DE102023207829B4 - Method and system for detecting anomalies in a set of signals - Google Patents

Abstract

Method for detecting anomalies in a set of signals, wherein the signal represents a quantity of a technical system to be checked, and wherein the method comprises:
- Providing a model (10) for mapping a fixed number of signals in a past time window to an output target, wherein the output target maps the signals in a future time step as future signals (ŷ), and providing a static limit (α) with respect to existing residuals (R) of the predicted future signals (ŷ) by the model (10),
- Capturing a set of current signals as well as the parameter values currently influencing the signal to be checked, from at least one measurement source of a technical system,
- Input of the current signals into the model (10) over the past time window to output a predicted future signal (ŷ),
- Generating a dynamic dynamic value (β), wherein the dynamic dynamic value (β) depends on a variation range of the current signals and/or a variation range of the currently influencing parameter sizes over the past time window, wherein the dynamic dynamic value (β) is set higher the more the signals and/or the influencing parameter sizes vary over the time window,
- Generating a dynamic limit value (α _n ) based on the dynamic value (β) and the static limit value (α),
- Determining a current residual (R) based on the predicted future signal (ŷ),
- Generating an anomaly alarm when the current residual (R) of the future signal (ŷ) predicted by the model (10) exceeds the dynamic limit (α _n ).

Description

The invention relates to a method and a device for detecting anomalies in a set of signals, wherein the signal represents a quantity of a technical system to be checked.

An industrial plant process in an automated system comprises a multitude of individual, interacting subprocesses. High productivity requires a reliable process flow. To achieve this, the industrial plant and its components undergo regular maintenance and servicing.

The DE 10 2012 009 657 B3 discloses a method and a system for detecting and identifying an oscillatory fault in a flight control system of an aircraft. DE 103 55 022 A1 discloses a method for monitoring a technical system. DE 199 398 72 A1 discloses a method and a device for sensor monitoring, in particular for an ESP system for vehicles. DE 10 2016 117 190 A1 discloses a method and a device for monitoring the state of at least one wind turbine and a computer program product

These maintenance works are usually carried out at regular intervals. Such work involves, for example, replacing components or even entire machine subsystems.

Despite these efforts, there is still a risk of disruptions occurring that could lead to production outages. Due to the complexity of the interlocking subprocesses, predicting potential future errors is difficult.

This often means that an impending failure of a critical system cannot be detected early enough. As a result, the system or its components can fail during operation.

Systems such as wind turbines use sensors to monitor their components for anomalies during operation. Each wind turbine within a wind farm is equipped with sensors that measure various parameters related to its operation, aiming to minimize premature failures and damage by identifying anomalies in the sensor data.

A wind turbine may need to be shut down to repair damage. This downtime prevents the turbine from producing energy. Therefore, it is a constant task to develop improved methods for predicting potential anomalies to prevent major damage. The same applies to other technical systems.

These monitoring devices/predictive methods should be suitable for recognizing the current condition of components/systems and, if necessary, providing indications of emerging or existing damage.

The DE 102016117190 A1 Disclosing a method for monitoring the state of at least one wind turbine, comprising: acquiring first measurement signals by one or more sensors, wherein the first measurement signals indicate one or more parameters relating to at least one rotor blade of the at least one wind turbine in a normal state; training a learning algorithm based on the first measurement signals of the normal state; acquiring second measurement signals by the one or more sensors; and detecting an indeterminate anomaly by the learning algorithm trained in the normal state when a current state of the wind turbine, determined based on the second measurement signals, deviates from the normal state.

It is therefore an object of the invention to provide a method and a device for detecting anomalies in a set of signals originating from a technical system as the source. Furthermore, it is an object to provide a generator with a gearbox and a corresponding method or device.

The problem is solved by a method having the features of claim 1 and a device having the features of claim 11. Furthermore, the problem is solved by a generator having the features of claim 13 and a wind turbine having the features of claim 15.

The dependent claims list further advantageous measures which can be suitably combined to achieve further advantages.

