indicators, or a prediction of a maintenance requirement, a fault or a failure of the Io-Link device.

[0006] Disclosure of the invention

[0007] The invention is based on the objective of enabling automated technical monitoring of a sensor or sensor arrangement comprising such a sensor as reliably as possible and with the least possible technical effort.

[0008] The invention, designed to solve this problem, is based on the idea of ​​providing or implementing AI-based detection of at least one sensor function or malfunctions directly at the sensor. This is based on the specific insight that a sensor transmits faulty data when malfunctioning, for example, if the internal electronics are defective, but the data receiver cannot detect an actual malfunction without a thorough plausibility check.

[0009] Therefore, it is specifically proposed that the sensor be able to independently detect when it is no longer functioning as intended or correctly. The proposed detection is thus based on the sensor's own self-awareness and therefore functions independently of environmental influences.

[0010] It should be noted that today's sensors or sensor arrangements (hereinafter referred to simply as "sensor(s)") are mostly designed as so-called "smart" sensors or "intelligent" sensors, since in addition to the actual measurement of quantities they also perform signal processing or...

[0011] These sensors enable onboard signal processing. They typically feature a microprocessor or microcontroller, which may also include DSP functionality. Furthermore, these sensors have standardized interfaces for communication with higher-level systems, such as an IO-Link system or another fieldbus system.

[0012] The invention proposes a computer-implemented method and a correspondingly configured sensor for simplified or, where possible, automated self-monitoring of the functional state of such an intelligent sensor, in particular based on artificial intelligence (Kl), by means of a suitable monitoring function or monitoring device.

[0013] By appropriately configuring the monitoring function or...

[0014] Furthermore, the monitoring device of such a sensor for a given technical application can ensure that the functionality of the sensor, especially with regard to the respective technical application or the respective technical use of the sensor, can be reliably checked "onboard".

[0015] In the computer-implemented method according to the invention for the automatable self-monitoring of at least one sensor function of a sensor, a monitoring function based on machine learning (ML) or artificial intelligence (AI) and present or set up in the sensor is provided in particular, wherein the method comprises the following steps:

[0016] Providing at least one technical behavior model trained by machine learning on an existing sensor of a given sensor type, which includes internal technical properties and possible external

[0017] includes the operating conditions of the sensor and which is adapted to a given technical purpose or a corresponding application scenario of the sensor;

[0018] Setting up the monitoring function in the sensor based on the technical behavior model, to detect a possible deviation of at least one of the internal technical properties of the sensor during the ongoing operation of the sensor, wherein the technical behavior model is based on external operating conditions of the sensor that depend on the respective technical purpose or application scenario, and to detect an anomaly of the at least one technical property of the sensor by means of the detected deviation of the at least one technical property of the sensor, wherein the monitoring function detects an anomaly based on the at least one technical behavior model adapted to the respective purpose or application scenario;

[0019] Issuance of a notification or warning signal by the sensor regarding a possible malfunction of the sensor concerning a detected anomaly.

[0020] The self-monitoring of at least one sensor function takes place within the sensor itself, specifically through self-perception. This monitoring approach is based on the premise that the necessary technical insights into the sensor's functional state, which depend in a relatively complex manner on the specific application or scenario and the corresponding potential external operating conditions, can only be obtained within the sensor itself. Using a suitably trained behavioral model, these complex relationships can be automatically taken into account during monitoring based on machine learning or computer science.

[0021] It should be noted that the user's specified technical purpose or application scenario for the sensor corresponds to a concrete, technical or physical application scenario to which the aforementioned behavioral model is adapted. Such an application scenario could be, for example, sensor-based pressure and temperature monitoring in a physical, chemical, or biochemical reactor or a corresponding production plant, or a so-called "condition monitoring" sensor for acquiring status data from an industrial manufacturing plant or an industrial production or processing process.

[0022] For a given sensor of a specific type, various possible application scenarios and the operating behaviors dependent on these scenarios and expected by the sensor can be taught or trained in a single behavioral model. This allows the behavioral model to cover different application scenarios or purposes of the sensor during monitoring. Based on such a trained behavioral model, the machine learning or computer intelligence (ML) system can then, through interpolation, also cover application scenarios for which the behavioral model was not originally (specifically) trained.

[0023] The internal technical properties of the sensor monitored during machine learning (ML) or artificial intelligence (AI)-based self-monitoring include, for example, current and power values, internal humidity levels, and / or internal temperatures. Furthermore, these sensor-internal properties also encompass the ratio or spatial distribution of corresponding measured values ​​when these physical quantities are measured at different locations within the sensor. This allows, in particular, the reliable detection of potential failures or overloads of individual technical components or assemblies within the sensor.

[0024] In contrast to the disclosure in DE10 2021 133 338 A1, the use of machine learning (ML) or cognitive technology (Kl) via a specified behavioral or monitoring model within the sensor itself enables automated monitoring, not solely based on measured current or power consumption values. This is because the reliability of the ML- or Kl-based self-monitoring proposed herein relies on considering external operating conditions determined by the respective application or scenario, as well as the relationship or spatial distribution of the aforementioned physical measured quantities within the sensor.

[0025] As described in more detail below, a "condition monitoring sensor" can take external operating parameters such as temperature, humidity, and ambient pressure into account when monitoring components or assemblies located within the sensor. The internal temperature and humidity at or around such a component have a significant impact on the overall operating condition of the sensor. For example, excessively high temperatures within the component / sensor are critical for the corresponding electronic components and have a major impact on their operating condition and lifespan.

[0026] According to a further aspect of the monitoring method according to the invention, it can be provided that the at least one trained technical behavior model is applied to a plurality or multiplicity of applications for the specified purpose or purpose.

[0027] The application scenario involves the application of a specified sensor type, whereby the monitoring function detects an anomaly based on at least one technical behavior model applied to the respective sensor and adapted to the respective application or application scenario.

[0028] In the computer-implemented method according to the invention for configuring a monitoring function of a sensor, it is particularly provided that at least one trained technical behavior model corresponding to a predetermined purpose or application scenario of a plurality or multitude of sensors of a predetermined sensor type is provided, wherein the at least one technical

[0029] The behavioral model relates to internal technical properties of the respective sensors and / or external operating conditions, depending on the specified purpose or application scenario of the sensor, and the configuration of the monitoring function is carried out using the technical behavioral model.

