Controller for controlling a technical system, and method for configuring the controller
A technical system controller is trained using a machine learning method. For this purpose, a chronological sequence of training data is detected for the machine learning method. The training data includes state data, which specifies states of the technical system, and control action data, which specifies control actions of the technical system. A chronological sequence of control action data is extracted specifically from the training data and is checked for a change over time. If a change over time is ascertained, a time window including the change is ascertained, and training data which can be found within the time window is extracted in a manner which is specific to the time window. The controller is then trained by the machine learning method using the extracted training data and is thus configured for controlling the technical system.
1. A method for configuring a control device for a technical system by a machine learning method, wherein
a) a temporal sequence of training data for the machine learning method is captured, wherein the training data comprise state data specifying states of the technical system and control action data specifying control actions of the technical system,
b) a temporal sequence of the control action data is specifically extracted from the training data,
c) the sequence of the control action data is checked for a change over time, and detection of a change over time results in a time window that comprises the change over time of the control action data being ascertained,
d) the training data situated within the time window that comprises the change over time of the control action data are extracted on a time-window-specific basis to provide extracted time-window-specific training data that is a filtered subset of the temporal sequence of training data, and
e) the control device is trained, on the basis of the extracted time-window-specific training data, by the machine learning method and thus configured to control the technical system, and
wherein training data situated outside the time window that comprises the change over time of the control action data are suppressed, rejected, and/or given a lower weighting during the training of the control device.
2. The method as claimed in claim 1 , wherein the technical system is a gas turbine, a wind turbine, a solar power station, an internal combustion engine, a production plant, a traffic control system, an energy control unit for a building, a power transmission system and/or a 3D printer.
3. The method as claimed in claim 1 , wherein the machine learning method uses or implements a reinforcement learning method, an artificial neural network, a recurrent neural network, a convolutional neural network, an autoencoder, a deep learning architecture, a support vector machine, a data-driven trainable regression model, a k nearest neighbor classifier, a physical model and/or a decision tree.
4. The method as claimed in claim 1 , wherein operating data of the technical system are captured in the course of operation and integrated into the sequence of the training data, and in that the control device continues to be trained, on the basis of the extracted training data, in the course of operation.
5. The method as claimed in claim 1 , wherein the check for a change in the sequence of the control action data over time involves a stipulated change pattern being sought in this sequence, and in that the change over time is detected as a result of the change pattern being found.
6. The method as claimed in claim 1 , wherein the check for a change in the sequence of the control action data over time involves a measure of change being ascertained on at least one of the basis of a comparison, a difference formation, a weighted difference formation and a difference quotient for control action data associated with various times, and in that a change over time is detected on the basis of the ascertained measure of change.
7. The method as claimed in claim 6 , wherein the control action data respectively comprise multiple components, and in that a component-specific measure of change is ascertained for each of the components of the control action data.
8. The method as claimed in claim 1 , wherein at least one of a rate and an amplitude of changes in the sequence of the control action data over time is ascertained, and in that a length of the time window and/or a threshold value for comparison with a measure of change for the changes over time is defined on the basis of at least one the ascertained rate and amplitude.
9. The method as claimed in claim 1 , wherein a target function is provided, which takes at least a portion of the state data as a basis for ascertaining a performance parameter to be optimized during operation of the technical system, and in that the control device is trained to optimize the performance parameter, on the basis of the portion of the state data that is contained in the extracted training data.
10. The method as claimed in claim 1 , wherein a share of the extracted training data in the training data is ascertained, and in that the ascertained share is taken as a basis for ascertaining a measure of quality for the training data and outputting the measure of quality to control the technical system.
11. The method as claimed in claim 1 , wherein the training data and/or the extracted training data are aggregated by an aggregation function, and in that the training is effected on the basis of the aggregated training data.
12. A control device for controlling a technical system, designed to carry out a method as claimed in claim 1 .
13. A computer program product, comprising a computer readable hardware storage device having computer readable program code stored therein, said program code executable by a processor of a computer system to implement a method designed to carry out a method as claimed in claim 1 .
14. A computer-readable storage medium having a computer program product as claimed in claim 13 .