Generalized pattern recognition for fault diagnosis in machine condition monitoring

• US 8,886,574 B2
• Filed: 06/12/2012
• Issued: 11/11/2014
• Est. Priority Date: 06/12/2012
• Status: Active Grant

• Current Assignee: Siemens AG
• Original Assignee: Siemens AG
• Inventors: Yuan, Chao, Chakraborty, Amit, Wiebking, Leif, Hackstein, Holger
• Primary Examiner(s): Gaffin, Jeffrey A
• Assistant Examiner(s): BUSS, BENJAMIN J
• Application Number:
  US13/494,234
• Publication Number:
  US 20130332773A1
• Time in Patent Office: 882 Days
• Field of Search: 706/12, 706/904
• US Class Current: 706/12
• CPC Class Codes: G05B 23/0205 by means of a monitoring system capable of detecting and responding to faults
• CPC Class Codes: G05B 23/0229 knowledge based, e.g. expert systems; genetic algorithms
• 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 20/00 Machine learning

Abstract

A generalized pattern recognition is used to identify faults in machine condition monitoring. Pattern clusters are identified in operating data. A classifier is trained using the pattern clusters in addition to annotated training data. The operating data is also used to cluster the signals in the operating data into signal clusters. Monitored data samples are then classified by evaluating confidence vectors that include substitutions of signals contained in the training data by signals in the same signal clusters as the signals contained in the training data.

Claims

1. A method of machine condition monitoring, comprising:

receiving, by a computer, historic operating data including operating data from O signals over time;

extracting I patterns x, each pattern x being extracted from operating data from an individual signal;

clustering the I patterns into K pattern clusters c_k based on similarities;

clustering the O signals into R signal clusters based on correlations among the operating data from the O signals;

receiving an annotated training data sample containing training data from N signals selected from the O signals, the training data having at least one marked failure time period;

creating a K×

N confidence matrix containing K confidence values for each of the N signals, each confidence value representing a confidence that a pattern x extracted from data in the marked failure time period of a signal belongs to one of the K pattern clusters;

training a classifier using the K×

N confidence matrix;

receiving, by a computer, a monitored data sample including monitored data from the O signals; and

classifying, by the classifier running on a computer, the monitored data sample as indicating or not indicating a failure based on confidence vectors computed for a plurality of sub-combinations of the monitored data sample, the sub-combinations each having data representing all of the N signals, at least one of the N signals being represented in at least one sub-combination by data from another signal in the same signal cluster c_k as the represented signal.

2. A method as in claim 1, wherein extracting I patterns from the individual signals in the operating data comprises using a sliding time window of fixed length.

3. A method as in claim 1, wherein clustering the I patterns into K pattern clusters and clustering the O signals into R signal clusters further comprises:

for each of the I patterns, extracting scalar features describing the pattern; and

using the scalar features to cluster the patterns and the signals.

4. A method as in claim 1, wherein clustering the I patterns into K pattern clusters is performed across operating data from all O signals.

5. A method as in claim 1, wherein clustering the I patterns into K pattern clusters is performed across operating data from signals in a same signal cluster.

6. A method as in claim 1, wherein clustering the I patterns into K pattern clusters comprises applying K-mean clustering to form the clusters.

7. A method as in claim 1, wherein clustering the O signals into R signal clusters further comprises:

calculating pair-wise signal correlation based on the historic operating data; and

clustering the signals using hierarchical clustering.

8. A method as in claim 1, wherein one of the K confidence values for a particular signal equals 1, and the remaining of the K confidence values equal zero.

9. A method as in claim 1, wherein, for each of the N signals, a sum of the K confidence values equals 1.

10. A method as in claim 1, wherein classifying the monitored data sample further comprises:

applying a greedy algorithm to identify a dominant confidence vector; and

determining whether an evaluation function value for the dominant confidence vector exceeds a threshold.