[0030] According to a further aspect of the configuration method according to the invention, it can be provided that the trained or learned technical behavior model is applied to the majority or multitude of sensors of the specified sensor type before their delivery.

[0031] According to a further aspect of the configuration method according to the invention, it can be provided that the internal technical properties of the respective sensor include the electrical current consumption and / or electrical power consumption and / or the temperature behavior during operation of the sensor, which depend on the specified purpose or application scenario.

[0032] According to a further aspect of the configuration method according to the invention, it can be provided that the external operating conditions, which depend on the specified purpose or application scenario, include humidity and / or ambient pressure.

[0033] According to a further aspect of a sensor designed or configured according to the invention, a monitoring function can be provided, for example, based on a factory-pre-trained behavioral model that maps internal operating characteristics or operational behavioral properties of the sensor in a digital twin of the sensor. Such operating properties can be provided, in particular, depending on one or more applications or scenarios of a given sensor or sensor of a corresponding sensor type. Such operating properties can, for example, be those required for a given application or scenario.The typical application scenario would involve the electrical current consumption and / or electrical power consumption and / or the electrical voltage behavior and / or the temperature behavior, and / or the current humidity and / or the ambient pressure of the sensor.

[0034] Such operating parameters can be continuously recorded during sensor operation using measuring elements integrated into the sensor, or at suitable or empirically determined time intervals. Additionally, the sensor data supplied by the sensor, in particular the data quality and / or the data rate, can also be considered in a corresponding behavioral model.

[0035] Furthermore, the behavioral model can consider not only monitoring-relevant relationships between the aforementioned physical quantities (e.g., current, voltage, and / or electrical power), but also malfunctions of at least one sensor function that are expected with respect to an anomaly affecting these quantities. For this purpose, the behavioral model underlying the machine learning (ML) or cognitive intelligence (KL) can also be trained with corresponding technical data or contain appropriately trained data that relates to relationships or correlations between possible malfunctions of a sensor and the underlying anomalies of the aforementioned physical quantities.

[0036] Regarding the consideration of data quality and / or data rate in the behavioral model, it should be noted that internally, an industrial sensor operates with measuring instruments or corresponding microchips, which can detect changes in at least one physical quantity relevant for monitoring. This change information is processed within the sensor and transmitted via an industrial network protocol, such as an IO-Link-based network. With an IO-Link network, a constant data rate is always available for transmission, which can be configured by an IO-Link master. However, internal sensor malfunctions can lead to changes or fluctuations in the data rate, which may result in the measured values ​​at the sensor output being outdated or erroneous.Using the behavioral model proposed herein, changes in the sensor's internal data rate can be automatically detected and, for example, communicated to a user of the sensor or the user can be warned accordingly.

[0037] During the measurement processes of the aforementioned physical quantities, the data quality or signal quality of the sensor signal can also be recorded or determined. According to the behavioral model disclosed herein, the signal quality thus recorded or determined can be taken into account when monitoring the sensor function.

[0038] In the computer-implemented method according to the invention for the automated creation of a monitoring function of a sensor concerned here for a given purpose or application scenario, it can also be provided that the method is carried out by means of a machine learning system, wherein the method comprises the following steps: communication between a user and the machine learning system via a computer interface, in order to acquire data relating to the given purpose or application scenario; training a behavioral model with respect to the aforementioned characteristic operating parameters on the basis of the acquired data or optimizing an existing behavioral model on the basis of the acquired data; releasing the trained behavioral model for the automated creation of the monitoring function.

[0039] The invention also relates to a computer-implemented method for using a behavioral model trained as described above to configure a monitoring function of a sensor concerned here for a given purpose or application scenario, wherein this method comprises the following steps: providing a user with access to a trained behavioral model by means of a computer processor; receiving data entered via a computer interface for the given purpose or application scenario; causing the trained behavioral model to analyze the entered data and provide instructions for generating a new or modified monitoring function based on the trained behavioral model.

[0040] According to a further aspect of the method according to the invention, it can be provided that the trained behavioral model is prompted to configure a monitoring function for a given purpose or application scenario based on machine-readable instructions and data entered by the user via the computer interface.

[0041] According to a further aspect of the inventive method and a correspondingly designed or configured sensor, a monitoring function configured for a specific application or scenario and integrated into the sensor can detect changes in at least one of the characteristic operating parameters during the sensor's operation. Based on such a detected change, an anomaly can be inferred, and a corresponding message, e.g., visual (by flashing) and / or audible, is then issued to the user. If the sensor is connected to an IO-Link system, a corresponding error signal can be transmitted, e.g., to the IO-Link master, a higher-level evaluation system, or a programmable logic controller (PLC).

[0042] The approach described herein for the automated configuration of a monitoring function of a sensor (or a corresponding sensor arrangement) has the advantage that it can reliably distinguish whether there is a technical failure of a sensor or whether external circumstances, e.g., errors in the application to be monitored, environmental conditions, or functional conditions or behavioral disturbances related to the respective purpose or application scenario, are the cause of a detected malfunction of the sensor.

[0043] This approach thus enables significantly greater reliability in detecting a sensor malfunction, i.e., a more precise prediction of malfunctions, resulting in shorter downtimes and reliably preventing incorrect measurements caused by the malfunction.

[0044] According to a further aspect of the inventive method and a correspondingly designed or configured sensor, the behavioral model can be taught normal states corresponding to the intended use or application scenario, and deviations from one or more of these normal states can be detected by means of a monitoring function provided in the sensor. Several such states can also be taught and classified accordingly. This allows functional states that were previously unknown at the time of teaching or training the behavioral model to be detected as anomalies during subsequent operation of the sensor.

[0045] According to a further aspect of the inventive method and a correspondingly designed or configured sensor, anomalies can be detected based on the deviation from empirically predetermined threshold values ​​for the aforementioned characteristic operating parameters and / or corresponding operating states.

[0046] According to a further aspect of the method according to the invention, and with a correspondingly designed or configured sensor, the relative distribution of the aforementioned operating parameters to one another, a specific temporal sequence, or the periodicity of these operating parameters can be used as a basis for functional or condition monitoring. It should be noted that such a determination would not be possible using only one of the aforementioned threshold monitoring methods.

