Principal component analysis based fault classification

• US 7,447,609 B2
• Filed: 12/31/2003
• Issued: 11/04/2008
• Est. Priority Date: 12/31/2003
• Status: Active Grant

• Current Assignee: Honeywell International Inc.
• Original Assignee: Honeywell International Inc.
• Inventors: Guralnik, Valerie, Foslien, Wendy K
• Primary Examiner(s): Lau; Tung S
• Assistant Examiner(s): Bhat; Aditya S
• Application Number:
  US10/750,222
• Publication Number:
  US 20050149297A1
• Time in Patent Office: 1,770 Days
• Field of Search: 702/179
• US Class Current: 702/179
• CPC Class Codes: G05B 23/0254 based on a quantitative model, e.g. mathematical relationships between inputs and outputs; functions: observer, Kalman filter, residual calculation, Neural Networks
• CPC Class Codes: G06F 18/232 Non-hierarchical techniques
• CPC Class Codes: G06F 18/2433 Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection

Abstract

Principal Component Analysis (PCA) is used to model a process, and clustering techniques are used to group excursions representative of events based on sensor residuals of the PCA model. The PCA model is trained on normal data, and then run on historical data that includes both normal data, and data that contains events. Bad actor data for the events is identified by excursions in Q (residual error) and T2 (unusual variance) statistics from the normal model, resulting in a temporal sequence of bad actor vectors. Clusters of bad actor patterns that resemble one another are formed and then associated with events.

Claims

1. A computer implemented method of identifying events in a process, the method comprising:

running a principal component analysis model on sensor data from the process;

calculating statistics related to the model;

determining if an event is occurring;

finding a nearest cluster of bad actors related to the event to identify the event;

storing the found nearest cluster of bad actors in a storage device; and

further comprising for new bad actors;

identifying a sequence of cluster matches;

correlating the sequence of cluster matches to known events;

determining if a cluster needs to be split when new bad actors are added; and

splitting the cluster into two clusters using a goodness of fit algorithm.

2. The method of claim 1 and further comprising:

determining if a new event category is encountered; and

broadening limits for the sequence of clusters.

3. A computer implemented method of identifying events in a process, the method comprising:

running a principal component analysis model on sensor data from the process;

calculating statistics related to the model;

determining if an event is occurring;

finding a nearest cluster of bad actors related to the event to identify the event; and

storing the found nearest cluster of bad actors in a storage device;

wherein a cluster is limited to a predetermined number of bad actors; and

wherein the predetermined number of bad actors is ten.

4. A computer implemented method of identifying events in a process, the method comprising:

running a principal component analysis model on sensor data from the process;

calculating statistics related to the model;

determining if an event is occurring;

finding a nearest cluster of bad actors related to the event to identify the event;

storing the found nearest cluster of bad actors in a storage device; and

using a feature scoring scheme to identify top contributors of bad actors;

wherein the feature scoring scheme is based on rank, value, and percent of contribution to a Q-residual sensor to identify a relative importance.

5. The method of claim 4, wherein the top-contributors are determined based on a majority percentage of the Q-residual.

6. The method of claim 4, where the top-contributors are determined based on only the contributors with absolute values that are drastically different from values of other contributors.

7. The method of claim 4 wherein the scoring scheme is based on predetermined limits.

8. The method of claim 7 wherein, the limits are computed statistically through change point detections.

9. A system for identifying events in a process, the system comprising:

means for running a principal component analysis model on sensor data from the process;

means for calculating statistics related to the model;

means for determining if an event is occurring;

means for finding a nearest cluster of bad actors related to the event to identify the event;

means for identifying a sequence of cluster matches;

means for correlating the sequence of cluster matches to known events;

means for determining if a cluster needs to be split when new bad actor(s) are added; and

means for splitting the cluster into two clusters using a goodness of fit algorithm.

10. The system of claim 9 and further comprising:

means for determining if a new event category is encountered; and

means for broadening limits for the sequence of clusters.

11. A system for identifying events in a process, the system comprising:

means for running a principal component analysis model on sensor data from the process;

means for calculating statistics related to the model;

means for determining if an event is occurring;

means for finding a nearest cluster of bad actors related to the event to identify the event; and

means for feature scoring to identify top contributors of bad actors in a cluster;

wherein the means for feature scoring is based on rank, value, and percent of contribution to a Q-residual sensor to identify a relative importance.

12. The system of claim 11, wherein the top-contributors are determined based on a majority percentage of the Q-residual.

13. The system of claim 11, where the top-contributors are determined based on only the contributors with absolute values that are drastically different from values of other contributors.

14. The system of claim 11 wherein the scoring scheme is based on predetermined limits.

15. The system of claim 14 wherein, the limits are computed statistically through change point detections.