Kernel-based fault detection system and method
An improved fault detection system and method is provided. The fault detection system and method combines the use of discrimination and representation based feature extraction to reliably detect symptoms of faults in turbine engines. Specifically, the fault detection system and method uses a kernel-based Maximum Representation Discrimination Features (MRDF) technique to detect symptoms of fault in turbine engines. The kernel-based MRDF system and method combines the use of discriminatory features and representation features in historical sensor data to facilitate feature extraction and classification of new sensor data as indicative fault in the turbine engine. Furthermore, the kernel-based MRDF technique facilitates the uncovering of nonlinear features in the sensor data, thus improving the reliability of the fault detection.
1. A fault detection system for detecting faults in a turbine engine, the fault detection system comprising:
a kernel-based Maximum Representation Discrimination Features (MRDF) feature extractor, the kernel-based MRDF feature extractor receiving measured turbine sensor data from the turbine engine, the kernel-based MRDF feature extractor using a nonlinear kernel function, the nonlinear kernel function comprising a nonlinear mapping of the measured turbine sensor data to a higher dimensional space, the kernel-based MRDF feature extractor using the nonlinear kernel function to extract features indicative of nonlinear correlations in the sensor data; and
a classifier, the classifier receiving the extracted features and classifying the extracted features to determine if a fault occurred in the turbine engine,
wherein the kernel-based MRDF feature extractor is developed using historical sensor data from known good and known bad turbine engines.
2. The system of claim 1 wherein the kernel-based MRDF feature extractor utilizes a kernel function with components developed using a generalized eigen-decomposition based on the historical sensor data.
3. The system of claim 1 wherein the kernel-based MRDF feature extractor simultaneously extracts features relating to representation of sensor data from good and bad turbine engines and relating to discrimination between the sensor data between good and bad turbine engines.
4. The system of claim 3 wherein the kernel-based MRDF feature extractor uses a weighting factor to weight features relating to representation and the features relating to discrimination.
5. The system of claim 1 wherein the classifier uses a closeness function to distinguish between extracted features corresponding to a good engine and extracted features corresponding to a bad engine.
6. An apparatus comprising:
a processor;
a memory coupled to the processor;
a fault detection program residing in the memory and being executed by the processor, the fault detection program including:
a kernel-based Maximum Representation Discrimination Features (MRDF) feature extractor, the kernel-based MRDF feature extractor receiving measured turbine sensor data from a turbine engine, the kernel-based MRDF feature extractor using a nonlinear kernel function, the nonlinear kernel function comprising a nonlinear mapping of the measured turbine sensor data to a higher dimensional space, the kernel-based MRDF feature extractor using the nonlinear kernel function to extract features indicative of nonlinear correlations in the sensor data; and
a classifier, the classifier receiving the extracted features and classifying the extracted features to determine if a fault occurred in the turbine engine,
wherein the kernel-based MRDF feature extractor is developed using historical sensor data from known good and known bad turbine engines.
7. The apparatus of claim 6 wherein the kernel-based MRDF feature extractor utilizes a kernel function with components developed using a generalized eigen-decomposition based on the historical sensor data.
8. The apparatus of claim 6 wherein the kernel-based MRDF feature extractor simultaneously extracts features relating to representation of sensor data from good and bad turbine engines and relating to discrimination between the sensor data between good and bad turbine engines.
9. The apparatus of claim 8 wherein the kernel-based MRDF feature extractor uses a weighting factor to weight features relating to representation and the features relating to discrimination.
10. The apparatus of claim 6 wherein the classifier uses a closeness function to distinguish between extracted features corresponding to a good engine and extracted features corresponding to a bad engine.
11. A program product comprising:
a) a fault detection program, the fault detection program including:
a kernel-based Maximum Representation Discrimination Features (MRDF) feature extractor, the kernel-based MRDF feature extractor receiving measured turbine sensor data from a turbine engine, the kernel-based MRDF feature extractor, the nonlinear kernel function comprising a nonlinear mapping of the measured turbine sensor data to a higher dimensional space, the kernel-based MRDF feature extractor using the nonlinear kernel function using a nonlinear kernel function to extract features indicative of nonlinear correlations in the sensor data; and
a classifier, the classifier receiving the extracted features and classifying the extracted features to determine if a fault occurred in the turbine engine; and
b) computer-readable signal bearing media bearing said program,
wherein the kernel-based MRDF feature extractor is developed using historical sensor data from known good and known bad turbine engines.
12. The program product of claim 11 wherein the kernel-based MRDF feature extractor utilizes a kernel function with components developed using a generalized eigen-decomposition based on the historical sensor data.
13. The program product of claim 11 wherein the kernel-based MRDF feature extractor simultaneously extracts features relating to representation of sensor data from good and bad turbine engines and relating to discrimination between the sensor data between good and bad turbine engines.
14. The program product of claim 13 wherein the kernel-based MRDF feature extractor uses a weighting factor to weight features relating to representation and the features relating to discrimination.
15. The program product of claim 11 wherein the classifier uses a closeness function to distinguish between extracted features corresponding to a good engine and extracted features corresponding to a bad engine.
16. A method of detecting faults in a turbine engine, the method comprising the steps of:
receiving measured turbine sensor data from the turbine engine;
extracting features indicative of nonlinear correlations in the sensor data from the sensor data using a kernel-based Maximum Representation Discrimination Features (MRDF) extraction technique using a nonlinear kernel function, the nonlinear kernel function comprising a nonlinear mapping of the measured turbine sensor data to a higher dimensional space; and
classifying the extracted features to determine if a fault occurred in the turbine engine,
wherein the kernel-based MRDF extraction technique is developed using historical sensor data from known good and known bad turbine engines.
17. The method of claim 16 wherein the kernel-based MRDF extraction technique utilizes a kernel function with components developed using a generalized eigen-decomposition based on the historical sensor data.
18. The method of claim 16 wherein the kernel-based MRDF extraction technique simultaneously extracts features relating to representation of sensor data from good and bad turbine engines and relating to discrimination between the sensor data between good and bad turbine engines.
19. The method of claim 18 wherein the kernel-based MRDF extraction technique uses a weighting factor to weight features relating to representation and the features relating to discrimination.