Method and device for detecting anomalies in sensor recordings of a technical system

Abstract

Method and device for detecting anomalies in sensor recordings of a technical system. Computer-implemented method for detecting anomalies in a plurality of sensor recordings of a technical system, the method comprising the steps of: • determining a first anomaly value, the first anomaly value characterizing whether an anomaly is present with respect to all of the plurality of sensor recordings; • determining a plurality of second anomaly values, wherein one second anomaly value out of the plurality of second anomaly values corresponds to one sensor recording of the plurality of sensor recordings and characterizes whether an anomaly is present with respect to other sensor recordings of the plurality of sensor recordings; • detecting one sensor recording of the plurality of sensor recordings if the first anomaly value characterizes the presence of an anomaly and the second anomaly value corresponding to this sensor recording does not characterize an anomaly and this second sensor value deviates from other second anomaly values of the plurality of second anomaly values by more than a predefined extent.

Description

Technical Field

[0001] The present invention relates to a method for detecting anomalies in sensor records of a technology system, an anomaly detection device, a method for training an anomaly detection device, a training device, a computer program, and a storage medium. Background Technology

[0002] Unpublished DE 10 2020 208 642.7 discloses methods and apparatus for detecting anomalies in technical systems. Summary of the Invention

[0003] Advantages of the invention A technical system can use sensors to sense parameters of the external physical environment or the system's operational status, and these sensors transmit corresponding sensor records to the system. However, sensors may transmit erroneous signals, for example, when they malfunction. To mitigate this problem, redundant sensors of the same type can be used, preferably.

[0004] It is desirable to monitor the sensor logs of multiple sensors in order to identify any abnormal behavior in the sensor logs, and if so, to identify which sensor(s) caused the abnormal behavior.

[0005] The computer-implemented method for detecting anomalies in multiple sensor records of a technical system according to this application can identify anomalies in multiple sensor records. If a sensor can be identified as abnormal based on its output signal, this has the advantage that the sensor can be automatically inspected or replaced. Alternatively or additionally, it is conceivable that the technical system advantageously, for a short period or permanently, disregards sensor records detected as abnormal for continued operation. This enables the technical system to operate correctly even in the presence of defective or unreliable sensors or erroneous or abnormal sensor records.

[0006] Disclosure of the invention In a first aspect, the present invention relates to a computer-implemented method for detecting anomalies in multiple sensor records of a technical system, the method comprising the following steps: • Identify a first outlier, which indicates whether an anomaly exists across all sensor records; • Identify a plurality of second outliers, wherein one of the second outliers corresponds to one of the plurality of sensor records and characterizes the presence of anomalies in the other sensor records among the plurality of sensor records with respect to that sensor record; • If a first outlier indicates the presence of an anomaly and a second outlier corresponding to that sensor record does not indicate an anomaly, and the second outlier deviates from other second outliers among the plurality of second outliers by a predefined degree, then an anomaly is detected in one of the plurality of sensor records.

[0007] Sensor recording can be understood as the result of measuring the physical surrounding environmental conditions or physical operating conditions of a technical system using appropriate sensors. Preferably, multiple sensor records are recorded during the operation of the technical system and at the same time.

[0008] In the context of this invention, if the occurrence of a sensor record is unreliable relative to other measurements of the sensor corresponding to that sensor record, then an anomaly exists in the sensor record or the sensor record is interpreted as an anomaly. For example, a sensor record may include values ​​that fall within a typical range during system operation. Values ​​outside the typical range for that value can be interpreted as unreliable or anomaly. If a sensor record includes at least one value outside the typical range for that value, then the sensor record may be interpreted as an anomaly in this case.

[0009] This method can be understood as follows: First, for multiple sensor records, a first outlier is used to determine whether an anomaly exists in the multiple sensor records. If an anomaly exists, multiple second outlier values ​​can be used to determine which sensor record caused the abnormal behavior. Preferably, this method is applied during the operation of the technical system so that anomalies can be detected during the operation of the technical system.

