Methematical Methods of Relative Engine Performance VOLPONI

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 27

Mathematical Methods of Relative Engine Performance Diagnostics

Author(s): Louis A. Urban and Allan J. Volponi


Source: SAE Transactions , 1992, Vol. 101, SECTION 1: JOURNAL OF AEROSPACE (1992),
pp. 2025-2050
Published by: SAE International

Stable URL: https://www.jstor.org/stable/44733159

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

SAE International is collaborating with JSTOR to digitize, preserve and extend access to SAE
Transactions

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
922048

Mathematical Methods of Relative Engi


Performance Diagnostics
Louis A. Urban and Allan J. Volponi
Hamilton Standard Div.ision of United Technologies Corp.

ABSTRACT intended as a working tool to allow those who were


necessarily performance specialists to get good first
The basic concepts and mathematics of a methodology that
approximation solutions to most performance questio
allows successful relative assessment of simultaneously heart of the book is a set of simultaneous differential equ
occurring multiple faults in gas turbine engine modules and that show how the measurable dependent parameters
sensors, are presented in detail. Included are a detailed temperatures, pressures, flows, spool speeds) respond to c
exposition of a thermodynamic/differential equation based in the generally immeasurable independent parameter
model of a representative type of gas turbine engine; discussion efficiencies, pumping capacities and nozzle sizes) and to c
of correction methods to minimize data scatter; the mathematical in the engine geometry, control or environment. The eq
design aspects of a demonstratedly successful approach to data are set up in a completely generalized internal performanc
filtering to deal with residual scatter caused by large or small to make them independent of engine exhaust geomet
sensor errors, data correction errors, or engine system modelling control mode, and for any chosen geometry, control and b
errors.
parameter they may be simply manipulated to yield
traditionally used and widely understood matrix of influ
INTRODUCTION
coefficients or partial derivatives (ref. 3).

Since the introduction of practical aircraft gas turbines over


Mathematically stated the matrices imply that
fifty odd years ago, considerable engineering effort has been
devoted to the art of engine thermo-mathematical modelling z = He χ (1)
with ever greater precision to fulfill the very real and compelling
needs of successful design, development and installed use. One
wherein z is a vector of engine
of these needs is the ability to perform engine performance
a vector of engine module param
diagnostics to the module level. influence coefficients or pa
interrelating the two sets.
The purpose of engine diagnostics is to lead to actions taken on
the basis of judgements made from directly measured When or looked at in this simple w
that the
inferentially calculated information. For the performance of mathematical converse
diagnostics of gas turbine engines to the module level, Gas Path
Analysis (GPA) was developed by the authors to provide the * = Η~λ ζ (2)
ability to determine the extent to which each engine module and
which
the attendant engine sensors deviate from any specified is to say that a vector of e
level and
determined
thus to isolate changing or abnormal modules and sensors (ref. 1). from the inverse o
coefficient matrix (which may b
BASIC CONCEPTS
times a vector of meaningfully r
deltas.
Gas Path Analysis (GPA), or more generically, simultaneous
multiple fault module and sensor diagnostics, traces itsInroots to
1969 serious effort began on d
Gas Turbine Engine Parameter Interrelationships (ref.concept,
2), a book and after verifying it
written by the one of the authors (Urban) in 1967. The book was
validity, a patent was applied for
2025

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
The initial applications were diagnostic systems for the FT4 and were essentially a compilation of small differences between large
T53 engines, followed over the next 20 some years with many numbers, the precision with which they depicted the intended
others, including several power generating land gas turbines, and phenomena depended on how many iterations the deck made, and
various models of the JT9D, CF6, TF39, F100, JT8D, J52, on the direction and magnitude of the implanted changes. To the
PW2000, PW2500, and PW4000 series aircraft engines. Users extent built in by its designer all runs of the cycle deck have to
have included engine manufacturers, the US Army, Navy and Air satisfy all the thermodynamic requirements of module matching
Force, US and non-US commercial airlines, and power utilities. (e.g., flow continuity, pressure rise and fall, compressor/turbine
power balance, nozzle flow parameters, etc.), and so too must the
To successfully implement a diagnostic system based on differences between runs. Simple checks of the old available
equation (2) several conditions must be rigorously met: tables showed that meeting these requirements was often limited
in accuracy to only the first significant figure and sometimes not
* The engine must be properly modelled, with baseline and even that. For example, if the cycle deck was set to iterate until it
correction procedures that allow accurate assessment of the achieved closure on the performance maps such that dependent
true measurement deltas. parameters are calculated to the nearest 1/10 of 1%, and the run
differences (fault coefficients) typically have magnitudes of the
* The influence coefficients must truly reflect how the order of 1/2 of 1%, then these fault coefficients are only known to
measurements respond to the potential module problems, to the nearest 20%. For these and other reasons it was decided that a
changes in the environment (inlet conditions), and to other different approach to engine modelling had to be used.
factors which might be present such as off-take bleed and
altitude Reynold's number. ENGINE MODELLING

* The module problems that exist must be among those that To gain the requisite precision and accuracy for the task at
are being sought. the preferred approach was to develop appropriate engine m
for each application by extension of the techniques of referen
* The measurements taken must be meaningfully related to Appendix 1 describes in detail the mathematical model fo
the problems sought and detectably respond to changes such engine, a twin spool mixed flow turbofan engine of th
therein. schematically shown (Appendix lFig. 1). The engine consis
a core engine having a high pressure compressor driven by
* The measurements taken must be noise free, or methods pressure turbine, a combustion chamber and an exhaust; a
used to properly account for the noise and eliminate it as an together fan and low pressure compressor coaxially driven
error source, (ref. 1) low pressure turbine; a duct for the fan discharge flo
extended for downstream mixing of the core and fan f
Fulfillment of these conditions, particularly the first two, offtake service bleeds at the HPC inlet, middle and discha
requires close cooperation with the engine manufacturer to handling bleed at the HPC inlet; high and low pressure tu
understand the engine's operating characteristics and define its cooling flows from the HPC discharge; and a final exh
numerics, and with the airframe manufacturer in the case of nozzle. The Appendix 1 equations shown are necessary
airborne applications to define the numerics of installation sufficient to define the engine numerics for GPA - or for mo
effects such as air conditioning off-take bleeds. Without other purpose. The equations are shown in their absolute
exception it has been our experience to encounter complete partial differentiated forms, and are dictated by
cooperation in supplying basic performance information, to the physical/thermodynamic laws for module matching, viz.,
extent that it was known and available, for any given engine.
* Conservation of mass, energy and momentum
Unfortunately from the viewpoint of system design and * Pressure rise requires pressure fall
development, it was also our experience that the information * Spool speed of driven is speed of driver
available, particularly the influence coefficients or partial * Power balance properly accounted
derivatives, were never in the form nor of the precision required. * Modules must operate on their maps
This was understandable, since the existing tables were usually
generated for much simpler purposes; e.g., to provide users or In general no matter what the engine configuration is,
other interested parties with approximate magnitudes and number of equations will always be equal to the number of
directions in which measurable parameters could be expected to for dependent parameters plus one. In this particular case
vary in response to certain singly occurring module or other are thus 47 partial differential equations in 48 depen
independent parameter variations. Simultaneous multiple fault unknowns and a plurality of independent knowns. This is
isolation was not a consideration. minimum required for diagnostics of the engine as descr
above; additional equations would be added if other factors
The usual method of generating the tables would be to run an present, such as variable compressor geometry, or if know
iterative cycle deck or module matching deck at an operating on other parameters was also desired, such as static pressur
point of interest to establish a base point, and then repeatedly every single engine station. Obviously, models similar in g
rerun the deck while perturbing the independent parameters of form but differing in specific content were also developed fo
interest one at a time by selected magnitudes and directions. The other types of engines and engine cycles we have worked o
arithmetic differences between these runs and the base point were better visualize the equations they may be presented in the
then compiled as the influence coefficient tables. Since the tables of Appendix 1 Fig. 2. Each row is one of the differe
2026

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
equations and each χ in the chart is a non-zero coefficient for the * Use the manufacturer's value if the difference was minor.
column-heading parameter in that equation. Fig. 2 is in the
general form
* Use the model value if the difference was greater than
A y = Β χ about 0.04

where the dependents are on the left and independents on the


right. To solve the set the first step is to decide what the baseline * Refine the value when a significant volume of data was
available.
parameter shall be (e.g., Nlc2 ) and transfer the column for that
parameter from the left to the right hand side matrix. The A
matrix is thus diminished by one and becomes an invertible The refinement procedure was to perform regression analyses
square A* matrix and the Β becomes an augmented B* matrix. of the corrected parameter deltas from baseline vs engine inlet
The solution then becomes
temperature and add the slope of the regression line (in % per %)
to the exponent assumed in the original data reduction. The
y = ( A* )~l Β * χ volume of data used was generally at least 50 operating points
over at least a 50 deg F band. The data could be from one or from
The resultant vectors of dependent parameter changes for several engines, all of the same quality, e.g., new or newly
changes in the specified independent parameters (e.g., changes in overhauled.
the dependents caused by a change in HPC efficiency level at
constant Nlc2 ) are by definition the fault coefficients, and the
resultant vector of dependent parameter changes with the A mild surprise in the early days was that for some high
baseline parameter (which is a vector of the augmented B* performance engines, the models indicated and real data
matrix) defines the steady state slopes of the dependents. confirmed the efficacy of making Theta based corrections to
some of the pressures and pressure ratios. The Theta exponents
With regard to the choice of baseline parameter, different were relatively small (magnitudes in the order of .05) but their
organizations favor different ones (e.g., EPR, Nlc2 , weighted use did help to reduce data scatter. At least for diagnostic
EPR) when discussing performance. For the purpose of purposes, some manufacturers have since followed suit. It should
diagnostics it matters not one whit which is chosen, since the be pointed out that any deviations from manufacturers' published
engine is the engine and it really doesn't care how you chose to correction procedures are only used for data manipulation within
look at it; if a module is deviant the match point is shifted the diagnostic programs and not for such things as guaranteed
affecting all parameters, and thermodynamic manipulation Thrust , SEC or EGT margins, which can become highly charged
through the diagnostic equations will, as it logically should, give emotional, political and economic issues.
the same answer on the modules no matter what is chosen for the
baseline (ref. 5). SOME DATA CORRECTIONS

