Artificial intelligence-based anomaly detection and prediction

1 . A system for rendering an explanation output for users regarding an anomaly predicted by an anomaly detection module on a basis of high frequency sensor data or values derived therefrom in an industrial production process, wherein the anomaly detection module predicts the anomalies when the anomaly detection module classifies sensor data or ranges of sensor data that describe a state of a machine, a component, and/or a product of the production process as different from the data that are normally expected, wherein the system comprising graphical processing unit and is configured to:

calculate an explanation for the predicted anomaly based on sensor data or ranges of sensor data processed by the anomaly detection module which uses a machine learning model, wherein the anomaly detection module uses the sensor data or ranges of the sensor data to predict the anomaly;

create an explanation mask for calculating the explanation, wherein the explanation mask alters the sensor data or range of sensor data in which the anomaly was predicted, such that the anomaly detection module classifies the sensor data masked with the explanation mask as normal;

optimize values in the explanation mask through iterative processing of the machine learning model for the respective masked sensor data; and

transmit the optimized explanation mask in the form of the explanation output rendered visually for users, wherein the user can identify which sensor data, ranges of sensor data, or values derived therefrom, are responsible for the anomaly predicted by the anomaly detection module, based on the optimized explanation mask.

2 . The system according to claim 1 , configured to visually render the optimized explanation mask, wherein the optimize explanation mask is a combination of initial sensor data, or values derived therefrom and multiplied or added with the optimized explanation mask; or an adjacent presentation of the initial sensor data or values derived therefrom and the optimized explanation mask, for the user.

3 . The system according to claim 1 , configured to optimize the values in the explanation mask under the boundary conditions, such that at least one of:

the change to the initial sensor data or values derived therefrom is minimized;

the change to the initial sensor data or values derived therefrom is focused on localized regions of the sensor data;

the change favors coherent regions for the sensor data or values derived therefrom over isolated sensor data; or

the changed values lie in a region of observable values that can be expected, or are physically possible.

4 . The system according to claim 3 , configured to distort initial sensor data or values derived therefrom, and/or the explanation mask with random values.

5 . The system according to claim 1 , wherein the explanation mask has the same dimensions as the initial sensor data or values derived therefrom, and the system is configured to add the explanation mask iteratively to the initial sensor data or values derived therefrom, and/or to multiply the explanation mask iteratively with initial sensor data or values derived therefrom.

6 . The system according to claim 1 , wherein the dimensions of the explanation mask is initially smaller than that of the initial sensor data or values derived therefrom, and the system is configured to determine an explanation mask using up sampling, which has the same dimensions as the initial sensor data or values derived therefrom.

7 . The system according to claim 1 , comprising

a data input module which acquires the sensor data or range of sensor data processed by the anomaly detection module, the machine learning model used by the anomaly detection module on the sensor data or ranges of sensor data, and the anomaly predicted by the machine learning model;

an evaluation module, which acquires the data obtained from the data input module, and the machine learning model that was obtained, and generates the explanation output on the basis thereof, wherein the evaluation module comprises a first transmitter module, which sends the explanation output to the user;

a memory module that comprises a receiver module, which receives an evaluation of the predicted anomaly made by the user based on the explanation output as a correctly or incorrectly predicted anomaly, wherein the memory module receives and stores the data obtained from the data input module, and the machine learning model that was obtained, the explanation output generated by the evaluation module, and the evaluation received from the receiver module; and

a second transmitter module, which sends the evaluation of the predicted anomaly made by the user as a correctly or incorrectly predicted anomaly to the anomaly detection module in order to improve the machine learning model.

8 . The system according to claim 1 , configured to compare a currently predicted anomaly in sensor data or ranges of sensor data for this current anomaly with sensor data for predicted anomalies stored in a memory module in the system, compare an explanation output for this current anomaly with the explanation outputs for predicted anomalies stored in the memory module, and/or compare a combination comprising the sensor data and the explanation mask for this current anomaly with a corresponding combination of predicted anomalies stored in the memory module in the explanation, wherein the system is configured to determine the comparison by executing an algorithm in which distances are implemented.

9 . The system according to claim 1 , configured to render initial sensor data in another data domain, and generate the explanation output in this data domain.

10 . The system according to claim 1 , configured to initiate scanning at a higher frequency for the sensor data, ranges of sensor data, or values derived therefrom, which are responsible for the anomaly predicted by the anomaly detection module.

11 . The system according to claim 1 , wherein the system is used during development of a model for an anomaly detection module, wherein it is ensured, on the basis of an explanation output from the system in a training process for a machine learning model used in the anomaly detection module, that no false, random, or irrelevant aspects of training data are learned.