[0010] Multiple sensor recording preferably allows for the detection of multiple sensors and the provision of these data to the technical system. Sensors installed within the technical system and those not installed within it can be used as sensors, such as those that allow external observation of the technical system.

[0011] Preferably, the first outlier can be a probability or probability density value. In this case, if the first outlier is lower than a predefined first threshold, an anomaly may exist in multiple sensor records.

[0012] Furthermore, preferably, the second outlier can be a probability or probability density value. In this case, if the second outlier is lower than a predefined second threshold, the sensor record corresponding to the second outlier can be interpreted as abnormal. Alternatively, the second threshold can be the first threshold.

[0013] Alternatively, the first outlier can be determined based on multiple sensor records using a first anomaly detection model. The first outlier can be determined by the first anomaly detection model, preferably using a machine learning method. The advantage of using a machine learning method is that it determines the first outlier more accurately than, for example, using a rule-based method. Therefore, this method can improve the accuracy of detecting anomalies in the technical system.

[0014] In particular, the machine learning method can be a normalizing flow. In this case, a normalizing flow can be configured to receive multiple sensor records as input and determine the output, where the output, or a portion of the output, is provided as a first outlier. The advantage of a normalizing flow for this purpose is that it can determine very precise probabilities or probability density values ​​regarding the occurrence of multiple sensor records, leading to more accurate anomaly detection.

[0015] Furthermore, it is conceivable that, in order to determine multiple second outliers for each sensor record, there exists a corresponding second anomaly detection model, which corresponds to the corresponding sensor record and is configured to determine the second outlier corresponding to the sensor record. Preferably, the second anomaly detection model is configured to determine the second outlier based on sensor records from among the multiple sensor records that do not correspond to the second anomaly detection model. The second outlier can be determined by the second anomaly detection model preferably using a machine learning method. The advantage of using a machine learning method is that it determines the second outlier more accurately than, for example, using a rule-based method. Accordingly, this method can improve the accuracy of detecting anomalies in the technical system.

[0016] In particular, the machine learning method can be a prescriptive process. In this case, the prescriptive process can be configured to receive sensor records from multiple sensor records, excluding the corresponding sensor records, as input and determine the output, wherein the output, or a portion of the output, is provided as a second outlier. The advantage of a prescriptive process for this purpose is that it can determine a very precise probability or probability density value for the occurrence of all sensor records from the multiple sensor records, leading to more accurate anomaly detection.

[0017] In another aspect, the present invention relates to a computer-implemented method for training anomaly detection equipment, wherein the method includes the following steps: • Provides multiple sensor records of the same or similar technical systems; • The first anomaly detection model is trained based on records from multiple sensors; • Train multiple second anomaly detection models, where one second anomaly detection model is trained for each of the other sensor records from the multiple sensor records for one of the sensor records.

[0018] In the provided steps, it is preferable that multiple sensor records are provided simultaneously by multiple different sensors of the same type. Each sensor record in the multiple sensor records can therefore be understood as a corresponding measurement at the same time.

[0019] Multiple sensor records can therefore be understood as training data for this method. Preferably, multiple training data are provided for training, and a first anomaly detection model and / or a second anomaly detection model are trained based on the multiple training data.

[0020] In the step of training the first anomaly detection model, it is preferable to train the first anomaly detection model such that it assigns values ​​to the sensor records among multiple sensor records, the values ​​representing the probability or probability density of the occurrence of the multiple sensor records. For this purpose, the first anomaly detection model can be trained using machine learning training methods, such as training with improved log-likelihood determined by the first anomaly detection model on the training data, for example, using stochastic gradient descent.

[0021] In the step of training multiple second anomaly detection models, each second anomaly detection model can be understood as corresponding to a sensor record. The second anomaly detection model can preferably be trained such that it assigns values ​​to other sensor records in the training data—that is, sensor records among multiple sensor records without a corresponding sensor record—values ​​representing the probability or probability density of the occurrence of the other sensor records. For this purpose, the second anomaly detection model can be trained using machine learning methods, preferably using the log-likelihood of other records determined by the second anomaly detection model, such as Stochastic Gradient Descent.