[0047] According to a further aspect of the inventive method and a correspondingly designed or configured sensor, several behavioral models can be used for functional monitoring or classification, preferably operating in parallel. These behavioral models can be generated using machine learning methods such as deep neural networks (i.e., tightly meshed networks, convolutional neural networks, feedback neural networks), or transformer networks. These behavioral models can be trained using decision tree-based methods, margin-based methods, clustering methods, ensemble methods, nearest-neighbor methods, and / or linear and / or nonlinear regression methods. For example, a behavioral model can be learned using these methods that understands the normal behavior of the parameters to be measured during normal operation.However, behavioral models of typical error cases can also be trained in the same way.

[0048] According to a further aspect of the inventive method and a correspondingly designed or configured sensor, a further aspect of the inventive self-monitoring or self-diagnosis of the sensor in question can be the prediction of a possible or imminent maintenance requirement, a fault, and / or even a complete failure of the sensor or a corresponding sensor arrangement. Such a prediction is possible with high reliability using machine learning and a suitably appropriate model.

[0049] It should be noted that, in principle, all current, voltage, or electrical power data of the sensor are suitable for monitoring based on a behavioral model using machine learning (ML) or computer intelligence (KL), where there is a relationship between the behavior of the current, voltage, or electrical power and the states, anomalies, faults, and / or deviations.

[0050] Furthermore, it should be noted that, particularly in sensor arrays with multiple sensors, the aforementioned relationships may not be obvious or intuitively recognizable. However, machine learning can also detect and identify complex relationships between such multiple data sources and the states, anomalies, errors, and / or deviations to be detected, which would not be directly apparent to a user.

[0051] The result is the monitoring of sensor functions that previously could only be implemented with additional hardware. Furthermore, compared to conventional setpoint monitoring, the ML- or KL-based self-monitoring according to the invention enables the monitoring, detection, and / or prediction of certain states, anomalies, errors, and / or deviations of the sensor function in question. It should also be noted that the aforementioned behavioral model can be pre-trained in an external computer system, e.g., in a decentralized "cloud"-based system. Such behavioral models can be pre-trained at the factory during the manufacturing of the sensors (or corresponding sensor arrangements) in question, and a corresponding monitoring function can be configured on them upon delivery or transferred to them at a later time via a software / data update.

[0052] According to a further aspect of the inventive method and a correspondingly designed or configured sensor, preprocessing can take place during the learning or training of a behavioral model. This preprocessing involves extracting features from the underlying data of specified operating parameters or transforming, filtering, aggregating, and / or otherwise preprocessing such data. Such preprocessing can be linked to a respective behavioral model.

[0053] Various methods are available for the automated extraction of such features from the underlying data relating to the corresponding sensor measurement signals. For example, the following mathematical approaches are possible:

[0054] Averaging;

[0055] Calculation of standard deviations, variances, median values, and minimum / maximum values ​​within specific empirically defined measurement ranges; determination of rates of change of measured values; calculation of the trend behavior of measured values; determination of autocorrelations, correlations, fluctuations, and frequency values ​​from a frequency analysis (e.g., using Fourier transform); determination of the spectral power density of corresponding sensor signals; determination of principal frequencies, bandwidths, and energy values ​​within empirically defined frequency bands; and the relationship of such technical characteristics to each other and the combination of several measured quantities.

[0056] According to a further aspect of the inventive method and a correspondingly designed or configured sensor, a behavioral model of the sensor in question can be updated during operation using measured values ​​of the operating parameters, which thus serve as training data. This allows the behavioral model to be continuously improved. Furthermore, the behavioral model can be adapted to changes in the application or intended use of the sensor and therefore remains up-to-date. When identifying potential malfunctions, statistical parameters of the operating parameters, such as the mean, standard deviation, variance, kurtosis, skewness, or median, can be determined.

[0057] According to a further aspect of the method according to the invention, and a correspondingly designed or configured sensor, various classes of algorithms, which are known per se, can be used for machine learning. Preferred classes are artificial neural networks, decision tree-based methods, margin-based methods, cluster methods, ensemble methods, nearest-neighbor methods; and / or linear and / or nonlinear regression methods.

[0058] According to a further aspect of the method according to the invention, and a correspondingly designed or configured sensor, pattern recognition can be provided using the measured values ​​for the aforementioned operating parameters, whereby patterns in these parameters and / or in parameters derived from them are recognized, and these patterns are linked to the states, anomalies, faults, and / or deviations based on the model. For example, patterns can be recognized in the distribution of the operating parameters, in the sequence of the parameters, in the periodicity of the parameters, etc.

[0059] Referring to the application scenarios described above, patterns required for pattern recognition can be determined, by means of which a malfunction of the sensor can be detected using the appropriately trained behavioral model via pattern recognition.

[0060] As a first embodiment, such patterns can be based on the technical finding that electrical currents on an electronic circuit board flow in a specific ratio to each other and with a certain time interval between them. If the sensor malfunctions, however, this ratio and / or the temporal pattern of the electrical currents is no longer correct.

[0061] As a second example in the field of sensor-based temperature monitoring or temperature changes, such patterns can also be based on the technical finding that, in a given sensor, the external temperature and the sensor's internal temperature tend to behave similarly or are correlated accordingly. Therefore, in the event of a sensor malfunction, the sensor's internal temperature will deviate from the external temperature. Corresponding behavioral patterns, which result from the external environmental conditions dependent on the application or scenario, thus appear different when the sensor is faulty. In other words, the patterns in the distribution of operating parameters can take into account possible correlations between the sensor's internal technical characteristics and potential external operating conditions.

[0062] According to a further aspect of the method according to the invention, as well as a correspondingly designed or configured sensor, a statistical test procedure can also be carried out using the measured values ​​for the operating parameters, in particular of different sensor types.

[0063] It should be noted that DE102021133338, in contrast to the present invention, involves the monitoring of an IO-Link system and / or at least one IO-Link device within the IO-Link system. In this process, the electrical current, voltage, and / or power are measured at at least one port of an IO-Link master within the IO-Link system. The electrical parameters are measured at a port of the IO-Link master, and thus directly within the IO-Link master itself. Furthermore, machine learning-based artificial intelligence (AI) is applied within the IO-Link master to execute one or more monitoring functions. Therefore, machine learning-based monitoring is performed directly by the IO-Link master. During the machine learning process, a previously trained behavioral model for the current, voltage, and / or power is applied.The behavioral model characterizes the relationship between current, voltage and / or electrical power and the states, anomalies, faults, deviations and / or maintenance indicators of the IO-Link system and / or the at least one IO-Link device.

