Technology for monitoring a switchgear cabinet

Abstract

The invention relates to technology for monitoring a switchgear cabinet (200). One aspect of the technology relates to a device (100) in the switchgear cabinet (200). The device (100) comprises a data detection module (102) which detects at least two data values of sensor data of the switchgear cabinet (200); a parametrisation module (104) which, in a training phase of the device (100), sets at least one parameter of a classification of nominal states of the switchgear cabinet (200) on the basis of the detected data values of the sensor data, the classification being implemented in the device (100); a switch (110) or a switching module (110) which is designed to change a rate at which the at least one parameter is set; and a monitoring module (106) which, in a monitoring phase of the device (100), determines on the basis of the detected data values of the sensor data whether a current state of the switchgear cabinet (200) is a nominal state according to the classification based on the at least one set parameter.

Description

[0001] Description

[0002] Technology for monitoring a control cabinet

[0003] The present invention relates to a technique for monitoring a control cabinet. In particular, a device in a control cabinet for monitoring it, a correspondingly equipped control cabinet, and a method for monitoring a control cabinet are disclosed.

[0004] In the prior art, so-called control cabinet monitors with programmable or parameterizable evaluation of measured variables are known for monitoring a control cabinet.

[0005] The conditions inside a control cabinet can be monitored with various sensors to detect faults and defects, impermissible operating conditions, unauthorized access, etc. Measuring the temperature inside the control cabinet is common. Temperature monitoring is described, for example, in document WO 2008 / 052 813 A1. Furthermore, the condition of the control cabinet door can be monitored with various sensors, such as ultrasonic distance sensors as described in document EP 4 1 13 163 A2.

[0006] The data collected by sensors is conventionally transmitted to a higher-level or central control system, usually a cloud system, to process the condition of the control cabinet or to allow user monitoring. However, this requires knowledge of the devices installed in the control cabinet and the local environmental conditions in order to interpret the collected data during monitoring. Furthermore, such network access to the control cabinet can represent a security vulnerability, thus counteracting the intended operational monitoring due to an increased risk of attack.

[0007] Another disadvantage of the current state of the art is that an installer - under

[0008] Considering the complex interrelationships between the control cabinet's environment, the installed functionality, and the specific on-site usage, the rules for safety monitoring must be determined and programmed. In practice, this necessitates extended periods of test operation after the control cabinet has been installed, before it can be used operationally. The time-consuming acquisition and complex evaluation of the measured values ​​during test operation is followed by the adjustment of the parameters of several configurable condition monitoring rules. These parameters must be adapted to the specific on-site conditions. Adjusting the parameters requires trained personnel and necessitates an evaluation unit with a suitable parameterization interface.

[0009] Another disadvantage of the conventional approach is that impermissible conditions can occur during later operation, which are not detected by the rule-based monitoring, as they were not foreseeable at the time of programming or parameterization and therefore could not be taken into account when selecting the rules and their parameters.

[0010] The object of the invention is to remedy this situation. This object is achieved by the features of the independent claims. Advantageous embodiments and further developments of the invention are specified in the dependent claims.

[0011] Exemplary embodiments of the invention, which can be optionally combined, are disclosed below with partial reference to the figures. In particular, features mentioned in the context of the device can also be implemented accordingly in the method, for example, by a step of providing the corresponding feature or by a step of executing a function of the device. Furthermore, the device can include any feature mentioned in the context of the method and can be configured to execute any step mentioned in the context of the method.

[0012] According to a first aspect, a device is provided in a control cabinet for monitoring the control cabinet. The device comprises a data acquisition module configured to acquire at least two data values ​​from sensor data of the control cabinet. The device further comprises a parameterization module configured to set at least one parameter of a classification of nominal states of the control cabinet during a learning phase of the device, based on the acquired data values ​​of the sensor data. The classification is implemented in the device. The device further comprises a switch or a switching module configured to change the rate at which the at least one parameter is set.The device further includes a monitoring module which is designed to determine, in a monitoring phase of the device, based on the acquired data values ​​of the sensor data, whether a current state of the control cabinet is a nominal state according to the classification dependent on the at least one set parameter.

[0013] In other words, a device is provided that monitors the current state of a control cabinet (for example, the operating conditions within a control cabinet) and evaluates data from various sensors to classify this current state. Exemplary implementations can perform the classification (i.e., including the evaluation) locally within the device. This locally performed evaluation encompasses not only monitoring the current state through classification but also adjusting the classification itself by parameterizing it; for example, an evaluation algorithm can automatically adapt to the prevailing conditions.

[0014] By having the device automatically set at least one classification parameter based on the acquired data values, exemplary embodiments of the device do not require user parameterization. This or other exemplary embodiments, due to the acquisition of data values ​​(by the device's data acquisition module), the evaluation of these values ​​for classification adjustment (by the device's parameterization module), and monitoring of the current state through classification (by the device's monitoring module), do not require a network connection, internet access, connection to a cloud system, or connection to a higher-level control system (e.g., evaluation unit).

[0015] In exemplary embodiments of the device, the parameterization module and / or the monitoring module (for example, the use of an algorithm for anomaly detection) can be executed entirely locally (e.g., on a microcontroller). The parameterization module can itself parameterize the classification applied by the monitoring module.

[0016] In exemplary embodiments of the device, the complex parameterization by the user and the corresponding interfaces that were conventionally required can be eliminated.

[0017] By combining the parameterization module and the monitoring module, exemplary embodiments of the device can determine error states at any time that a manually programmed evaluation of measured variables overlooks, since these states were not foreseen during programming.

[0018] Thanks to the parameterization module and the monitoring module in the device, exemplary embodiments of the device do not rely on a cloud system or other higher-level units for evaluating the data values. No data values ​​need to be transmitted to such systems. A network connection with corresponding bandwidth and security measures is therefore unnecessary.

[0019] Thus, exemplary embodiments of the device (i.e., a monitoring device or control cabinet monitor) can monitor conditions within a control cabinet without requiring any parameterization by the user. The evaluation of the measured values ​​can be performed entirely locally within the device, eliminating the need for a connection to higher-level evaluation units or cloud systems.

[0020] Determining whether the current state is a nominal state or not can involve classifying the current state. The classification can be a function of the current state. The at least one parameter can be a parameter of the classification; that is, the classification is a function of the current state that is set (e.g., fixed, defined, or modifiable) by the at least one parameter. The result of the classification can include classifying the current state as a nominal state or a non-nominal state. While the terms "classify" and "classification" in the narrower sense form a pair of meanings analogous to function and function value, these terms can be considered synonymous in the present description of technical phenomena.

[0021] The device can include a processing circuit, such as a microprocessor with memory, containing program code for the data acquisition module, the parameterization module, and / or the monitoring module. For example, the parameterization module and the monitoring module can include GPUs (Graphics Processing Units) designed to process data in parallel and accelerate matrix operations in an energy-efficient manner, optionally to implement the two modules (i.e., the self-parameterizing classification) as a support vector machine. The classification can be implemented directly in the device using machine learning models (ML models), for example, without requiring significant computing power (TinyML) for the self-parameterization of the classification and without evaluation and machine learning via a network connection.