[0022] Furthermore, it is conceivable that multiple sensor records are provided by a second technical system, which is structurally identical or at least similar to the technical system, for example, where the technical system uses or includes sensors with the same structure. For instance, the technical system could be a specific type of machine that generates multiple samples. In this case, it is conceivable that at least one of these samples is used to record multiple sensor records, and the multiple sensor records are then used to train an anomaly detection device, which is used to detect anomalies for other samples among the multiple samples.

[0023] Furthermore, it is conceivable to determine the sensor records of a prototype of the technical system during the development process of the technical system. In the context of this invention, the prototype can be understood as resembling the technical system, and the sensor records thus determined are used to train the technical system. Attached Figure Description

[0024] Embodiments of the invention are described in detail below with reference to the accompanying drawings. In the drawings: Figure 1 The structure of the anomaly detection device is schematically shown; Figure 2 The structure of a control system for manipulating actuators by utilizing anomaly detection equipment is schematically shown; Figure 3 An embodiment for controlling a robot that is at least partially autonomous is illustrated schematically; Figure 4 An embodiment for controlling a manufacturing system is illustrated schematically; Figure 5 An embodiment for controlling an access control system is illustrated schematically; Figure 6 An embodiment for controlling and monitoring systems is illustrated schematically; Figure 7 An embodiment for controlling a personal assistive device is illustrated schematically; Figure 8 An embodiment for controlling a medical imaging system is illustrated schematically; Figure 9 The structure of the training device used to train the anomaly detection device is schematically shown. Detailed Implementation

[0025] Figure 1 The diagram shows the detection recorded by three sensors (x) a x b x c An anomaly detection device (70) records an anomaly in a sensor within a sensor. The anomaly detection device (70) receives a first input signal (x). 70 The first input signal includes records from these three sensors. The first input signal (x) 70 The data is fed to a first anomaly detection model (72), wherein the first anomaly detection model is configured to determine a first probability density value, the first probability density value representing: how likely it is that three sensors record (x) a x b x c They appear together. The first anomaly detection model (72) is the standard procedure in this embodiment.

[0026] The first anomaly detection model is based on the first input signal (x) 70 A first probability density value is determined and forwarded to a first comparison unit (74). The first comparison unit (74) compares the first probability density value with a first threshold (T). U Comparison. For a first probability density value greater than or equal to a first threshold (T)... U In the case of input signal (x), the first comparison unit (74) transmits the first judgment value to the judgment unit (76), the first judgment value representing the input signal (x) 70 In, that is, in the records of three sensors (x) a x b x c There is no anomaly in the combination of ). Otherwise, the first comparison unit (74) selects the first judgment value, such that the first judgment value represents the combination of the three sensor records (x) a x b x c An anomaly exists in the combination of ). Optionally, the first comparison unit (74) may also transmit the probability density value determined by the first anomaly detection model (72) to the judgment unit.

[0027] In addition, the input signal (x) 70 The signal (x) is transmitted to the distribution unit (71), which receives the input signal (x) from the distribution unit. 70 The corresponding sensor record (x) is determined in the data. a x b x c Next, the corresponding sensor records (x) a x b x c The data is transmitted to one of the three second anomaly detection models (73a, 73b, 73c). The second sensor records (x...) b ) and third sensor recording (x c The signal was fed to the first of the three second anomaly detection models (73a), and the first sensor recorded (x) a ) and third sensor recording (x c The signal was fed to the second of the three second anomaly detection models (73b), and the first sensor recorded (x). a ) and second sensor recording (x b The data is fed to the third of the three second anomaly detection models (73c). Each of the second anomaly detection models (73a, 73b, 73c) is used to record (x) sensor data that is fed to the corresponding anomaly detection model. a x b x cDetermine the probability density value, which represents the likelihood that the transmitted sensor record (x) will occur. a x b x c ).

