System state monitoring using recurrent local learning machine

• US 7,403,869 B2
• Filed: 04/29/2005
• Issued: 07/22/2008
• Est. Priority Date: 11/04/2002
• Status: Active Grant

• Current Assignee: SmartSignal Corporation (Emerson Electric Company)
• Original Assignee: SmartSignal Corporation (Emerson Electric Company)
• Inventors: Xu, Xiao, Wegerich, Stephan W.
• Primary Examiner(s): Barlow, Jr.; John E
• Assistant Examiner(s): Sievers; Lisa C.
• Application Number:
  US11/119,322
• Publication Number:
  US 20050278143A1
• Time in Patent Office: 1,180 Days
• Field of Search: 702/183
• US Class Current: 702/183
• CPC Class Codes: G05B 13/0275 using fuzzy logic only
• CPC Class Codes: G05B 17/02 electric
• CPC Class Codes: G05B 23/0245 based on a qualitative model, e.g. rule based; if-then decisions

Abstract

Reference data observations for given system are used to develop a reference set of such observations. Subsequent observations (comprising, in one embodiment, current observations) are then used to facilitate selection of portions of this reference set to yield a resultant set of observations that serves as a model. This model can then be used in comparison to actual system performance to detect, for example, a trend towards a faulty condition. Pursuant to a preferred approach, the model is recomputed from time to time and, pursuant to a particular embodiment, is recomputed with each new set of current observations.

Claims

1. A method for monitoring a system instrumented with one or more sensors that measure a plurality of variables characterizing performance of the system, said system having at least one additional varying parameter descriptive of its circumstances, comprising the steps of:

acquiring a current observation of the plurality of variables;

generating an estimate of an expected value for at least one of said plurality of variables;

comparing at least one said estimate to its corresponding measured variable; and

storing in computer-readable form in a memory device an indication of the performance of said system based at least in part on a result of the comparing step;

wherein said step of generating an estimate comprises;

providing a reference set of observations characteristic of expected performance of said system, each such reference set observation comprising at least a subset of the plurality of variables, and each such reference set observation further being associated with a value of the at least one additional varying parameter;

providing a measure of the at least one additional varying parameter relating to the acquired current observation;

selecting from said reference set a subset of observations that have an associated value of the at least one additional varying parameter sufficiently close to the measure of the at least one additional parameter relating to the acquired current observation;

forming a model set of observations from the selected subset of observations;

computing said estimate using the acquired current observation and the model set of observations in a similarity-based model; and

repeating these steps for generating an estimate for at least some of subsequently acquired current observations.

2. A method according to claim 1 wherein said step of computing said estimate in a similarity based-model utilizes a similarity operation wherein like elements of two observations are differenced, the absolute value of the difference is divided by an expected range for the element, and the resulting quantity is subtracted from 1;

and the similarity of the two observations is determined as the average of all such similarities determined for elements of the two observations.

3. A method according to claim 1 wherein said step of computing said estimate in a similarity based-model utilizes a similarity operation wherein the norm of the difference of two observations is added to 1;

and the result thereof is inverted.

4. A method according to claim 1 wherein said step of selecting from said reference set a subset of observations comprises selecting a predetermined number of reference set observations that have an associated value of the at least one additional varying parameter closest to the measure of the at least one additional parameter relating to the acquired current observation.

5. A method according to claim 1 wherein said step of selecting from said reference set a subset of observations comprises selecting those reference set observations that have an associated value of the at least one additional varying parameter lying within a predetermined range of the measure of the at least one additional parameter relating to the acquired current observation.

6. A method according to claim 1 wherein said step of forming a model set of observations from the selected subset of observations further comprises choosing a smaller subset of the selected subset of observations that have sufficient similarity to the acquired current observation, and forming the model set from the smaller subset.

7. A method according to claim 1 wherein said step of computing said estimate in a similarity based-model comprises using as input to the similarity-based model a modified form of the acquired current observation wherein an element corresponding to at least one of the plurality of variables is removed, and the removed variable is estimated by the similarity-based model.

8. A method according to claim 1 wherein said step of comparing at least one said estimate to its corresponding measured variable comprises differencing the estimate and its corresponding measured variable to generate a residual that is compared to a threshold for a determination at least in part of the presence of a fault in the performance of said system.

9. A method according to claim 1 wherein said step of comparing at least one said estimate to its corresponding measured variable comprises differencing the estimate and its corresponding measured variable to generate a residual, and wherein successive residuals are tested with a statistical hypothesis test for a determination at least in part of the presence of a fault in the performance of said system.