[0022] The sensor data can be measured quantities.

[0023] The device (in particular the data acquisition module) can be connected to one or more sensors. One or more sensors can be located in or on the data acquisition module and / or in or on the control cabinet. The sensor(s) can be internal and / or external sensors related to the data acquisition module or the control cabinet. Alternatively or additionally, the sensor(s) can be configured as environmental sensors (which, for example, acquire data about the environment of the control cabinet) and / or operational sensors (which, for example, acquire data about the operation of the control cabinet).

[0024] The device can also be electrically connected to a power source via a conductive connection or inductively coupled. The power source can be located in the control cabinet, for example, on a mounting rail next to the device. Alternatively or additionally, the power source can be another module of the device. The power source can be configured to supply electrical power to the data acquisition module, the parameterization module, the monitoring module, and / or the sensor(s).

[0025] The nominal state can be a state in which the control cabinet functions properly and / or the recorded data values ​​are as expected and / or consistent with each other and / or meaningful for an operating state of the control cabinet.

[0026] The nominal state can be defined by at least one (set) parameter (P1, ...) of the device. For example, the monitoring module can determine, based on the acquired data values ​​(D1, D2, ...) of the sensor data, whether a current state of the control cabinet is a nominal state according to the classification dependent on the at least one set parameter (P1, ...):

[0027] Classification: Nominal state (D1 , D2, ..., P1 , ...) is true or false.

[0028] Thus, the nominal state can presuppose a relation or correlation between at least two data values.

[0029] Optionally, at least one parameter (P1 , ...) can specify a multidimensional range (B_nominal) in which a combination of the recorded data values ​​(D1 , D2, ...) corresponds to the nominal state, for example:

[0030] Classification as nominal state if a data vector (D1 , D2, ...) of the data values ​​lies in the range B_nominal(P1 , ...).

[0031] The at least one parameter (P1, ...) can, for example, include an upper limit and / or a lower limit for the sensor data values ​​in the nominal state. The at least one parameter can be a limit for individual data values, for example:

[0032] Classification as nominal state if P1 < D1 < P2 and P3 < D2 < P4.

[0033] Alternatively or additionally, at least one parameter can be a limit value for at least one linear function of the data values, for example:

[0034] Classification as a nominal state if ((D1 , D2, ...), A) < P1 , where (..., A) is the dot product with a vector A. Alternatively or additionally, the at least one parameter can be a limit value for at least one nonlinear function f of the data values ​​D1 , D2, ..., for example:

[0035] Classification as nominal state if f (D 1 , D2, ... ) < P 1 .

[0036] The device can be designed to be detachably attached to a mounting rail of the control cabinet.

[0037] Parameterization can refer to setting at least one parameter of the classification. The parameterization module can enable the device to perform the parameterization itself (for example, autonomously and / or offline) (self-parameterization).

[0038] The sensor data, which must contain at least two data values, can include at least one data value from environmental data of the control cabinet and / or at least one data value from operating data of the control cabinet.

[0039] In other words, the sensor data can include environmental data relating to the ambient state of the control cabinet and / or operational data relating to the operating state of the control cabinet. The operating state of the control cabinet can refer to the operating state of other devices installed within the control cabinet. The ambient state of the control cabinet can refer to one or more intensive thermodynamic state variables of the environment of the control cabinet, such as air temperature, relative humidity, or oxygen content. Here, the "environment" of the control cabinet, the "environmental data" of the control cabinet, and the "ambient state" of the control cabinet can refer to the external environment of the control cabinet and / or the internal environment of the control cabinet, in particular to environmental conditions that exist only inside the control cabinet, only outside the control cabinet, or due to fluid communication of a medium (e.g.,The "environment" can refer to any influence on the devices installed in the control cabinet, such as external electric, magnetic, or electromagnetic fields. This includes the air between the inside and outside of the control cabinet, which can be measured or measured as part of the measurement process. Alternatively or additionally, the "environment" can refer to any influence on the devices installed in the control cabinet, such as external electric, magnetic, or electromagnetic fields.

[0040] Environmental and / or operational data can be acquired by one or at least two sensors in or on the control cabinet. The environmental sensor(s) can be configured to acquire data about the control cabinet's environment, e.g., from inside and / or outside the cabinet, and / or to detect the cabinet's closed / closed status. The operational sensor(s) can acquire operational data from the control cabinet (for example, from a specific device installed within the cabinet), such as one or more cabinet currents, voltages, or power consumption.

[0041] The parameterization module can be configured to repeatedly set at least one parameter. Alternatively or additionally, the monitoring module can be configured to repeatedly determine whether the current state of the control cabinet is a nominal state. Alternatively or additionally, the parameterization module and the monitoring module can be configured to perform the setting of the at least one parameter and the determination of whether the current state of the control cabinet is a nominal state alternately or simultaneously.

[0042] "Repeated" can refer to periodic, continuous, or event-driven occurrences.

[0043] A rate of repeated adjustment of at least one parameter can be an example of a device's learning rate. Repeated parameter adjustments by the parameterization module enable continuous adaptation of the monitoring system to changing operating and environmental conditions. The parameterization module can dynamically adjust the classification to new states, providing flexibility and increased accuracy in anomaly detection. By repeatedly determining whether the nominal state is present, the monitoring module can detect fault conditions in real time.

[0044] Running the training and monitoring processes in parallel allows the device to continuously learn and adapt without interrupting the monitoring functions. This ensures continuous monitoring while the classification remains up-to-date or becomes more accurate. Alternating execution allows for specific phases dedicated to learning and monitoring. This can help to efficiently utilize the device's system resources and optimize the performance of the device's microcontroller.

[0045] The parameterization module and / or the monitoring module can perform a moving average calculation of the data values ​​and / or histogram-based statistics.

[0046] The parameterization module can set (e.g., update) at least one parameter as a moving average of the sensor data based on the data values. The monitoring module can determine, based on a deviation between the recorded sensor data values ​​and the parameter as a moving average, whether the nominal state exists (e.g., if the deviation is less than a relative threshold) or not (e.g., if the deviation is greater than the relative threshold).

[0047] Alternatively or additionally, the parameterization module can set at least one parameter as a histogram-based statistic based on the data values ​​(e.g., update it). For example, the at least one parameter can specify how frequently a combination of data values ​​is recorded. The combination can be a combination of (e.g., simultaneously measured) data values ​​for different sensor data (e.g., from different sensors or for different measured variables, such as power consumption and temperature) and / or a combination of a temporal sequence of (e.g., periodically measured) data values ​​for the same sensor data (e.g., from the same sensor or for the same measured variable).

[0048] Alternatively or additionally, the parameterization module can execute one of the following machine learning (ML) methods to set at least one parameter. The monitoring module can then execute the classification resulting from the ML method.