• - Bereitstellen eines Modells zur Abbildung von einer festen Anzahl von in einem zurückliegenden Zeitfenster liegenden Signalen auf ein Ausgabeziel, wobei das Ausgabeziel die Signale in einem zukünftigen Zeitschritt als Zukunftssignale abbildet sowie bereitstellen eines statischen Grenzwertes in Bezug auf vorhandene Vorhersagefehler (Residuen) der vorhergesagten Zukunftssignale durch das Modell,
• - Erfassen eines Satzes von aktuellen Signalen als auch von, dass zu überprüfende Signal, aktuell beeinflussenden Parametergrößen aus zumindest einer Messquelle eines technischen Systems,
• - Eingabe der aktuellen Signale in das Modell über das zurückliegende Zeitfenster zur Ausgabe eines vorhergesagten Zukunftssignals,
• - Generieren eines dynamischen Dynamikwertes, wobei der dynamische Dynamikwert von einer Variationsbreite der aktuellen Signale und/oder einer Variationsbreite der aktuell beeinflussenden Parametergrößen über das zurückliegende Zeitfenster abhängt, wobei der dynamische Dynamikwert umso höher gesetzt wird, je höher die Signale und /oder die beeinflussenden Parametergrößen über das Zeitfenster variieren,
• - Generieren eines dynamischen Grenzwertes anhand des Dynamikwertes und des statischen Grenzwertes,
• - Ermitteln von einem aktuellen Residuum anhand des vorhergesagten Zukunftssignals,
• - Generieren eines Anomalienalarms, wenn das aktuelle Residuum des durch das Modell vorhergesagten Zukunftssignals den dynamischen Grenzwert überschreitet.

The problem is solved by a method for detecting anomalies in a set of signals, where the signal represents a quantity of a technical system to be checked, and the method comprises:

• - Providing a model for mapping a fixed number of signals in a past time window to an output target, wherein the output target maps the signals in a future time step as future signals, and providing a static limit value with respect to existing prediction errors (residuals) of the predicted future signals by the model,
• - Capturing a set of current signals as well as the parameter values currently influencing the signal to be checked, from at least one measurement source of a technical system,
• - Input of current signals into the model over the past time window to output a predicted future signal,
• - Generating a dynamic dynamic value, wherein the dynamic dynamic value depends on a range of variation of the current signals and/or a range of variation of the currently influencing parameter sizes over the past time window, whereby the dynamic dynamic value is set higher the greater the variation of the signals and/or the influencing parameter sizes over the time window,
• - Generating a dynamic limit value based on the dynamic value and the static limit value,
• - Determining a current residual based on the predicted future signal,
• - Generating an anomaly alert when the current residual of the future signal predicted by the model exceeds the dynamic threshold.

An anomaly alarm is a deviation alarm that indicates anomalies.

The residuals are determined based on the measured signals and the correlating future signals determined by the model.

The signal to be checked can also be a group of signals. Furthermore, the dynamic value and the dynamic limit value can be determined at discrete time intervals or continuously.

The currently influencing parameters, for example, in predicting the temperature in a wind turbine, could be the speed. These influencing parameters must be defined in advance with respect to the future signal(s) to be predicted.

The technical system can be a single component, an arrangement of interacting components, or an entire plant.

The signals/parameter sizes are usually detected by one or more sensors.

In particular, the signals can be formed as a sequence of discrete values.

Signals can be, for example, the parameters that characterize the system/component. Signals can refer to parameters such as temperature, rotational speed, torque, etc., depending on the component or system.

The anomaly alert can be displayed and saved, for example, so that a deviation graph can be created over time.

The model, for example a prediction model, can be stored on a computer on which the procedure is carried out and the signals are transmitted.

A time window, for example, is a period of time over a specific period.

An anomaly or outlier refers to a data point whose properties deviate significantly from the norm and often represent the (future) failure of the corresponding component/system or provide insights into the expected service life.

Residuals can be viewed as the difference between the actual observed signals and the future signals predicted by the model.

The static limit value is set once in relation to the signal, for example, based on the distribution of the prediction errors (residuals).

In model-based anomaly detection, a specific parameter is monitored (vibration, temperature, etc.).

A residual value is accepted and considered normal if it is below a predefined limit; otherwise, a warning is triggered. According to the invention, it was recognized that it is disadvantageous if the limit is statically defined and the effects of dynamics in the technical system under test are not taken into account.

This is now solved using the method according to the invention. A dynamic dynamic value is generated, wherein the dynamic dynamic value depends on the range of variation of the current signals and/or the corresponding parameter values over the past time window. Furthermore, a dynamic limit value is generated based on the dynamic value, for example by multiplying the dynamic value by the static limit value. The dynamic dynamic value thus takes into account the current dynamics of the system in the defined past time window.

The dynamic limit is a scalar value.