As previously alluded to, numerical execution requires The process of going from measured data to net measure
complete definition of engine internal temperatures, pressures deltas for module diagnostics is illustrated by Fig. 1. At a g
and flows at the operating points of interest, and the engine operating point one starts out with two bits of information:
module performance maps. The former are generally available to the engine manufacturer or user experience says the nom
those with a need to know, and the latter, if not directly available, measured value should be; and what the observed value ac
can be inferentially extracted from the manufacturer's fault is. The difference between the two consists of a host of fa
coefficient tables by straightforward, if tedious procedures. only one of which is the sought for net delta caused by de
Assuming that the operating point has been accurately defined modules. The nominal baseline by definition states how
and the differential based model is calculated to double precision nominal or average engine will behave in a loss-fr
(a trivial matter with modern computers), the resulting fault environment with nominal exhaust areas and measure
coefficients and outputs will be accurate to that same extent and perfect instrumentation (ref. 1). Corrections must be ma
checks of the previously cited module matching considerations installation losses associated with the test cell or the aircraft
will be precise to at least four significant figures. nacelle, and for instrumentation calibration errors. These are
established from engine steady state data over a multiplicity of
An interesting sidelight is that the Tarn and Pam fault operating points spread evenly over the power spectrum, upon
coefficients for temperatures, pressures and flows, when initial installation of the engine in its normal operational
expressed in percent of point values, are in effect the exponents of environment and with the normal subsequently used
the Theta and Delta corrections to the corresponding corrected instrumentation. Since errors in estimation of all the many
parameters. Conventional simplified dimensional analysis tells potential factors contribute to scatter in the final answer, it is
us that generally the Theta exponent for pressures should be zero, critically important to make all corrections with the finest
for temperatures unity and for flows 1/2. Most engine practical precision. The influence coefficients arrived at from the
manufacturers today recognize that this is not precisely so and above described differential based model are used, as
use values that differ from these for temperatures and for fuel appropriate, to make corrections for the effects of non-standard
flow. In some cases the exponents implied by the models did not exhaust nozzles; and with flight data, for the effects of air
agree with these manufacturers' values and the general approach conditioning bleed, and for the effects of Reynold's number
was to variation with flight condition.
2027

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
I"" Δ DUE TO DEVIANT MODULES - - φ ļ
meas Δ DUE TO DEVIANT MODULES ļ

œ^ECTED GAS PATH Δ "UNSOUGHT PROBLEMS"


GAS PATH

PARAMETER

FAN NOZZLE -

_ JL 1 NOMINAI

Δ CORE NOZZLE ^ -

85% MC MAX CONT. T/O

Figure 1 Calculating M

Bleed corrections are made by


High Bypass Low Bypass know
the absolute bleed flow Turbofan
per air-con
Turbofan
of altitude; from M dM/dY
the flight dM/dY data re
operating point and pack status is;
the airflow of theFan How 3.0is;
engine 2.5calcula
bleed from the engine; then
LPC Flow L5 L5 adjust
by the product of the percent b
influence coefficients
HPC Flow
for
03 06
bleed.
Fan Efficiency 6.0 5.0
Reynold's number LPC corrections
Efficiency Ī5 Ī0 are
operating point the Rei (Reynold
effects on HPC Efficiency
pumping capacity,05 L0 effi
areas of all the HPT Efficiency
modules LO L2
(ref. 6); th
parameters by
LPT Efficiency 05 07
the model influenc
module effects. These module laps
HPT Area L6 L6
from the following empirical exper
LPT Area 06 LO
Fan Nozzle Area 3.0

Final Nozzle Area 0.7 2.0


dM = ^ dY
and finally the dependent paramete
= 3.44[ (0.7 - te/)1·79 - (0.7 - teW)179]
dP = ( ^ ^ dY ] = ( Infi. Coeff

As with any empirically derived


wherein RcIref is the Rei for the baseline (e.g., for 35000 ft. 0.85 alterable with the user's experience
Mach) and NIref is the N1 at some reference power condition, of differing cycles.
e.g., 85% power. Some typical empirical values for dM/dY are
After all corrections have been m
of each parameter is still composed
due to deviant modules, which is t
interest; and the portion due to ins
and the presence of other unsou
2028

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
errors, baseline errors, correction errors or other real but The columns represent engine component efficiency and flow
unsought engine errors (ref. 1). Early experience showed that the parameter deltas while the rows represent deltas in the
desired signal (net delta) to unwanted noise (balance of gross (measurable) dependent variables. This linear model allows one
deltas) ratio is about one-to-one, and that the use of gross deltas to forecast expected changes (%) in spool speeds, temperatures,
would lead to unacceptable data scatter and inability to pressures, etc., for given changes in module performance at a
confidently detect real module changes. Accordingly, after specific operating point. Since the number of independents
evaluating alternative approaches, in 1975 we introduced into ( N=10 ) is smaller than the number of dependents (Af=72), a least
our methodology and programs the use of a minimum error squares solution could be obtained which would yield the inverse
estimation technique which is a specialized adaptation of relationship X-De Ζ where De is a Ν χ M matrix such that
Kaiman filtering or optimal estimation theory. DeHe=I (identity). Given a baseline reference, actual measured
data and the means to calculate the % deltas (Z) between them, an
DEVELOPMENT OF THE MATHEMATICAL MODEL estimate for module performance change can be calculated.

Use of the methodology so far described results in Although


systemsthis
orappears to be an attractive approach, it falls short
programs which provide a relative performance of providing
analysis, consistent results in practice. Measurement
the
intent of which is to assess changes in module performance data normalization errors, sensor bias, and
nonrepeatabilities,
relative to some fixed reference or baseline. To further describe modelling inaccuracies all contribute to error in the calculated
this in some detail, the first step in the modelling process is to measurement delta (Z) which in turn will propagate the error to
determine a representation interrelating changes in the the estimated component performance (X) (ref. 11). The efficacy
independent engine parameters to changes in the dependent of a linear diagnostic system will depend heavily on the extent to
variables. In general, this relationship is nonlinear. However, in which it can isolate and accommodate unanticipated
order to make the system mathematically tractable, it is common perturbations in the input parameters. Before discussing how we
practice to utilize a linear approximation. For example, if can amend the linear model to accommodate error, let us first
z=/(jc,y), we could represent this function in terms of a Taylor look briefly at the possible sources of error with which our system
expansion about some reference point (Xo,y0) must contend.

To proceed from raw data to performance deltas, a number of


Z = /(*.30 = /(*<>. 3Ü + ^ ( * - *o ) steps must be taken. Referring to Figure 1 , it can be seen we must
have available, for each measured engine parameter, the
+ ^ ( y - y„ ) + Higher Order Terms (3) expected measurement level over the operating regime of interest
(baseline), and a method of reducing the raw measurement to
those conditions assumed for the baseline in order for a
Assuming small changes in χ and y, the higher order terms are comparison to be made and a delta to be computed. The baselin
close to zero and can be dropped resulting in a linear relationship. usually takes the form of a set of gas generator curves as
Applying some simple algebra to the eq.3 would yield function of some independent parameter such as engine pressu
ratio ( EPR ) or corrected spool speed ( Nlc2 ) at a standard da
condition. The data normalization for the raw inputs consists,
ÚL ~ Γ lï°' (ŚL' 1 dx + , (M (df' dy Figure 1 suggests, of a myriad of corrections and adjustment
zo ~ U/ 'dx)o Xo + , l zj 'dy)c y o (4) [ref. 1,10]. These include corrections for ambient condition
(inlet temperature, pressure, and humidity), exhaust nozz
The bracketed terms are evaluated at some steady state geometry, Test Cell geometry (if data is collected at a test stand
operating point and are referred to as engine fault coefficients. and Mach no., Reynold's number index and engine service blee
Generalizing to M dependent and Ν independent variables this corrections if the data is acquired during flight. The complex
becomes a vector matrix relationship Ζ = HeX, where He is an nature of this normalization process with its empirical factors and
ΜχΝ matrix of fault coefficients (FCM). An example of such a model dependent numerics coupled with measurement device
matrix for the mixed flow twin spool turbofan as described in nonrepeatabilities, assures us of some degree of concomita
Appendix 1. is depicted in Figure 2 below. error.