[0049] Exemplary ML methods include nearest neighbor classification, density-based spatial cluster analysis with noise, support vector machine, isolation forest, or deep learning methods for a neural network model, optionally for a model with long and short time interval memory (Long Short-Term Memory or LSTM), for a recurrent neural network, for an auto-encoder, or for a generative-adversarial network.

[0050] The model involving long-term and short-term memory can also be called the Long-Short-Term Memory model. The "Isolation Forest" method can also be referred to as "Isolation Forest" in technical terms.

[0051] The hardware of the parameterization module and the monitoring module can include a shared (artificial) neural network. At least one parameter can comprise the parameters of the neural network (for example, neural network weights). The neural network can be (for example, according to the aforementioned machine learning method) a recurrent neural network (for example, for the model with long- and short-term memory or for the auto-encoder) or a generative-adversarial network.

[0052] A generative-adversarial network, often referred to as a GAN (from "Generative Adversarial Network"), can be a special type of neural network consisting of two competing models: a generator and a discriminator. These two models compete to produce synthetic data that is indistinguishable from real data.

[0053] The device and / or the parameterization module and / or the monitoring module can use a classification (preferably based on machine learning). This classification can indicate whether the current state of the control cabinet corresponds to its nominal state. Based on the parameterization module, the classification can improve during operation of the control cabinet and / or adapt to changes in the control cabinet (for example, its use or the devices installed within it).

[0054] In one variant of each embodiment, the training data can include simulation data.

[0055] The implemented machine learning (ML) methods can be pre-trained with training data (for example, "out of the box," i.e., in a delivered and not yet installed state). For instance, at least one parameter can include predetermined weights of the neural network. This training data or these predetermined parameters can represent typical conditions and / or signal waveforms in control cabinets. In other words, the device can already be pre-trained "out of the box" to, for example, shorten the initial learning phase and / or the initial processing time of the parameterization module. Preferably, the pre-training also includes simulation data as training data.

[0056] The set parameter can be a true subset of all classification parameters (for example, all model parameters or all neural network weights). In other words, self-parameterization, i.e., the learning process after the device is commissioned, can set (for example, adjust) a true subset of all classification parameters (i.e., not all parameters, but only a subset of the parameters). This can reduce the device's computing power requirement (for local self-parameterization). For example, the parameters of only one layer or individual layers of the neural network can be set.

[0057] The switch or switching module can also be configured to activate the learning phase or to switch between the learning phase and the monitoring phase.

[0058] For example, the rate of setting at least one parameter can only differ from zero during the learning phase.

[0059] The monitoring phase of the monitoring module can be executed alternatively (i.e., complementarily) or simultaneously (i.e., in parallel) to the learning phase of the parameterization module. In the former case, the self-parameterization of the parameterization module can be limited to a learning phase that precedes the monitoring phase. After the learning phase, at least one parameter (e.g., the classification model) can remain unchanged (i.e., frozen). The device can then switch to pure monitoring mode during the monitoring phase.

[0060] The rate of setting the at least one parameter (i.e., a learning rate) can include a step length of a change in the at least one parameter, or a frequency of change in the at least one parameter, or a proportion of the learning phase to the operating time of the device, or a ratio of the duration of the learning phase to the duration of the monitoring phase.

[0061] The switch can be a mechanical switch or an electronic switch (for example, a function of a user interface of the device). The switch can be, for example, a push button, a bistable rocker switch, a DIP switch, a slide switch, or a rotary switch. The switch can be accessible on a device housing inside the control cabinet. Alternatively or additionally, the switch (as the switching module) can automatically perform switching operations and / or receive commands from a user, for example, voice commands or via near-field communication (e.g., NFC) or short-range radio communication (e.g., Bluetooth).

[0062] The switching module can trigger (e.g., to switch between the learning and monitoring phases) when a criterion is met. For example, if the classification accuracy is less than a predetermined threshold during the monitoring phase, or if the learning rate (during parallel self-parameterization and monitoring) is greater than a predetermined threshold, the monitoring phase can be terminated and / or the learning phase activated or switched to. This prevents error-prone monitoring and / or allows for responses to modifications in the control cabinet or changes in its usage patterns. Alternatively or additionally, if the learning rate is less than a predetermined threshold or the classification accuracy is greater than a predetermined threshold during the learning phase, the learning phase can be terminated and / or the monitoring phase activated or switched to.This saves energy and prevents unwanted drift in the classification (e.g., as a result of a gradually deteriorating condition of the control cabinet).

[0063] For example, the switching module can (automatically) end the learning phase depending on the progress of the ML process (i.e., the machine learning process) (e.g., after processing a predetermined number of the acquired data values ​​(e.g., per unit of time, i.e., when a processing rate is reached or exceeded), or when one of the positive quality criteria has reached or exceeded a threshold, or when one of the negative quality criteria (e.g., a loss function) has reached or fallen below a predetermined threshold). Alternatively or additionally, the switching module can (automatically) end the learning phase after a predetermined period (e.g., one operating day, one week, or one year).

[0064] The training phase can be activated (and optionally the monitoring phase terminated) if a classification quality criterion falls below a predetermined quality threshold, for example, if the classification error rate exceeds a predetermined (acceptable) error level. This applies accordingly to a positive quality criterion or a negative quality criterion (i.e., a criterion where a smaller, not necessarily negative, numerical value indicates better quality).

[0065] Determining whether the current state (e.g., operating state) of the control cabinet corresponds to the nominal state or not can be subject to the four possible cases of a truth matrix: True positive with a rate r p (i.e., the control cabinet is nominal, and the classification, which depends on at least one parameter being set, has determined this), false negative with a rate f n(i.e., the control cabinet is nominal, but the classification, which depends on at least one parameter being set, has incorrectly determined it to be non-nominal), false positive with a rate f p (i.e., the control cabinet is not nominal, and the classification, which depends on at least one parameter being set, has incorrectly determined it to be nominal), and correctly negative with a rate r n (i.e., the control cabinet is non-nominal, and the classification, which depends on at least one parameter being set, has correctly determined this to be non-nominal).

[0066] The positive quality criterion can include the sensitivity (or true positive rate) of the classification. The sensitivity indicates the probability that a truly nominal control cabinet is correctly classified as nominal, for example, according to: True nominal (positive test result) = rp + ln

[0067] The negative quality criterion can include a false-negative rate of classification. The false-negative rate indicates the probability that a truly nominal (e.g., functioning) control cabinet is incorrectly classified as non-nominal (e.g., defective), for example, according to

[0068] / PC (actually defective / negative test result) = — — . ' P+fn

[0069] The positive quality criterion can include the specificity (or true-negative rate) of the classification. Specificity indicates the probability that a truly non-nominal (e.g., non-functional) control cabinet is correctly classified as non-nominal, for example, according to P(truly non-nominal|negative test result) =

[0070] ■ n+ /

[0071] The negative quality criterion can include a false-positive rate of classification. The false-positive rate indicates the probability that a non-nominal control cabinet is incorrectly classified as nominal, for example, according to P(actually functional! positive test result) = fp r» +fp