[0064] Brief description of the drawings

[0065] Exemplary embodiments of the invention are shown in the drawings and are explained in more detail in the following description.

[0066] Fig. 1 shows a sensor arrangement of individual sensors for condition monitoring of a plant component according to the prior art; Fig. 2 shows an exemplary diagram of a current measured in a sensor of a sensor arrangement shown in Fig. 1 as a characteristic operating parameter of the sensor;

[0067] Figs. 3A - 3D show flowcharts or block diagrams of the inventive method for artificial intelligence (Kl)-based self-monitoring of the functional state of a sensor of a sensor arrangement shown in Fig. 1, using different examples of a correspondingly trained technical behavior model of a sensor concerned here;

[0068] Fig. 4 shows an embodiment of the teaching or training of a technical behavior model using an artificial intelligence-based machine learning system, based on a combined flow / block diagram;

[0069] Figures 5A - 5C illustrate the structure of an encoder / decoder architecture relevant here for generating configuration data of a technical monitoring function relevant here, based on a technical behavioral model.

[0070] Examples of the invention

[0071] Fig. 1 shows a sensor arrangement according to the prior art, as an example of an arrangement of two vibration sensors 100, 105 and an inductive velocity sensor 110, as well as a device known per se for its configuration or parameterization. The sensor arrangement 100-110 shown here serves for vibration monitoring of a system or machine (not shown here) and is intended to trigger an alarm when a predetermined vibration level is exceeded.

[0072] In the illustrated embodiment, the three sensors 100, 105, 110 transmit high-frequency analog signals or values ​​to an electronic diagnostic unit 130, e.g., an industrial PC, via respective analog communication links 115, 120, 125. The diagnostic unit 130 further processes the raw data 115, 120, 125 supplied by the sensors 100, 105, 110 to determine the possible presence of a specified alarm situation. The diagnostic unit 130 then transmits the aggregated data via a corresponding Ethernet or Internet Protocol (IP)-based communication link 135, 140 to higher-level systems, in this embodiment to the IT network of a cloud computing platform 145 and to a programmable logic controller (PLC) 150.

[0073] In the sensors affected here, the electrical current or power consumption of the sensor can change due to heating or aging of the electronic components, e.g., through the resulting drying out of electrolytic capacitors. Even this typically very slow change can be detected using the monitoring function disclosed herein. Furthermore, a prediction of any necessary maintenance can be made, thus enabling predictive maintenance planning.

[0074] Furthermore, in sensors for displacement and distance measurement, which operate inductively, magnetostrictively, or optoelectronically, the current consumption is often non-linearly related to the displacement or distance when the current for excitation of the sensor element is automatically regulated, for example, to compensate for the damping of the measuring path. This non-linear dependency can be incorporated into a behavioral model disclosed herein.

[0075] Figure 2 shows a graph of the current measured over time t in an exemplary sensor of a sensor arrangement shown in Figure 1 for a periodic process in a plant. Anomalies 20 to 23 relating to the data quality of the sensor data supplied by the sensor are marked. While anomaly 20 manifests as a sharp peak ("outlier"), anomalies 21, 22, and 23 represent changes in the curve. Conventionally, threshold monitoring is performed, in which the measured current I is compared with a threshold value, shown here as SW for comparison. It is evident that in this example, only anomaly 20, the "outlier," can be detected using threshold monitoring.

[0076] Figures 3A and 3B show two flowcharts of the inventive method for intelligent self-monitoring of the functional state of a sensor which in the present embodiment is connected to an IO-Link system via communication technology.

[0077] Fig. 3A shows in particular the learning and training phase for a technical, in this case electrical, behavioral model M. In this embodiment, the current I, the voltage U, and / or the electrical power P are measured and recorded to serve as training data for the behavioral model. The corresponding operating state Z, i.e., the corresponding characteristic operating parameters of the sensor during the measurement, is also recorded. The training data I, U, and P are then fed to a preprocessing unit 40. During preprocessing 40, features can be extracted from the data, or the data can be transformed, filtered, aggregated, and / or otherwise preprocessed.

[0078] The applied preprocessing is linked to the respective behavioral model M and implemented together with the behavioral model M in sensor 1. Subsequently, the behavioral model M is trained (or taught) 41, in which the relationship between the training data, i.e., the current I, the voltage U, the electrical power P and the corresponding operating state Z, is learned through machine learning.

[0079] Machine learning is based, for example, on pattern recognition, in which patterns are recognized in the behavior of the electrical quantities I, U, and P and linked to the operating states Z. In this embodiment, this linkage is achieved through machine learning using an artificial neural network.

[0080] Figure 3B shows an evaluation during the operation of the sensor. A measurement 42 of the electrical quantities, i.e., the current I, the voltage U and / or the electrical power P, is performed at corresponding ports 11 of the sensor 1. The current I is determined in the sensor 1 by the voltage drop across a small sensing resistor and subsequent A / D conversion of the voltage drop into a digital value.

[0081] Alternatively, electronic current measuring components can be used to determine the current I from a magnetic field measurement. The electrical power P is calculated from the product of the current I and the voltage U. The electrical quantities I, U, and P measured during operation are fed to a preprocessing unit 43. The preprocessing unit 43 is carried out in the same way as the preprocessing unit 40.

[0082] Subsequently, the current operating state Z is classified by applying the behavioral model M and the electrical quantities measured during operation—namely, the current I, the voltage U, and the electrical power P—using machine learning. Here, too, the machine learning is based, for example, on pattern recognition, in which patterns are recognized in the course of the measured electrical quantities I, U, and P, and the operating state Z is classified based on the behavioral model M. In the present embodiment, this is achieved through machine learning using an artificial neural network. Alternatively, a statistical test procedure can be performed. The behavioral model M can be updated during its application by the electrical quantities I, U, and P measured during the sensor's operation.