[0028] The first of these three second anomaly detection models (73a) will target the second sensor record (x) b ) and third sensor recording (x c The probability density value determined by the first of the three second anomaly detection models (73a) is submitted to the second comparison unit (75a). The second comparison unit determines whether the probability density value determined by the first of the three second anomaly detection models (73a) is greater than or equal to the second threshold (T). a If the determined probability density value is greater than or equal to the second threshold (T) a Then, the second comparison unit (75a) transmits the second judgment value to the judgment unit (76), and the second judgment value represents the value recorded by the second sensor (x). b ) and / or third sensor recording (x c There is no anomaly in the second sensor record (x). Otherwise, the first comparison unit (75a) selects the second judgment value, such that the second judgment value represents the value recorded by the second sensor (x). b ) and / or third sensor recording (x c An anomaly exists in the model. Optionally, the second comparison unit (75a) may also transmit the probability density value determined by the first (73a) of the three second anomaly detection models to the judgment unit (76).

[0029] The second of these three second anomaly detection models (73b) will target the first sensor record (x) a ) and third sensor recording (x c The probability density value determined by the second of the three second anomaly detection models (73b) is submitted to the third comparison unit (75a). The third comparison unit determines whether the probability density value determined by the second of the three second anomaly detection models (73b) is greater than or equal to the third threshold (T). b If the determined probability density value is greater than or equal to the third threshold (T) b Then, the third comparison unit (75b) transmits the third judgment value to the judgment unit (76), the third judgment value representing that, as recorded by the first sensor (x) b ) and / or third sensor recording (x c There is no anomaly in the data. Otherwise, the third comparison unit (75b) selects a third judgment value, such that the third judgment value represents the data recorded by the first sensor (x). b ) and / or third sensor recording (x cAn anomaly exists in the model. Optionally, the third comparison unit (75b) may also transmit the probability density value determined by the second (73b) of the three second anomaly detection models to the judgment unit (76).

[0030] The third of these three second anomaly detection models (73c) will target the first sensor record (xa) and the second sensor record (x). b The determined probability density value is submitted to the fourth comparison unit (75c). The fourth comparison unit determines whether the probability density value determined by the third of the three second anomaly detection models (73c) is greater than or equal to the fourth threshold (Tc). If the determined probability density value is greater than or equal to the fourth threshold (Tc), then the probability density value is determined to be greater than or equal to the fourth threshold (Tc). c Then, the fourth comparison unit (75c) transmits the fourth judgment value to the judgment unit (76), the fourth judgment value representing that, as recorded by the first sensor (x) a ) and / or second sensor recording (x b There is no anomaly in the data recorded by the first sensor (x). Otherwise, the fourth comparison unit (75c) selects a fourth judgment value, which represents the value recorded by the first sensor (x). a ) and / or second sensor recording (x b An anomaly exists in the model. Optionally, the fourth comparison unit (75c) may also transmit the probability density value determined by the third (73a) of the three second anomaly detection models to the judgment unit (76).

[0031] The three second anomaly detection models (73a, 73b, 73c) in this embodiment are the standard procedures.

[0032] The judgment unit (76) determines whether an anomaly exists based on the transmitted judgment value, and if so, records the anomaly in the sensor (x). a x b x c The anomaly is determined in which sensor record is present. Optionally, the determination unit (76) may also additionally determine this based on the transmitted probability density value. If the first determination value indicates that no anomaly exists, the anomaly detection device (70) outputs an anomaly detection output (y). 70 If the first judgment value indicates an anomaly, the judgment unit determines at least one judgment value from a plurality of second, third, and fourth judgment values, wherein the at least one judgment value indicates that no anomaly exists. Anomaly detection output (y) 70 In this case, the decision unit (76) selects the method that makes its representation recorded in the sensor (x) a x b x c An anomaly exists in the sensor data, and the sensor records were not used to determine the determined judgment value.