[0072] The positive quality criterion can include a positive predictive value (also known as relevance, effectiveness, accuracy, or "Positive Predictive Value," PPV) of the classification. The positive predictive value indicates the proportion of control cabinets correctly classified as nominal out of the total number of control cabinets classified as nominal, for example, according to P(actually defective (positive test result) = — . rp + fp

[0073] The negative quality criterion can include a false discovery rate (FDR) of the classification. The FDR indicates the proportion of control cabinets that are actually non-nominal and those incorrectly classified as nominal, relative to the total number of cabinets classified as nominal, for example, according to P(actually functional! positive test result) = - . i~p+fp

[0074] The positive quality criterion can include a negative predictive value (NPV) of the classification. The negative predictive value indicates the proportion of control cabinets correctly classified as non-nominal out of all control cabinets classified as non-nominal, for example, according to (actually functional (negative test result) = r). + . n + fn

[0075] The negative quality criterion can include a false omission rate (FOR) of the classification. The false omission rate indicates the proportion of enclosures that are actually nominal and those incorrectly classified as non-nominal out of all non-nominal enclosures, for example, according to [reference to relevant standard / guideline].

[0076] ^(actually defective| negative test result) = — rn fn

[0077] In some implementations, no "ground truth" is available for training. Therefore, during the learning phase, it is unknown whether the current state is actually the nominal state or not. Furthermore, a performance function is required for training a machine learning (ML) method. Depending on the type of method used, different "labels" can be used for this purpose.

[0078] Some training methods, such as statistically based procedures, assume that the nominal states are dominant and that the few non-nominal states can be neglected during the learning phase (i.e., assumption: everything is nominal).

[0079] In other methods, such as predictive methods, the goal is to predict future measurements from the past. Historical data can then be used for training (e.g., not the nominal yes / no state, but "raw data"). For classification, the deviation between the prediction and the measured value (e.g., against a threshold) is then evaluated.

[0080] In other methods, such as reconstruction-based methods (e.g., autoencoders), the dimension of the data is reduced to a latent feature space from which the input data is then reconstructed. The reconstruction error forms the basis of the power function. For classification, the reconstruction error must then be re-evaluated (e.g., against a threshold).

[0081] In some embodiments, the device can detect that a specific pattern of reconstruction errors frequently correlates with actual anomalies (e.g., existing defects or impending failures), thus improving the positive predictive value. This enables not only the detection of outliers during operation but also the prediction and planning of maintenance work to avoid costly downtime and increase operational reliability.

[0082] The transition from the learning phase to the monitoring mode can be triggered by an external trigger, e.g. manual operation of the switch, optical detection of a one- or two-dimensional code, or receipt of a control command via a communication interface of the switching module.

[0083] Setting the at least one parameter (e.g., the machine learning process) can involve a variation in the learning rate over time. For example, a high learning rate can be used at the beginning of operation of the device or control cabinet so that the at least one parameter quickly adapts to the prevailing conditions, such as a specific configuration or application of the control cabinet. As learning progresses, the learning rate can be reduced to prevent the system from learning to gradually develop fault conditions, which would counteract monitoring.

[0084] For setting at least one parameter, e.g., for a machine learning process (technically: "training"), especially for fine-tuning, all recorded data values ​​(e.g., measured values) from the sensor data can be used. Alternatively, only individual data values ​​can be selected. This reduces the computing power required for training the model. The selection of data values ​​for the learning process can be based on a fixed grid (e.g., using every 100th measured value). The selection can also be made by a preceding machine learning model (e.g., a trained algorithm or artificial neural network) (e.g., a clustering algorithm or a convolutional network), with the goal of selecting those data values ​​that contain the most information for the training progress.

[0085] In addition to setting at least one parameter (for example, the self-parameterizing algorithm), the device can perform further evaluations of the data values ​​(e.g., measured values). These further evaluations may also require parameterization (e.g., comparisons of the measured values ​​with lower and upper limits). The monitoring module can be configured to determine the current state according to the classification based on a combination of the acquired sensor data values.

[0086] The at least two environmental and / or operational data values ​​can be acquired (e.g., received) by two or more sensors (in or on the control cabinet). The classification can encompass a multidimensional field spanned by the at least two data values, in which disjoint regions correspond to the nominal and non-nominal states. Using a combination of data values ​​reduces the probability of false alarms. Individual data values ​​that are not significant in the overall picture are less likely to lead to a false detection of a non-nominal state. This improves the reliability and autonomy of the monitoring device.

[0087] The operating data can comprise at least two data values. Alternatively or additionally, the at least two data values ​​of the operating data can each include timestamps, or the data values ​​of the operating data can comprise a time series. Alternatively or additionally, the operating data can consist of measured variables acquired within the control cabinet. Alternatively or additionally, the operating data can include at least one of the following measured variables of the control cabinet: a current through a component of the control cabinet, a current input or power input of the control cabinet, a current output or power output of the control cabinet, and an open and / or closed position of a door or access panel of the control cabinet.

[0088] The closed state or the open / closed position can be detected by a sensor located inside the control cabinet, for example, a rotary position sensor that detects the rotation of the door, or a distance sensor that detects the distance to the inside of the door or an access panel. Alternatively or additionally, an overpressure (e.g., relative to the air outside the control cabinet) can be maintained inside the control cabinet. The closed state can be monitored by a pressure sensor. The environmental data can comprise at least two data values. Alternatively or additionally, the at least two environmental data values ​​can each include timestamps, or the environmental data values ​​can comprise a time series. Alternatively or additionally, the environmental data can consist of measurements taken outside the control cabinet.Alternatively or additionally, the environmental data may include one or more measured variables of the control cabinet, for example, an air temperature in or around the control cabinet and / or an air humidity in or around the control cabinet and / or a concentration of smoke particles in the air of the control cabinet and / or an oxygen, nitrogen or carbon dioxide content in the air in the control cabinet and / or an overpressure in the control cabinet and / or a noise level inside or outside the control cabinet, a power level of a pattern of acoustic signal (determined by at least one parameter) inside or outside the control cabinet or mechanical vibrations (for example, their frequency or amplitude or power level) inside or outside the control cabinet.

[0089] The at least two data values ​​of the operational data or environmental data can have different timestamps and / or be received from different data sources and / or be different physical measured quantities.

[0090] The device may further include an output module designed to output the specified current state, the received data values, the set at least one parameter and / or a deviation from the nominal state.

[0091] The output module can be configured for optical or acoustic output.

[0092] The output module can have a network interface (e.g., wired or wireless) and / or be in data communication with an external control unit. The external control unit can be a user device, such as a tablet computer or mobile phone. For example, the aforementioned device information or other data from the control cabinet (e.g., control instructions for switching the learning or monitoring phase on or off) can be exchanged with the user device via short-range communication. The user device can then function as a switch and / or output interface for the device (e.g., via an application running on the user device).

[0093] According to a second aspect, a control cabinet is provided. The control cabinet includes at least one sensor that outputs at least two data values ​​from the sensor data of the control cabinet. The control cabinet further includes a device according to the first aspect for monitoring the control cabinet, which is configured to determine, based on the output data values ​​of the sensor data, whether a current state of the control cabinet is a nominal state.