2029

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
ETAF G AM F ETACL GAMCL ETACH GAMCH ETAT A5 ETATL A6
GAMT2 0.0785 0.3594 0.1579 -0.0225 0.0618 0.0668 0.0769 -0.0276 0.2756 -0.1517
N1C2 0.1138 -0.8844 0.2326 -0.0606 0.0871 0.0941 0.1083 -0.0389 0.4099 -0.2315
N2C2 0.0099 -0.1998 -0.1645 0.0524 0.1782 -0.2291 0.2264 -0.1317 0.0513 0.1752
T3C2 0.0854 -0.5755 -0.2652 0.2598 -0.0714 -0.0772 -0.0888 0.0319 0.3027 -0.3329
PS3C2 0.2309 -1.3944 0.2404 0.9199 -0.4750 -0.5135 -0.5908 0.2120 0.8036 -1.2244
T4C2 0.0045 -0.3277 -0.3233 0.0547 -0.1111 -0.1452 0.3000 -0.3780 0.0516 0.2456
PS4C2 0.0196 -0.7734 0.0493 0.3774 0.0918 0.0702 0.1846 -0.9820 0.1507 0.0539
WFC2 -0.2389 -0.1438 -0.3880 -0.3272 -0.2079 -0.2243 -0.2661 0.1001 -0.6335 0.2630
T7C2 -0.1733 0.3980 -Ό.5048 -0.5194 -0.2839 -0.3069 -0.3529 0.1265 -0.7969 0.3135
FNC2 -0.0121 0.3663 0.0143 -0.1731 -0.0194 -0.0209 -0.0241 0.0086 0.0492 -0.0634
PT25C2 -0.0187 0.6214 0.0046 -0.2965 -0.0413 -0.0447 -0.0514 0.0184 0.0501 -0.0903
T25C2 -0.1803 -0.2120 0.0969 -0.0976 0.0257 0.0278 0.0320 -0.0115 0.1811 -0.1175

Figure 2: Twin Spool Mixed Flow Turbofan Engine Fault Coef

coefficients
Although there are several approaches possible (He),
toa mitigate
linear approximation
the relating Xs and Ζ can be
effects of measurement error, the method derived (see weAppendix 2) yielding a set of sensor fault
will discuss
involves the inclusion of individual error variables in the linear coefficients Hs. The general form of this MxNs matrix for the
model; one for each of the gas path parameters under model in Figure 2 is given below in Figure 3.
consideration. In the model depicted in Figure 2, we have Λί=72
main gas path parameter deltas, GAMT2, N1C2, N2C2 , etc. Each The linear model given for our diagnostic system now becomes
of these deltas, by virtue of the data normalization process, is a ι Xe
function of inlet temperature (72), inlet pressure (P2) and the
Z = HeXe+HsXs= 'He L ! Hs] J
independent base parameter ( EPR ). Thus we must provide for a
L ! J L Xs
total of 72+3=75 error parameters. We have the following
measurement and fault vectors where Η is an Μ χ (Ne+Ns) matrix obtained by concatenating the
engine and sensor fault matrices and X is a combined vector of
Ne=10 Ns=15 M=12 engine performance and measurement error deltas. Since the
Engine Faults Sensor Faults Measurement number of unknowns (Ne-¥Ns = Ne+M+3) is greater than the
(Xe) (Xs) Delta
number of equations ( M ) in eq.5, we have an underdetermined
(Z)
system which has no unique solution. Generalized or pseudo
ETAF GAMT2err GAMT2
inverses may be employed to generate a solution to eq.5,
GAMF Nlerr N1C2 however, this technique affords little control over which of the
ETACL N2err N2C2 infinitely many possible solutions is chosen. Since the model
represents a physical entity, a knowledgeable analyst would have
GAMCL T3erT T3C2
the ability to rate each solution as to its suitability. For example, a
ETACH PS 3 err PS3C2 solution which suggests that a compressor's efficiency was
GAMCH T4err T4C2 increasing while its pumping capacity was decreasing after
several thousand flight hours, would not be as likely as a solution
ETATH PS4err PS4C2
indicating that both compressor functions were deteriorating
Ā5 W F err WFC2 together. The need to limit the solution to a set of likely events led
ETATL T7err T7C2 early researchers to consider algorithms which employed
Ā6 FNerr FNC2 probabilistic methods. This required a slight reformulation of the
model (5) as follows:
PT25err PT25C2
Ζ = ΗΧ + Θ (6)
T25err T25C2
where Θ was a normally distr
T2err
the nonrepeatabilities of the
P2err mean and covariance R. The
P7err now be re-interpreted as inclu
described above other than
where Z = f(XeXs). It should be noted that the elements
covariance term R ofis the
an Xs
Μ χ
vector are measurement errors attributed to the raw data could be calculated from
parameters, whereas the elements of the Z vector are deltasprecisions.
of For example,
corrected quantities. Thus, for example, the measurement delta
for NI C2 will be influenced by errors in raw N1,T2,P2 , and Ρ 7 R = Hs DIAG[ OqAM
since N1C2 =f(Nl,T2) = NU JĪ2 and NlC2baSe = f(EPR) =
f(P7/P2). Perturbations in any of these four parameters will affect > ^ΡΓ25 » &T2 >
the value of the N1C2 delta in Z. As with the engine fault

2030

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
PS2err Nlerr N2err T3err PS3err T4err PS4err WFerr T7err FNerr PT25err T25err T2err PT2err PT7err

GAMT2 -1.56 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.9940 -0.4340

N1C2 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 -0.50 0.7040 -0.7040

N2C2 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 -0.50 0.2822 -0.2822

T3C2 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 -1.00 0.4307 -0.4307

PS3C2 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 -0.0726 -0.9274

T4C2 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 -1.00 0.5542 -0.5542

PS4C2 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.3106 -1.3106

WFC2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.00 -0.62 1.0899 -2.0899

T7C2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 -1.00 0.7845 -0.7845

FNC2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 0.00 0.4617 -1.4617

PT25C2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 0.00 0.00 -0.2904 -0.7096

T25C2 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.00 -1.00 0.4639 -0.4639

Figure 3: TSMF Turbofan Sensor Fault Coeffieicn

The solution to system (6) is given by the estimate X De Nex M Engine Diagnostic Matrix

Ds NsxM Sensor Diagnostic Matrix


X = X + D[Z-HX] (la)
where X = Nxl vector of a-priori estimates for engine and
measurement error fault deltas. = P0HT [HP0HT + R]-1 (lb)

GAMT2 N1C2 N2C2 T3C2 PS3C2 T4C2 PS4C2 WFC2 T7C2 FNC2 PT25C2 T25C2

ETAF 0.1490 0.0190 0.0385 0.1927 -0.0151 0.0807 -0.0985 -0.2102 0.0156 0.2105 0.2447 -0.5050

G A MF 0.1831 -0.4145 -0.0181 0.0015 -0.0457 0.0186 -0.0775 -0.1227 -0.0140 0.2468 0.2795 -0.0170

ETACL 0.0782 0.0136 -0.0135 -0.2693 0.1131 -0.1011 0.0624 -0.0586 -0.0838 0.0401 -0.0417 0.1299

GAMCL 0.1463 -0.5093 0.0887 0.1524 0.1767 -0.0565 0.2181 -0.0686 -0.2056 0.0701 -0.2346 -0.0312

J)e ETACH -0.0119 0.1810 0.7430 0.4070 -0.3403 -1.0351 0.2128 0.01
GAMCH 0.0033 0.2219 -1.3177 0.2554 -0.2560 0.0541 0.1694 0.0066 -0.1037 0.0342 -0.0848 0.1115

ETATH -0.0043 0.1170 0.3950 -0.3102 -0.1278 0.5985 -0.0694 -0.1046 -0.2780 0.0811 0.0161 -0.0193

A5 -0.0237 0.1484 0.0420 0.0822 0.0425 -0.1693 -0.3794 0.1049 -0.0646 0.0772 -0.0819 0.0237

ETATL 0.0613 0.1528 0.0654 0.0982 0.0552 0.0881 -0.1128 -0.1668 -0.1596 0.1018 0.2367 0.0970

A6 -0.0320 -0.0424 0.0829 -0.1403 -0.0652 0.1312 0.0364 0.0531 -0.0013 -0.0274 -0.1142 -0.0544

PS2err -0.2354 0.0568 -0.0034 0.0322 -0.0295 0.0186 -0.0292 -0.0114 0.0059 0.1450 -0.0356 0.0121

Nlerr -0.0277 0.4165 -O.0246 -0.1548 -0.0065 -0.0031 -0.0778 -0.0465 0.0604 -0.0121 0.0824 -0.1270

N2err -0.0003 -0.0502 0.3956 -0.0732 0.0613 -0.0347 -0.0453 -0.0038 0.0190 -0.0058 0.0217 -0.0257

T3err -0.0112 -0.1106 -0.0706 0.4924 -0.0814 -0.0737 0.0145 0.0009 -0.0158 0.0124 0.0261 -0.0908

PS3err 0.0282 -0.0794 -0.1750 -0.2253 0.2172 0.1354 -0.0469 -0.0050 0.1008 -0.0272 0.0332 0.0235

T4err -0.0116 -0.0444 -0.2790 -0.0746 0.0415 0.4651 -0.0837 -0.0054 -0.0258 0.0037 0.0099 -0.0570

PS4err 0.0204 -0.1112 -0.0725 0.0290 -0.0207 -0.1134 0.3037 -0.0502 0.1040 -0.0706 0.0496 0.0038

Ds WFerr 0.0126 -0.1049 -0.0252 0.0029 -0.0231 -0.0411 -0.0793 0.4377


T7err -0.0065 0.1364 -0.0198 -0.0498 0.0364 -0.1599 0.1643 -0.2771 0.4893 -0.0247 -0.0479 -0.0363

FNerr -0.0929 -0.0159 0.0067 0.0229 -0.0027 0.0252 -0.0650 -0.1026 -0.0144 0.5348 -0.2672 -0.0087

PT25err 0.0228 0.1083 0.0277 0.0481 0.0161 0.0091 0.0456 0.0010 -0.0279 -0.2672 0.5587 -0.0098

T25err -0.0085 -0.1814 -0.0206 -0.1815 0.0333 -0.0573 0.0038 -0.0332 -0.0230 -0.0095 -0.0106 0.6853

T2err 0.0283 0.0377 -0.0303 -0.2043 0.0378 -0.0836 0.0458 0.0470 -0.0377 0.0216 -O.0441 -0.1388

PT2err 0.2167 0.0621 0.0024 0.0113 -0.0154 0.0248 -0.0285 0.0309 0.0485 0.0066 -0.1426 0.0007

PT7err 0.0286 -0.0256 -0.0062 0.0045 -0.0181 0.0289 -0.0894 -0.0738 0.0211 -0.1624 -0.1148 0.0014

Figure 4 Diagnostic Matrix ( D )


