Method and computer system for generating a decision logic for a controller
Machine learning, ML, using labeled data is performed to thereby train a ML model to generate decision logic for a controller of an IACS, in particular a power system controller. The ML model has ML model inputs and ML model outputs. The ML model is decomposed into a plurality of computational sub-processes, wherein intermediate signals output by one of the computational sub-processes are input into a consecutive one of the computational sub-processes. Information on the trained ML model is generated based on the intermediate signals and is output for interpretation, verification, visualization, and/or inspection of at least one of the computational sub-processes.
1. A method of generating a decision logic for a controller of an Industrial Automation Control System, IACS, in particular for a controller for power system protection or power system control, the method comprising:
performing machine learning, ML, using labeled data to thereby train a ML model to generate the decision logic, wherein
the ML model has ML model inputs and ML model outputs, and
the ML model is decomposed by introducing one or more checkpoints into the ML model to divide the ML model into a plurality of computational sub-processes, wherein, at each checkpoint, intermediate signals output by one of the computational sub-processes are input into a subsequent one of the computational sub-processes, and wherein, at at least one checkpoint, a total number of the intermediate signals is equal to or less than three or when the total number of the intermediate signals is greater than three, the at least one integrated circuit is adapted to perform a projection of a decision boundary into a two-or three-dimensional space;
generating and outputting a visual representation of values of the intermediate signals of at least one of the computational sub-processes;
generating the decision logic for the controller using the ML model;
deploying the decision logic to the controller for execution; and
executing, by the controller, the decision logic during operation of the power system.
2. The method of claim 1 , further comprising:
controlling an optical output device to output the visual representation.
3. The method of claim 2 , wherein the visual representation depends on at least one decision boundary or numerical function of the decision logic as a function of the intermediate signals.
4. The method of claim 2 , wherein the decision logic is a classifier that determines which one of at least two different control actions is to be performed, and wherein the visual representation visualizes decision boundaries separating the control actions.
5. The method of claim 1 , wherein the three or less intermediate signals are three or less real-valued scalars.
6. The method of claim 2 , wherein the method further comprises performing a linear or nonlinear projection of a decision boundary into a two-or three-dimensional space, wherein the visual representation visualizes the projection of the decision boundary.
7. The method of claim 6 , further comprising
selectively modifying the ML model, including removing and/or replacing at least part of a computational graph, and
repeating the supervised ML for the modified ML model.
8. The method of claim 1 , wherein the ML is performed iteratively by learning in stages.
9. The method of claim 1 , wherein the ML model receives plural sets of ML model inputs, wherein the method further comprises generating aggregated information from the plural sets of ML model inputs and using the aggregated information in combination with the intermediate signals to generate the visual representation.
10. The method of claim 9 , wherein the plural sets of ML model inputs are grouped according to time or according to signal type.
11. The method of claim 9 , wherein in dependence on a type of the decision logic, the plural sets of ML model inputs are grouped by signal type to provide the input of the ML model or grouped by time to provide the input of the ML model.
12. The method of claim 1 , wherein the ML model includes at least one artificial neural network.
13. The method of claim 1 , wherein the decision logic is a power system protection logic.
14. The method of claim 1 , wherein the ML model outputs correspond to control actions for immediate trip, trip with delay, and restrain, and wherein the ML model inputs include at least one of the following:
3-phase admittances;
3-phase impedances;
phasors, including voltages and currents;
raw voltages and/or currents;
superimposed voltages and/or currents;
voltages and/or currents filtered using a digital filter;
alpha-beta transformed voltages and/or currents;
DQ-transformed voltages and/or currents;
Fourier transformed voltages and/or currents;
wavelet transformed voltages and/or currents.
15. The method of claim 1 , wherein a first one of the plurality of computational sub-processes comprises a first artificial neural network, and wherein a second one of the plurality of computational sub-processes comprises a second artificial neural network that is different from the first artificial neural network.
16. The method of claim 1 , wherein the decision logic is a decision logic for a digital protection relay.
17. A computer system for generating a decision logic for a controller of an Industrial Automation Control System, IACS, in particular for a controller for power system protection or power system control, the computer system comprising:
at least one integrated circuit adapted to:
perform machine learning, ML, using labeled data to thereby train a ML model to generate the decision logic, wherein
the ML model has ML model inputs and ML model outputs, and
the ML model is decomposed by introducing one or more checkpoints into the ML model to divide the ML model into a plurality of computational sub-processes, wherein, at each checkpoint, intermediate signals output by one of the computational sub-processes are input into a subsequent one of the computational sub-processes, and wherein, at at least one checkpoint, a total number of the intermediate signals is equal to or less than three or when the total number of the intermediate signals is greater than three, the at least one integrated circuit is adapted to perform a projection of a decision boundary into a two-or three-dimensional space;
generate and output a visual representation of values of the intermediate signals of at least one of the computational sub-processes;
generate the decision logic for the controller using the ML model; and
deploy the decision logic to at least one integrated circuit of the controller that is adapted to execute the decision logic during operation of the power system; and
an output interface adapted to output the visual representation.
18. An Industrial Automation Control System, IACS, in particular a power system, comprising:
a controller adapted to execute a decision logic to decide which one of a plurality control actions must be taken; and
a computer system for generating the decision logic, the computer system comprising:
at least one integrated circuit adapted to:
perform machine learning, ML, using labeled data to thereby train a ML model to generate the decision logic, wherein
the ML model has ML model inputs and ML model outputs, and
the ML model is decomposed by introducing one or more checkpoints into the ML model to divide the ML model into a plurality of computational sub-processes, wherein, at each checkpoint, intermediate signals output by one of the computational sub-processes are input into a subsequent one of the computational sub-processes, and wherein, at at least one checkpoint, a total number of the intermediate signals is equal to or less than three or when the total number of the intermediate signals is greater than three, the at least one integrated circuit is adapted to perform a projection of a decision boundary into a two-or three-dimensional space;
generate and output a visual representation of values of the intermediate signals of at least one of the computational sub-processes;
generate the decision logic for the controller using the ML model; and
deploy the decision logic to the controller; and
an output interface adapted to output the visual representation.