[0033] Optionally, more than one sensor records (x a x b x c It can also be output by anomaly detection (y) 70 This is characterized as an anomaly. Additionally, optionally, it is only when recorded by the sensor corresponding to (x) a x b x c When the probability density value determined by the second anomaly detection model (73a, 73b, 73c) exceeds the probability density value determined by another second anomaly detection model (73a, 73b, 73c) to a predefined extent, the sensor records (x) a x b x c Only then can it be output by anomaly detection (y) 70 This is characterized as an anomaly.

[0034] In this embodiment, limiting recording to three sensors is an exemplary selection. With necessary modifications, the anomaly detection device (70) can also be used for the input signal (x). 70 Recorded by several other sensors in the system.

[0035] Figure 2 An actuator (10) is shown in the environment in which it interacts with the control system (40). The surrounding environment (20) is detected by multiple sensors (30) at preferred regular time intervals. The sensors (30) include imaging sensors such as, for example, cameras. Alternatively, the sensors (30) may also include other types of sensors, such as sensors capable of measuring the physical surrounding environmental conditions or operating conditions of the control system (40).

[0036] The signals from the sensors (S) are transmitted to the control system (40). The control system (40) then receives a series of sensor records (S). Based on this, the control system (40) determines a control signal (A), which is transmitted to the actuator (10).

[0037] The control system (40) receives a series of sensor records (S) from the sensor (30) in an optional receiving unit (50), which converts the series of sensor records (S) into a series of input images (x). 60 Input image (x) 60 For example, it could be part of a camera's sensor recording or further processing, which is included in the sensor record (S). Input image (x) 60 This includes individual frames of the video recording. In other words, the input image (x) is determined based on the sensor recording (S). 60 The input image (x) of this sequence60 The image is fed to an image classifier (60), which in this embodiment is a neural network.

[0038] The control system (40) includes an anomaly detection device (70). In this embodiment, this is configured to identify anomalies in the sensor records (S) of the imaging sensor (30). In other embodiments, it is conceivable that the anomaly detection device (70) is configured to detect anomalies in the sensor records (S) of other types of sensors. This series of sensor records (S) is converted by the receiving unit (50) into a series of input signals (x) for the anomaly detection device. 70 If the anomaly detection device is configured to identify anomalies in the sensor record of the imaging sensor, then the input signal (x) 70 )Optionally, it may also include an input image (x 60 ).

[0039] The image classifier (60) preferably uses the first parameter ( The parameterization is performed, with the first parameter stored in and provided by the parameter memory (P). The anomaly detection device (70) preferably uses a second parameter ( The parameter is parameterized, and the second parameter is also stored in the parameter memory and provided by the parameter memory.

[0040] The image classifier (60) extracts data from the input image (x). 60 Determine the output parameter (y) in ) 60 The output parameter represents the input image (x) 60 The category of anomaly detection device (70) is determined by the input signal (x). 70 Determine the anomaly detection output (y) in ) 70 ).

[0041] Output parameter (y) 60 ) and anomaly detection output (y 70 The signal is transmitted to a conversion unit (80), which determines a control signal (A) based on this signal. The control signal is then transmitted to an actuator (10) to control the actuator (10) accordingly. Output parameter (y) 60 This includes information about the object in the corresponding input image (x). 60 It can be identified on the screen.

[0042] The actuator (10) receives a control signal (A), and the actuator (10) is correspondingly controlled and performs a corresponding action. In this case, the actuator (10) may include (but not necessarily structurally integrated) a control logic device, which determines a second control signal from the control signal (A) and uses the second control signal to control the actuator (10).

[0043] If the anomaly detection output (y) 70 If an anomaly is detected, the control signal (A) can be selected to limit the possible actions of the actuator (10). If no anomaly is detected, it is conceivable that the possible actions are not limited based on the anomaly, but rather based on the surrounding environment (20) of the control system (40) as determined by the image classifier (60). Furthermore, it is conceivable that, in the event of an anomaly, at least a portion of the sensor record (S) is transmitted to the manufacturer or operator of the control system (40).

