METHOD AND SYSTEM FOR PREDICTING TURBOMACHINERY FAILURE EVENTS EMPLOYING GENETIC ALGORITHM

• US 20090100293A1
• Filed: 10/16/2007
• Published: 04/16/2009
• Est. Priority Date: 10/16/2007
• Status: Active Grant

• Current Assignee: GE Infrastructure Technology, LLC (General Electric Company)
• Original Assignee: General Electric Company
• Inventors: Kiehl, Thomas R., Hoogs, Bethany K., Interrante, John A., LaComb, Christina A., Senturk-Doganaksoy, Deniz
• Granted Patent: US 7,627,454 B2
• Time in Patent Office: Days
• US Class Current: 714/26
• CPC Class Codes: G05B 23/0229 knowledge based, e.g. expert systems; genetic algorithms

Abstract

A method for predicting or detecting an event in turbomachinery includes the steps of obtaining operational data from at least one machine and at least one peer machine. The operational data comprises a plurality of performance metrics. A genetic algorithm (GA) analyzes the operational data, and generates a plurality of clauses, which are used to characterize the operational data. The clauses are evaluated as being either “true” or “false”. A fitness function identifies a fitness value for each of the clauses. A perturbation is applied to selected clauses to create additional clauses, which are then added to the clauses group. The steps of applying a fitness function, selecting a plurality of clauses, and applying a perturbation can be repeated until a predetermined fitness value is reached. The selected clauses are then applied to the operational data from the machine to detect or predict a past, present or future event.

Claims

1. A method for predicting or detecting an event in turbomachinery comprising:

obtaining operational data from at least one machine, said operational data comprising a plurality of performance metrics associated with the operation of said at least one machine, said performance metrics being associated with a plurality of time periods;

obtaining peer operational data from at least one peer machine, said peer operational data comprising a plurality of performance metrics associated with the operation of said at least one peer machine;

determining if said at least one peer machine has experienced said event or has not experienced said event;

employing a genetic algorithm to analyze said operational data and said peer operational data comprising;

generating a plurality of clauses, said clauses used to characterize said operational data, each of said clauses comprising a plurality of alleles, said alleles comprising a count of time periods, at least one performance metric, a comparison operator, a threshold value, and a positive fraction;

evaluating said plurality of clauses as being either “

true”

or “

false”

, a “

true”

evaluation being obtained if for any given number of time periods equal to said count of time periods, at least a fraction of said time periods equal to said positive fraction contain a performance metric which satisfies said comparison operator with respect to said threshold value, and a “

false”

evaluation obtained otherwise;

applying a fitness function to identify a fitness value for each of said clauses, said fitness value determined by the degree to which each of said clauses evaluates as “

true”

when applied to said at least one peer machine for which it is known said event has occurred, and which each of said clauses evaluates as “

(false”

when applied to said at least one peer machine for which it is known said event has not occurred;

selecting a plurality of said clauses having a (greater fitness value than other clauses, those clauses having a greater fitness value forming a selected clauses group;

applying a perturbation to said alleles of said selected clauses to create additional clauses and adding said additional clauses to said selected clauses group;

repeating, at least one of, said applying a fitness function step, selecting a plurality of said clauses step and applying a perturbation step until a predetermined fitness value is reached for said selected clauses;

applying said selected clauses to the operational data from said at least one machine to determine whether the operational data indicates a past, present or future event.

2. The method of claim 1, wherein the event comprises a failure event.

3. The method of claim 2, wherein said at least one machine and said at least one peer machine are chosen from the group comprising:

an engine, a turbomachine, a turbine, a compressor, and a generator.

4. The method of claim 3, wherein the step of obtaining operational data and the step of obtaining peer operational data comprise monitoring the operation of said machines in real-time using a plurality of sensors.

5. The method of claim 1, wherein said performance metrics comprise operating parameters of said at least one machine and said at least one peer machine.

6. The method of claim 1, wherein said fitness value is determined by considering the time periods for which a pattern is detected and a pattern'"'"'s nearness to the time periods in which the event occurred.

7. The method of claim 1, wherein said fitness value is based on the number of times that a pattern is detected for a peer machine that experienced the event vs. a peer machine that did not experience the event.

8. The method of claim 1, wherein said alleles are generally comprised of the form:

<

tag>

<

greater than or less than>

<

value>

for <

n>

out of <

m>

time periods;

wherein,

“

tag”

, is an exceptional anomaly score or an operational metric;

“

greater than or less than”

, is a choice between a “

greater than”

mathematical operator or “

less than”

mathematical operator;

“

value”

, is a numerical or alphanumerical value;

“

n” and

“

m”

, are numerical values representing a specific quantity of time periods.

