Turbine Engine DiagnosisOn the Reconstruction of Performance Parameters

from Engine and FADEC Measurements

RSL Electronics LTD.

November 2005

IAF

AH-64Hermes 450 UAVCH-53UH-60

Snecma

CFM56

The Challenge

Diagnosis of a twin-spool turbofan, based on pressure and temperature measurements at different stations, FADEC variables (VBV, VSV, WFM, etc.), environmental conditions.

Technique: reconstruction of station performance measures (efficiency measures, flow measures) at said stations.

The Information

Black box (OEM-supplied) thermodynamic steady-state simulator

Limited test-cell measurements of engine, used for parameter range estimation.

The Simulator

performance measures

environment conditions

Fan Eff.

Fan Flow…

LPT FlowAmb. Tmp.…

control input N1

T+P S12…

VBV…

station measurementscontroller output

The Mission

Producing a simulator reconstructor (The Perdictor)

The Simulator

performance measures

environment conditions

Fan Flow…

LPT FlowAmb. Tmp.…

control input N1

T+P S12…

VBV…

Fan Eff. Fan Eff.Fan Flow

…

LPT Flow

performance measures

The Predictor

The Goal

The Predictor as a diagnostic tool for the real engine

environment conditions

…

Amb. Tmp.…

control input N1

T+P S12…

VBV…

The Predictor

Fan Eff.Fan Flow

LPT Flow

performance measures

…

Is this a well-posed problem?

1. Existence ?

2. Uniqueness ?

3. Continuity ?

(Hadamard, 1902, 1923;

Tikhonov & Arsenin, 1977;

Morozov, 1993;

Kirsch, 1996;

Haykin, 1999)Generally: NO!In practice:

•Prior knowledge about simulation type/problem nature

•Restraints on parameter combinations/distributions

•Parameters limited to localized working conditions

For Practical Purposes: WE’LL TRY!

There exist more well-posed

(Kandariya Visvanatha Mahadeva Temple, Khajoraho, India, 10th-11th century AD)

But it works!

examples..

Predictor Implementation

The Predictor (non-linear regression) was realized on feed-forward neural networks.

Different predictor/NN for each performance parameter.

Fan Eff. Predictor

Fan Flow Predictor

LPT Eff. Predictor

Estimated Fan Eff.

Estimated Fan Flow

Estimated LPT Eff.

Contol Input

Stations P+T

FADEC Params

Env. Conditions

Data Sets

Two sets for training and testing of The Predictor.

Sets to span spaces to be encountered: Performance parameters (engine integrity

states) in all stations/cocktails thereof

Engine operating states

Environmental conditions

Working with a simulator, manufacturing synthetic data with no limits on size, distribution (independence) and quality.

Data Sets

Data sets were produced using extensive execution of the simulator under controlled input conditions, as to span the working conditions to be encountered.

Parameter Space

GeneratorThe

Simulator

Engine Control + Env. Params

Performance Params

P+T, FADEC Params

In

Out

Data Sets

Test

set

Training set

Predictor Training/Testing

Test

set

Training set

A Predictor (e.g. LPT

Eff.)

Data Sets

Train Env+P+T+FADEC Params

Train Performance Params

Predicted Performance Params

Test Env+P+T+FADEC Params

Test Performance Params +

-

Σ Test Error

Train Loop

Test Loop

Data Sets Distribution Strategy

Environment and engine control (e.g. N1) inputs were chosen to be uniformly distributed in ranges obtained from test-cell data.

0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.60

200

400

600

800

1000

1200ZTAMBA

e.g. ambient temperature histogram

Data Sets Distribution Strategy

Performance parameters (efficiencies and flows) were chosen from two distributions, representing either “normal behavior” or “faulty behavior”

“normal behavior” consisted of most values near their nominal expected value (scaler = 1.0), with exponential distribution with parameter μ= 0.001 to the “left” of 1, accounting for less than ideal performance.

“faulty behavior” was chosen from a uniform distribution in the scaler range 0.95 – 1.0.

Modeled parameter was uniform at interval 0.95-1.0 according to target reconstruction range.

Data Sets Distribution Strategy

e.g. Fan isentropic efficiency HISTOGRAM

0.95 0.955 0.96 0.965 0.97 0.975 0.98 0.985 0.99 0.995 10

0.5

1

1.5

2

2.5

3

3.5

4x 10

4 ZSE13D

scaler values

Fig. 2.

ZSE13D histogram, from Ptest

There’s a “tail” here, when

parameter was faulty

Normal cases 0.95 0.955 0.96 0.965 0.97 0.975 0.98 0.985 0.99 0.995 1

200

400

600

800

1000

1200

1400

1600

1800

ZSE13D

See fault distribution strategy presented below

(Pf = 0.5)

Fault Distribution Strategy

Test set fault distribution assumed fault occurrence independence, resulting in sets with decreasing frequency of occurrence of multiple simultaneous faults.

The number of simultaneous faulted parameters (not including the modeled parameter) was chosen at each data point, such that if a probability for a single faulted parameter is Pf, then the probability for n faults was Pf

n. Pessimistic values used for Pf were 0.1 and 0.5.

Fault Distribution Strategy

Fault distribution HISTOGRAM for Pf = 0.5

0 5 10 15 20 250

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5x 10

4

Number of simultaneous faults

His

togra

m c

ount Average fault

number is 1 (for Pf

= 0.1 average is ~0.11)

Results

e.g. Fan flow scaler (Pf = 0.5)

-6 -4 -2 0 2 4 6

x 10-3

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

Ttest

- Tnet

his

togra

m c

ount

0.5 P_FAULT SW12RA STD DIFF: 0.00036467 [0.036467% res.]

90th Percentile Error: 0.050%

Graph shows reconstructed parameter value VS original value

Results

e.g. Booster isentropic eff. scaler (Pf = 0.5)

-4 -3 -2 -1 0 1 2 3

x 10-3

0

5000

10000

15000

Ttest

- Tnet

his

togra

m c

ount

0.5 P_FAULT SE23DA STD DIFF: 0.00019 [0.019% res.]

90th Percentile Error: 0.025%

Graph shows reconstructed parameter value VS original value

Results

e.g. LPT flow scaler (Pf = 0.5)

-0.04 -0.03 -0.02 -0.01 0 0.01 0.02 0.03 0.04 0.050

2000

4000

6000

8000

10000

12000

Ttest

- Tnet

his

togra

m c

ount

0.5 P_FAULT SW49RA STD DIFF: 0.0036329 [0.36329% res.]

90th Percentile Error: 0.377%

Graph shows reconstructed parameter value VS original value

Conclusions

The regression attempt has produced results whose 90th percentile deviation averaged0.156%, exceeding requirements for detecting plausible fault performance degradations of 1% –2%.

The less precise regression has been achieved for LPT efficiency and flow parameters

This feasibility study proved the capability of utilizing machine learning techniques to reproduce actual engine status from accessible measurements, at least in the simulator context.

ANY QUESTIONS?