MFCC and CELP to detect turbine engine faults

• US 8,655,571 B2
• Filed: 06/23/2011
• Issued: 02/18/2014
• Est. Priority Date: 06/23/2011
• Status: Expired due to Fees

• Current Assignee: Raytheon Technologies Corporation d/b/a RTX
• Original Assignee: United Technologies Corporation (Raytheon Technologies Corporation d/b/a RTX)
• Inventors: Geib, Andrew F., Kuo, Chung Chieh, Gawecki, Martin, Tsau, EnShuo, Kang, Je Won, Scheid, Paul Raymond
• Primary Examiner(s): Nguyen, Tan Q
• Application Number:
  US13/167,519
• Publication Number:
  US 20120330495A1
• Time in Patent Office: 971 Days
• Field of Search: 701/111, 701/115, 701/102, 701/33.4, 701/29.6, 701/32.9, 701/100, 702/56, 731/19.A, 73/18.2
• US Class Current: 701/100
• CPC Class Codes: F01D 21/003 Arrangements for testing or measuring for measuring vibrations G01H
• CPC Class Codes: F05D 2260/80 Diagnostics
• CPC Class Codes: F05D 2270/807 Accelerometers
• CPC Class Codes: F05D 2270/81 Microphones
• CPC Class Codes: G05B 2219/24042 Signature analysis, compare recorded with current data, if error then alarm
• CPC Class Codes: G05B 23/0221 Preprocessing measurements, e.g. data collection rate adjustment; Standardization of measurements; Time series or signal analysis, e.g. frequency analysis or wavelets; Trustworthiness of measurements; Indexes therefor; Measurements using easily measured parameters to estimate parameters difficult to measure; Virtual sensor creation; De-noising; Sensor fusion; Unconventional preprocessing inherently present in specific fault detection methods like PCA-based methods
• CPC Class Codes: G05B 23/0229 knowledge based, e.g. expert systems; genetic algorithms

Abstract

A fault detection and diagnosis method for a gas turbine engines comprises collecting a sensor signal from an acoustic or vibrational sensor at the gas turbine engine, pre-processing the sensor signal to remove predictable background, and extracting a feature set from the sensor signal using Mel-Frequency Cepstral Coefficients (MFCC) algorithms and/or Code Excited Linear Prediction (CELP) algorithms. Fault and non-fault states are reported based on comparison between the feature set and a library of fault and non-fault feature profiles corresponding to fault and non-fault states of the gas turbine engine.

Claims

1. A fault detection and diagnosis system for a gas turbine engine, the system comprising:

a sensor positioned on the gas turbine engine; and

a health monitoring unit which receives and processes a sensor signal from the sensor, the health monitoring unit comprising;

a digital storage medium containing a library of fault and non-fault feature profiles mapping fault states and non-fault states of the gas turbine engine, respectively, to combinations of Mel-Frequency Cepstral Coefficient (MFCC) frequency bank energies and/or Code Excited Linear Prediction (CELP) polynomial coefficients; and

a processor which extracts a feature set from the sensor signal using at least one of MFCC algorithms and CELP algorithms, and detects and diagnoses faults in the gas turbine engine by matching the feature set to one or more of the fault or non-fault feature profiles.

2. The fault detection system of claim 1, wherein the processor extracts the feature set from the sensor signal using both MFCC and CELP algorithms.

3. The fault detection system of claim 2, wherein a first subset of features is extracted from the sensor signal using MFCC algorithms, a second subset of features is extracted from the sensor signal using CELP algorithms, and the first and second subsets of features are concatenated to form the feature set.

4. The fault detection system of claim 1, wherein the sensor comprises an acoustic pressure sensor.

5. The fault detection system of claim 1, wherein the sensor comprises a vibrational accelerometer.

6. The fault detection system of claim 1, wherein a plurality of sensors are positioned on the gas turbine engine, all of which produce sensor signals which are received and processed by the health monitoring unit using MFCC and CELP algorithms.

7. The fault detection system of claim 6, wherein the plurality of sensors comprises both acoustic pressure sensors and vibrational accelerometers.

8. A fault detection and diagnosis method for a gas turbine engine, the method comprising:

collecting a sensor signal from an acoustic or vibrational sensor at the gas turbine engine;

pre-processing the sensor signal to remove predictable background;

extracting a feature set from the sensor signal using Mel-Frequency Cepstral Coefficients (MFCC) algorithms and Code Excited Linear Prediction (CELP) algorithms;

comparing the feature set to a library of fault and non-fault feature profiles that map fault and non-fault states of the gas turbine engine to combinations of MFCC frequency bank energies and/or CELP polynomial coefficients; and

reporting fault and non-fault states of the gas turbine engine based on comparison between the feature set and the feature profiles.

9. The fault detection and diagnosis method of claim 8, wherein comparing the feature set to a library of fault and non-fault feature profiles comprises:

identifying a non-fault state if the feature set matches any non-fault feature profile; and

identifying a fault state if the feature set matches the corresponding a fault feature profile.

10. The fault detection and diagnosis method of claim 8, further comprising:

scheduling maintenance for the gas turbine engine based on the diagnosed fault state or the reported non-fault state.

11. The fault detection and diagnosis method of claim 8, wherein extracting the feature set requires no knowledge of engine speed.

12. The fault detection and diagnosis method of claim 8, wherein extracting the feature set comprises:

Fourier-transforming the sensor signal;

applying a series of M mel-scale filter functions to the Fourier-transformed sensor signal to produce a series of M frequency banks; and

designating the energy within each frequency bank as a feature in the feature set.

13. The fault detection and diagnosis method of claim 12, wherein the mel-scale filter functions are overlapping triangular filter functions.

14. The fault detection and diagnosis method of claim 8, wherein extracting the feature set comprises:

fitting the sensor signal to an N-dimensional CELP polynomial; and

designating the pitch and the coefficients of the N-dimensional CELP polynomial as features of the feature set.

15. The fault detection and diagnosis method of claim 8, further comprising:

reducing the feature set before comparing the feature set to a library of fault and non-fault profiles.

16. The fault detection and diagnosis method of claim 8, further comprising:

down-sampling the sensor signal before extracting the feature set, so as to produce a desired number of samples per second.

17. The fault detection and diagnosis method of claim 8, wherein extracting the feature set from the sensor signal comprises extracting M MFCC features, N CELP coefficients, and 1 pitch from each sample to form an X-dimensional feature set for each sample, with X=M+N+1.

18. The fault detection and diagnosis method of claim 8, further comprising:

aggregating a plurality of results from comparing feature sets to the library of faults and non-fault profiles, to produce a synthesized fault diagnosis.

19. The fault detection and diagnosis method of claim 8, wherein the fault and non-fault feature profiles map characteristics of feature sets onto states of the gas turbine engine.

20. The fault detection and diagnosis method of claim 19, wherein the fault and non-fault feature profiles are produced by extracting a feature set from a sensor signals corresponding to known fault and non-fault conditions.