[0094] According to a third aspect, a method for monitoring a control cabinet is provided. At least two sensor data values ​​from the control cabinet are acquired. During a learning phase of the device, at least one parameter of a classification of nominal states of the control cabinet is set based on the acquired sensor data values. The classification is implemented in the device. A switch or switching module is provided, the switch or switching module being configured to change the rate of setting the at least one parameter. During a monitoring phase of the device, it is determined, based on the acquired sensor data values, whether a current state of the control cabinet is a nominal state according to the classification dependent on the at least one set parameter.

[0095] The second and third aspects can be further developed through any feature, or corresponding procedural step, of the first aspect.

[0096] The invention and its technical context are explained in more detail below with reference to the figures. It should be noted that the invention is not limited by the embodiments shown. In particular, unless explicitly stated otherwise, it is also possible to extract aspects of the concepts illustrated in the figures and combine them with other elements and findings from the present description and / or figures. It should be emphasized that the figures, and especially the depicted dimensions, are only schematic. Identical reference numerals denote identical or interchangeable features, so that explanations from other figures may be consulted for supplementary information.

[0097] They show:

[0098] Figure 1 shows a schematic block diagram of an embodiment of a device in a control cabinet for monitoring the control cabinet.

[0099] Figure 2 shows a schematic block diagram of an embodiment of a control cabinet.

[0100] Figure 3 shows an exemplary flowchart of a process in the control cabinet during the monitoring phase and the learning phase according to one embodiment.

[0101] Figure 4 shows a schematic view of an exemplary switch of the device,

[0102] Figure 5 shows an exemplary flowchart of the switching between the phases of the method and the device in the control cabinet; Figure 6 shows examples of a variable learning rate according to an embodiment of the method and the device in the control cabinet.

[0103] Figure 7 shows a schematic diagram of a field of recorded data values ​​as part of the training data for the learning phase.

[0104] Figure 8 shows a schematic diagram of the field with an exemplary state classification as a result of the learning phase, and

[0105] Figure 9 shows a schematic diagram of the field with an application of the set state classification to the current state during the monitoring phase.

[0106] Figure 1 shows a schematic block diagram of an embodiment of a device in a control cabinet for monitoring the control cabinet. The device is generally designated by reference numeral 100.

[0107] The device 100 comprises a data acquisition module 102, which is configured to acquire at least two data values ​​from sensor data in the control cabinet. The data acquisition module 102 can be in data communication with one or more sensors.

[0108] The device 100 further comprises a parameterization module 104, which is configured to set at least one parameter of a classification of nominal states of the control cabinet 200 during a learning phase of the device 100 based on the received data values ​​of the sensor data.

[0109] The device 100 further comprises a monitoring module 106, which is configured to determine, in a monitoring phase of the device 100, based on the received data values ​​of the sensor data, whether a current state of the control cabinet 200 is a nominal state according to the classification dependent on the at least one set parameter.

[0110] The device 100 can operate in two phases: a learning phase and a monitoring phase. The learning phase can precede the monitoring phase.

[0111] Alternatively or additionally, the training phase can partially overlap with or occur simultaneously with the monitoring phase. In this context, "training" (or simply "learning") can refer to setting at least one classification parameter.

[0112] The device 100 can learn in advance, relearn, stop learning, or learn continuously, for example, in response to a classification accuracy calculated by the parameterization module 104. During the learning phase, the device 100 can use training data, such as at least two sensor data values, or simulation data for learning. The device 100 can use one or more machine learning algorithms during the learning phase.

[0113] The device 100 optionally includes an output module 108, which is configured to output the determined current state, the received data values, the set at least one parameter and / or a deviation from the nominal state.

[0114] The output module 108 can output an acoustic and / or optical message and / or a message to a user interface, for example, a message to an application on a mobile device or computer device, etc. Advantageously, this output can take place via a unidirectional interface so that the interface does not pose an attack risk to the device 100.

[0115] The device 100 optionally includes a switch 110 or a switching module 110, which is configured to allow a user to switch between the learning phase and the monitoring phase, or to switch automatically (for example, event-driven in response to the calculated quality), or (for example, in the case of asynchronous or time-decoupled learning and monitoring phases) to switch the learning (i.e., setting the at least one parameter) on or off, or to switch the monitoring (i.e., determining whether a nominal state exists) on or off.

[0116] One or all modules of the device 100 can be in data communication with one or all modules of the device 100.

[0117] Figure 2 shows a schematic block diagram of an embodiment of a control cabinet, generally designated by reference numeral 200. The control cabinet 200 comprises at least one sensor 202, which outputs at least two data values ​​from sensor data. The sensor 202 can be an environmental sensor and / or an operational sensor. The at least two data values ​​can comprise a time series for a measured quantity.

[0118] The sensor(s) 202 can be located in and / or on the control cabinet 200.

[0119] The control cabinet 200 further comprises a device 100 for monitoring the control cabinet 200, which, according to the first aspect, is designed to determine, based on output data values ​​of the sensor data, whether a current state of the control cabinet 200 is a nominal state.

[0120] The acquisition (e.g., measuring and storing) of the sensor data values ​​(e.g., sensor readings) typically takes place in the device 100, which is mounted in the control cabinet 200 for this purpose. The device 100 can also perform other functions besides acquiring the data values ​​(e.g., measured values). The sensors 202 can be internal (i.e., installed in the device 100) or external (i.e., connected via cable or wireless connections).

[0121] The evaluation of the data values ​​(e.g., measured values, especially sensor readings) can be performed locally in the device 100 that acquired the data values ​​(directly or indirectly). Alternatively, the acquired measured values ​​can also be transmitted to a higher-level evaluation unit (e.g., a controller or a cloud system) and evaluated there, preferably while anonymizing the deployment location. In this way, the data values ​​can contribute to the creation of a predictive maintenance plan.

[0122] Depending on whether the learning phase or monitoring phase is activated, the evaluation can include the function of module 104 or 106.

[0123] When evaluating data values ​​(e.g., measured values), the current state is classified using configurable rules (function of module 106). A simple rule might, for example, state that a specific measured value must not fall below a lower threshold and must not exceed an upper threshold. If the rule is violated, an impermissible state is detected and a reaction is triggered (e.g., a message to personnel, a warning light, a switching contact, etc.).

[0124] Furthermore, this classification is parameterized by the device 100 (function of module 104).

[0125] Exemplary embodiments of the device 100 thus overcome certain disadvantages of the prior art mentioned above, in that one embodiment of the monitoring device 100 (e.g., the control cabinet monitor) executes a self-parameterizing evaluation algorithm. The control cabinet monitor acquires the available measured values ​​of the environmental conditions via internal and / or external sensors and processes them with an anomaly detection algorithm. Such an algorithm is capable of distinguishing between normal and abnormal measured value inputs. Normal measured value inputs occur with a certain frequency during normal operation. Abnormal measured value inputs do not occur, or only rarely, during normal operation (i.e., the nominal state). Their occurrence therefore indicates an impermissible condition of the switchgear.

