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

An anomaly detection device configured to detect anomalies in a plurality of sensor recordings of a technical system, device configured to:

• ascertain a first anomalous value which, with regard to all sensor recordings of the plurality of sensor recordings, characterizes whether or not an anomaly is present, the first anomalous value being based on a probability density value which indicates how probable a joint occurrence of the plurality of sensor recordings is;
• ascertain a plurality of second anomalous values, each of the second anomalous values from the plurality of second anomalous values corresponding to a respective sensor recording of the plurality of sensor recordings, each of the second anomalous values, with regard to the respective sensor recording and under the condition of an occurrence of the other sensor recordings, characterizing whether or not an anomaly is present in the respective sensor recording, each of the second anomalous values being based on a probability density value which indicates how probable an occurrence of the respective sensor recording is given an occurrence of the other sensor recordings;
• and
• detect an anomaly in a specific sensor recording of the plurality of sensor recordings when the first anomalous value characterizes the presence of an anomaly, and the second anomalous value corresponding to the specific sensor recording characterizes an anomaly, and the second anomalous value corresponding to the specific sensor differs beyond a predefined extent from the other second anomalous values of the plurality of second anomalous values.

A computer-implemented method for training an anomaly detection device, the method comprising the following steps:

• providing a plurality of sensor recordings of the technical system or of a technical system having the same design or of a similar technical system;
• training the first anomaly detection model based on the plurality of sensor recordings, the first anomaly detection model being trained in such a way that the first anomaly detection model ascertains a first anomalous value which, with regard to all sensor recordings of the plurality of sensor recordings, characterizes whether or not an anomaly is present, the first anomalous value being based on a probability density value which indicates how probable a joint occurrence of the plurality of sensor recordings is;
• and
• training a plurality of second anomaly detection models, each of the second anomaly detection models configured for a respective sensor recording of the plurality of sensor recordings and being trained in such a way that the second anomaly detection model ascertains a second anomalous value for the respective sensor recording which characterizes whether or not an anomaly is present in the respective sensor recording under the condition of an occurrence of the other sensor recordings of the plurality of sensor recordings, the second anomalous value being based on a probability density value which indicates how probable an occurrence of the respective sensor recording is given an occurrence of the other sensor recordings.

A training device configured to train an anomaly detection device, the device configured to:

• provide a plurality of sensor recordings of the technical system or of a technical system having the same design or of a similar technical system;
• train the first anomaly detection model based on the plurality of sensor recordings, the first anomaly detection model being trained in such a way that the first anomaly detection model ascertains a first anomalous value which, with regard to all sensor recordings of the plurality of sensor recordings, characterizes whether or not an anomaly is present, the first anomalous value being based on a probability density value which indicates how probable a joint occurrence of the plurality of sensor recordings is;
• and
• train a plurality of second anomaly detection models, each of the second anomaly detection models configured for a respective sensor recording of the plurality of sensor recordings and being trained in such a way that the second anomaly detection model ascertains a second anomalous value for the respective sensor recording which characterizes whether or not an anomaly is present it the respective sensor recording under the condition of an occurrence of the other sensor recordings of the plurality of sensor recordings, the second anomalous value being based on a probability density value which indicates how probable an occurrence of the respective sensor recording is given an occurrence of the other sensor recordings.

A non-transitory machine-readable memory medium on which is stored a computer program for training an anomaly detection device, the computer program, when executed by a computer, causing the computer to perform the following steps:

• providing a plurality of sensor recordings of the technical system or of a technical system having the same design or of a similar technical system;
• training the first anomaly detection model based on the plurality of sensor recordings, the first anomaly detection model being trained in such a way that the first anomaly detection model ascertains a first anomalous value which, with regard to all sensor recordings of the plurality of sensor recordings, characterizes whether or not an anomaly is present, the first anomalous value being based on a probability density value which indicates how probable a joint occurrence of the plurality of sensor recordings is;
• and
• training a plurality of second anomaly detection models, each of the second anomaly detection models configured for a respective sensor recording of the plurality of sensor recordings and being trained in such a way that the second anomaly detection model ascertains a second anomalous value for the respective sensor recording which characterizes whether or not an anomaly is present it the respective sensor recording under the condition of an occurrence of the other sensor recordings of the plurality of sensor recordings, the second anomalous value being based on a probability density value which indicates how probable an occurrence of the respective sensor recording is given an occurrence of the other sensor recordings.