9. The method of claim 1, wherein said alleles are generally comprised of the form:

<

tag>

<

greater than or equal to or less than or equal to>

<

value>

for <

n>

out of <

m>

time periods;

wherein,

“

tag”

, is an exceptional anomaly score or an operational metric;

“

greater than or equal to or less than or equal to”

, is a choice between a “

greater than or equal to”

mathematical statement or “

less than or equal to”

mathematical statement;

“

value”

, is a numerical or alphanumerical value;

“

n” and

“

m”

, are numerical values representing a specific quantity of time periods.

10. The method of claim 1, wherein said genetic algorithm utilizes a fitness function generally of the form:

$\mathrm{pospercent}=100<br><br>\frac{t<br><br><br><br>p}{t<br><br><br><br>o}$ $\mathrm{negpercent}=100<br><br>\frac{f<br><br><br><br>p}{f<br><br><br><br>o}$ $\mathrm{fitness}=1000<br><br><br><br>{\left(\mathrm{pospercent}+1\right)}^{\frac{1}{\mathrm{negpercent}<br><br>+70}}-1<br><br>\text{)}.$

wherein, tp and to indicate the number of true positives and the total number of positive cases respectively and fp and fo represent the number of false positives and the total number of false cases respectively.

11. A method of predicting or detecting the occurrence of an event for an entity, using a genetic algorithm, comprising:

obtaining operational data for the entity, the operational data comprising measurements of a plurality of performance metrics associated with the operation of the entity, each of said performance metrics being associated with at least one of a plurality of time periods;

obtaining peer operational data for a plurality of peers of the entity for which it is known whether or not the event has occurred for each of the peers, the peer operational data comprising measurements of peer performance metrics;

employing a genetic algorithm to analyze said operational data and said peer operational data comprising;

generating a plurality of clauses, said clauses used to characterize said operational data, each of said clauses comprising a plurality of alleles, said alleles comprising a count of time periods, at least one performance metric, a comparison operator, a threshold value, and a positive fraction;

evaluating said plurality of clauses as being either “

true”

or “

false”

, a “

true”

evaluation being obtained if for any given number of time periods equal to said count of time periods, at least a fraction of said time periods equal to said positive fraction contain a performance metric which satisfies said comparison operator with respect to said threshold value, and a “

false”

evaluation being obtained otherwise;

applying a fitness function to identify a fitness value for each of said clauses, said fitness value determined by the degree to which each of said clauses evaluates as “

true”

when applied to said entity for which it is known said event has occurred, and which each of said clauses evaluates as “

false”

when applied to said peers for which it is known said event has not occurred;

selecting a plurality of said clauses having a greater fitness value than other clauses, those clauses having a greater fitness value forming a selected clauses group;

applying a perturbation to said alleles of said selected clauses to create additional clauses and adding said additional clauses to said selected clauses group;

repeating, at least one of, said applying a fitness function step, selecting a plurality of said clauses step and applying a perturbation step until a predetermined fitness value is reached for said selected clauses;

applying said selected clauses to the operational data from said entity to determine whether the operational data indicates a past, present or future event.

12. The method of claim 11, wherein the event comprises a failure event and the entity is chosen from the group comprising:

an engine, a turbomachine, a turbine, a compressor, and a generator.

13. The method of claim 12, wherein the performance metrics and the peer performance metrics comprise operating parameters of said entity and said peers, respectively.

14. The method of claim 13, wherein the step of collecting operational data on the entity comprises monitoring the operation of said entity in real-time using a plurality of sensors.

15. The method of claim 13, wherein said fitness value is determined by considering the time periods for which a pattern is detected and said pattern'"'"'s nearness to the time periods in which the event occurred.

16. The method of claim 13, wherein said fitness value is based on the number of times that a pattern is detected for a peer that experienced the event vs. a peer that did not experience the event.

17. The method of claim 11, wherein the event comprises a fraudulent financial transaction and the entity comprises a business.

18. The method of claim 17, wherein said peers comprise businesses in a similar Standard Industry Classification as said entity.

19. The method of claim 17, wherein said operational data and said performance metrics comprise financial data.

20. The method of claim 17 wherein the step of obtaining operational data comprises extracting the operational data from publicly available balance sheet information.