[0126] The term anomaly can be synonymous with a non-nominal state, i.e., a negation of the nominal state or a deviation from the nominal state. Various types of anomalies can be identified using a classification of the control cabinet's current state, for example, an anomaly detection algorithm. These include, in particular, the following anomalies.

[0127] Point anomalies: This non-nominal state corresponds to individual data values ​​that deviate significantly from the distribution of normal data values. The set parameter (e.g., mean, spread or variance, or higher moments or cumulants) can represent the distribution of the nominal state.

[0128] Context-related anomalies: This non-nominal state corresponds to data values ​​that can only be classified as anomalies in a specific context (e.g., determined by the temporal progression of the recorded data values), while they would be classified as normal in a different context. The ability to detect context-dependent anomalies without having to define them beforehand is a major advantage of anomaly detection compared to rule-based evaluation of the measured values.

[0129] Collective anomalies: This non-nominal state corresponds to groups (i.e., combinations) of data values ​​that are only recognized as an anomaly due to their collective occurrence. The combined occurrence is to be evaluated as an anomaly, whereas individual such data values ​​within the group do not constitute an anomaly (e.g., not necessarily an anomaly).

[0130] One advantage of an embodiment of the present invention is that the classification depends on the set parameter, for example, implementing the anomaly detection algorithm in such a way that it can adapt to the prevailing conditions during operation. That is, the classification learns during operation what data values ​​of the nominal state (i.e., "normal" measured value inputs) look like. When the device 100 is first commissioned at the installation site, the algorithm has no information about the prevailing operating conditions there. However, depending on the anomaly detection algorithm used, it may be advantageous to use a pre-trained model. This means that the device 100 is already programmed during production with "basic knowledge" that reflects typical conditions and signal waveforms in electrical control cabinets.This reduces the need for input data and computing time for adaptation to local conditions.

[0131] Alternatively or additionally, the conditions in the control cabinet can initially be estimated (e.g., as basic knowledge) using expert knowledge and / or based on a parts list for the control cabinet. At least one parameter can be predetermined, e.g., by expert knowledge and / or according to an entry in a database, which is determined by the parts list for the control cabinet. In such cases, according to one embodiment of the invention, subsequent monitoring and corresponding parameterization, preferably during operation, are necessary.

[0132] Figure 3 shows an exemplary schematic flow diagram of a method 300 which can be carried out by an embodiment of the device 100 in the control cabinet 200 during the monitoring phase and the learning phase.

[0133] During the learning phase, device 302 (e.g., receives) 100 data values ​​from the sensor data and, in step 304, sets at least one parameter of a classification of nominal states of the control cabinet 200 based on the acquired data values ​​from the sensor data.

[0134] Simultaneously (if a previous parameter setting exists) or subsequently, in step 306, the device 100 determines, based on the acquired sensor data values, whether a current state of the control cabinet 200 is a nominal state according to the classification dependent on the at least one set parameter. During the learning phase, the device 100 can store the state and one or more associated classification parameters. During the learning phase, the device 100 can repeat setting 304 and, if necessary, update one or more classification parameters.

[0135] In a monitoring phase of the device 100, 306 the device 100 can determine, based on the received data values ​​of the sensor data, whether a current state of the control cabinet 200 is a nominal state according to the classification dependent on the at least one set parameter.

[0136] Normally, one or more parameters are not updated during the monitoring phase, unless the monitoring phase and the learning phase are active simultaneously.

[0137] In Figure 3, a partial process flow is shown with dashed lines when only the learning phase is active. A partial process flow flow when only the monitoring phase is active is shown with dotted lines.

[0138] Figure 4 shows an exemplary switch 110 that can be used in any embodiment of the device 100. The exemplary switch 110 of the device 100 is a mechanical switch 110. Each embodiment of the switch 100 can have at least two or all of the following three switching states: (exclusive) learning phase; (combined) learning and monitoring phase; and (exclusive) monitoring phase.

[0139] The phases of device 100 can be changed manually by the user and / or automatically (e.g., by software) controlled by a switching module. For example, the learning phase can be changed to a monitoring phase after a certain amount of training data and / or a certain decrease in the learning curve or classification accuracy, or another trigger. In another example, the monitoring phase can be changed to a learning phase after exceeding a certain number of errors and / or a certain monitoring time, or another trigger, or the learning phase can be activated in parallel with the monitoring phase.

[0140] Alternatively or additionally, the learning phase can be automatically terminated after a predefined period (e.g., one week, one year). This prevents errors from creeping into the nominal classification later on.

[0141] Alternatively or additionally, the learning phase can be automatically terminated depending on the learning progress (e.g., after processing a predetermined number of data values ​​or after reaching a predetermined value of a quality criterion or loss function).

[0142] The transition from the learning phase to the monitoring phase can be triggered by an external trigger, e.g. manual operation via a button or similar, by digital input or by a control command via a communication interface.

[0143] In one variant of the embodiment, the device 100 can be provided with a switch (push button, DIP switch, rotary switch, etc.) with which it can be switched between the learning and monitoring phases.

[0144] Figure 5 shows an exemplary schematic flow diagram of the switching between the phases of the device 100 in the control cabinet 200. This shows the special case of the method 300 with mutually exclusive learning and monitoring phases.

[0145] While in this special case a learning rate (i.e., a measure of the change in the at least one parameter, the classification, or its quality) is only different from zero during the learning phase, in one variant of each embodiment the learning rate can be constant over time, for example, for parallel operation of modules 104 and 106. Figure 6 schematically shows this case as a double-dotted line. Figure 6 also schematically shows a diagram for a variable learning rate according to an embodiment of the device 100 in the control cabinet 200. Figure 6 shows four different learning rate profiles as examples, which can vary depending on the implementation of step 304 (for example, depending on the learning algorithm used).

[0146] The learning rate can, for example, be purely time-controlled (dashed line in Figure 6).

[0147] The learning process can be controlled by varying the learning rate over time. A high learning rate can be used, particularly at the beginning of operation, to allow the algorithm to quickly adapt to the prevailing conditions. As learning progresses, the learning rate can be reduced to prevent the model from becoming overly reliant on gradually developing errors.

[0148] A particularly advantageous special case of the purely time-controlled learning rate is exponentially decreasing (dotted line in Figure 6).

[0149] Furthermore, the learning rate can be event-driven (solid line in Figure 6). For example, if the classification accuracy exceeds a first threshold, the learning rate can be continuously reduced. If the classification accuracy exceeds a second threshold that is higher than the first, the classification can remain unchanged.

[0150] All available measurements can be used for the learning process. Alternatively, only individual measurements can be selected. This reduces the computational power required for training the model. Measurement data can be selected according to a fixed grid (e.g., using every 100th measurement). Selection can also be based on machine learning models (e.g., clustering algorithms) with the goal of selecting the measurements that contain the most information for training progress.