[0044] Alternative or additional sites may include anomaly detection output (y 70 The image is also transmitted to the receiving unit (50). The receiving unit can then, for example, receive the image from the input image (x) in case of an anomaly in the sensor record (S). 60 In the determination of the input image (x), other sensor records are selected. 60 ).

[0045] In other embodiments, the control system (40) includes a sensor (30). In other embodiments, the control system (40) may alternatively or additionally include an actuator (10).

[0046] In other preferred embodiments, the control system (40) includes at least one processor (45) and at least one machine-readable storage medium (46) on which instructions are stored, which, when executed on the processor (45), cause the control system (40) to perform the method according to the invention.

[0047] In an alternative implementation, a display unit (10a) is provided alternatively or additionally to the actuator (10). In the event of an anomaly, the display unit (10a) may be manipulated, for example, to display the presence of the anomaly to the user or operator of the control system.

[0048] Figure 3 The control system (40) is shown as a way to control a robot that is at least partially autonomous, in this case, a motor vehicle (100) that is at least partially autonomous.

[0049] The sensor (30) may be, for example, a plurality of video sensors preferably disposed in the vehicle (100). The receiving unit (50) can then use at least one sensor record (S) from the video sensor as an input image (x). 60 Forward to the image classifier (60).

[0050] The image classifier (60) is set to classify the input image (x) 60 Identify objects in ( ).

[0051] The actuators (10) preferably disposed in the motor vehicle (100) may be, for example, the brakes, drives or steering devices of the motor vehicle (100). A control signal (A) may be determined such that one or more actuators (10) are manipulated to prevent the motor vehicle (100) from colliding with objects identified by the image classifier (60), especially when the objects are objects of a particular category such as pedestrians.

[0052] Alternatively, the at least partially autonomous robot may also be other mobile robots (not shown), such as robots that move by flying, floating, diving, or walking. The mobile robot may, for example, be a lawnmower or a cleaning robot that is at least partially autonomous. In this case, control signals (A) may also be determined, and the mobile robot's actuators and / or steering mechanisms are manipulated to prevent the at least partially autonomous robot from colliding with objects identified by an artificial neural network (60), for example.

[0053] Alternatively or additionally, the display unit (10a) can be controlled using a control signal (A) and, for example, display the determined area. It is also possible, for example, in a motor vehicle (100) with a non-automatic steering system, that the display unit (10a) is controlled using the control signal (A) such that when it is determined that the motor vehicle (100) is about to collide with one of the reliably identified objects, the display unit outputs an optical or acoustic warning signal.

[0054] Figure 4 One embodiment is shown in which a control system (40) is used to operate the production machine (11) of the production system (200) by controlling the actuator (10) that operates the production machine (11). The production machine (11) may be, for example, a machine for stamping, sawing, drilling and / or cutting.

[0055] The sensor (30) may include, for example, an imaging sensor that detects, for example, the characteristics of the production products (12a, 12b). It is possible that the production products (12a, 12b) are movable. It is possible that the actuator (10) controlling the production machine (11) is manipulated based on the detected correlation of the production products (12a, 12b) so that the production machine (11) correspondingly performs subsequent processing steps for the correct production products (12a, 12b). It is also possible that the production machine (11) is adapted to the same production steps for processing subsequent production products by identifying the correct characteristics of the same production products (i.e., no erroneous correlations) in the production products (12a, 12b).

[0056] Furthermore, the production machine (11) may include sensors (30) capable of measuring the physical surrounding environmental conditions or operating status of the production machine (11). The sensor records (S) of these sensors (30) can be transmitted to an anomaly detection device (70) to determine whether the production machine (11) is operating under normal environmental conditions and in normal operating condition. If an anomaly is detected by the anomaly detection device (70), the operation of the production machine (11) can be stopped, for example, or scheduled for automatic maintenance.