This means that with nearly constant signals/parameter sizes (low variance, i.e., low dynamics in the system), the dynamic limit value or dynamic value can be chosen to be low, for example, it can have a low value, and with strongly fluctuating signals/parameter sizes (high variance/high dynamics in the system), the dynamic value is higher the more the signals/parameter sizes vary over the time window.

This enables faster detection of anomalies. This means that the fault detection time, i.e., the point in time at which a fault is detected, is improved. Compared to the static limit value, this time is significantly shifted forward. The time required by the inventive method to detect an (incipient) failure/damage is thus considerably reduced, which impacts the overall performance of the system/component. The component in question can now, for example, be replaced during maintenance before damage or a short-term failure occurs.

The method according to the invention thus takes into account the dynamics of the input signals or other system variables.

The method according to the invention thus makes it possible to detect anomalies before they endanger the availability of the component / technical system as real malfunctions.

The inventive method allows for significantly lower threshold values for anomaly detection during phases in which the signals or input signals do not change or change only slightly (low dynamic range). This enables much faster anomaly detection. Conversely, a higher threshold value is accepted during phases in which the input signals change significantly (high dynamic range).

In further development, the model is designed as a machine learning function approximated by historical signals. Based on signals that characterize a normal state of the technical system, the machine learning function is trained to map to a future signal using a fixed number of signals from the past time window. Thus, the change in the signal with respect to a future time step can be predicted by inputting past signals.

Furthermore, the model is trained as an artificial intelligence method, specifically an artificial neural network. This allows for the generation of a good predictive model. This also enables the consideration of noisy signals. In particular, recurrent neural networks are used because of their ability to learn long- and short-term dependencies, which require little training effort.

In particular, the model is trained using signals without anomalies. These can be real-world data or simulation data.

In further training, the anomaly alarm is only triggered after a predefined number of consecutive exceedances. "Consecutive" can also mean that these exceedances occur within a specified time period. This prevents, for example, a false alarm from being triggered by a single faulty sensor reading/evaluation.

In further development, the dynamic value assumes a scalar value between a value c and 1 (one), where 1 (one) is the maximum value and c is greater than zero. The dynamic limit is then calculated by multiplying the static limit and the dynamic value. This development prevents the dynamic limit from being set to zero.

In further training, the dynamic value is set to one if a predetermined variation range of the signals and/or parameter values is exceeded. This means that the static limit value is used if the signals fluctuate by a certain amount.

In further development, the dynamic value is configured to assume only discrete values in the range between 0 and 1 (one) as its maximum value, depending on the variation range of the signals and/or the parameter sizes. This allows for different variation ranges, i.e., Strong fluctuations can be simplified and assigned to a dynamic value.

Further training involves using a statistical method or model from the fields of correlation analysis, regression analysis, simulation model, experimental method, or sensitivity analysis to determine the range of variation of the signals and/or parameter values within the given time window. This allows for a good representation of the dynamics of the input signals/parameter values.

The statistical method can be selected depending on the signal being tested.

In further training, the signals are transmitted from a technical system as the source. This could be, for example, a component/system etc. of a vehicle, wind turbine, or industrial plant.

To improve the accuracy of the future signals predicted by the model, the current signals can, for example, be preprocessed. Preprocessing can, for instance, reduce noise in the signals, which can affect the accuracy of the model's predictions.

Furthermore, the problem is solved by a device for detecting anomalies in a set of signals, wherein the signal represents a quantity of a technical system to be checked, wherein in particular the signal is designed as a sequence of discrete values, wherein a storage unit with a model is provided, wherein the model is configured to map a fixed number of signals lying in a past time window to an output target, wherein the output target accomplishes a mapping of the signals in a future time step as future signals, wherein furthermore a static limit value is provided in the storage unit, which is defined by the model with respect to existing residuals of the predicted future signals, wherein furthermore a computing unit is provided, which is configured to acquire a set of current signals as well as parameter quantities currently influencing the signal to be checked from at least one measurement source.
wherein the computing unit is further designed to input the current signals into the model over the past time window and to output a predicted future signal,
wherein the computing unit is configured to generate a dynamic dynamic value, wherein the dynamic dynamic value depends on a variation range of the signals and/or a variation range of the currently influencing parameter sizes over the past time window, wherein the dynamic dynamic value is set higher the more the signals and/or the influencing parameter sizes vary over the time window, and wherein the computing unit is configured to generate a dynamic limit value based on the dynamic value and the static limit value.
comprising a comparison unit for determining a current residual based on the predicted future signal and for generating an anomaly alarm when the current residual of the future signal predicted by the model exceeds the dynamic limit.