[0083] As an ML or KL approach, for example, an ML or KL system described below with reference to Fig. 4 can be provided using a neural network for the automated ML or KL-supported learning or training of a technical behavior model of a sensor concerned here.

[0084] Figure 3C illustrates such a ML or KL system 200 with a behavioral model 205 disclosed herein during the training phase.

[0085] The training is performed for a specific sensor 210 of a predefined sensor type using an operating / control module 215, which operates or controls the ML / AI system 200 based on the behavioral model 205 and also performs the training of the ML / AI system 200, for example, on the following input variables or input data: a) Sensor-internal information 220:

[0086] - Sensor type;

[0087] - Spatial arrangement or distribution of measuring ranges for e.g. electrical current / power consumption, temperature within the sensor;

[0088] - Standard values ​​for the data rate or data quality in the sensor's standard operation. b) External information 225:

[0089] - Intended use (a multitude of possible use scenarios);

[0090] - external operating conditions of the sensor, which are possible or to be expected in various application scenarios.

[0091] The training covers, for example, the following technical relationships or...

[0092] Correlations between the different input variables were taken as a basis: 230:

[0093] - Possible or expected external operating conditions for the respective application;

[0094] - possible or expected internal operating conditions of the sensor depending on the external operating conditions;

[0095] - Depending on certain behaviors of sensor-internal operating parameters, possible or expected malfunctions of the sensor. Such behaviors can relate to sensor-internal operating parameters as well as the aforementioned temporal or spatial sequences or periodicities of measured internal operating parameters, internal data rates or data quality supplied by the sensor, and corresponding behavioral patterns.

[0096] It should be noted that such training can be performed not only for a single or specific sensor type, but also for a multitude of sensors of different types. A machine learning / AI system trained in this generic way, with a behavioral model revealed herein, can then be implemented in a large number of such sensors without sensor-specific adjustments.

[0097] Such a generic approach is made possible in particular by the fact that a large number of possible application scenarios as well as the aforementioned technical relationships or correlations between the different input variables can be trained during the training.

[0098] Fig. 3D illustrates the operation of a sensor 210 concerned here, in which a correspondingly trained ML or KL system 200 with a behavior model 205 disclosed herein is implemented or set up.

[0099] In a hypothetical, generically trained machine learning / AI system, the specific sensor type and its intended use are first entered as input variables into the machine learning / AI system 200 or a specified operating / control module 215 250. The other required sensor-internal input variables, such as the spatial arrangement or distribution of the specified measurement areas within the sensor, as well as the specified standard values ​​for the data rate or data quality in the sensor's standard operation, are derived, for example, from the entered sensor type. Alternatively or additionally, these sensor-internal operating variables can also be entered into the machine learning / AI system 200, for example, based on a corresponding technical data sheet 255 of the sensor 210 260.

[0100] After the input of 250, 260 of the aforementioned data or information, the ML / AI system 200 can be put into operation 265 and performs the now automated monitoring 270 of the functional state of the sensor 210 in the present application by continuously recording 275 of the aforementioned physical operating variables or measured variables.

[0101] In the ML / AI system 200, the 275 currently recorded measured values, or corresponding digitized measurement data, are examined using the existing behavioral model 205 to determine whether certain trained behaviors or behavioral patterns can be derived from these measurement data that indicate one or more malfunctions of the sensor 210. Depending on the result of this examination 280, a warning signal can be issued locally or via a communication network (e.g., an IO-Link network) 285. The output 285 of such a warning can be made dependent on one or more empirically defined threshold values ​​for the deviation of a specified behavior of the sensor 210.

[0102] In the following, with reference to the ML or AI technology shown in Figures 3A - 3D, three exemplary applications or application scenarios of a sensor concerned here are described, and how irregularities regarding the functional state of a sensor concerned here can be reliably detected using a behavioral model trained accordingly for machine learning (ML) or artificial intelligence (AI).

[0103] The first application scenario involves using the sensor in high-temperature environments, such as in metal processing, around melting furnaces, or in foundries. High ambient temperatures pose a challenge for the internal electronics and the corresponding circuit boards of such a sensor. The electronic microchips arranged on the circuit boards within the sensor only function correctly within a certain temperature range, and excessively high temperatures can damage them or cause malfunctions. By monitoring the electrical currents, voltages, and / or power values ​​at multiple points on the circuit board, as well as the relationship between these measurements, the ML (machine reader) can...The invention aims to detect irregularities in the behavioral model and reliably warn the user of a malfunction or even an impending sensor failure. Simply considering the limit values ​​of such electrical quantities is insufficient, as the invention is based on the understanding that reliable malfunction detection is only possible by considering the relationship between the measured values ​​and by using a suitably trained behavioral model.

[0104] In the first use case, the reliability of sensor malfunction detection can be further improved by considering, either alternatively or additionally, the relationship between internal measurements and external operating conditions during monitoring. This approach is based on the understanding that external operating conditions have a significant impact on internal measurements. A sign of a malfunction can be a deviation of the measured internal values ​​from expected values ​​due to prevailing external conditions. A suitable or appropriately trained behavioral model based on a machine learning approach can detect such deviations, or even details thereof, and warn the user of a sensor malfunction.

[0105] In the first use case, such a deviation could be, for example, a measured internal temperature that is too high compared to the current external or ambient temperature. A corresponding behavioral model, trained for machine learning or computer science, expects a certain internal temperature based on the external temperature. If the internal temperature becomes too high due to overheating of the electronics, this can be detected by the behavioral model.

[0106] The sensor's second application scenario corresponds to a location with relatively harsh environmental conditions, such as on a machine tool using cooling lubricants. These lubricants are known to be multi-component mixtures designed to provide both cooling and lubrication during metal machining. The cooling effect effectively dissipates the heat generated during machining and forming processes. However, moisture ingress can pose a risk to the sensor's electronics, potentially leading to inaccurate readings or complete sensor failure. By measuring humidity internally within the sensor, a suitably trained behavioral model can detect such moisture ingress at a very early stage and issue a corresponding warning to the user.

[0107] The third application scenario again involves the use of the sensor in a machine tool. In this scenario, too, certain external operating conditions of the sensor exist, which can influence its function and thus affect the internal operating parameters. However, these influences are initially assumed to be normal.

