Method and architecture for AI-assisted supervision for a controlled system
Embodiments described herein provide a supervisor for fault management at a production system. During operation, the supervisor can obtain a set of sensor readings and a state of the production system. A respective sensor reading is an output of a sensor in the production system. The supervisor can then determine, using an artificial intelligence (AI) model, whether the set of sensor readings accommodates a fault associated with a corresponding sensor. Subsequently, the supervisor can determine an action that mitigates an effect of the fault and modify the set of sensor readings based on the action. Here, the modified set of sensor readings is used by a controller that controls the production system.
1. A method for fault management at a production system, the method comprising:
obtaining, by a computer system, a set of sensor readings and a state of the production system, wherein a respective sensor reading is an output of a sensor in the production system;
providing the set of sensor readings and the state of the production system to a Reinforcement Learning (RL) agent trained based on data indicating physical states of the production system;
determining, using the RL agent, whether the set of sensor readings includes at least a faulty reading produced by a corresponding faulty sensor;
determining an action that adjusts a value of the faulty reading; and
modifying the set of sensor readings based on the action, wherein the modified set of sensor readings is used as an input to a controller that controls the production system.
2. The method of claim 1 , further comprising identifying the faulty sensor;
wherein adjusting the value of the faulty reading further comprises adding a value to a sensor reading of the faulty sensor.
3. The method of claim 1 , wherein determining the action further comprises:
determining a reference signal for the production system, wherein the reference signal indicates a target behavior of the production system; and
determining the action such that the controller maintains the target behavior of the production system.
4. The method of claim 3 , wherein the state includes an adjusted state generated by the controller for maintaining the target behavior of the production system, and wherein the adjusted state is generated based on a previously modified set of sensor readings.
5. The method of claim 1 , wherein the production system comprises a plurality of sensors measuring a same element; and
wherein the method further comprises combining respective sensor readings from the plurality of sensors based on a fusion scheme.
6. The method of claim 1 , wherein training of the RL agent is agnostic to system dynamics of the production system.
7. The method of claim 1 , wherein the RL agent is trained to learn a policy that includes detecting a fault associated with the faulty sensor at a time of occurrence and incorporating the action.
8. The method of claim 7 , wherein the policy further includes one or more of:
determining a magnitude of the fault; and
identifying the faulty sensor.
9. The method of claim 7 , wherein the fault includes one or more of:
a time-varying fault that appears and disappears during operation of the production system; and
a dynamic fault that changes during operation of the production system.
10. A non-transitory computer-readable storage medium storing instructions that when executed by a computer cause the computer to perform a method for fault management at a production system, the method comprising:
obtaining a set of sensor readings and a state of the production system, wherein a respective sensor reading is an output of a sensor in the production system;
providing the set of sensor readings and the state of the production system to a Reinforcement Learning (RL) agent trained based on data indicating physical states of the production system;
determining, using the RP agent, whether the set of sensor readings includes at least a faulty reading produced by a corresponding faulty sensor;
determining an action that adjusts a value of the faulty reading; and
modifying the set of sensor readings based on the action, wherein the modified set of sensor readings is used as an input to a controller that controls the production system.
11. The non-transitory computer-readable storage medium of claim 10 , wherein the method further comprises identifying the faulty sensor;
wherein adjusting the value of the faulty reading further comprises adding a value to a sensor reading of the faulty sensor.
12. The non-transitory computer-readable storage medium of claim 10 , wherein determining the action further comprises:
determining a reference signal for the production system, wherein the reference signal indicates a target behavior of the production system; and
determining the action such that the controller maintains the target behavior of the production system.
13. The non-transitory computer-readable storage medium of claim 12 , wherein the state includes an adjusted state generated by the controller for maintaining the target behavior of the production system, and wherein the adjusted state is generated based on a previously modified set of sensor readings.
14. The non-transitory computer-readable storage medium of claim 10 , wherein the production system comprises a plurality of sensors measuring a same element; and
wherein the method further comprises combining respective sensor readings from the plurality of sensors based on a fusion scheme.
15. The non-transitory computer-readable storage medium of claim 10 , wherein training of the RL agent is agnostic to system dynamics of the production system.
16. The non-transitory computer-readable storage medium of claim 10 , wherein the RL agent is trained to learn a policy that includes one or more of:
detecting a fault associated with the faulty sensor at a time of occurrence;
incorporating the action;
determining a magnitude of the fault; and
identifying the faulty sensor.
17. The non-transitory computer-readable storage medium of claim 16 , wherein the fault includes one or more of:
a time-varying fault that appears and disappears during operation of the production system; and
a dynamic fault that changes during operation of the production system.
18. A computer system, comprising:
a storage device;
a processor;
a non-transitory computer-readable storage medium storing instructions, which when executed by the processor causes the processor to perform a method for fault management at a production system, the method comprising:
obtaining a set of sensor readings and a state of the production system, wherein a respective sensor reading is an output of a sensor in the production system;
providing the set of sensor readings and the state of the production system to a Reinforcement Learning (RL) agent trained based on data indicating physical states of the production system;
determining, using the RL agent, whether the set of sensor readings includes at least a faulty reading produced by a corresponding faulty sensor;
determining an action that adjusts a value of the faulty reading; and
modifying the set of sensor readings based on the action, wherein the modified set of sensor readings is used as an input to a controller that controls the production system.
19. The computer system of claim 18 , wherein the method further comprises identifying the faulty sensor;
wherein adjusting the value of the faulty reading further comprises adding a value to a sensor reading of the faulty sensor.
20. The computer system of claim 18 , wherein a fault associated with the faulty sensor includes one or more of:
a time-varying fault that appears and disappears during operation of the production system; and
a dynamic fault that changes during operation of the production system.