[0151] In addition to the self-parameterizing algorithm, the device can perform 100 further evaluations of the measured values. These further evaluations may also require parameterization (e.g., comparison of the measured values ​​with lower and upper limits).

[0152] The algorithm (e.g., model) for anomaly detection can be based on a statistical model (e.g., moving average models, histogram-based models), a machine learning model (e.g., "knearest neighbor", DBSCAN, support vector machines, isolation forest, etc.), or a deep learning model (e.g., long-short-term memory models, recurrent neural networks, auto-encoders, generative adversarial networks, etc.).

[0153] The classification (for example, the model used) can be preset "at the factory" with parameters (e.g., based on training data) that reflect typical conditions and signal waveforms in control cabinets. The training data can be based on real measurements and / or simulations.

[0154] The predefined classification (e.g., the pre-trained model) can be compressed after training to reduce computation time and memory requirements when running it on Device 100. Various methods can be used to compress the pre-trained models, often referred to in the literature as "TinyML" (for "Tiny Machine Learning"). For example, the model parameters can be quantized (e.g., using 16-bit fixed-point values ​​instead of 32-bit floating-point values ​​per parameter). If neural networks are used, "pruning" can be employed (i.e., omitting edges and neurons whose influence on the model's output values ​​is negligible). Furthermore, knowledge distillation can be used.This approach involves first training larger, more complex models (for example, with a greater number of parameters) and then transferring the essential learned patterns to a smaller model. This makes it possible to execute procedure 300, in particular step 304, on a microcontroller.

[0155] Several model variants can be implemented, each pre-trained with different training data. After commissioning device 100, it can then be evaluated which model variant best suits the prevailing conditions and selected for evaluation.

[0156] The parameterization (i.e., the self-parameterization through step 304 of the setting) of the procedure 300, for example, the learning process after commissioning of the device 100, can be carried out in such a way that not all parameters of the classification (e.g., ML model) are adjusted, but only a subset of the parameters (e.g., in the case of a neural network, only individual layers). This can reduce the computing power required.

[0157] In one variant of each embodiment, for example as shown schematically in Fig. 5, the self-parameterization of the device 100 can be limited to a learning phase only. After the learning phase, the classification (e.g., the model) is frozen and the device 100 switches to pure monitoring mode.

[0158] The at least two data values ​​can comprise a time series of one or more measured variables.

[0159] Figure 7 schematically shows a field 700 of a combination of two sensor data 702, here using the example of the two measured quantities current and temperature. The data values ​​704 are shown in the field as a trajectory of the control cabinet's state. These data values ​​serve as training data during the learning phase.

[0160] The exemplary horizontally plotted data value of an operating sensor 202 in Figure 7 is a measured value of the current in amperes, and the exemplary vertically plotted data value is a measured value of an environmental sensor 202, namely the measured temperature in degrees Celsius. This can correspond to step 302 of the method.

[0161] As can be seen in Figure 7, the data values ​​704 imply sub-regions of the field 700 as nominal states. Figure 8 shows an example of a state classification as a result of the training phase. The exemplary classification algorithm distinguishes 5 patterns of acceptable regions (e.g., clusters) in the two-dimensional field 700 of the training data as nominal states 802. This can correspond to step 304 of the procedure.

[0162] Figure 9 schematically illustrates the application of condition classification during the monitoring phase. This can correspond to step 306 of the procedure.

[0163] Figure 9 shows an example of how the current state is classified as acceptable (nominal state 802) or unacceptable (error state 902) depending on the recorded data values, if the data values ​​lie in a range outside the nominal state.

[0164] A resource-efficient implementation of modules 104 and 106 of the device 100 enables the execution of method 300, in particular step 304 of setting the at least one parameter of the classification, on a microcontroller. Variants of an exemplary implementation are described below:

[0165] Hardware integration of the Data Acquisition Module 102 can include:

[0166] • Sensor connection: Connecting sensors to the microcontroller via analog inputs or digital interfaces such as l 2 C, SPI or UART.

[0167] • Examples of sensors: temperature, humidity, current, voltage and vibration sensors.

[0168] A software implementation of the data acquisition module 102 may include:

[0169] • Data acquisition loop: Periodic reading of sensor values ​​at defined time intervals.

[0170] • Preprocessing: Optional filtering or normalizing of the sensor data to reduce noise and prepare the data for the ML model.

[0171] • Buffer memory: Storage of current and previous data values ​​in a ring buffer to account for temporal relationships. To execute the parameterization module 104 resource-efficiently on a microcontroller of the device 100, one of the following techniques can be used in step 304.

[0172] • Moving statistical models: o Calculation of the moving mean and standard deviation for each sensor data set. o Threshold-based detection: Setting tolerance limits (e.g., mean ± n standard deviation) that serve as classification parameters.

[0173] • Incremental k-means clustering: o Real-time clustering of data points to identify typical operating states. o Adjustment of cluster centers for new data points with low computational overhead.

[0174] • Simple, low-complexity autoencoder: o Neural network with a hidden layer, trained to reconstruct input data. o Anomaly detection through reconstruction errors: High error rates indicate unknown (abnormal) states.

[0175] The resource requirements of the microcontroller, which preferably executes both modules 104 and 106, can be reduced by any one of the following techniques:

[0176] • Model quantization:

[0177] Reducing the data width of parameters and calculations to 8-bit or 16-bit integers to reduce storage requirements and computing power.

[0178] • Pruning of models:

[0179] Removing unimportant weights in neural networks to reduce the number of calculations.

[0180] • Use of fixed-point arithmetic:

[0181] Avoidance of floating-point operations in favor of faster integer calculations.

[0182] • Use of "TensorFlow Lite for Microcontrollers": o Lightweight ML framework designed for microcontrollers with few kilobytes of RAM. o Support for the algorithms and optimizations mentioned.

[0183] • Memory management: o Limited storage of historical data, e.g., through fixed-size ring buffers. o Efficient data structures to minimize RAM consumption.

[0184] • Learning rate and learning phase: o Dynamic adjustment of the learning rate: High learning rate at the beginning, which decreases during the learning phase. o Phase control: Switching between the learning and monitoring phases according to defined criteria (e.g., time elapsed or stabilization of the model parameters).

[0185] An implementation of the monitoring module 106 can be used for a

[0186] Anomaly detection includes the following sub-steps of step 306:

[0187] • Real-time evaluation of the sensor data based on the parameters (e.g., of the model) of the classification set (i.e., "learned") in parameterization module 104.

[0188] • Calculation of deviations between current data values ​​and expected nominal values.

[0189] A decision logic for whether the nominal state exists or not can include the following sub-steps of step 306.

[0190] • Threshold check: Comparison of deviations with predefined limits.

[0191] • Classification of the current state: Determining whether the state is nominal or abnormal.

[0192] Output module 108 can be used in response to the detection of an anomaly.

[0193] (i.e., a non-nominal state) perform one of the following steps.

[0194] • Alarm triggering: Activating LEDs, buzzers, or other signaling devices. • Communication: Optional sending of warning messages via interfaces such as UART (Universal Asynchronous Receiver Transmitter), CAN (Controller Area Network), PROFINET (Process Field Network), or wireless connections (e.g., Bluetooth Low Energy).

