As part of the NASA Aviation Safety Program, a unique model-based diagnostics method that employs neural networks and genetic algorithms for aircraft engine performance diagnostics has been developed and demonstrated at the NASA Glenn Research Center against a nonlinear gas turbine engine model. Neural networks are applied to estimate the internal health condition of the engine, and genetic algorithms are used for sensor fault detection, isolation, and quantification. This hybrid architecture combines the excellent nonlinear estimation capabilities of neural networks with the capability to rank the likelihood of various faults given a specific sensor suite signature. The method requires a significantly smaller data training set than a neural network approach alone does, and it performs the combined engine health monitoring objectives of performance diagnostics and sensor fault detection and isolation in the presence of nominal and degraded engine health conditions.
Aircraft engine performance is diagnosed by estimating a set of internal engine health parameters from available sensor measurements. The following relationship between the engine health parameters and the sensed parameters can be used to express the general approach:
y = f ( p , operating condition) + w
where y is a vector representing the sensed parameters (gas path temperatures and pressures, spool speeds, fuel flow, and variable geometry), p is a vector of engine health parameters (component efficiencies and flow capacities), f (·) is a nonlinear function of p and the engine operating condition, and w is a vector representing measurement inaccuracies. System nonlinearities and potential sensor measurement inaccuracies make this estimation problem very challenging.
Hybrid engine health estimation architecture.
Long description
The hybrid engine health estimation architecture, as shown in the figure,
is composed of a bias data set, a neural network estimator, an engine
model, and the genetic algorithm optimization technique. Engine output data
are based on the current operating condition and engine health
parameters. These sensed parameters are corrupted by a white noise vector
v
and a bias vector
b
. To make the problem manageable, we assumed that at
most one sensor could be biased at a time. The bias data set, which is
composed of a large number of bias vectors, is defined a priori and is used
by the genetic algorithms in the search for a bias vector 
that matches well with an actual bias contained in the measurement vector. The neural network estimator is trained offline with noise-corrupted, but
bias-free, sensor measurements to estimate engine health parameters 
. The neural network will perform sufficiently well in estimating health
parameters as long as the sensor measurements do not contain any bias. For
a given set of estimated health parameters and sensor bias, the
engine model is executed and its output data are evaluated against the
physical sensor measurements. The bias data set, the neural network estimator,
and the engine model are coordinated by the genetic algorithms in
the search for an optimal solution. After the search process,
the searched bias vectors are ranked according to their
corresponding fitness value, which is a value indicating the agreement between
the measured and predicted engine output parameters. A ranked list
of several plausible fault candidates can help to avoid false alarms
or missed detections.
The table shows an example of the technique's estimation performance applied to a military twin-spool turbofan engine simulation, which was used to represent both the engine and the engine model shown in the figure. Here, 12 sensed engine values were used to estimate the 9 engine health parameters listed in the table. In this case, a 9.5s bias was modeled in the sensed fuel flow value. Without bias detection, the estimation errors of some engine health parameters are higher than 20 percent, and one is as high as 120 percent. With bias detection, the estimator is able to correctly identify and quantify the bias in the fuel flow. This results in greatly improved health parameter estimation accuracy with all estimation errors at 15 percent or less.
| Health parameter | Actual condition, percent |
With bias detection | Without bias detection | ||
|---|---|---|---|---|---|
| Estimated condition, percent |
Error, percent |
Estimated condition, percent |
Error, percent |
||
| Fan efficiency | -2.900 | -2.788 | -3.876 | -2.950 | 1.722 |
| Fan flow | -1.800 | -1.811 | 0.596 | -1.819 | 1.076 |
| Booster flow | 0 | 0 | ------- | -0.134 | ------- |
| High-pressure compressor efficiency | -2.300 | -2.172 | -5.578 | -2.305 | 0.234 |
| High-pressure compressor flow | -1.900 | -2.027 | 6.658 | -1.497 | -21.213 |
| High-pressure turbine efficiency | -1.400 | -1.614 | 15.254 | -1.715 | 22.516 |
| High-pressure turbine flow | 1.000 | 0.875 | -12.484 | 2.201 | 120.045 |
| Low-pressure turbine efficiency | -2.000 | -2.197 | 9.857 | -2.303 | 15.146 |
| Low-pressure turbine flow | 2.100 | 2.083 | -0.819 | 2.393 | 13.942 |
Simon, Donald L.: A Hybrid Neural Network-Genetic Algorithm Technique for Aircraft Engine Performance Diagnostics. AIAA-2001-3763 (NASA/TM-2001-211088, 2001.
QSS contact: Takahisa Kobayashi, 216-433-3739,
Takahisa.Kobayashi@grc.nasa.gov
Army contact: Donald L. Simon, 216-433-3740,
Donald.L.Simon@grc.nasa.gov
Authors: Takahisa Kobayashi and Donald L. Simon
Headquarters program office: OAT
Programs/Projects: AvSP
Last updated: June 2002
Responsible NASA Official:
Gynelle.C.Steele@nasa.gov
216-433-8258
Point of contact for NASA Glenn's Research & Technology reports:
Cynthia.L.Dreibelbis@nasa.gov
216-433-2912
SGT, Inc.
Web page curator:
Nancy.L.Obryan@nasa.gov
216-433-5793
Wyle Information Systems, LLC