[0057] Figure 5 An embodiment is shown in which a control system (40) controls an access control system (300). The access control system (300) may include physical access control, such as a door (401). The sensor (30) may be, for example, a video sensor configured to detect persons in front of the access control system (300). The detected input image (x) can be interpreted using an image classifier (60). 60 If multiple persons are detected simultaneously, their identities can be reliably determined through the relationships between them (i.e., objects), for example, by analyzing their movements. The actuator (10) can be a lock that releases or retains the access control according to a control signal (A), for example, opening the door (401) or not opening the door. For this purpose, the control signal (A) can be selected based on the interpretation of the object identification system (60), for example, based on the determined identity of the person. Logical access control can also be used instead of physical access control.

[0058] The image from sensor (30) can be used as an input signal (x) 70 The signal is transmitted to the anomaly detection device (70). If the anomaly detection device (70) can detect the anomaly, it can at least temporarily block and / or automatically arrange for a security guard to conduct an inspection.

[0059] Figure 6 An embodiment is shown in which the control system (40) is used to control the monitoring system (400). This embodiment is similar to... Figure 5 The embodiment shown differs in that a display unit (10a) is provided instead of the actuator (10), which is controlled by the control system (40). For example, the identity of the object recorded by the sensor (30) can be reliably determined by the artificial neural network (60) so as to infer, for example, which objects are suspicious, and then select the control signal (A) so that the object is highlighted in color by the display unit (10a).

[0060] Figure 7One embodiment is shown in which a control system (40) is used to control a personal assistive device (250). The sensor (30) is preferably a video sensor that receives an image of the user's (249) posture.

[0061] Based on the sensor records (S) of the sensor (30), the control system (40) determines the control signal (A) of the personal assistive device (250), for example, by means of gesture recognition performed by a neural network. The determined control signal (A) is then transmitted to the personal assistive device (250), and the personal assistive device is thus operated accordingly. The determined control signal (A) can be selected in particular to correspond to a desired operation guessed by the user (249). The guessed desired operation can be determined based on the gesture recognized by the artificial neural network (60). The control system (40) can then select the control signal (A) to transmit to the personal assistive device (250) based on the guessed desired operation and / or select the control signal (A) to transmit to the personal assistive device based on the guessed desired operation (250).

[0062] The corresponding operation may include, for example, the personal assistive device (250) retrieving information from a database and reproducing the information in a manner acceptable to the user (249).

[0063] Instead of personal assistive devices (250), household appliances (not shown) may also be provided, especially washing machines, stoves, ovens, microwave ovens or dishwashers, for corresponding operation.

[0064] Figure 8 One embodiment is shown in which a control system (40) controls a medical imaging system (500), such as an MRT device, X-ray device, or ultrasound device. Sensors (30), for example, can be provided by imaging sensors, and the control system (40) manipulates the display unit (10a). For example, an image classifier (60) can determine whether an area recorded by at least one imaging sensor is conspicuous, and then select a manipulation signal (A) such that the area is highlighted in color by the display unit (10a).

[0065] The image from sensor (30) is used as the input signal (x) 70 The information is transmitted to the anomaly detection device (70). In the event of an anomaly, this can be displayed on the display unit (10a) and / or the system maintenance (500) can be automatically scheduled.

[0066] Figure 9An embodiment of a training system (140) is shown, which is configured to train an anomaly detector (70). For training, a training data unit (150) accesses a computer-implemented database (St2), wherein the database (St2) includes at least one training dataset (T), wherein the training dataset (T) comprises tuples (x) recorded by the sensor. i ).

[0067] The training data unit (150) determines at least one tuple (x) of sensor records in the training dataset (T). i And the tuple (x) i The data is transmitted to the anomaly detection device (70). The anomaly detection device (70) uses a first anomaly detection model and multiple second anomaly detection models to determine a first anomaly or multiple second anomaly values.

[0068] The first outlier and multiple second outliers are output as ( ) is transmitted to the change unit (180).

