Dynamic, resilient virtual sensing system and shadow controller for cyber-attack neutralization
An industrial asset may have monitoring nodes (e.g., sensor or actuator nodes) that generate current monitoring node values. An abnormality detection and localization computer may receive the series of current monitoring node values and output an indication of at least one abnormal monitoring node that is currently being attacked or experiencing a fault. An actor-critic platform may tune a dynamic, resilient state estimator for a sensor node and output tuning parameters for a controller that improve operation of the industrial asset during the current attack or fault. The actor-critic platform may include, for example, a dynamic, resilient state estimator, an actor model, and a critic model. According to some embodiments, a value function of the critic model is updated for each action of the actor model and each action of the actor model is evaluated by the critic model to update a policy of the actor-critic platform.
1. A system to protect an industrial asset, comprising:
a plurality of monitoring nodes, each monitoring node generating a series of current monitoring node values over time that represent a current operation of the industrial asset;
an abnormality detection and localization computer to receive the series of current monitoring node values and to output an indication of at least one abnormal monitoring node that is currently being attacked or experiencing a fault;
a controller of the industrial asset; and
an actor-critic platform, coupled to the abnormality detection and localization computer and the controller, to tune a dynamic, resilient state estimator for a sensor node and to output tuning parameters for the controller that improve operation of the industrial asset during the current attack or fault, including:
the dynamic, resilient state estimator,
an actor model that tunes the parameters for the controller as appropriate, and
a critic model, wherein a value function of the critic model is updated for each action of the actor model and each action of the actor model is evaluated by the critic model to update a policy of the actor-critic platform.
2. The system of claim 1 , wherein the controller is based on substantially real-time optimization.
3. The system of claim 1 , wherein the controller comprises an adaptive model predictive controller.
4. The system of claim 1 , wherein the abnormal monitoring node is associated with an actuator or controller.
5. The system of claim 1 , wherein the actor-critic platform uses reinforcement learning to reconstruct a full state using information from monitoring nodes that are not abnormal.
6. The system of claim 5 , wherein the reinforcement learning is implemented via a deep Q network or any other type of reinforcement learning.
7. The system of claim 1 , wherein the dynamic, resilient state estimator is trained off-line using a high-fidelity model of the industrial asset.
8. The system of claim 7 , wherein the dynamic, resilient state estimator is updated on-line via continuous learning during operation of the industrial asset.
9. The system of claim 1 , wherein the tuning parameters are associated with at least one of: (i) optimization constraints, (ii) set points, and (iii) weights.
10. The system of claim 1 , wherein the actor-critic platform uses reinforcement learning, implemented via a deep Q network or any other type of reinforcement learning, to reconstruct a full state using information from monitoring nodes that are not abnormal, the critic model updates weights of the deep Q network, and the actor model updates parameters of the controller.
11. The system of claim 10 , wherein the parameters of the controller updated by the actor model are associated with at least one of: (i) weighting matrices, (ii) constraint types, (iii) constraint values, and (iv) key plant model parameters.
12. The system of claim 1 , wherein the dynamic, resilient state estimator is associated with at least one of: (i) an indexed selector, (ii) bumpless transfer control, and (iii) a proportional-integral-derivative controller, (iv) a switched dynamic control, (v) a smooth transition controller, and (vi) an adaptive controller.
13. The system of claim 1 , wherein the industrial asset is associated with at least one of: (i) a turbine, (ii) a gas turbine, (iii) a wind turbine, (iv) an engine, (v) a jet engine, (vi) a locomotive engine, (vii) a refinery, (viii) a power grid, (ix) a dam, and (x) an autonomous vehicle.
14. A computerized method to protect an industrial asset associated with a plurality of monitoring nodes, each monitoring node generating a series of current monitoring node values over time that represent current operation of the industrial asset, comprising:
receiving, by an abnormality detection and localization computer, the series of current monitoring node values;
outputting, by the abnormality detection and localization computer, an indication of at least one abnormal monitoring node that is currently being attacked or experiencing a fault;
tuning, by an actor-critic platform coupled to the abnormality detection and localization computer and a controller, a dynamic, resilient state estimator for a sensor node; and
outputting, by the actor-critic platform, tuning parameters for a controller that improve operation of the industrial asset during the current attack or fault,
wherein the actor-critic platform includes the dynamic, resilient state estimator, an actor model, and a critic model such that a value function of the critic model is updated for each action of the actor model and each action of the actor model is evaluated by the critic model to update a policy of the actor-critic platform.
15. The method of claim 14 , wherein the actor-critic platform uses reinforcement learning to reconstruct a full state using information from monitoring nodes that are not abnormal.
16. The method of claim 15 , wherein the reinforcement learning is implemented via a deep Q network or any other type of reinforcement learning.
17. The method of claim 14 , wherein the dynamic, resilient state estimator is trained off-line using a high-fidelity model of the industrial asset.
18. The method of claim 17 , wherein the dynamic, resilient state estimator is updated on-line via continuous learning during operation of the industrial asset.
19. The method of claim 14 , wherein the tuning parameters are associated with at least one of: (i) optimization constraints, (ii) set points, and (iii) weights.
20. A non-transitory, computer-readable medium storing instructions that, when executed by a computer processor, cause the computer processor to perform a method to protect an industrial asset associated with a plurality of monitoring nodes, each monitoring node generating a series of current monitoring node values over time that represent current operation of the industrial asset, the method comprising:
receiving, by an abnormality detection and localization computer, the series of current monitoring node values;
outputting, by the abnormality detection and localization computer, an indication of at least one abnormal monitoring node that is currently being attacked or experiencing a fault;
tuning, by an actor-critic platform coupled to the abnormality detection and localization computer and a controller, a dynamic, resilient state estimator for a sensor node; and
outputting, by the actor-critic platform, tuning parameters for a controller that improve operation of the industrial asset during the current attack or fault,
wherein the actor-critic platform includes the dynamic, resilient state estimator, an actor model, and a critic model such that a value function of the critic model is updated for each action of the actor model and each action of the actor model is evaluated by the critic model to update a policy of the actor-critic platform.
21. The medium of claim 20 , wherein the actor-critic platform uses reinforcement learning, implemented via a deep Q network or any other type of reinforcement learning, to reconstruct a full state using information from monitoring nodes that are not abnormal, the critic model updates weights of the deep Q network, and the actor model updates parameters of the controller.
22. The medium of claim 21 , wherein the parameters of the controller updated by the actor model are associated with at least one of: (i) weighting matrices, (ii) constraint types, (iii) constraint values, and (iv) key plant model parameters.