[0108] A machine learning (ML) or computer intelligence (KL) system, as shown in Figures 3A-3D, can learn a corresponding behavioral model based on the external environmental conditions typically prevailing in this application scenario, or train an existing standard model accordingly. Using this learned or trained behavioral model, which is configured or implemented in the sensor, the sensor can be monitored to detect, determine, or ascertain the aforementioned irregularity.

[0109] Such an irregularity can occur when the temperature or temperature profile in the sensor deviates significantly from the temperature range or profile expected based on the behavioral model for the given application scenario. A corresponding warning signal can then be sent to the user if this deviation exceeds an empirically defined threshold.

[0110] Alternatively or additionally, the humidity or humidity profile in the sensor can also be used as a basis, whereby the deviation is again detected or determined based on humidity values ​​expected in the present application scenario.

[0111] Alternatively or additionally, the electrical power consumption of the electronics arranged in the sensor can also be used as a basis to detect or determine its deviation from a power consumption typical for the present application scenario of the sensor.

[0112] It should be noted that the reliability or quality of the monitoring result can be improved by monitoring at least two of the aforementioned physical or electrical quantities simultaneously.

[0113] Furthermore, the reliability of sensor monitoring can be improved by using measured values ​​for the aforementioned operating parameters to identify patterns in these parameters and / or in derived parameters, and then linking these patterns to the states, anomalies, faults, and / or deviations based on the model. For example, patterns can be identified in the distribution of the operating parameters, in the sequence of the operating parameters, in the periodicity of the operating parameters, etc.

[0114] Possible, well-known methods for pattern recognition include, for example, "supervised" learning algorithms, e.g., decision trees, or "unsupervised" learning algorithms, e.g., K-means clustering and reinforcement learning algorithms such as Q-learning.

[0115] Suitable neural networks for machine learning are based on autoencoder models. These models compress data while preserving its key features. It should also be noted that the reliability and quality of the monitoring results can be improved by training patterns on at least two of the aforementioned physical or electrical quantities and using them as the basis for monitoring.

[0116] In summary, it should be noted that the approach proposed here, which is based on the self-perception of a sensor involved, is only made possible by the ML- or KL-based consideration of the described, additional operational (technical) information provided by the sensor.

[0117] The ML- or KL-based monitoring using a behavioral model disclosed herein has the particular advantage that, based on a suitably trained behavioral model, the functional monitoring of a sensor in question can be carried out automatically and yet very reliably in many different application scenarios, even covering or taking into account application scenarios not explicitly trained. Therefore, when using such a sensor, the specific application scenario planned or intended by the respective user does not need to be known in advance, as the behavioral model can be adapted to the respective specific application scenario during operation.

[0118] The machine learning system 400 according to the invention, shown in Fig. 4, is based on an ML- or AI-based learning approach 405, by means of which data 420 relevant to a given application or application scenario 410 of an exemplary sensor arrangement 415 are determined for training purposes. Furthermore, suitable initial training data 425 for training the technical behavior model are generated from the data 420 thus determined for the application or application scenario 410 by means of data preprocessing. In this exemplary embodiment, the CM vibration sensors 300-310 represent IO-Link devices connected to an IO-Link communication system, but can also be connected to another preferably industrially suitable communication system.

[0119] The computer interface 430, based on Kl, serves in particular to digitally record and temporarily store the technical information or boundary conditions 435 required for the correspondingly automatable configuration of a monitoring function or monitoring device of the sensor arrangement 415 for the described characteristic operating parameters with regard to the intended purpose or application scenario 410. The configuration is carried out by user 445 through interactions 440 via the computer interface 430 in natural language. The purpose or application scenario 410 is entered by the user 445, preferably including relevant technical information regarding the purpose or application scenario. This technical information regarding the specified purpose or application scenario...Deployment scenarios can be processed using data technology based on associated context / metadata.

[0120] The machine learning system 400 uses the data 435 acquired through user input 440 to train a behavioral model 450 stored in a database or data storage for the given sensor arrangement 415, e.g., for one or more of the previously described CM sensors 300-310, and the respective application or application scenario 410. This pre-trained technical behavioral model 450, which includes at least one behavioral component described below, is made available to the user 445 for interaction 440 via the computer interface 430.

[0121] Using second training data generated on the basis of updated data concerning the technical information / boundary conditions 435 for the described characteristic operating parameters, the existing behavioral model 450 can be retrained. The retrained behavioral model can then be enabled by the user 445 via the computer interface 430 for the automated configuration of the monitoring function 455 of the sensor arrangement 415 for the respective application or application scenario 410.

[0122] The current behavioral model can now be used to analyze the data 435 entered by the user 445 concerning the technical boundary conditions of the characteristic operating variables and to provide suitable instructions for generating a modified configuration of the monitoring function 455 based on the current behavioral model 460.

[0123] The monitoring function 455 provided in the sensor arrangement 415 can now be automatically configured for the specified purpose or application scenario based on the newly generated or modified parameter set. For this purpose, the machine learning system 400 can be prompted by a suitable user 445 input request to the behavioral model to provide machine-readable instructions based on a query submitted to the behavioral model by the user 445 via the computer interface 430. The input request can include one or more operating instructions relating to one or more characteristic operating parameters of the sensor arrangement 415, corresponding to the intended purpose or application scenario. These operating instructions can include machine-readable instructions for monitoring the functional state of the sensor arrangement 415.

[0124] In the monitoring approach according to the invention, the user 445 advantageously does not have to deal with individual technical operating parameters and their technical significance for the operational or functional reliability of the sensor arrangement 415, as is common in the prior art. Therefore, a manual (i.e., non-automatable) configuration of a monitoring function concerned here, as practiced in the prior art, is not necessary.

[0125] Figures 5A to 5C illustrate the known structure of an encoder / decoder architecture relevant here for generating configuration data for a monitoring function or monitoring device relevant here, based on a behavioral model of a sensor or sensor arrangement relevant here.

