Method and system of monitoring prognostics

• US 7,917,335 B2
• Filed: 11/18/2008
• Issued: 03/29/2011
• Est. Priority Date: 09/26/2005
• Status: Active Grant

• Current Assignee: Lockheed Martin Corporation (Martin Marietta Corporation)
• Original Assignee: Lockheed Martin Corporation (Martin Marietta Corporation)
• Inventors: Turischev, Alexey, Harrison, Gregory A.
• Primary Examiner(s): Raymond; Edward
• Application Number:
  US12/272,926
• Publication Number:
  US 20090070282A1
• Time in Patent Office: 861 Days
• Field of Search: 702182-185, 702/188
• US Class Current: 702/183
• CPC Class Codes: G05B 23/024 Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks
• CPC Class Codes: G06N 3/02 Neural networks

Abstract

A neural network learns the operating modes of a system being monitored under normal operating conditions. Anomalies can be automatically detected and learned. A control command can be issued or an alert can be issued in response thereto.

Claims

1. A method comprising:

sensing periodic time based indicia;

sensing a plurality of sets of conditions associated with operation of a selected system at spaced apart timed intervals;

generating representations of the sensed conditions;

processing the representations along with at least some of the time based indicia, with at least one neural network with the network exhibiting a plurality of learned states with each state corresponding to a respective set of sensed conditions; and

combining results of the processing, and responsive to the combining,

producing at least one of an alarm indicator, or, a control signal,

where processing includes comparing time evolution of sensed conditions with learned time evolution of operation of the selected system.

2. A method as in claim 1 where the sensed state of the sensors at respective intervals changes over time to create a trajectory of system states through time that is learned by the neural network.

3. A method as in claim 2 where the trajectory of states through time adapts in response to repeated training with new time-based training data to learn a system state trajectory that comprises a path that describes an average state trajectory within selected ranges, thereby creating a footprint in the memory of the system state trajectory through time.

4. A method as in claim 3 which includes, a learned state trajectory path able to monitor a state trajectory of a system in operation, and providing an indication of deviation from the previously learned trajectory path.

5. A method as in claim 4 which includes detecting and monitoring anomalies, reporting the anomalies to a user, and causing control actions to take place.

6. A method as in claim 5 which includes displaying the current set of system state parameters and their expected ranges as time progresses.

7. A method as in claim 6 which includes learning trajectories of aircraft in various flight patterns and indicating deviances from the expected flight plan.

8. A method as in claim 6 that includes learning expected time-based behavior of a plurality of ocean states at distributed locations and detecting anomalies in ocean rhythms that may be indicative of tsunami conditions or other unexpected ocean conditions.

9. A method comprising:

providing a learning mode to train at least one neural network in the expected,

normal, operation of a movable vehicle;

sensing a plurality of vehicle inputs, the values of which are indicative of normal operation of the vehicle, and responsive thereto establishing an expected, normal operational profile of the vehicle in the at least one neural network;

providing a detection mode for the at least one neural network; and

sensing the plurality of vehicle inputs, the values of which are indicative of real-time vehicle operation and, responsive thereto, determining, using the at least one neural network and the previously established normal operational profile, if the present values are indicative of normal operation of the vehicle.

10. A method as in claim 9 which includes generating an anomaly indicator where the present values are not indicative of normal operation of the vehicle.

11. A method as in claim 10 which includes providing periodic time based indicia.

12. A method as in claim 11 which includes combining periodic time based indicia with vehicle inputs in the learning mode.

13. A method as in claim 12 which includes basing the normal operational profile, at least in part, on the combined periodic time based indicia.

14. A method as in claim 13 which includes training a plurality of neural networks in the expected, normal, operation of a movable vehicle.

15. A method as in claim 14 which includes allocating members of the plurality of vehicle inputs to different members of the plurality of neural networks.

16. A method as in claim 15 which includes generating at least one vehicle control signal in response to having generated the anomaly indicator.

17. A method as in claim 9 which includes generating at least one vehicle control signal in response to determining that the present values are indicative of normal operation of the vehicle.

18. A method as in claim 17 which includes providing periodic time based indicia and combining periodic time based indicia with vehicle inputs in the learning mode.

19. A method comprising:

providing a learning mode to train at least one neural network in the expected,

normal, behavior of a selected entity;

sensing a plurality of time varying inputs associated with the expected behavior of the entity, and responsive thereto establishing an expected, normal behavior profile of the entity in the at least one neural network;

providing a sequential plurality of time based indicia;

providing a detection mode for the at least one neural network; and

sensing the plurality of time varying inputs, the values of which are indicative of real-time behavior of the entity, and the time based indicia, and, responsive thereto, determining, using the at least one neural network and the previously established normal behavior profile, if the present values are indicative of expected, normal behavior of the entity.

20. A method as in claim 19 which includes generating an anomaly indicator where the present values are not indicative of normal behavior of the entity.

21. A method as in claim 19 which includes combining periodic time based indicia with entity inputs in the learning mode.

22. A method as in claim 21 which includes basing the normal behavior profile, at least in part, on the combined periodic time based indicia.

23. A method as in claim 22 which includes training a plurality of neural networks in the expected, normal, behavior of the entity.

24. A method as in claim 23 where the entity is selected from a class which includes at least one of a movable vehicle, a body of water, or a gas filled region.