2031

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
The term P0 appearing in (7b) is an Ν χ Ν positive two equations allow the state and its associated statistics to be
semi-definite matrix (N=Ne+Ns) which functions as a weighting extrapolated in time from knowledge of the system dynamics
factor to skew the solution in an acceptable direction. A sample (<P(k)) and uncertainties ( Q(k )). The remaining equations govern
diagnostic matrix D for the example we have been considering the state estimation and optimal gain calculation. The system we
appears in Figure 4 below. have described in equations 7, can be obtained by setting <P(k)=I,
Q(k)=0 and constraining P(k) to be constant ( =P0 ) for all epoches
This matrix provides a set of inverse relationships, expressing k. The fault coefficients H(k) and measurement covariances R(k)
module performance shifts and measurement error in terms of are functions of discrete time k , only to the extent that they are a
measurable parameter shifts. In addition, the predictor/corrector function of some independent parameter ( EPR ) which may be
form of equation (7a) allows optimal usage of any a-priori changing. The recursive nature of the Kaiman filter equations
estimates of engine/sensor degradation that might be available. may be exploited in the diagnostic system if a suitable
Λ

degradation model for module performance as a function of time


Before discussing in more detail, the selection of P0 and X and
can be derived. Such relationships have been derived empirically
their impact on the final solution, we will digress briefly to
for several classes of commercial turbofan engines. This gives
discuss the genesis of this technique and explore its natural
us, in effect, the state transition matrix <P(k) to be used in the
generalization.
foregoing equations, and a performance trend system suitable for
flight acquired data analysis could be constructed. Simulation
This approach is a special case of a discrete time Kaiman filter
studies performed in the early 80's revealed that a fully recursive
which provides optimal estimates for the states of a system at
filter was computational burdensome and required alternate
each time epoch k. A readable account of Kaiman Filtering can be
filter formulation to avoid numerical instabilities and filter
found in [ref. 7]. The system equations governing the filter are
divergence. A hybrid system which produced results
commensurate with that of the Kaiman filter (in this application)
but was immune to these computational difficulties was
State Extrapolation :
constructed by suppressing the covariance update and
X(k I k - 1) = <P{k) X{k - ' ' k - 1) extrapolation. This produced a system with a fixed gain
(diagnostic matrix D as in equation 7b) and a state extrapolation
Co variance Extrapolation :
where the engine/sensor performance deltas from the previous
P(k I k - 1) = <P(k) P(k - 1) <P(k)T + Q(k)
analysis point were projected in time via the degradation model
to be used as a-priori estimates for the current analysis point.
Kaiman Gain :
Sample simulation runs, for the mixed flow turbofan, are
D(k) = P(k I k - 1) H(k)T [ H(k) P(k I k - 1) H(k)T + R(k) ]_1
illustrated in Figures 5. In this chart, we have implanted
degradation at rates equal to that assumed by the filter (7a) as well
State Update mģ
as deterioration at accelerated rates of 2x and 5x the model rate. It
X(k I k) = X(k' k - 1) + D(k) [ Z(k) - H(k) X(k I k - 1)] is clear that the filter works fairly well when we already know the
answer, but not as well when we do not, which is often the case.
Co variance Update :
The usefulness of this approach as a diagnostic tool will depend
P(k) = [I - D(k) H(k)' P(k I k - 1)
on fidelity of the assessments. Faulty modules/sensors must be
isolated and false alarms (misassessments) kept to a minimum. In
the next section we will explore this concept more closely by
with initial conditions Χ (0/-1 ), P(0)=Po known and Q(k), R(k),
<P(k), and H(k) assumed known for all discrete time k. The first
considering the accuracy limitations inherent in this approach.

2032

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
Filter Response to Varying Degradation Rates

Fig 5a: Normal Degradation Fig 5b: 2x Normal Degradation Fig 5c: 5x Normal Degradation

ACCURACY OF THE BASIC LINEAR DIAGNOSTIC


ex - χ - χ = χ - χ - Dz + DHx
SYSTEM
= (/ -DH) (x - χ) - ΟΘ

The examples of Fig.5 illustrate that the Kaiman filter and its
hybrid cousins, work well when you already know the answer.
Since Θ is a random vector, the error term is also random ve
Since the actual degradation dynamics of gas turbines
with is not
mean and covariance as follows
known to any great degree of precision, the accuracy of such a
diagnostic system is a matter of concern. The magnitude
E[ex]of
= (I -DH) (x-x)
errors, produced by this approach depends on many factors and is
quite complicated. We explore this question by returning to the
V[ex]=DRDT (8)
simplest case, that of a test-stand analysis consisting ofIt
a single
is clear from (8) that the exp
data point snapshot. The system is described by equationswhen
6 and 7. I=DH. The latter is not tru
Misassesments in the engine/sensor performance deltas
systemare of equations under con
described by the error term ex = x-x which becomes
error term will depend dire
2033

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
estimate for deterioration ( χ ) matches the actual degradation x. For the particular diagnostic system in our example, this becomes
With χ held constant, the actual error ex from any given analysis
will also depend on the level of noise ( Θ ) present in the
measurements ( Ζ ). Thus, the actual error depends on x, H, R and 0.0026 < < 1-8736
II χ II
P0 ( since D=f(HtP0tR) ). As an illustration, a simulation of 20
noisy measurement vectors representing each of three different
implant faults was presented to the diagnostic system of Figure 4 Equation 9b provides a bound for the average normalized error,
and equations 7 with the a-priori estimate χ set at zero (i.e. the actual error being slightly more (or less) depending on the
engine status assumed to be nominal). The results appear in signature of the measurement noise vector theta. The
Figure 6 below. combination of engine/sensor faults which will admit the least
error is given by the (Ne+Ns) eigenvector associated with the
The RSS of the implanted faults and the mean error for each minimum eigenvalue whereas the worst case possible is in the
analysis appears at the bottom of the column. These represent the direction of the max eigenvalue. All other faults being
norm (or magnitude) of the vectors χ and χ - E[ χ ] = E[ ex ] somewhere in between. This bound is valid for a system
respectively. Since the mean error magnitude will increase consisting of Ne+Ns faults and M measurements and will vary as
proportionately to the magnitude of the implant vector (eq.8), a the matrix D varies. The matrices H and R are fixed (at a
good comparative metric is the normalized error WerrorW/WxW. This particular engine operating point), leaving P0 as the sole vehicle
metric changes slightly between the test cases selected, for tuning the diagnostic system to perform in an optimal fashion.
beginning at .5776 with a degraded HPC, decreasing to .5519 Optimality for such systems is defined as the ability to assess with
when an additional HPT fault is added and finally increasing to acceptable accuracy those faults or combination of faults that are
.5693 when a fuel flow measurement bias is added to both faults. considered reasonable or likely to occur in the given engine
These slight variations are indicative of the general statement operational environment under consideration. Test stand
that error does not directly depend on the number of faults applications, for example, may have both pre and post overhaul
present. Indeed it is possible to generate two single faults whose engines to analyze which would offer engine fault combination
combined error is less than either taken individually. For a fixed possibilities different from those which would be anticipated in
diagnostic system ( D ) with no a-priori information ( i.e. χ = 0) the long term degradation of an in-flight analysis. This
the error incurred in estimating an unknown fault X is related to subjectivity has made the exercise of tuning more of a black art
the eigenvalues λ of the matrix (I-DH)(I-DH)T. In particular, the than an exact science. The diagnostic system designer typically
squared normalized error is a weighted average of these strives for some level of uniformity among the class of faults to be
eigenvalues. This relationship gives us bounds for the normalized assessed; with engine fault assessment as one concern and
errors achievable by the system as follows: measurement (or sensor) error detection as another. Since the
inception of this technique in the mid 70's, the clarion call has
been to tune the system to provide high accuracy in both engine
Xa'2 xì and sensor faults alike. Unfortunately, with unknowns
Il ex II2 _ / _ ν1 ι
H χ II2 ~ ~ ' outnumbering equations approximately 2-1, the system behaves
like an Ν dimensional seesaw (Figure 7). When emphasis is
i ' 1 (9α) directed toward one fault it is accompanied by an attendant loss in
another. If all (single engine and sensor) faults are considered
equally important, the best the designer can achieve is a balanced
π- Il e, II β - (9b) seesaw which assesses all Ne+Ns singletons at the same
*' M,N II * II
approximate level of accuracy. What then, would be a reasonable
level to expect?