[0126] Fig. 5A illustrates an embodiment of an encoder architecture. The encoder comprises an encoder input 678, one or more encoder blocks 674, 614, and an encoder output 676. In this embodiment, the encoder output 676 comprises two processing levels: a linear layer 616 and a so-called "softmax" layer 618. Such a linear layer is known to connect each input neuron to each output neuron of the neural network. A "softmax" layer is commonly used in the final layer of a neural network model for classification tasks and converts raw output values ​​(so-called "logits") into probabilities by forming the exponential function of each output and normalizing these values ​​by dividing by the sum of all values.This allows a vector of K real values ​​to be transformed into a vector of K real values ​​whose sum equals 1. The input values ​​can be positive, negative, zero, or greater than one, with the "softmax" function converting them into values ​​between 0 and 1, so that these values ​​can be interpreted as probability values.

[0127] The encoder architecture shown here can be derived from the encoder-decoder architecture known in the prior art (see Fig. 5C). The encoder architecture shown can include an additional encoder output to connect the encoder block directly to the decoder of an encoder-decoder architecture, as shown in Fig. 5C.

[0128] The input data is received at the encoder input 678, which can apply a so-called "embedding" 602 to the input data. Applying the input embedding 602 can refer to forwarding the input data through an embedding layer. Furthermore, the encoder input 678 can apply a position encoding 604. Applying the position encoding 604 can refer to adding a position factor to the embedded input data. Preferably, the input data can specify a sequence of elements, where the position factor can indicate the position of the elements within the sequence.

[0129] The embedded input data can be processed by one of the encoder blocks 674 or 614 (optional block) shown. The embedded input data can be provided via a special data connection to a normalization process 608 (hereinafter referred to as "layer normalization") performed for the respective levels or layers of the underlying software architecture. Such levels / layers include, for example, the application layer (or user interface) and a data layer intended for training a technical behavior model of a sensor or sensor array.

[0130] For the embedded input data, a so-called "multi-head seif attention" mechanism 606 can be implemented. This mechanism allows an underlying behavioral model to focus on different parts of the input data, similar to how humans pay attention to specific words when understanding a sentence. This mechanism enables the behavioral model to determine which parts of the input are relevant for a given task, making it highly flexible and powerful. Such a process step is thus used to improve the expressiveness and modeling capabilities of the neural network underlying the ML or AI system. This process step can also be understood as a filter applied to the embedded input data.By applying the filter to the embedded input data, the elements associated with the embedded input data that contribute to the output data to be generated can be identified. Therefore, the filter can represent the degree of contribution of the elements associated with the embedded input data to the output data to be generated. In addition to the multi-head self-attention layer 606, a so-called feed-forward layer 610 and a corresponding layer normalization 612 can be provided for the embedded input data. Using such feed-forward layers, a forward-directed neural network is formed in a manner known per se, consisting of an input layer, one or more hidden layers, and an output layer. The data thus flows only in one direction, from the input layer through the hidden layers to the output layer.

[0131] Fig. 5B illustrates an embodiment of a decoder architecture. The decoder comprises a decoder input 684, one or more decoder blocks 680, 632, and a decoder output 692. The present decoder architecture can be derived from the encoder-decoder architecture known in the prior art and shown in Fig. 5C. The decoder architecture can correspond to the decoder architecture associated with the encoder-decoder architecture, irrespective of whether one or more hidden states are preserved by the encoder of the decoder. A multitude of such decoder architectures are available in the prior art.

[0132] The decoder input 684 can apply the input data embedding 620 and the position encoding 622 analogously to the input embedding 602 and the position encoding 204, as described in the corresponding context of Fig. 5A. The decoder block 680 can include the corresponding layer normalizations 626, a masked multi-head self-attention layer 624, a so-called feed-forward layer 628, and a corresponding layer normalization 630, which are provided via a special data connection of the respective layer normalization 626.

[0133] Furthermore, the masked multi-head self-attention 624 can be applied to the embedded input data, where the masked multi-head self-attention 624 is essentially the same as that shown in Fig. 5A. However, it includes additional masking of a portion of the embedded input data associated with elements that appear later in the sequence than the element to be generated. Additionally or alternatively, the portion of the input data associated with elements that appear later in the sequence than the element to be generated may not be received or received and / or may not be converted into the embedded input data. Therefore, the encoder can be configured for classification tasks, while the decoder can only be configured for text generation. Fig. 5C illustrates the interaction between a described encoder and a decoder.

[0134] The encoder-decoder can comprise encoder input 688, one or more encoder blocks 686, 664, decoder input 694, decoder block 690, and decoder output 692. Encoder input 688 can correspond to encoder input 278 of Fig. 5A. One or more encoder blocks 686, 664 can correspond to one or more encoder blocks 674, 614 of Fig. 5A. Decoder input 294 can correspond to decoder input 684 of Fig. 5B. The decoder block 690 can include a masked multi-head self-attention layer 670, a layer normalization layer 672, a previously mentioned feed-forward layer 638, and a layer normalization layer 640. These can be configured analogously to the masked multi-head self-attention layer 624, the layer normalization layers 626 and 630, and the feed-forward layer 628, as described in the context of Fig. 5B. The decoder block 690 can further include a multi-head self-attention layer 650 and a layer normalization layer 648.

[0135] Analogous to the description in Fig. 5B, the context tensor can be obtained from the masked multi-head self-attention 670 and the layer normalization 672. The layer normalization 648 can be applied to the context vector obtained from the multi-head self-attention 650 and the hidden states of one or more encoder blocks 686, 664. The context vector resulting from the layer normalization 648 can be processed, analogous to the description in Fig. 5B, via a feed-forward layer 638 using a layer normalization 640. The context vector resulting from the layer normalization 640 can be provided to further decoder blocks 642, analogous to the decoder block 690. The context vector obtained from the one or more decoder blocks 690, 642 can be provided to the decoder output 692. Decoder output 292 can correspond to decoder output 682 of Fig. 5B.

[0136] With the described architecture, the encoder-decoder can receive input data at encoder input 688 and one or more encoder blocks 686, 664, decoder block 690, and decoder output 692, and process this data as described. Based on the input data, the encoder-decoder can sequentially generate output data. The output data generated in this sequential manner can be provided to and / or processed by decoder input 694, one or more decoder blocks 690, 642, and decoder output 692. Preferably, a sequence can be provided to encoder input 688. After generating at least a portion of the output data, decoder input 694 can be supplied with at least a portion of the elements of the already generated output data.This allows the next elements of the output data to be generated with higher accuracy by taking into account both the input data and the generated output data.