[0195] In each embodiment of the device 100, a suitable microcontroller can be used as the hardware platform, e.g.:

[0196] • ARM Cortex-M4 or Cortex-M7 with DSP support.

[0197] • Espressif ESP32 for additional WLAN / Bluetooth functionality during signal output.

[0198] To improve the energy efficiency of device 100 and reduce waste heat in control cabinet 200, a power-saving mode can be used. The microcontroller can be put into a sleep mode for microseconds when none of the modules 102, 104, and 106 require processing. Alternatively or additionally, the microcontroller can be operated at a lower clock frequency to save energy when power is sufficient. For example, the clock frequency and the learning rate can be controlled inversely.

[0199] A user interface of the device 100 can include the switch 110 for phase control, for example a physical push button to switch between learning and monitoring phase and / or LED indicators to provide feedback on the current operating status and / or phase.

[0200] The device, especially module 104, can be efficiently programmed (for example, in C or C++) without unnecessary abstraction layers. Data acquisition for module 102 can be interrupt-driven. Sensors 202 only transmit data values ​​when relevant changes occur. In each module, the calculation of mathematical functions can be accelerated by using lookup tables, i.e., pre-calculated values. Reference symbol list

[0201] 100 Device

[0202] 102 Data Acquisition Module

[0203] 104 Parameter module

[0204] 106 Monitoring module

[0205] 108 Output module

[0206] 110 switches or switching modules

[0207] 200 control cabinet

[0208] 202 Sensor, for example operating sensor or environmental sensor

[0209] 300 procedures

[0210] 302 Recording Data Values

[0211] 304 Setting the parameters of a classification

[0212] 306 Determine whether condition is nominal according to classification

[0213] 700 Field of the recorded at least two data values

[0214] 702 sensor data

[0215] 704 data values

[0216] 802 Nominal state, e.g. permissible state

[0217] 902 Non-nominal state, e.g. anomaly or fault state

Claims

1. Device (100) in a control cabinet (200) for monitoring the control cabinet (200), comprising: a data acquisition module (102) configured to acquire at least two data values ​​from sensor data of the control cabinet (200); a parameterization module (104) configured to set at least one parameter of a classification of nominal states of the control cabinet (200) during a learning phase of the device (100) based on the acquired data values ​​of the sensor data, wherein the classification is implemented in the device (100); a switch (110) or a switching module (110) configured to change a rate of setting the at least one parameter;and a monitoring module (106) configured to determine, in a monitoring phase of the device (100), based on the acquired data values ​​of the sensor data, whether a current state of the control cabinet (200) is a nominal state according to the classification dependent on the at least one set parameter.

2. Device (100) according to claim 1, wherein the at least two data values ​​of the sensor data comprise at least one data value of environmental data of the control cabinet (200) and / or at least one data value of operating data of the control cabinet (200).

3. Device (100) according to claim 1 or 2, wherein the parameterization module (104) is configured for repeatedly setting the at least one parameter, and / or wherein the monitoring module (106) is configured to repeatedly determine whether the current state of the control cabinet (200) is a nominal state, and / or wherein the parameterization module (104) and the monitoring module (106) are configured to perform the setting of the at least one parameter and the determination of whether the current state of the control cabinet (200) is a nominal state alternately or simultaneously.

4. Device (100) according to any one of claims 1 to 3, wherein the parameterization module (104) and / or the monitoring module (106) performs a moving average of the data values ​​and / or histogram-based statistics; or wherein the parameterization module (104) performs one of the following machine learning (ML) methods for setting the at least one parameter, and the monitoring module (106) performs the classification resulting from the ML method for determining the following: Nearest neighbor classification, density-based spatial cluster analysis with noise, Support vector machine, Isolation forest, or Deep learning methods for a neural network model, optionally for a long- and short-term memory model, for a recurrent neural network, for an auto-encoder, or for a generative-adversarial network.

5. Device (100) according to one of claims 1 to 4, wherein the switch (110) or the switching module (110) is further configured to switch on the learning phase or to switch between the learning phase and the monitoring phase.

6. Device (100) according to one of claims 1 to 5, wherein the rate of setting the at least one parameter is different from zero only during the learning phase.

7. Device (100) according to claims 1 to 6, wherein the monitoring module (106) is configured to determine the current state according to the classification based on a combination of the acquired data values ​​of the sensor data.

8. Device (100) according to any one of claims 1 to 7, wherein the operating data comprises at least two data values, optionally wherein the at least two data values ​​of the operating data each comprise timestamps or the data values ​​of the operating data comprise a time series, and / or wherein the operating data are measured variables acquired within the control cabinet (200), and / or wherein the operating data comprise at least one of the following measured variables of the control cabinet (200): a current through a component of the control cabinet (200), a power input of the control cabinet (200), a power output of the control cabinet (200), and an open and / or closed position of a door or inspection opening of the control cabinet (200).

9. Device (100) according to any one of claims 1 to 8, wherein the environmental data comprises at least two data values, optionally wherein the at least two data values ​​of the environmental data each including timestamps or the environmental data values ​​comprise a time series, and / or wherein the environmental data are measured quantities acquired outside the control cabinet (200), and / or wherein the environmental data include at least one of the following measured quantities of the control cabinet (200): an air temperature inside or outside the control cabinet (200), a relative humidity inside or outside the control cabinet (200), a concentration of smoke particles in the air inside the control cabinet (200), an oxygen content, nitrogen content or carbon dioxide content in the air inside the control cabinet (200), an overpressure inside the control cabinet (200), and a noise level or mechanical vibrations inside or outside the control cabinet (200).

10. Device (100) according to one of claims 8 or 9, wherein the at least two data values ​​of the operating data or environmental data have different timestamps and / or are acquired from different data sources and / or indicate different physical measured quantities.

11. Device (100) according to one of claims 1 to 10, wherein the device further comprises an output module (108) configured to output the determined current state, the received data values, the set at least one parameter and / or a deviation from the nominal state.

12. Control cabinet (200), comprising: at least one sensor (202) configured to output at least two data values ​​from sensor data of the control cabinet (200); and Device (100) according to one of claims 1 to 11 for monitoring the control cabinet (200), which is configured to determine, based on the output data values ​​of the sensor data, whether a current state of the control cabinet (200) is a nominal state.

13. Method (300) for monitoring a control cabinet (200), comprising: Acquiring (302) at least two data values ​​from sensor data of the control cabinet (200); in a learning phase of the device (100), setting (304) at least one parameter of a classification of nominal states of the control cabinet (200) based on the acquired data values ​​of the sensor data, wherein the classification is implemented in the device (100); Providing a switch (110) or a switching module (110) configured to change a rate of setting (304) of the at least one parameter; and in a monitoring phase of the device (100), determining (306) based on the acquired data values ​​of the sensor data whether a current state of the control cabinet (200) is a nominal state according to the classification dependent on the at least one set parameter.

Images