2034

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
ESTIMATE χ ESTIMATE χ ESTIMATE χ
IMPLANT
Χ
MEAN
X
I Š!Ď! IMPLANT I Š!Ď! IMPLANT
MEAN Χ
MEAN ļ STL
ETAF 0.0272 0.1438 ETAF 0.0167 0.1438 ETAF -0.2986 0.1438

GAMF -0.0020 0.1473 GAMF 0.0312 0.1473 GAMF -0.1528 0.1473

ETACL -0.0858 0.1085 ETACL -0.0925 0.1085 ETACL -0.1804 0.1085

GAMCL 0.1148 0.1773 GAMCL 0.0418 0.1773 GAMCL -0.0612 0.1773

ETACH -1.00 -0.7570 0.2081 ETACH -1.00 -0.860$ 0.2081 ETACH -1.00 -0.8382 0.2081

GAMCH -2.00 -0.8735 0.2625 GAMCH -2.00 -0.8590 0.2625 GAMCH -2.00 -0.8492 0.2625

ETATH -0.2041 0.2021 ETATH -1.00 -0.8462 0.2021 ETATH -l.OO -1.0031 0.2021

"Ā5 0.0448 0.1211 ĀŠ Õ5Õ 0.3957 0.1211 Ā5 ÕÍÕ 0.5530 0.1211


ETATL -0.0783 0.1001 ETATL -0.1463 0.1001 ETATL -0.3965 0.1001

~A6 -0.0205 0.0632 A6 -0.1746 0.0632 A6 -0.0950 0.0632


PS2SEN 0.0140 0.1551 PS2SEN 0.0229 0.1551 PS2SEN 0.0058 0.1551

NI SEN -0.0935 0.0814 N1SEN -0.0838 0.0814 Ν ISEN -0.1535 0.0814

N2SEN 0.2153 0.0768 N2SEN 0.1976 0.0768 N2SEN 0.1919 0.0768

T3SEN -0.0382 0.1281 T3SEN 0.0177 0.1281 T3SEN 0.0191 0.1281

PS3SEN 0.3799 0.0781 PS3SEN 0.5726 0.0781 PS3SEN 0.5651 0.0781

T4SEN 0.1582 0.0767 T4SEN 0.0872 0.0767 T4SEN 0.0791 0.0767

PS4SEN -0.0505 0.1029 PS4SEN -0.1498 0.1029 PS4SEN -0.2251 0.1029

WFSEN 0.0070 0.2647 WFSEN 0.1046 0.2647 WFSEN ĒĪ0 0.7611 0.2647
T7SEN 0.1502 0.1984 T7SEN 0.2403 0.1984 T7SEN -0.1754 0.1984

FNSEN 0.0020 0.1611 FNSEN 0.0035 0.1611 FNSEN -0.1504 0.1611

PT25SEN 0.0401 0.1047 PT25SEN 0.0125 0.1047 PT25SEN 0.0140 0.1047

T25SEN -0.0751 0.1186 T25SEN -0.0624 0.1186 T25SEN -0.1122 0.1186

T2SEN -0.0468 0.0838 T2SEN -0.0281 0.0838 T2SEN 0.0424 0.0838

PT2SEN -0.0141 0.1797 PT2SEN -0.0233 0.1797 PT2SEN 0.0230 0.1797

PT7SEN -0.0867 0.1746 PT7SEN -0.0762 0.1746 PT7SEN -0.1869 0.1746

11x11 2.236 11x11 2.5 11x11 2.916

llmean errori I 1.2915 II mean errarli 1.3798 llmean errorll 1.6598

llerrorll/llxll 0.5776 llerrorll/llxll 0.5519 llerrorll/llxll 0.5693

Figure 6 Simulation Results Using No Α-Prior Information

A partial answer can be obtained


we assume that the average level
approximately the same as the a
infinitely possible combination
distribution law, then the expect
the mean of the square roots of the eigenvalues of
(l-DH)(l-DH)T. It can be shown, that a lower bound for this error

is given by J 1 - (M/N) . Thus, for the example we have been


considering with M-12 and N=10+15=25, the average error will
be > V 0.52 = 0.72 tuning efforts notwithstanding. This
certainly would not cause the would-be designer to be imbued
with overwhelming optimism. Nevertheless, the situation is not
as bleak as it might appear either, since tradeoffs between faults
can be made and better accuracy achieved for a selected subset.

Indeed this has been the case in our example, where 1λΑνο = 0.77
and the error for the combination of faults reported in Figure 6
were all at the 0.56 level.

Figure 7 Tuning the


2035 System

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
Λ

The desire for increased accuracy in engine fault assessments component, then an alternate estimate for XS(I) was computed by
combined with the need to remain robust with respect to the following formula and the estimation process repeated with
measurement error has led designers to incorporate additional this new estimate appearing as an a-priori value.
analysis and/or form special purpose hybrid systems. Several
enhancements used by the authors are presented in the sequel.
ak [X-sW - Xs0LD(k)]
Ic
%SNEw(I) = ^SOLD(I) ~ ~
ENHANCEMENTS TO THE BASIC MODEL Σα1
k

ak = [ D s H s ] k ¡ = k, Ith element of the matrix product D s H s


Two simple additions to the basic linear model which can
impact the overall accuracy of the system are the inclusion of
a-priori estimates of performance change and a facility for large
measurement error accommodation.

The first was introduced in the previous section and is an


inherent part of the predictor/corrector form of the linear
estimator. Performance degradation models for several classes
of turbofan engines have been developed from empirical studies
(ref. 8). Such models were incorporated into Kaiman filter based
diagnostic programs developed by one of the authors (Volponi)
and engineers at Pratt & Whitney Aircraft in the early 80's. The
models provide an expected level of change in engine
performance as a function of flight cycles and serves to push the
analysis in the proper direction. As a generality, the closer the
model is to the actual degradation of the engine under scrutiny,
the greater the accuracy of the estimates. However, even a rough
guess in the proper direction is beneficial. This can be illustrated
by returning to the example in Figure 5b which in effect is using
an a-priori model 1/2 that of the implanted faults. Comparing
this to an analysis with no a-priori model ( i.e. χ =0 ) in Figure 8,
demonstrates the advantage.

Another area in which some improvement can be made is in


how we process measurement error. In the preceding, we treated
sensor faults (Xs) in the same regard as engine faults (Xe), i.e., as
another set of states to be estimated. If we change our
requirements slightly and view the sensor faults as a detection
mechanism whose purpose is to absorb small measurement
perturbations and flag large excursions then the tuning can be
slanted more toward the engine faults. If only small errors are
present, such an analysis would provide greater accuracy. If on
the other hand a large perturbation were present, the linear gain
would magnify the engine fault misassessment and produce an
unacceptable analysis even if the sensor fault was detected and
flagged. Unless a means could be found to recover from such
detected errors, this method would yield no useful information on
engine performance until such time that the source of error was
corrected and more reliable measurements were obtained. In the
early 80's, a simple algorithm was devised to accommodate
single large measurement errors (ref. 9), thus allowing a palatable
engine assessment to be made.

The approach consisted of establishing a set of threshold values


for each of the Ν s sensor faults and comparing the estimated Fig 8: 2x Normal Degradation with Χ = 0
faults (Xs ) to these limits to determine if a large error was
present. If such a condition was detected, say in the Ith
2036

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
jL

Engine ! Sensor - Compare

f xe ^ Fault A's ensor s

Α-Priori J _ χ a = χ + d ( Ζ- Η Χ ) Jk-*, To Threshold ~ no


Estimates ļ xs w Values yes

Compute New
Α-Priori Value for

( 2α* X, (k) - Xs (k))

Σ 2 MAX (ts / Threshold)


ak = [ Ds Hs ] ijth

Figure 9 Large Error Recovery Logic

This relationship is derived in ref. 10 and represents the best is possible for the recursion to diverge, thus making it prudent to
a-priori estimate in the sense that total assessed measurement limit the number of allowable iterations.
error is minimized. A flow chart of the procedure appears in
Figure 9. As a simple illustration, consider a measurement delta An alternate approach which circumvents the need to iterate to
vector representative of a +1% change in HPT efficiency some tolerance, is based on an augmented linear model which
(column 7 in Figure 2). The snapshot analysis of this vector using would generate an expanded diagnostic matrix such as the one
system (7a, 7b) with xbar=0 appears in the 1st column of Figure depicted in Figure 1 1 . These matrices can be computed at various
10. If we superimpose a 2% bias in Wf (Fuel Flow) and operating points and include columns for all of the measured
re-analyze, we generate estimates with and without the recovery parameters. This allows any parameter to be selected as the base
logic. These appear in columns 2 and 3 of Figure 10. Comparing parameter. The resultant diagnostic matrix is produced by
column 2 row by row with column 1, it can be seen that the deleting the corresponding column. This method is a
inclusion of the Habías detracts from the ability of the system to generalization of a technique reported by one of the authors
assess the implanted HPT fault and contributes to the general (Urban) in ref. 5 for square invertible systems having the same
level of misassessment in other engine/sensor faults. A form as equation 1. The use of an augmented matrix approach
comparison of columns 1 and 3 reveals the benefits of the requires that special attention be given to data normalization,
recovery process wherein virtually all of the the misassesments baseline generation, and model description and is outside the
due to the sensor bias are removed. scope of the present discussion.

The astute reader has probably already realized that the Summary
technique described above would have difficulty in recovering
from an error in the base parameter. For example, if EPR = P7/P2 In conclusion, the basic concepts and mathematics of a
were the base power setting parameter, than a large perturbation methodology that allows relative assessment of simultaneously
in either P7 or P2 would produce a large error in EPR which in occurring multiple faults in gas turbine engine modules and
turn would affect all the model numerics including the sensors, have been presented in detail. Included are a detailed
calculation of the measurement deltas themselves. Remarkably, exposition of a thermo-mathematical/differential equation based
experience has demonstrated that the diagnostic matrix (D) in the model of a representative type of gas turbine engine, description
linear model can still be used effectively to detect the EPR of correction methods to minimize data scatter, and the
problem even though the matrix itself is a function of EPR. The mathematical design aspects of a demonstratedly successful
road to recovery , however, is somewhat longer. One method that approach to data filtering to deal with residual scatter caused by
has proven reasonably effective is to use the above technique tolarge or small sensor errors, or data correction or engine system
generate an initial best guess for a new EPR value and reiterate modelling errors. Since prior papers by the authors (see refs.) and
the entire analysis in a recursive fashion until error estimates others are available that have documented program results using
stabilize, i.e. the difference between the kth estimate and the the methodology, the emphasis in this presentation has instead
k-lst estimate is below some preselected tolerance level. One been on the more theoretical design considerations. Sufficient
problem with this approach is that under certain conditionsdetail has been included to bare most of the major aspects of our
(presence of large errors in other parameters in addition to EPR) it systems and to provide a guide to those interested in either
2037

