Fault detection system and method using approximate null space base fault signature classification

• US 20070005528A1
• Filed: 02/23/2006
• Published: 01/04/2007
• Est. Priority Date: 05/31/2005
• Status: Active Grant

• Current Assignee: Honeywell International Inc.
• Original Assignee: Honeywell International Inc.
• Inventors: Mylaraswamy, Dinkar, Menon, Sunil, Mukherjee, Joydeb, Kini, Venkataramana
• Granted Patent: US 7,233,932 B2
• Time in Patent Office: Days
• US Class Current: 706/15
• 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: G05B 23/0281 Quantitative, e.g. mathematical distance; Clustering; Neural networks; Statistical analysis
• CPC Class Codes: G06F 18/2135 based on approximation criteria, e.g. principal component analysis

Abstract

A system and method for fault detection is provided. The fault detection system provides the ability to detect symptoms of fault in turbine engines and other mechanical systems that have nonlinear relationships between two or more variables. The fault detection system uses a neural network to perform feature extraction from data for representation of faulty or normal conditions. The values of extracted features, referred to herein as scores, are then used to determine the likelihood of fault in the system. Specifically, the lower order scores, referred to herein as “approximate null space” scores can be classified into one or more clusters, where some clusters represent types of faults in the turbine engine. Classification based on the approximate null space scores provides the ability to classify faulty or nominal conditions that could not be reliably classified using higher order scores.

Claims

1. A fault detection system for detecting faults in a turbine engine, the fault detection system comprising:

a neural network system, the neural network system adapted to receive sensor data from the turbine engine and generate a plurality of approximate null space scores that represent discrimination features in the sensor data; and

a discriminant based classifier, the discriminant based classifier adapted to receive the approximate null space scores and classify the approximate null space scores to determine a likelihood of fault in the turbine engine.

2. The system of claim 1 wherein the neural network system comprises a chain of neural networks, each of the chain of neural networks adapted to generate a score from the sensor data, and wherein a subset of the chain of neural networks are adapted to generate the plurality of approximate null space scores.

3. The system of claim 1 wherein the neural network system comprises a chain of neural networks, and wherein each of the chain of neural networks includes an encoding neural network and a corresponding decoding neural network, wherein each encoding neural network creates a score from the sensor data, each score comprising a reduced feature space representation of the sensor data;

and wherein each decoding neural network receives the score from the corresponding encoding neural network and creates an estimate of the sensor data, wherein the estimate of the sensor data is passed to the next encoding network in the chain of neural networks.

4. The system of claim 1 wherein the discriminant based classifier comprises a plurality of discriminant functions, each discriminant function capturing characteristics of a boundary between score clusters.

5. The system of claim 4 wherein the plurality of discriminant functions comprises discriminant functions developed using an artificial neural network.

6. The system of claim 4 wherein the discriminant based classifier determines which side of the boundary between score clusters the approximate null space scores are on to classify the approximate null space scores.

7. The system of claim 4 wherein the discriminant based classifier determines a new cluster exists when a magnitude of a distance from the projected point in the approximate null space to two or more classes are not statistically significantly different.

8. A method of detecting fault in a turbine engine, the method comprising the steps of:

receiving sensor data from the turbine engine;

generating a plurality of approximate null space scores from the sensor data, the plurality of approximate null space scores representing discrimination features in the sensor data; and

classifying the null space scores based on discrimination to determine a likelihood of fault in the turbine engine.

9. The method of claim 8 wherein the step of generating a plurality of approximate null space scores comprises generating the plurality of approximate null space scores using a chain of neural networks.

10. The method of claim 9 wherein each of the chain of neural networks comprises step of generating a plurality of approximate null space scores comprises an encoding neural network and a corresponding decoding neural network.

11. The method of claim 8 wherein the step of classifying the null space scores based on discrimination to determine a likelihood of fault in the turbine engine comprises using a plurality of discriminant functions, each of plurality of discriminant functions capturing characteristics of a boundary between score clusters.

12. The method of claim 11 wherein the plurality of discriminant functions comprises discriminant functions developed using an artificial neural network.

13. The method of claim 8 wherein the step of classifying the null space scores based on discrimination to determine a likelihood of fault in the turbine engine comprises determining which side of a boundary between score clusters the approximate null space scores are on.

14. A program product comprising:

a) a fault detection program, the fault detection program including;

a neural network system, the neural network system adapted to receive sensor data from a turbine engine and generate a plurality of approximate null space scores that represent discrimination features in the sensor data; and

a discriminant based classifier, the discriminant based classifier adapted to receive the approximate null space scores and classify the approximate null space scores to determine a likelihood of fault in the turbine engine; and

b) computer-readable signal bearing media bearing said program.

15. The program product of claim 14 wherein the neural network system comprises a chain of neural networks, each of the chain of neural networks adapted to generate a score from the sensor data, and wherein a subset of the chain of neural networks are adapted generate the plurality of approximate null space scores.

16. The program product of claim 14 wherein the neural network system comprises a chain of neural networks, and wherein each of the chain of neural networks includes an encoding neural network and a corresponding decoding neural network, wherein each encoding neural network creates a score from the sensor data, each score comprising a reduced feature space representation of the sensor data;

and wherein each decoding neural network receives the score from the corresponding encoding neural network and creates an estimate of the sensor data, wherein the estimate of the sensor data is passed to the next encoding network in the chain of neural networks.

17. The program product of claim 14 wherein the discriminant based classifier comprises a plurality of discriminant functions, each discriminant function capturing characteristics of a boundary between score clusters.

18. The program product of claim 17 wherein the plurality of discriminant functions comprises discriminant functions developed using an artificial neural network.

19. The program product of claim 17 wherein the discriminant based classifier determines which side of the boundary between score clusters the approximate null space scores are on to classify the approximate null space scores.

20. The program product of claim 17 wherein the discriminant based classifier determines a new cluster exists when a magnitude of a distance from the projected point in the approximate null space to two or more classes are not statistically significantly different.