11. A method of machine condition monitoring, comprising:

receiving, by a computer, historic operating data including operating data from O signals over time;

extracting I patterns x, each pattern x being extracted from operating data from an individual signal;

clustering the I patterns into K pattern clusters c_k based on similarities;

clustering the O signals into R signal clusters based on correlations among the operating data from the O signals;

receiving an annotated training data sample containing training data from N signals selected from the O signals, the training data having at least one marked failure time period;

creating a K×

N confidence matrix containing K confidence values for each of the N signals, each confidence value representing a confidence that a pattern x extracted from data in the marked failure time period of a signal belongs to one of the K pattern clusters;

training a classifier using the K×

N confidence matrix;

receiving, by a computer, a monitored data sample including monitored data from the O signals; and

classifying, by the classifier running on a computer, the monitored data sample as indicating or not indicating a failure based on confidence vectors computed for a plurality of sub-combinations of the monitored data sample, the sub-combinations each having data representing all of the N signals, at least one of the N signals being represented in at least one sub-combination by data from another signal in the same signal cluster c_k as the represented signal;

wherein each confidence value representing a confidence that pattern x belongs to pattern cluster c_k is defined by;

12. A non-transitory computer-usable medium having computer readable instructions stored thereon for execution by a processor to perform a method of machine condition monitoring, the method comprising:

receiving historic operating data including operating data from O signals over time;

extracting I patterns x, each pattern x being extracted from operating data from an individual signal;

clustering the I patterns into K pattern clusters c_k based on similarities;

clustering the O signals into R signal clusters based on correlations among the operating data from the O signals;

receiving an annotated training data sample containing training data from N signals selected from the O signals, the training data having at least one marked failure time period;

creating a K×

N confidence matrix containing K confidence values for each of the N signals, each confidence value representing a confidence that a pattern x extracted from data in the marked failure time period of a signal belongs to one of the K pattern clusters;

training a classifier using the K×

N confidence matrix;

receiving a monitored data sample including monitored data from the O signals; and

classifying, by the classifier, the monitored data sample as indicating or not indicating a failure based on confidence vectors computed for a plurality of sub-combinations of the monitored data sample, the sub-combinations each having data representing all of the N signals, at least one of the N signals being represented in at least one sub-combination by data from another signal in the same signal cluster c_k as the represented signal.

13. A non-transitory computer-usable medium as in claim 12, wherein extracting I patterns from the individual signals in the operating data comprises using a sliding time window of fixed length.

14. A non-transitory computer-usable medium as in claim 12, wherein clustering the I patterns into K pattern clusters and clustering the O signals into R signal clusters further comprises:

for each of the I patterns, extracting scalar features describing the pattern; and

using the scalar features to cluster the patterns and the signals.

15. A non-transitory computer-usable medium as in claim 12, wherein clustering the I patterns into K pattern clusters is performed across operating data from all O signals.

16. A non-transitory computer-usable medium as in claim 12, wherein clustering the I patterns into K pattern clusters is performed across operating data from signals in a same signal cluster.

17. A non-transitory computer-usable medium as in claim 12, wherein clustering the I patterns into K pattern clusters comprises applying K-mean clustering to form the clusters.

18. A non-transitory computer-usable medium as in claim 12, wherein clustering the O signals into R signal clusters further comprises:

calculating pair-wise signal correlation based on the historic operating data; and

clustering the signals using hierarchical clustering.

19. A non-transitory computer-usable medium as in claim 12, wherein each confidence value representing a confidence that pattern x belongs to pattern cluster c_k is defined by:

20. A non-transitory computer-usable medium as in claim 12, wherein one of the K confidence values for a particular signal equals 1, and the remaining of the K confidence values equal zero.

21. A non-transitory computer-usable medium as in claim 12, wherein, for each of the N signals, a sum of the K confidence values equals 1.

22. A non-transitory computer-usable medium as in claim 12, wherein classifying the monitored data sample further comprises:

applying a greedy algorithm to identify a dominant confidence vector; and

determining whether an evaluation function value for the dominant confidence vector exceeds a threshold.