An anomaly alarm indicates an anomaly.

The device is specifically designed to carry out the process. In particular, all advantages or advantageous designs can also be transferred to the device.

A computing unit can be, for example, a computer with the appropriate software. The comparator can be a software module.

Further development can include a display unit to indicate when a limit is exceeded and an anomaly alarm occurs. This can, for example, be displayed as a graph over time, allowing for the monitoring of anomalies and enabling intervention at an appropriate time, such as by replacing a component.

Furthermore, the problem is solved by a generator with a gearbox, the generator comprising a method and/or a device as described above, wherein the signals are characteristic parameters of the gearbox or the generator.

Furthermore, the gearbox can have at least one bearing, with the signals including at least one bearing temperature.

Furthermore, the task is solved by a wind turbine with a generator as described above.

The gearbox's function is to convert the relatively slow speed of the main shaft into the higher speed required by the generator to produce electrical energy. The constantly changing wind speed and the torque acting on the wind turbine These factors lead to dynamic temperature behavior in the gearbox bearings.

A wind turbine incorporating a generator with a method/device according to the invention can thus quickly detect anomalies in the temperature of bearings in a wind turbine gearbox/generator. This is necessary because, from the moment a failure begins, the condition of the system deteriorates until mechanical damage or a functional impairment occurs.

Such a wind turbine can prevent damage to the gearbox. Further features and advantages of the present invention will become apparent from the following description with reference to the accompanying figures.

• 1: ein erfindungsgemäßes Verfahren,
• 2: ein Generator einer Windkraftanlage,
• 3: die Variation /Schwankungen von zwei Parametergrößen,
• 4: ein Diagramm mit einem statischen Grenzwert,
• 5: die grafische Darstellung eines Fehlerprozesses,
• 6: zeigt das erfindungsgemäße Verfahren im Allgemeinen.

The figures show schematically:

• 1 : a method according to the invention,
• 2 : a generator of a wind turbine,
• 3 : the variation/fluctuations of two parameter sizes,
• 4 : a diagram with a static limit value,
• 5 : the graphical representation of an error process,
• 6 : shows the method according to the invention in general.

1 The invention shows a method for detecting anomalies in a set of signals originating from a technical system as the source. The signals are configured as a sequence of discrete values, which were, for example, recorded by a sensor of a technical system/component etc. to be tested.

In a first step S1, a model 10 ( 6 ) provided in relation to a signal to be checked.

The Model 10 ( 6 The function is designed to map a fixed number of signals from a past time window to an output target, where the output target represents a mapping of the signals in a future time step as future signals ŷ. The signals could, for example, be the temperature or vibration of a bearing in a wind turbine or a wind turbine gearbox 4, which are to be monitored.

The Model 10 ( 6 ) was trained using historical observations regarding the relevant parameter or signal in order to predict the value of the relevant signal in the next future step. For this purpose, model 10 ( 6 ) as a learning function approximated by historical signals, designed to perform a mapping to a future signal ŷ based on the fixed number of signals in the past time window.

This involves the model 10 ( 6 ) a representational learning function based on signals that characterize a normal state of the technical system, trained to create a mapping to a future signal based on the fixed number of signals in the past time window.

Thus, the model 10 ( 6 ) as a predictive model to predict the signals in a future time step. For this purpose, model 10 ( 6 ) is trained exclusively on normal signal data (without anomalies) and learns which values occur under normal conditions. Model 10 ( 6 ) thus predicts a future signal ŷ based on input signals that reflect the operation of the component/technical system or, in this case, the bearing in its normal state.

This allows the warehouse to be monitored.

Furthermore, the Model 10 ( 6 ) Prediction errors (residuals) are generated with respect to the predicted future signals.

A static limit value α can now be determined based on these residuals R.

The model 10 ( 6 ) as an artificial neural network. This allows even noisy signals to be taken into account. In particular, recurrent neural networks are used because of their ability to learn long- and short-term dependencies, which require less training effort.

In a second step S2, a set of current signals is acquired as input signals from at least one measurement source of the technical system. These current signals lie within a past time window, which is also the basis for model 10 ( 6 ) has been trained.