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
diagnostic system design or just in the more limited aspects of rnid-70's. Also, thanks to Joseph M. Kos of HSD for his help and
engine modelling, data correction, or data filtering. contributions in the initial development of the methodology, and
to GE and P&WA for their cooperation on programs, and for their
References endorsement and use of the concepts. Finally, we would like to
acknowledge the HSD Systems Design CAE group for their
1 . Urban, L. Α., Gas Path Analysis - A Tool for Engine assistance, with special thanks to Candy Gallo and Lori Williams
Condition Monitoring , 33rd Annual Flight Safety for their patience and skill in preparing this document.
Foundation Seminar, Christchurch, New Zealand, Sept.
1980.
2. Urban, L. Α., Gas Turbine Engine Parameter Estimated Fault χ
Interrelationships , HSD UTC, Windsor Locks, Ct., 1st ed. Without Recovery
1967, 2nd ed. 1969.
3. Urban, L. Α., Gas Path Analysis Applied to Turbine Engine
Implanted 1+1% ETATH +1 % ETATH +1% ETATH
Fault X +2% Wf bias +2% Wf bias
Condition Monitoring , AIAA/S AE Paper 72-1082, Nov.
ETAF 0.0294 -0.3880 0.0362
1972.
4. Urban, L.A., Gas Turbine Engine Analyzer , U.S. Patent No. GAMF 0.0000 -0.2453 0.0039

3,731,070, May 1973. ETACL -0.0055 -0.1232 -0.0036


5. Urban, L. Α., Parameter Selection for Multiple Fault
GAMCL -0.0147 -0.1535 -0.0124
Diagnostics of Gas Turbine Engines , ASME Paper
74-GT-62. ETACH 0.0867 0.1259 0.0861

6. Urban, L.A., An Approach to the Analysis of Altitude GAMCH -0.0449 -0.0365 -0.0451
Performance Problems of Axial Flow Turbojet Engines , ETATH 0,5943 0,3426 0.5983
Masters Thesis, Columbia
"A5 -0.1055 0.0932 -0.1087
University School of Engineering, Jan. 1954.
7. Gelb, A. et al., Applied Optimal Estimation , M.I.T. Press, ETATL 0.0718 -0.2543 0.0771
Cambridge, Mass., 1974. "Ā6 0.0801 0.1946 0.0782
8. Sallee, G.P., Performance Deterioration Based on
PS2SEN 0.0014 -0.0214 0.0017
In-Service Engine Data; JT9D Jet Engine Diagnostics
N1SEN 0.0182 -0.0728 0.0197
Program , NASA CR- 159525, Nov. 1973.
9. Volponi, A.J., A Large Measurement Error Recovery N2SEN 0.0225 0.0173 0.0226
Algorithm for Gas Turbine Module Performance Analysis , T3SEN -0.0345 -0.0363 -0.0345
Hamilton Standard Report no. 3430, April 1982.
PS3SEN -0.1258 -0.1490 -0.1255
10. Volponi, A. J., Gas Path Analysis - An Approach to Engine
Diagnostics , Time-Dependent Failure Mechanisms and T4SEN 0.0364 0.0295 0.0365
Assessment Methodologies, Proceedings of the 35th PS4SEN -0.0155 -0.1175 -0.0138
Symposium, Mechanical Failures Prevention Group,
IMÜÜ -0.0368 0,8347 tMU
Gaithersburg, Md., April 20-22, 1982, Cambridge
T7SEN -0.0978 -0.6623 -0.0887
University Press, 1983.
11. Abernethy, R.B., Uncertainty in Gas Turbine FNSEN 0.0166 -0.1868 0.0199
Measurements , AIAA Paper 73-1230, Nov. 1973. PT25SEN 0.0033 0.0057 0.0033

T25SEN -0.0043 -0.0712 -0.0032


Acknowledgements:
T2SEN -0.0187 0.0734 -0.0202
The authors gratefully remember the late Donald A. Prue, Sr.
PT2SEN 0.0129 0.0760 0.0118
V.P. H.S.D., for providing the opportunity, freedom and
PT7SEN 0.0181 -0.1277 0.0204
encouragement over the years to pursue this work. Also, thanks to
Dr. Gerwin Dienger of Lufthansa, and particularly to TuT™"""™ 1 .65 0.47
Sven-Goran Danielsson of SAS for his faith, technical courage
and cooperation in developing important major programs in
llexll/llxll "Õ47 Õ74 021
the

Figure 10 Analysis with Large Error Recovery

2038

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
EPR GAMT2 N1C2 N2C2 T3C2 PS3C2 T4C2 PS4C2 WFC2 T7C2 FNC2 PT25C2 T25C2

ETAF 0.1065 0.1490 0.0190 0.0385 0.1927 -0.0151 0.0807 -0.0985 -0.2102 0.0156 0.2105 0.2447 -0.5050

GAMF 0.0667 0.1831 -0.4145 -0.0181 0.0015 -0.0457 0.0186 -0.0775 -0.1227 -0.0140 0.2468 0.2795 -0.0170

ETACL 0.0445 0.0782 0.0136 -0.0135 -0.2693 0.1131 -0.1011 0.0624 -0.0586 -0.0838 0.0401 -0.0417 0.1299

GAMCL 0.1692 0.1463 -0.5093 0.0887 0.1524 0.1767 -0.0565 0.2181 -0.0686 -0.2056 0.0701 -0.2346 -0.0312

ETACH 0.0965 -0.0119 0.1810 0.7430 0.4070 -0.3403 -1.0351 0.2128 0.0151 -0.0757 0.0268 -0.0585 0.0795

GAMCH 0.1157 0.0033 0.2219 -1.3177 0.2554 -0.2560 0.0541 0.1694 0.0066 -0.1037 0.0342 -0.0848 0.1115

ETATH 0.1351 -0.0043 0.1170 0.3950 -0.3102 -0.1278 0.5985 -0.0694 -0.1046 -0.2780 0.0811 0.0161 -0.0193

A5 0.1760 -0.0237 0.1484 0.0420 0.0822 0.0425 -0.1693 -0.3794 0.1049 -0.0646 0.0772 -0.0819 0.0237

ETATL -0.0350 0.0613 0.1528 0.0654 0.0982 0.0552 0.0881 -0.1128 -0.1668 -0.1596 0.1018 0.2367 0.0970

A6 0.0572 -0.0320 -0.0424 0.0829 -0.1403 -0.0652 0.1312 0.0364 0.0531 -0.0013 -0.0274 -0.1142 -0.0544

PS2eiT -0.0683 -0.2354 0.0568 -0.0034 0.0322 -0.0295 0.0186 -0.0292 -0.0114 0.0059 0.1450 -0.0356 0.0121

NI err -0.0299 -0.0277 0.4165 -0.0246 -0.1548 -0.0065 -0.0031 -0.0778 -0.0465 0.0604 -0.0121 0.0824 -0.1270

N2err -0.0250 -0.0003 -0.0502 0.3956 -0.0732 0.0613 -0.0347 -0.0453 -0.0038 0.0190 -0.0058 0.0217 -0.0257

T3err 0.0038 -0.0112 -0.1106 -0.0706 0.4924 -0.0814 -0.0737 0.0145 0.0009 -0.0158 0.0124 0.0261 -0.0908

PS3err -0.0884 0.0282 -0.0794 -0.1750 -0.2253 0.2172 0.1354 -0.0469 -0.0050 0.1008 -0.0272 0.0332 0.0235

T4err 0.0061 -0.0116 -0.0444 -0.2790 -0.0746 0.0415 0.4651 -0.0837 -0.0054 -0.0258 0.0037 0.0099 -0.0570

PS4crr -0.1490 0.0204 -0.1112 -0.0725 0.0290 -0.0207 -0.1134 0.3037 -0.0502 0.1040 -0.0706 0.0496 0.0038

WFerr -0.1943 0.0126 -0.1049 -0.0252 0.0029 -0.0231 -O.0411 -0.0793 0.4377 -0.2771 -0.1762 0.0018 -0.0525

T7err 0.0557 -0.0065 0.1364 -0.0198 -0.0498 0.0364 -0.1599 0.1643 -0.2771 0.4893 -0.0247 -0.0479 -0.0363

FNerr -0.2490 -0.0929 -0.0159 0.0067 0.0229 -0.0027 0.0252 -0.0650 -0.1026 -0.0144 0.5348 -0.2672 -0.0087

PT25err -0.1760 0.0228 0.1083 0.0277 0.0481 0.0161 0.0091 0.0456 0.0010 -0.0279 -0.2672 0.5587 -0.0098

T25err 0.0023 -0.0085 -0.1814 -0.0206 -0.1815 0.0333 -0.0573 0.0038 -0.0332 -0.0230 -0.0095 -0.0106 0.6853

T2err 0.0044 0.0283 0.0377 -0.0303 -0.2043 0.0378 -0.0836 0.0458 0.0470 -0.0377 0.0216 -0.0441 -0.1388

PT2err -0.1169 0.2167 0.0621 0.0024 0.0113 -0.0154 0.0248 -0.0285 0.0309 0.0485 0.0066 -0.1426 0.0007

PT7err 0.5793 0.0286 -0.0256 -0.0062 0.0045 -0.0181 0.0289 -0.0894 -0.0738 0.0211 -0.1624 -0.1148 0.0014

Figuře 11 Augmented Diagnostic Matrix

2039

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
Ζ
£
o
pa
tf
ρ
Η Ρ
ļ-H tf
£ Ρ
Χ o ο
ΗΗ Ρ ΗΗ
tf
Q tf
rv

Ζ Q
U χ
ω Χ ΝΗ
On ΗΗ
Q
Oh S Ζ
ω
«3 Ρ
α.
ο tf
ο <c
tf
co

s
£
Η

2040

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
C/5
a

2

2
O
su
s
8
J

s
tí CS
2 UJ
W 0¿
U D
2 O

5 tí

o
α Χ
6
D S
Η 2

£ g

§ <

α
UJ
Χ
S
-j

C/5

2041

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
. - to
. ¡ξ «V,

~§P~«
t¿ Sc £
.«Λ ·■
A
η

S *F

«ii
η Ο δ δ ι ® ® '

li

J2»**
^ ·?*
a si
' 11
11 £
___
^ λ>1£ U1 ÛT ¿' K Ur
u. 3 > ¡jT ^ ^ O ^ O?
"
" ìr
k, fe fe
">. i">.
£ „„» »
S: S:

* * -Γ2- g
i 'e «í
«i ® iT δ
i 11 .. ε «· 8
■ U" I * ^ 1
.. ■ 'C U" ,s I c- + «r
..
> 55
^ b£
e- "
*
*
1
Hn
·**
-, -
- sł
s
^ 31 V.3 " I k u. ι m ι

11
?Pli
?Pli ^
11 11
řř%£
ïf
11
bri
ς
k ^ it; k í
«O k «o Uj it; ζΓ k «CL ^

I
I
1„
co
Z
0
H- I S -> _?■

ϋ
a.lele
8 ,Ü
,Üřt> «ik1^·
1^·
1a
řt>
5 a. «Τ> ^ II
w

Ε ^
Iί"
Λ» "*Ρ
<·? J 11 £1 ■«.
fe

§ ΕSi S fe Λ
fe Κ^^ fe
Κ ί"
i iAK
I V?
AK ^ ^
<·? ϊΓ
ϊΓ §>¿£1
g g "I 11 r. ^ Ö fe V? + fe ,M| l·» „ _ fe <-«|M
11 ^ ,M| „ <-«|
I».
H I fe I4· r. è ' § /

Ê ^ î f è f s I § / feļ-~
1 »V " *!f ! j I Λ* ■ "t I Τ "
0

1 ί fk-im«
·? líit " 5*-8 " ^" íV
" ick 1 ^+. fir 'ι -i -τ ¡¡ ~»k
+ ι 'ι !»
S
_)
o
S ·< ·? k-im« ϊ-Μ 5 Í " · ick řf J " * 1 lie ^ . »? + ι I *1-" 'ι S?? - - %ΐ" 6"cř? ~»kK
i " I? ^ /. Ufi* 1 V Ημΐ Β ïi^ ί !
0
(Vi
ce
I| ee κκ 4ί
fekfek
4ί gs ^JJ^ ^^ ^fec-h
1 bĄ?§ e2 fek
e ν fek
¡-ι.«¡-ι.«
ξ c ιΒb"|c
ι ¡IvI ¿¿ "IV
8 I ££ *1"
*1" äk
äk
1 es to "*r «ο vo Γ-

2042

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
i
η

¥
κ, ί ^ C I«"
«?"^
"η ¿ 5
^ J' ^
S

* Ì
-a«8 ι ι
-«υ+ '« « ^^ε
'« ^ «"'i«
ε «Γ 'i«
k

-řšM κ" <υ


κ" <υ ε ε^ ^8iRī 8 -s.-s.-s.bsí -ř
-s. b -ř-fa-
-fa- Ü-^ s lié S>>
s S>> ļfi off
* off "
"ϊ"%
«J ^ * s«* η
¿? ° ri° »,¿>
ε «ρ * 5 *η
<■> JTβ"
« a ε β " e &
¿' ö ε S kj ί Ö JI <SQ. g kj 3
ö kj Ö <SQ. kj

•+¿
c
0

1 W it II
W
hH II
II "
"

1 *>"
W «B AV
ξ 2 1? «l*· H ^ I«
2 O «l*· H îç - I«
0 » ■ Η Β , +
*|£ ö α «τ · - --
Ē«g « *|£ öl
feV - g fer - - " ö
1 i ~'f ι > ^£ļs ^ ' vk
0 i§ «?
H kj + >ssξ^>^§II**«
·, āā >+ ξg I**«
·, ' a o," .w*
g S'®
*3 jš „~gl1 a S'® o,"
Ε + û ® jš gl et g
Q Ε b kj "Ί-" Í¿Un- I ξ ^ -í ¡í û ® jš et M g „ Λ
Q 1 b g «á? g ·Γ ķķ. ι I i·? ICI£ ^ V J
2 ^7^Γ-|ΐγ" I?? »v |¥ řř Síí^p
gg*
g* SS¿¿ηη t«I«I *·*·> >f rfl ' o I?? oO **>££-
^ rfl O fc |¥ÖÖM
í M JJQ Qν &&*1?++
I g* i S S η ž > * ^ «* rfl o b O I fc r g ι M I J j jļ Q ^ & +
^ » Os o ~ 2
Η

2043

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
I !
η Π

it it
« îe £
< <^
^ H <

tí ^ "El -F (C° «C0 tr^ 5 < sT"El

r -.
c- s - u * -
*. _
*.
* ^_^ ** '* ¿ a * ** C^ ^ ¿
"g ł^IłT "^5 fe .* vi 2* 1 "**
J "g "^5 f .* κ vi Μ!·? 2*
-'s - * 1 S
¡I "■--' * - Ii ^ §
fc5 *
fc S 3! 7■?
1 » »b b
S VVSi** tì
tì .?.?; ■
8
**■> 3 - S ^ sa?* Si ' t,
ļ«;5^ Χ
ki t# f-* e

II
ίΙ Γ ·*ΐ ! ř î' If Γ ^
II 11 ** l' „ II II R* II
*· ^ 5 'L ö * δ * 3 11
*· ? i ^ i 5 * εE|
δ *^i5** *
^ ^-β»
3 -> iĘ
ö

ni
II

o II
~|cr sf ik
~|cr I

«Γ|β·* β?Γ* t'


τ I·. fe Κ ^
η ¡ fe »Γ
-"U řk ° 1
?
? I^ ik-
w
-"Urϊ«řk « *1*
è sv-^
*1*4-
4- iü
Τ
ä=u ^ I 5 1 fe
C/D ^ I ^ μΡ|«ο ^3 ä=u ^ ^ ^ ^ 5 ^ ,
C/D
O Ζ ~~ ^I SμΡ|«ο ^£' Ir 'i ^">tfL* ^ ¿
su~ ι'
O gΖfe
fe*■ S ?k H H gg£k
glgl £tfL*
¿ ¿v su 1 ~fel··
L '¿>
< ^ ¡- ---- 1 s 1 + „I pi βΓ ř-T J"
s
w <1 --i
a- ^^ ιflip's?
fa c r-¡- ----
"*' >|3 1 ;-" s 1
1 I

<! ^ Lb* »? ^ a- ^ £!S


o <! Lb* ^ ík * »? ¿ ïrl»? ~5 ^ £!S -í ¿3. fek* _, "f^ìrTÌ Î Jb 's ' f |g ^ 's* jf
M .-.L·» u *" - -ς * ι

Β M ^-r- .-.L·» r ìi* u *" ί - »! -ς ί * - sv ι


H S fek 1 + J j + I." si V ^ « έ- ^ o ^
^ S I - - fek - ~ feV "? + ? j ^ I." » fe"k
0 ES g ¡ § - - ^ - gf ~ feV ^ hq ζ "? 1 ? r ^ ^ g » fe"k -i ^ -f f ^ 1 ^ tfk -*i^'
ES § I < ^ ^ hq ζ 1 ^ ^ ^ ^ ~ 1 χ

1ΒES1 1i"i-"
"i-"° §-^-¡1
-^-¡1 I ■
< ^ Iř
^ ! hq ¡74 ζIři
tfv-i Iři 1ff i®
^ i®I ^·+
·+ ^! ¿¿"Cic
řř ^iļ ,ř
f ~ ļ¿ ¿
1 χf fet?
i iL. r if, ΐ·ΐ .
s b-3 i "í ^ ^b^sl Λ^ι § u *íi¡ o 5 s's f 1 ι L? 7|i. ί S - 8 fe Ses g ϊ:Ξς ί 1 r ļ pía I «ii g
b-3

I i ik i1ř| ι 1 í|ř ef. s ť í-T'-rs ^


8 b-3 1-
en ><t «o *o ®2
H

2044

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
I δ ^ ì 8 ^ -
δ 5 ii s 8 g η - S
" §
Čt LSLS
"*« "*«
C« 8 " 8 IS 8
aT-Īp sT-*> č-κ aT 4

. «Π m r ir» «Λ r - L* ^ ^
^PCk £§.r
řs . «Π £ sTfe m r h ir» *· «Λ h r S * - L* h ^ *·
Il I! Il l'

J"1 J"1
w V» "» S w ft V

II

ί" ΐ
l,E 4? Il

f ^ ^
ř tí 0ri| «
1 Η 5 ^ Η
JP „i
^ JP «λ«λ ^ *·
·?« .4" ? Η ^ s JP h «λ +
__ SS fe h ł+ ? Il £ ^
'"ř
'"ř *°
3 ^Qι *1-
*° 13 fj»Ifi
- oT^
fe'k ι *1- »I
O jt ^ &
ä O 4 jt , Q I ; I
ψ IfeV , » + ¡?i .
Ο

Λ I'ĒĒ
? Î <* fi·
' 1?
* ** * ?* J?'*
+<*+ * fel8-"
+τ+ *τ *|«r
* --
·, ii
D ■* fcTlt-T <° * * bTlř-Γ ' + + 1 si oo
§ D 4Ϊ- ■* fcTlt-T <° ■ * ¿ * utì bTlř-Γ ' -
W* *
Z δ· I« «
¡6 ».
». ·~ *"Í
·~ 3 i ,«*ctì
£ -ι
^ '„ I, ^ka
I ^kaβ.|
β.| a
a
ã Z0
Λ * l"s* δ· *
¡6 ! * *
3 τ * i*
* ,« ř'*
* ctì *i
§ -ι §4;
„ 4* , i' «τ