10. A method for monitoring a system instrumented with one or more sensors that measure a plurality of variables characterizing performance of the system, comprising the steps of:

acquiring a current observation of the plurality of variables;

generating an estimate of an expected value for at least one of said plurality of variables;

comparing at least one said estimate to its corresponding measured variable; and

storing in computer-readable form in a memory device an indication of the performance of said system based at least in part on a result of the comparing step;

wherein said step of generating an estimate comprises;

providing a reference set of observations characteristic of expected performance of said system, each such reference set observation comprising at least a subset of the plurality of variables;

selecting from said reference set a subset of observations that have a similarity to the acquired current observation above a specified threshold;

forming a model set of observations from the selected subset of observations;

computing said estimate using the acquired current observation and the model set of observations in a similarity-based model; and

repeating these steps for generating an estimate for at least some of subsequently acquired current observations.

11. A method according to claim 10 wherein said step of computing said estimate in a similarity based-model utilizes a similarity operation wherein like elements of two observations are differenced, the absolute value of the difference is divided by an expected range for the element, and the resulting quantity is subtracted from 1;

and the similarity of the two observations is determined as the average of all such similarities determined for elements of the two observations.

12. A method according to claim 10 wherein said step of computing said estimate in a similarity based-model utilizes a similarity operation wherein the norm of the difference of two observations is added to 1;

and the result thereof is inverted.

13. A method according to claim 10 wherein said step of selecting from said reference set a subset of observations that have a similarity to the acquired current observation above a specified threshold comprises generating a similarity value between a reference set observation and the acquired current observation, using only a subset of the corresponding elements of the two observations.

14. A method according to claim 10 wherein said step of computing said estimate in a similarity based-model comprises using as input to the similarity-based model a modified form of the acquired current observation wherein an element corresponding to at least one of the plurality of variables is removed, and the removed variable is estimated by the similarity-based model.

15. A method according to claim 10 wherein said step of comparing at least one said estimate to its corresponding measured variable comprises differencing the estimate and its corresponding measured variable to generate a residual that is compared to a threshold for a determination at least in part of the presence of a fault in the performance of said system.

16. A method according to claim 10 wherein said step of comparing at least one said estimate to its corresponding measured variable comprises differencing the estimate and its corresponding measured variable to generate a residual, and wherein successive residuals are tested with a statistical hypothesis test for a determination at least in part of the presence of a fault in the performance of said system.

17. An apparatus for monitoring a system instrumented with one or more sensors that measure a plurality of variables characterizing performance of the system, comprising:

means for acquiring a current observation of the plurality of variables;

a memory for storing a reference set of observations characteristic of expected performance of said system, each such reference set observation comprising at least a subset of the plurality of variables;

a processor-executed module for selecting a subset of said reference set of observations having sufficient similarity to the acquired current observation and forming a model set of observations from the selected subset of observations;

a processor-executed module for generating an estimate of an expected value for at least one of said plurality of variables using the acquired current observation and the model set of observations in a similarity-based model;

a processor-executed module for comparing at least one said estimate to its corresponding measured variable to form a residual and determining an indication of the performance of said system based at least in part thereon; and

a processor-executed module for examining the acquired current observation and for augmenting the reference set of observations stored in said memory with the data of the acquired current observation when at least one value from the acquired current observation is outside of a specified range, by adding the acquired current observation to the reference set of observations.

18. An apparatus according to claim 17 wherein said processor-executed module for selecting a subset of said reference set selects a specified number of reference set observations to constitute the model set, that have the highest similarity to the acquired current observation.

19. An apparatus according to claim 17 wherein said processor-executed module for selecting a subset of said reference set selects those reference set observations to constitute the model set, that have a similarity to the acquired current observation above a specified threshold.

20. An apparatus according to claim 17 wherein said processor-executed module for augmenting the reference set of observations examines further removes an observation from the reference set of observations upon adding the acquired current observation.

21. An apparatus according to claim 17 wherein said a processor-executed module for comparing at least one said estimate to its corresponding measured variable further comprises a diagnostic module for examining patterns of residuals formed from differencing estimates and corresponding measured variables for at least a subset of the plurality of variables, and determines a presence of a failure mode of said system as an indication of the performance of said system.

22. A apparatus according to claim 17 wherein said processor-executed module for selecting a subset of said reference set uses a similarity value between a reference set observation and the acquired current observation, using only a subset of the corresponding elements of the two observations.

23. An apparatus according to claim 17 wherein said processor-executed module for generating an estimate of an expected value for at least one of said plurality of variables computes said estimate in a similarity based-model using as input to the similarity-based model a modified form of the acquired current observation wherein an element corresponding to at least one of the plurality of variables is removed, and the removed variable is estimated by the similarity-based model.

24. An apparatus according to claim 17 further comprising a processor-executed module for rendering a web browser viewable display of at least one of the set of said estimate, said residual and said indication of the performance of said system.

25. An apparatus according to claim 17 further comprising a processor-executed module for making accessible to other software programs at least one of the set of said estimate, said residual and said indication of the performance of said system.