The signal represents a parameter of the technical system that needs to be checked. For example, the signal could be the temperature of a bearing in a wind turbine, particularly a gearbox. Model 10 ( 6 ) is trained based on the normal state with respect to the parameter to be checked, in this case the temperature in the bearing. Furthermore, it can also The parameter sizes influencing the signal to be checked are recorded, if any exist.

In a third step S3, the current signals are used as input signals for model 10 ( 6 ) about the past time window to predict future signals. This allows model 10 to act ( 6 ) as a predictive model; the future signals ŷ represent the predicted signals or their values at a predicted time step. Model 10 ( 6 ) reflects the future signals ŷ in normal operation, according to the training.

In a fourth step S4, a dynamic limit value α _n is generated based on a dynamic value β and the static limit value α.

The dynamic value β lies in the range between c and 1 (one). The dynamic limit α _n is formed by multiplying the static limit α and the dynamic value β, such that for a fixed time duration (t, t + m), where t equals time, the following holds: $${\alpha }_n=\beta *\alpha ,\left(c\le \beta \le 1\right),$$ with α as a static limit
where β is restricted downwards by the constant c and is set, for example, to a value that reflects the model inaccuracy.

In this case, c must not be set to zero, so that α _n is always greater than zero.

The dynamic value β describes the dynamics of the elements in model 10 ( 6 ) currently incoming input signals as well as the parameter sizes or signal sizes influencing the predicted future signal ŷ, i.e. the fluctuations of those signals in the past time window which are included/taken into account in the predicted future signals ŷ.

In those phases in which the input signals do not change or the influencing parameter sizes do not change (low dynamics, low fluctuations), the dynamic value β and thus the dynamic limit value α _n is set low.

This means that the dynamic value β is closer to or near c, thereby reducing the dynamic threshold α _n accordingly. This allows for significantly faster detection of anomalies in the signals, which may indicate a malfunction, damage, or incipient damage to the component or system generating the signals.

Thus, the dynamic value β represents the actual dynamics of the system/component. The parameter c is a fixed lower limit.

In phases where the input signals or the influencing parameter values change significantly in the preceding time window (high dynamics, strong fluctuations), a larger dynamic value β and thus a higher dynamic limit α_{_n} is accepted. The dynamic limit α _{_n} can then equal the static limit α.

The dynamic value β can take on discrete scalar values between c and 1, where 1 is the maximum value and the minimum value is greater than 0. Once a predetermined fluctuation in the signals is reached, the dynamic value β is set to one. This means that once a certain fluctuation in the signals/parameter values is reached, the static limit α is applied.

This allows different ranges of variation, i.e., fluctuations of varying strength, to be assigned to a single dynamic value β in a simplified manner.

The determination of the range of variation, i.e., the fluctuations of the signals within the given time window, can be achieved using a statistical method or model from the fields of correlation analysis, regression analysis, simulation modeling, experimental methods, or sensitivity analysis. This allows for a good representation of the dynamics of the input signals. The statistical method can be selected depending on the signal being analyzed.

Furthermore, in a fifth step S5, an anomaly alarm is generated if the residuals R of the model 10 ( 6 ) predicted future signals exceed the dynamic threshold α _n , with an anomaly alarm indicating anomalies.

In this system, an anomaly alarm is only triggered after a predefined number of consecutive exceedances. "Consecutive" can also mean that these exceedances occur within a specified time period. This prevents, for example, a false alarm from being triggered due to a faulty sensor reading or incorrect evaluation.

To improve the accuracy of the predicted future signals ŷ by model 10 ( 6 For example, the signals can be preprocessed. Preprocessing can, for instance, reduce noise in the Signals are reduced, which affects the accuracy of the prediction by model 10 ( 6 ) can have an effect.

2 shows a generator 5 of a wind turbine 1 in connection with an application of the method.

In a wind turbine, the kinetic energy of the wind is first converted into mechanical and then into electrical energy.

The wind rotates the blades 6 around a rotor 2, which is connected to the main shaft 3. The function of the gearbox 4 is to convert the relatively slow speed of the main shaft 3 into the higher speed required by the generator to produce electrical energy. The constantly changing wind speed and the torque acting on the wind turbine 1 result in dynamic temperature behavior in the bearings of the gearbox bearing 7 and/or the rotor bearing 8.

The wind turbine 1 comprises a device 9 according to the invention with a method according to the invention. This has a model 10 ( 6 ) which has been trained on corresponding input signals from wind turbine 1.