[0069] The output determined based on the expected first outlier and multiple expected second outliers ( ) and the expected output (y i ), and thus the new model parameters for the first anomaly detection model and / or multiple second anomaly detection models are determined by the changing unit (180). In this embodiment, it is conceivable that the anomaly detection model is a standard procedure. In this case, the changing unit (180) can determine the new model parameters using gradient ascent methods such as Stochastic Gradient Descent or Adam. ).

[0070] The determined new model parameters ( The parameters are stored in the model parameter memory (St1).

[0071] In other embodiments, the described training is repeated iteratively for a predefined number of iterations until a determined output is achieved. The difference between the model parameter (yi) and the expected output (yi) is below a predefined threshold. In at least one iteration of the iteration, the new model parameters (yi) determined in the previous iteration... ) is used as the model parameter (Φ) of the anomaly detection device.

[0072] Furthermore, the training system (140) may include at least one processor (145) and at least one machine-readable storage medium (146) containing instructions that, when executed by the processor (145), cause the training system (140) to perform a training method according to one aspect of the invention.

[0073] The term "computer" includes any device used to process pre-given computational rules. These rules can exist in software, hardware, or a hybrid of both.

Claims

1. A computer-implemented method for detecting multiple (x) in a technical system (100, 200, 250, 300, 400, 500). 70 A method for detecting anomalies in sensor records, the method comprising the following steps: • Identify a first outlier, the first outlier relating to multiple (x) 70 All sensor records in the sensor logs (x a x b x c ) Indicate whether there is an anomaly; • Identify multiple second outliers, wherein one of the multiple second outliers is related to the multiple (x) 70 Sensor recording (x) a x b x c A sensor record (x) a x b x c Corresponding to and regarding the sensor record (x) a x b x c ) represents the multiple (x 70 Other sensor records in the sensor logs (x) a x b x c Are there any anomalies in )? • If the first outlier indicates the presence of an anomaly and is consistent with the sensor record (x) a x b x c If the corresponding second outlier does not represent an anomaly and the deviation of this second outlier from other second outliers among the plurality of second outliers exceeds a predefined degree, then the plurality of (x) outliers are detected. 70 Sensor recording (x) a x b x c A sensor record (x) a x b x c ) is an exception.

2. The method according to claim 1, Wherein the first outlier represents the plurality of (x) 70 Sensor recording (x) a x b x c The probability or probability density value of the occurrence of (x) and / or the second outlier characterize the plurality of (x) 70 Sensor recording (x) a x b x c In addition to the sensor records (x) corresponding to the second outlier a x b x c The probability or probability density value that appears outside of ().

3. The method according to claim 2, The first outlier is determined using the first outlier detection model (72).

4. The method according to any one of claims 1 to 3, The plurality of second outliers are determined using a plurality of second outlier detection models (73a, 73b, 73c), wherein each of the plurality of second outlier detection models (73a, 73b, 73c) is associated with a sensor record (x). a x b x c ) corresponds to and determines the sensor record (x) a x b x c The second outlier corresponding to ).

5. The method according to claim 3, The first anomaly detection model (72) includes a standard process model and / or the second anomaly detection model (73a, 73b, 73c) includes a standard process model.

6. An anomaly detection device (70), The anomaly detection device is configured to perform the method according to any one of claims 1 to 5.

7. A computer-implemented method for training an anomaly detection device (70), The anomaly detection device (70) is configured to perform the method according to any one of claims 3 to 5, wherein the method includes the following steps: • Provides multiple sensor records (x) of technical systems (100, 200, 250, 300, 400, 500) or technical systems with the same structure. i ); • Based on multiple (x) i ) Sensor recordings were used to train the first anomaly detection model (72); • Train multiple second anomaly detection models (73a, 73b, 73c), among which multiple (x i One sensor record in the sensor recording is based on multiple (x) i Other sensor records in the sensor logs are used to train each second anomaly detection model (73a, 73b, 73c).

8. A training device (140) configured to perform the method according to claim 7.

9. A computer program product having a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 5 or according to claim 7.

10. A machine-readable storage medium (46, 146) storing a computer program thereon, the computer program including instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 5 or according to claim 7.

Images