Claims

Patent claims 1. Computer-implemented method for automatable self-monitoring of at least one sensor function of a sensor, wherein a monitoring function based on machine learning (ML) or artificial intelligence (AI) and implemented in the sensor is provided, and wherein the method comprises the following steps: Providing at least one technical behavior model trained by machine learning on an existing sensor of a given sensor type, which includes internal technical properties and possible external operating conditions of the sensor and which is adapted to a given technical purpose or a corresponding application scenario of the sensor; Setting up the monitoring function in the sensor based on the technical behavior model, to detect a possible deviation of at least one of the internal technical properties of the sensor during the ongoing operation of the sensor, wherein the technical behavior model is based on external operating conditions of the sensor that depend on the respective technical purpose or application scenario, and to detect an anomaly of the at least one technical property of the sensor by means of the detected deviation of the at least one technical property of the sensor, wherein the monitoring function detects an anomaly based on the at least one technical behavior model adapted to the respective purpose or application scenario; Issuance of a notification or warning signal by the sensor regarding a possible malfunction of the sensor concerning a detected anomaly.

2. The method according to claim 1, wherein the at least one trained technical behavior model is applied to a plurality or multiplicity of sensors of the specified sensor type intended for the specified purpose or application scenario, wherein the monitoring function detects an anomaly on the The basis of at least one technical behavior model is installed on the respective sensor and adapted to the respective purpose or application scenario.

3. Method according to claim 1 or 2, wherein for a sensor of a predetermined sensor type, various application scenarios and the operating behaviors of the sensor that depend on and are expected from these application scenarios are taught or trained in a behavioral model.

4. Method according to one of claims 1 to 3, wherein the internal technical characteristics of the respective sensor and / or the external operating conditions are one or more of the following, depending on the intended purpose or The characteristic operating parameters that depend on the application scenario include: the electrical current consumption, the electrical power consumption, the electrical voltage behavior, the temperature behavior, the humidity, the ambient pressure, the data quality and / or the data rate of the sensor data supplied by the sensor.

5. Method according to claim 4, wherein the technical behavioral model includes monitoring-relevant relationships between the characteristic operating parameters.

6. Method according to claim 4 or 5, wherein the internal technical properties of the sensor are based on the ratio or spatial distribution of the characteristic operating parameters at different locations within the sensor.

7. Method according to one of the preceding claims, wherein an anomaly is detected based on the deviation from empirically predetermined threshold values ​​for the aforementioned characteristic operating parameters and / or corresponding operating conditions.

8. Method according to one of the preceding claims, wherein the relative distribution of said characteristic operating variables to each other and / or a specific temporal sequence and / or the periodicity of the characteristic operating variables are used as a basis for monitoring.

9. Method according to one of the preceding claims, wherein at least two technical behavioral models are used as a basis for monitoring, which are active in parallel or are used as a basis for monitoring.

10. Method according to one of the preceding claims, wherein the at least one behavioral model is updated by values ​​of the characteristic operating variables measured during the operation of the sensor as such training data.

11. Method according to any of the preceding claims, wherein artificial neural networks, decision tree-based methods, margin-based methods, cluster methods, ensemble methods, nearest-neighbor methods; and / or linear and / or nonlinear regression methods are used as classes of algorithms for machine learning.

12. Method according to one of the preceding claims, wherein pattern recognition is provided using measured values ​​of at least one characteristic operating parameter, wherein patterns in the operating parameters and / or in parameters derived therefrom are recognized and these patterns are linked to the states, anomalies, faults, and / or deviations on the basis of the technical behavior model, wherein patterns are recognized in the distribution of the operating parameters, in the sequence of the operating parameters or in the periodicity of the operating parameters.

13. Method according to claim 12, wherein the patterns take into account possible correlations between the internal technical properties and possible external operating conditions of the sensor in the distribution of operating parameters.

14. Method according to one of the preceding claims, wherein a statistical test procedure is performed using the measured values ​​for the at least one characteristic operating parameter.

15. Method for configuring a monitoring function of a sensor, wherein at least one trained technical behavior model corresponding to a given purpose or application scenario of a plurality or multitude of sensors of a given sensor type is provided, wherein the at least one technical behavior model relates to internal technical properties of the respective sensors which depend on the given purpose or application scenario of the sensor, and wherein the configuration of the monitoring function is carried out using the technical behavior model.

16. Method according to claim 15, wherein the trained or learned technical behavior model is applied to the majority or multitude of sensors of the specified sensor type before their delivery.

17. Sensor with a monitoring function which is configurable by means of at least one technical behavior model trained according to one or more of the preceding claims, wherein the at least one technical behavior model maps characteristic operating parameters and / or operational behavioral properties of the sensor in a digital twin of the sensor in which the respective sensor is set up.

18. Computer-implemented method for the automated creation of a self-monitoring function of a sensor concerned here for a given purpose or application scenario, wherein the method is carried out using a machine learning system and wherein the method comprises the following steps: Communication between a user and the machine learning system via a computer interface to collect data regarding the specified purpose or application scenario; Training a technical behavioral model with respect to the aforementioned characteristic operating parameters using the collected data, or optimizing an existing behavioral model using the collected data; Releasing the trained behavioral model for the automated creation of the monitoring function.

19. Method according to claim 18, wherein normal states corresponding to the intended use or application scenario are learned in the behavioral model and deviations from one or more of these normal states are detected by means of the monitoring function provided in the sensor.

20. Method according to claim 18 or 19, wherein preprocessing is carried out to train a behavioral model, wherein features are extracted from available data relating to the characteristic operating parameters and / or the available data are pre-transformed and / or filtered and / or aggregated and / or otherwise pre-processed.

21. Computer-implemented method for using a trained behavioral model to configure a monitoring function of a sensor concerned here for a given purpose or application scenario, wherein the method comprises the following steps: Providing a user with access to a trained behavioral model via a computer processor; receiving data entered via a computer interface for the specified purpose or application scenario; To cause the trained behavioral model to analyze the input data and provide instructions to generate a new or modified monitoring function based on the trained behavioral model.

22. Method according to claim 21, wherein the trained behavioral model is caused to configure a monitoring function for a given purpose or application scenario based on machine-readable instructions and data entered by the user via the computer interface.

Images