3 shows the variation/fluctuations of two parameter sizes relating to the signal to be checked, here the temperature of the gearbox bearing 7 of the wind turbine 1 over time in a diagram.

The parameters are the velocity G and the load L acting on wind turbine 1. The velocity G (top diagram) and load L (bottom diagram) are shown over time in a diagram.

Furthermore, a diagram below shows the dynamic value β in percent. This varies depending on the fluctuation of the input signals, here the velocity G and load L.

The dynamic value β describes the dynamics of the elements in model 10 ( 6 ) currently incoming input signals and parameter sizes, i.e. the fluctuations of those signals in the past time window which are included in the predicted future signals ŷ (temperature of the bearing) or which have an influence on the future signal ŷ (temperature of the bearing) to be predicted.

The fourth diagram shows the residuals R of the predicted future signals ŷ (temperature of the bearing) in relation to the actual measured temperatures over time.

Furthermore, the diagram shows the dynamic limit α _{_n} , which is obtained by multiplying the static limit α by the dynamic value β. If the residuals R are below the dynamic limit α _{_n} , no anomaly alarm is triggered.

This means that the following applies: $$y\le \widehat{y}\pm {\alpha }_n$$ with ŷ the through the model 10 ( 6 ) predicted future signals, y the corresponding measured actual signals and α _n the dynamic limit.

It can be seen that in those phases in which the input signals or the influencing parameter sizes do not change (low dynamics, low fluctuations), the dynamic value β and thus the dynamic limit value α _n for the detection of anomalies is set significantly lower.

This means that the dynamic value β is closer to or near c, thereby reducing the dynamic threshold α_{_n} accordingly. This allows for significantly faster detection of signal anomalies that indicate a malfunction, damage, or incipient damage to the component or system generating the signals. This enables earlier detection of signal anomalies.

In phases where the input signals or the influencing parameter values change significantly in the preceding time window (high dynamics, strong fluctuations), a larger dynamic value β and thus a higher dynamic limit α_{_n} is accepted. The dynamic limit α _{_n} can then equal the static limit α.

Furthermore, a stricter threshold α _{_n,w} can be generated. This can be achieved by adding a constant to the dynamic threshold α _{_n} . An alarm is triggered, for example, if several consecutive future signals exceed the dynamic threshold α _{_n} . This can be done, for example, by displaying and highlighting the signal on a display unit.

For example, if the stricter limit value α _n,W is exceeded, even a single exceedance can trigger an alarm.

In contrast, in 4 The static limit value α and a stricter static limit value _αW are displayed. This does not change, even under high dynamic conditions, for example, with regard to wind turbine 1. Anomalies are therefore only detected later.

5 Figure 1 shows a graphical representation of a fault process. Both the method and the device 9 serve to detect anomalies in the signals, for example, the temperature signals of the bearings. This can prevent damage to the gearbox, for instance. Timely fault detection is crucial for this. Therefore, the point at which a fault is detected should be as early as possible. From point P_a, at which the failure begins, the condition of the system deteriorates until point P_v, at which mechanical damage or a functional impairment occurs.

The inventive method and device 9, together with the dynamic threshold _αn, enable significantly faster fault detection at an earlier time P_e. The inventive method and device 9 take into account the current dynamics of the system in order to adjust the dynamic threshold _αn . The lower dynamic threshold _αn allows for faster detection of incipient failures.

The method and device 9 can be applied to any technical system/arrangement or component which is connected, for example, to a sensor/actuator for signal generation.

6 Figure 9 shows the method and device according to the invention in general. Here, _x1 , ..., _xn represent the signals lying in a past time window with respect to the signal to be checked and the parameter sizes influencing the signal to be checked from at least one measuring source of a technical system, which are in particular detected by sensors.

The input signals and parameter sizes are used to feed into the provided model 10 and into the generation of a dynamic limit value α _n .

Based on the provided model 10, the future signals ŷ are now predicted.

If the comparison of the predicted future signal ŷ with the actually measured signal y is below the calculated dynamic limit α _n, no alarm is triggered.

If this value is above the dynamic threshold α _n , an alarm is triggered.