Hg fe 0
g¿te^
a ►£ 'Bl?14 ? "te»u
1
+ ~lr'~ ^ ^|tí ^ ^ - s λ
ķ- Il 1 Λ 1 £ 11 ai ^ li" - - + >· I? ,-5 §* "¡«
§ ķ- «1 Il 7^* Λ £ ? t* ai «ί li" ? - 4·" - >· I? ř ,-5 4- .-
« * u ř Sf¿ ·■ ř í :r * ~ ,4 i-a
« § § * 5 ~ ι 5«l«í ¿tí 4 ^ ^ 4 ^ 2 « -
s I ^ rfevAsJ¿~ ι ** Šf V 55 ι 11 ¿5 1 !
i s lii^lX+î rfevAsJ¿~ ? * V ilk ί ι ¡**^ ¿5 " i ■
Ι is s fí ¿i if j 1 i " È -"»ιρ
s?5 Is+9hryi*-ļ
g s * r ?+ I*eu"
I 1I "ii ii? SI "ci-?
^ !|?«*
^ ι♦·.
H 2 ^ SS jq

2045

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
ts

11 cr-
-I - 1 *

Ji '.ι CC
.<* σν βο ft, οο
h *. θ· *
ii ii 11

ν»
ν»* ^^ Í
Í i"β-
"β-

C
3 II
Ψ Αι
£
ΪΛ
fTk
Q '
'
-,|<Ν
^
0 1?" ^|βΓ
§ ~ ¿ -4 ä"
o> w
o> te"'N"
^k- t ^^ tj ι
j ι
11^ h?
11 τι
h?Β tí*·
tí*· *-s
IgL
§ ! - Ae ί cl* IgL ^
§<*1
O § $<*
* ==§ * $
> Μ_ (í „ As
- ν ! ! -
O^ g
Ε Ν ^7 V-
^ - I*k _ §S ah! " == ί c> *_ ^ τ Μ_
t §s Λ- 8* TT - -τ- *jķ I ss Λ- " , i i ; c τ, Λ«
0μI 3 »'s--îtî
^k^kοο 88 -'
^ S?
« k
« k** SS ^
di ^k
-* -'c "Si?
* *^k» cIs'-r-
* >■s*g, »< I «ϊ~¿Ir
-r- 1 s § « ΐ T'
1 μ g 3 4- s- " " ^k : ¡4- 8 ^ " « - ¡ S » t § ^ ^ % _f c τ j« I I -r- í í ib s I g ^ < ¿Ir l jii
ωω E¡a¡a "S "^,ar
g -ι-
" ccI gI^,οΓ
feh"' ^c«ι
cc tI feh"
s 11«ι>s S§ _,
% >¿h
I + ~+"+I I* *"ļ 11 +"
h ++ ^^i si"
S si"xx »
* ¡ S ^,ar : ik -ι- L" ^,οΓ x¡ ^c I . τ § ι ■ ¿ _, ¡ I ; + ¿¡ +" I ¿ :, +" ¡y i ; ^ »
8 1 1-· i i·"-- f I -ií A" I ·-- k' ¡ 51 4 ■*·· ¿¡ : ϊ ι ■ ¡ --i ^ τ
8 I _ll_ř:ll_±l
^ S » 8
Η

2046

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms

S*
3 +3a***
' + ·-
Il H*"»
U Q h)

S r »
4 ^Ίι-Γ t) 1^ 11 t) 1^ Ml
«o S , r» ^ a^°
< «o < κ , m JT
.τ.τ
ι s <r
^ «Γ
s ^m I ®®
« m ®®
* £ ? « -I« 3·
Il * + ,

il-" -L · _ flf «"


a | r 11 % - - ' +
CT s«8 j- - «sic - ""jR t
8 + ? $ <·« « + 1 T"i .r
w 1 s" ^ 1
C/5 s" ^
- JS
« ty
ty **·
Z ž|i s <? " 1 1 γρ ü "t' *
2 SR
2 2 SR
31 a 31 ž|i
» fe f-s1 <? .»s
ι p. ïi fe f- 1
2 G< ΒG
& %%S fa
.a "£
-4· *'» s-
.aΡ fe
ii 1 f-
-S Ae *' s-
S ¡ ii -4· ' J Τ ■
Ië -Ł
ι Λ¿ *·"¿ 1 sl-
f/ . ìifSP! ! 1ä
§ R : ' ί ii I if ^ * ^ g J· J *■ „«ι» fe h Λ? 1
R ^ ^ ^ ^ ί I ^ ^ * g J· *■ fe h j 1
řT3HS3^ à^ k?
H ^ ^S5|δ S ^ "*L
Ià Ē ^* ^riSee
»-3* ■*: ri
So ^00
^ ^SS έSo
§ ^-~*~ «-Irl ,rI S
1¿ co i-*; «ζ »- ι«-Irl
§ - ,rI
η +uT S
ΚΟ1¿ΚS ν
S co
^ i-*;
ν ' 1 Ι fe + ! «ζ *?L? k ? 3 ■*: ^ * i SS ? 00 s Λ ι § S " η ^ J3· ΟΙ + Ι«? ^
ΟrV
J rV I rs
. ^rs
ν η Pí ·*: a S O m coi S «o ^ O ■*< ;?

gMM rVSdisco0
. t Sdisco0 rs ru ru
η ¡ ° Pífclco
·*: ř Κ a S Q°6 SSlTI
II OQΐ mSfc>-SSlTI
coi Β S Jk Β
° «o °¡ 5 ^5ϊ Oeo
1 ~ ■*<
~ «:A'
A' Sä 1
;? 1
I M¡5 ι i
¡5 ¿ Sdisco0 i*ι* > fclco^+ lf + ' i ° i II' » 8iQ S "1 »I ΒË-*°+
I SSlTI : fi- ° s ì 5 i eo ~ - /,- A'+«ρ
O
OJ δ ι ^ 51,
§ Jfķ Jfķ¡ 51,
S Is"1ilSgIs§il*g*>Ë-*°+
?k g- ·* ιs 2*> ?k i · 2 ' ,s o +
Ρ* fe 5 £§ S S
ce a; 3 Q ÎJ Ę

5«? S rí Sì

2047

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
o

II

s?
"-«ta
+

¿Ti t-

lit ¿Ti t-
m

N £
*
■Í" 1

PI β
%k β
r7~"


s"55P
-r 1

1 Λ»
GO 1
y-y es es
00 00

2 * =■ -ã^
η J? · tek
■«í ^ -> I
ρ II -> -Η I
a - +
U C c s»

5 U ^ C c it " s»
g % *" " - - ^
o
CQι ι
s «3
"s ξ fl^ Άτ
^ ^ |* |*
£ fl ® «« O« II I,
g ÎS
ÎS i>| «Λ«Λ
i>| Η
H Η βCS ^
" " βο II ^^ ^" ^^'*
" * βο
1 g2r'^Π:
Sh
'*co
co«
S u. «CS ^ «O
1 gt*J %ßΚ1Κ
<ο^ 1 §*3
*§ -I-
-I- +
I* bebe
, I*
+ , CSlQ
eil+ fl 5 3
0 S I u. « sKlsrk «O r' ¡ Π: JE <ο^ '
tí Í5 « Ν § - " tek č Ą ļs *
ä *5 ii « Ν Ρ " + Ą * . ι π I r ΙΦ r«| à: . ri I*

1 ä i *5 íí ii « 5|¿ ¿ Ρ ^ '* ft I + ř Ą 1 . ι ¡"h^" π I 1 C1!® r«| I à: f . ä?


é xasxas
S ^ Χ' ? '* Λ· Ι Ώ f ? ft j li > Λ«Ώ
s -a & 1 s fli
Λ° s
I xas 8 s a ? s I S i
H
|j s S 3 S $ S
2048

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
/- '

β
O

1 a ♦

§
fa a

, ο β
0 ® * Il "
W Š if " ' ai fel

W ě s * *
> >
fc E Hfs-čk
J3»" ο ι *li iks E f
1 > rlřii^ E J3»" !ř fs- fc f
¿ . ■ I ; : ι t f|£ I fc í , Λ" .' η
Igg §ss ♦řř %
irir ιI ■fIŤ ■
¡v i .¡v
? ¿ ¡. I .! i·. Λ"g. §! ■. ·'¡ř|1■a 1aI ;;řlř'l'l

i fe J I ìàì'fèl I ř Λ 'T* S t»
fe J

Ķ S

2049

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms
APPENDIX 2

Derivation of Sensor Fault Coefficient Matrix ( Hs )

If we let T2 (deg R) and P2 (psia) denote measured engine inlet temperature and pressure respectively an
be the ratios T2/5 18.67 and P2/14.696, we can describe the standard day corrected value P* of an arbitra
by

Ρ* = (Al)
Θα ôb

where a and b are suitably chosen exponents. Likewise the corrected form X* of the base power parameter X may be
given by

x'-sh (A2>

The percent delta change AP fro

1 + 100
+ + + ^
Ρ * (
REF (A3)
( (A3) }
}
Taking logs and differentiating equations A1-A3 yields

dP* _ dP „(d&' ~ b u ( dô'


Ύ*~Ύ~α dP* _ dP „(d&

Τ*" _ Ύ~ „ α'Θ) (d&' ~ ^ ñ U7 (Α5)


dX* _ dX „ (d&' ~ ^ ñ (dô'

dP * d Ρ ref/ Ρ REF ^ d(AP)


Ρ * Ρ REF ^ 100 <Α6)

where the approximate equality in A6


condition Pref = f(X*) we may also write

d P ref = dPKEF /Pref dX* = „


Pref = dX*/X* Χ* = „ X* (A7>
ss being the steady state slope of the gas generator curve at the opera
provides the relationship

d{AP) _ dp _ _ ¿¡Θ , (o M dò _ „dX ÍA81


100 Ρ 0 d Χ

which will generate an M

2050

This content downloaded from


160.75.86.201 on Sat, 21 Nov 2020 10:54:01 UTC
All use subject to https://about.jstor.org/terms

You might also like