Reference symbol list

1

Wind turbine

2

rotor

3

Main wave

4

transmission

5

generator

6

wing

7

gearbox bearing

8

Rotor bearings

9

device

10

Model

β

Dynamic value

an

dynamic limit

α

static limit value

ŷ

Future signals

R

residuum

y

measured signal

Claims (15)

A method for detecting anomalies in a set of signals, wherein the signal represents a quantity of a technical system to be checked, the method comprising: - providing a model (10) for mapping a fixed number of signals in a past time window to an output target, wherein the output target maps the signals in a future time step as future signals (ŷ), and providing a static limit value (α) with respect to existing residuals (R) of the predicted future signals (ŷ) by the model (10); - acquiring a set of current signals as well as of parameter quantities currently influencing the signal to be checked from at least one measurement source of a technical system; - inputting the current signals into the model (10) over the past time window to output a predicted future signal (ŷ); - generating a dynamic dynamic value (β), wherein the dynamic dynamic value (β) is determined by a range of variation of the current signals and/or a range of variation of the currently influencing parameter sizes depend on the past time window, whereby the dynamic dynamic value (β) is set higher the more the signals and/or the influencing parameter sizes vary over the time window, - Generating a dynamic limit value (α _n ) based on the dynamic value (β) and the static limit value (α ), - Determining a current residual (R) based on the predicted future signal (ŷ ), - Generating an anomaly alarm if the current residual (R) of the future signal (ŷ) predicted by the model (10) exceeds the dynamic limit value (α _n ). Procedure according to Claim 1 , characterized in that the model is designed as a learning function approximated by historical signals, wherein the model (10) trains the learning function based on signals which characterize a normal state of the technical system to perform a mapping to a future signal (ŷ) on the basis of the fixed number of signals in the past time window. Procedure according to Claim 2 , characterized in that the model (10) is designed as an artificial intelligence procedure. Method according to one of the preceding claims, characterized in that the anomaly alarm is only triggered when a predefined number of consecutive exceedances occur. Method according to one of the preceding claims, characterized in that the dynamic value (β) assumes a scalar value between a value c and one, wherein one is the maximum value and c is greater than zero, and wherein the dynamic limit value (α _n ) is formed by multiplying the static limit value (α) with the dynamic value (β). Procedure according to Claim 5 , characterized in that from a predetermined range of variation of the signals and/or the parameter sizes the dynamic value (β) is set to one. Method according to one of the preceding claims, characterized in that a statistical method from the field of correlation analysis, regression analysis, simulation model, experimental method or sensitivity analysis is used to determine the range of variation of the signals and/or the parameter sizes in the given time window. Procedure according to Claim 7 characterized in that the statistical method is used depending on the signal to be checked. Method according to one of the preceding claims, characterized in that the signals are transmitted from a technical system and/or component as the source. Method according to one of the preceding claims, characterized in that the current signals are pre-processed. Device (9) for detecting anomalies in a set of signals, wherein the signal represents a quantity of a technical system to be checked, characterized in that a storage unit with a model (10) is provided, wherein the model is configured to map a fixed number of signals in a past time window to an output target, wherein the output target maps the signals in a future time step as future signals (ŷ), wherein a static limit value (α) is further provided in the storage unit, which is defined by the model (10) with respect to existing residuals (R) of the predicted future signals (ŷ), wherein a processing unit is further provided, which is configured to acquire a set of current signals as well as parameter quantities currently influencing the signal to be checked from at least one measurement source, wherein the processing unit is further configured to input the current signals into the model (10) over the past time window and to output a predicted future signal (ŷ), wherein the processing unit is configured to The generation of a dynamic dynamic value (β) is configured, wherein the dynamic dynamic value (β) depends on a variation range of the signals and/or a variation range of the currently influencing parameter sizes over the past time window, wherein the dynamic dynamic value (β) is set higher the more the signals and/or the influencing parameter sizes vary over the time window, and wherein the computing unit is configured to generate a dynamic limit value (α _n ) based on the dynamic value (β) and the static limit value (α), and comprising a comparison unit for determining a current residual (R) based on the predicted future signal (ŷ) and for generating an anomaly alarm when the current residual (R) of the future signal (ŷ) predicted by the model (10) exceeds the dynamic limit value (α _n ). Device (9) according Claim 11 , characterized in that a display unit is provided to indicate the exceedance when an anomaly alarm occurs. Generator (5) with a gearbox (4), the generator (5) comprising a method according to one of the preceding claims and/or a device according to one of the preceding claims, wherein the signals are characterizing parameters of the gearbox (4) or of the generator (5). Generator (5) after Claim 13 , characterized in that the gearbox (4) has at least one gearbox bearing (7) and the signals include at least one bearing temperature. Wind turbine (1) with a generator (5) according to Claim 13 or 14 .